Datasets:

m42-health
/

HC4

doi stringlengths 3 59	license stringclasses 95 values	pmc stringclasses 0 values	pubmed_id stringclasses 0 values	subset stringclasses 1 value	text stringlengths 301 1.8M
10.1101/2022.05.12.491637	cc-by	null	null	openalex	Background Shotgun metagenomic sequencing is increasingly popular in taxonomic and resistome-profiling of polymicrobial samples due to its agnostic nature and data versatility. However, caveats include highcost, sequencing depth/sensitivity trade-offs, and challenging bioinformatic deconvolution. Targeted PCR-based profiling optimises sensitivity and cost-effectiveness, but can only identify prespecified targets and may introduce amplification biases. Ultra-high multiplexed PCR is a potential compromise. As no comprehensive comparative evaluation exists, we evaluated performance of each method in taxonomic/resistome-profiling of a well-defined DNA mock sample and seven "realworld" wastewater samples. Results We tested three sequencing approaches (short-read shotgun metagenomics, Illumina Ampliseq TM ultra-plexed AMR Research Panel, 16S rRNA gene sequencing) with seven bioinformatic pipelines (ResPipe, Illumina DNA Amplicon App, One Codex Metagenomic-/Targeted Loci classification and Ampliseq TM Report, DADA2, and an in-house pipeline for AmpliSeq data [AmpliSeek]). Metagenomics outperformed 16S rRNA gene sequencing in accurately reconstituting a mock taxonomic profile and optimising the identification of diverse wastewater taxa, while 16S rRNA gene sequencing produced more even taxonomic outputs which may be quantitatively inaccurate but enhance detection of low abundance taxa. Shotgun metagenomic and AmpliSeq sequencing performed equally well in profiling AMR genes present in a mock sample, but AmpliSeq identified more genes in more complex, "real-world" samples, likely related to sensitivity of detection at the metagenomic sequencing depth used. Conclusions A complementary sequencing approach employing 16S rRNA gene or shallow-metagenomic sequencing for taxonomic profiling, and the AmpliSeq AMR panel for high-resolution resistome profiling represents a potential lower-cost alternative to deep shotgun sequencing and may also be more sensitive for the detection of low-prevalence AMR genes. However, our evaluation highlights that both the sequencing and bioinformatics approach used can significantly influence results; for AmpliSeq AMR gene profiling, we developed AmpliSeek which outperformed the other pipelines tested and is open source. Sequencing approach and bioinformatic pipeline should be considered in the context of study goals and sample type, with study-specific validation when feasible.Background Shotgun DNA metagenomic characterisation of polymicrobial samples is increasingly used in both clinical and environmental studies, facilitating agnostic sequencing of all DNA present in a sample and enabling flexible comparisons with reference databases to determine sample composition 1 . For example, the RefSeq 2 and CARD databases 3 can be used for taxonomic and antimicrobial resistance (AMR) gene (i.e. "resistome") profiling respectively. This approach has underpinned multiple studies characterising microbiome-disease associations 4 , evaluating community diversity and anthropogenic impacts 5 , and investigating AMR 6, 7 . There is also growing interest in the use of shotgun metagenomics to profile wastewater for population-level surveillance of AMR (i.e. wastewaterbased epidemiology [WBE]) 8, 9 . However, shotgun metagenomics can be expensive, and bioinformatic deconvolution of the data challenging, especially when using short-read sequencing and trying to characterise the genetic context of specific loci, such as AMR genes 10 . Sequencing depth impacts sensitivity to detect rare targets of interest, and therefore, most studies involve a sequencing depth-sensitivity trade-off where uncommon sequences may lack coverage or be missed completely 10 . Additionally, for many studies, only a fraction of the metagenome (e.g. bacterial sequences) will be of interest, and in this context, much of the wider metagenome represents wasted sequencing effort; for example human DNA in studies analysing clinical samples for pathogen diagnostics 1 . Amplicon-based approaches enrich specific nucleic acid templates during library preparation to increase sensitivity for relevant but less abundant target sequences at reduced cost. A commonly used single target for profiling bacterial communities is the 16S rRNA gene which possesses both highly conserved regions facilitating the use of universal primers, and hypervariable regions which can be used to discriminate between bacterial taxa 11, 12 . Increasingly, targeted ultra-highly multiplexed PCR panels have been developed to enable amplicon-based evaluation of the presence/absence and diversity of specific features of a metagenome associated with a key phenotype, for example AMR gene diversity , or of a subset of organisms associated with disease, for example respiratory viruses. One example is Illumina AmpliSeq TM which can be used with both Illumina-and community-curated panels (primer pools). These large panels enable capture of a broader range of targets than universal primers or multiplex PCR, whilst theoretically optimising sensitivity and sequencing resource, and are attractive to researchers as they come with defined laboratory protocols and user-friendly bioinformatic pipelines. However, despite internal validation, Illumina-curated panels have demonstrated mispriming events leading to false positive calls of targets; the community-curated panels have largely not undergone any internal validation. Additionally, targeted methods may be prone to amplification biases, and by their nature only focus on specific features of the wider metagenome. Differences in reference database curation and mapping approaches between available bioinformatic pipelines may also influence results 16, 17 . Combining targeted approaches such as 16S rRNA gene sequencing and ultra-highly multiplexed PCR/amplicon sequencing of AMR gene targets represents a potential alternative to deep shotgun metagenomics for profiling species and AMR gene diversity in polymicrobial samples. We therefore assessed the performance of shotgun metagenomics, versus 16S rRNA gene sequencing and the AmpliSeq for Illumina Antimicrobial Resistance Research Panel (henceforth "AmpliSeq") in reconstituting the true taxonomic composition and resistome of a well-defined DNA mock microbial community. We also compared several bioinformatics approaches to characterising species and AMR gene profiles, including an in-house approach to AmpliSeq data analysis (AmpliSeek). With performance on the mock community as a reference, we then applied the best approach to seven untreated wastewater samples to quantify performance in the context of "real-world" sample complexity. We aimed to highlight potential limitations of each method and recommend potential use cases. Methods DNA mock community preparation and wastewater samples Metagenomic, 16S rRNA gene and AmpliSeq sequencing were all conducted on aliquots from the same mock DNA sample (Fig. 1 ). The mock was prepared by combining the ZymoBIOMICS Microbial Community DNA Standard (Zymo Research Corporation, Irvine, USA) with three bacterial isolate DNA extracts chosen to enrich the mock sample for clinically relevant AMR genes (Table S1 ); these bacteria had been characterised with whole genome sequencing as part of a previous study evaluating human bloodstream infections 18 . Wastewater samples comprised seven metagenomic DNA extracts from wastewater influent collected as part of local surveillance in Oxfordshire, UK, 2019, and stored as pellets at -80C. Metagenomic DNA was extracted using the PowerSoil kit (QIAGEN, Hilden, Germany) according to manufacturer protocols. Sequencing Shotgun metagenomic sequencing on mock DNA community was conducted by Novogene Co (Beijing, China) on the NovaSeq6000 (Illumina, San Diego, USA), generating 150 bp paired-end (PE) reads with a target sequencing depth of 20 million reads (6Gb). AmpliSeq was conducted using the community-designed Illumina for AmpliSeq AMR Research Panel and the Library PLUS kit with manufacturer guidelines; libraries were pooled and sequenced on the Illumina MiniSeq (150 bp PE reads) with replicate libraries prepared for the mock sample. 16S rRNA gene sequencing utilised 515F-806R primers for library preparation and libraries were sequenced on the Illumina MiSeq (250 bp PE reads) as previously described 19 . Wastewater samples underwent sequencing as described above, however, shotgun sequencing was conducted by the Wellcome Trust Centre for Human Genetics (Oxford, UK) with a depth of ~75 million reads (~23Gb) per sample based on previous deep sequencing and rarefaction analyses demonstrating this was the minimum sequencing depth required to capture most AMR gene diversity in this sample type from the same sampling site 10 . Software We tested seven bioinformatic pipelines (Fig. 1 , Table S2 ). Two pipelines were compared for metagenomic-based taxonomic profiling (ResPipe v1.4.0 and One Codex Metagenomic classification v8/13/2021), two for 16S rRNA-based taxonomic profiling (DADA2 v1. 16 [assignTaxonomy] and One Codex Targeted Loci classification v4/15/2021), and three for AmpliSeq-based resistome profiling (Illumina BaseSpace DNA Amplicon App v0.7.12, One Codex AmpliSeq Report v1/17/2019 and an inhouse BBTools wrapper pipeline [AmpliSeek] developed as part of this study available at https://github.com/KaibondChau/ampliseek). AmpliSeek performs trimming (BBDuk2) and merging (BBMerge) of reads produced by AmpliSeq sequencing before stringent mapping to AmpliSeq panel target sequences (BBMapSkimmer). Reference databases used by each pipeline are detailed in (Table S2 ). Geneious Prime v2021.2.2 was used for detailed in silico characterisation of the mock AMR profile by visualising the relevant AMR reference sequences and reads mapping to these sequences for each methodological approach (Fig. S1 ). Taxonomic profile analysis We used two previously described error metrics 20 to quantify differences between theoretical and pipeline-reported distribution of bacterial genera present in the mock sample, a modified mean absolute proportion error (MAPE) based on theoretical vs reported read differences, and Bray-Curtis dissimilarity 21 (BC) which further considers the total richness present. Both MAPE and BC scores range between 0 and 1, where 0 represents identical composition. For ease of interpretation, scores are presented as 1-MAPE and 1-BC, where higher values represent better accuracy. Since no theoretical truth existed when characterising wastewater, diversity metrics were used to compare differences between sequencing-pipeline combinations. Chao1 was used as an estimate of taxonomic richness by representing total count of unique observed taxa. Pielou's evenness was used to assess how similarly represented observed taxa were in abundance estimates (constrained to [0,1] with low values representing unbalanced community estimates where few taxa constitute the majority of abundance). Shannon index was used to quantify both richness and evenness as a measure of overall community complexity where higher values represent increased entropy and therefore complexity. A two-tailed Welch paired t test (allowing for different variances in the two samples) was used to determine significance between differences in diversity metrics. Sensitivity and specificity calculations for AMR gene detection The mock DNA community was annotated with AMR target sequences present in the metagenomic and AmpliSeq reference databases using 100% identity threshold to determine a "strict" in silico truth. These annotations were compared against pipeline calls using scoring thresholds defined by each pipeline (i.e. thresholds at which AMR targets are called present). For AmpliSeek, scoring thresholds were optimised to maximise sensitivity and specificity based on the "strict" in silico truth using receiver operating characteristic curves and Youden's index 22 (Fig. S2 ). Established stringent scoring thresholds for AMR target presence were used for ResPipe (lateral coverage=1) 10 and the One Codex AmpliSeq Report (coverage >= 85% and identity >= 95%) 23 . Sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) for AMR target detection were calculated by categorising AMR genes called by each method as true positive, true negative, false positive and false negative, when compared with the in silico truth (see above). By adapting this in silico truth with respect to the reference AMR gene database used for each method, we fairly compared overall methodological performance (i.e. the CARD v3.0.3 database of 2605 sequences for metagenomic AMR gene profiling approaches and the AmpliSeq reference panel of 815 [AmpliSeek] or 814 [One Codex] amplicon sequences which target 478 AMR genes). However, AmpliSeek scoring thresholds were optimised to the mock dataset and therefore AmpliSeek performance is not directly comparable to the DNA amplicon app and One Codex Ampliseq Report where scoring thresholds were not optimised to the mock dataset. Mock DNA sequences were also annotated at 98% and 90% identity thresholds to determine "intermediate" and "relaxed" in silico truth respectively to investigate AmpliSeq false positives, i.e. to include more AMR gene matches in the truth set as present where these were similar to a gene in the reference database. Results Taxonomic profiling All bacterial genera present in the mock DNA community were correctly detected by each sequencing method-pipeline combination, but the relative abundance of different genera was variably under-and overestimated for each combination (Fig. 2A ). This was most notable for 16S rRNA gene-based classification of Klebsiella which was underestimated by One Codex Targeted Loci classifier and overestimated by DADA2. Bacillus and Lactobacillus were also inaccurately estimated for One Codex Targeted Loci and Metagenomic estimates respectively. Regardless of pipeline, shotgun metagenomics more accurately reconstituted true taxonomic composition over 16S rRNA gene-sequencing (Fig. 2B ; mean 1-MAPE=0.77 vs 0.54, and mean 1-BC=0.87 vs 0.75). Differences in performance between pipelines within each dataset was small, with ResPipe (1-MAPE:0.79; 1-BC:0.88) marginally outperforming One Codex metagenomic classifier (0.75; 0.86) for shotgun sequencing data, and DADA2 (0.55; 0.77) outperforming the One Codex Targeted Loci classifier 0.52; 0.73) for 16S rRNA gene sequencing data (Fig. 2B ). AMR gene profiling Annotation of the mock DNA sequences with approach-specific reference databases identified 77 CARD and 43 AmpliSeq target sequences present at 100% similarity. These were all detected with 100% sensitivity by all respective methods/pipelines (Fig. 3A ), except the DNA Amplicon App, which performed poorly (22/43 known AmpliSeq targets detected; 51%). For precision, ResPipe (specificity=0.99; PPV=0.72) performed best, followed by AmpliSeek (0.98; 0.69), DNA Amplicon App (0.98; 0.65) and One Codex AmpliSeq Report (0.94; 0.49) (Fig. 3B, 3C ). However, notably, we found that most false positive hits obtained (67/75; 89%) arose from <=2% nucleotide variation from reference AMR sequences (AmpliSeek [18/19], DNA Amplicon App [12/12] and One Codex AmpliSeq Report [37/44]) (Fig. 4 ). Replicate AmpliSeq libraries produced consistent results with no impact on performance metrics for all pipelines tested (Fig. S3 ). AmpliSeq false positives were investigated by comparing pipeline calls to "intermediate" and "relaxed" mock in silico truth (see Methods). We considered AmpliSeq reference sequences present in the mock with 98-100% identity as target variants (Fig. 4 ) and assessed these separately to account for the limited sequence diversity amongst reference sequences for specific targets. An additional 37 reference sequences were identified in the mock at 98-100% identity. When these target variants were included in overall performance metrics, One Codex AmpliSeq Report retained 100% sensitivity with increased specificity and PPV (0.99; 0.92). However, AmpliSeek and DNA Amplicon App were less able to detect target variants, with overall sensitivities of 0.76 and 0.43 respectively. However, AmpliSeek sensitivity improved to 0.90 (Fig. 4 -grey fill) when including less stringent scoring thresholds (i.e. "potentially detected" calls) (Fig. S1 ). Real-world wastewater samples Taxonomic profiles Consistent with findings from our mock microbial community benchmarks, taxonomic profiles of "real-world" wastewater samples differed significantly between 16S rRNA and shotgun sequencing, with large differences in relative abundance and detection of core taxa (Fig. 5 ). For metagenomics, Proteobacteria were estimated as most abundant across all samples, with relative abundance approximately two-fold higher than reported by 16S rRNA gene sequencing. Conversely, relative abundance of phyla detected by 16S rRNA gene sequencing across all samples were reduced in metagenomic estimates (Firmicutes, Bacteriodota) or entirely missed/found at <1% relative abundance (Fusobacteriota, Campylobacteriota). Interestingly, this was not the case for Actinobacteriota which was estimated with higher relative abundance than 16S rRNA gene-based estimates by both shotgun pipelines. 16S rRNA gene-based estimates of phylum abundance were more even, with variable detection of Chloroflexi, Patescibacteria, Planctomycetota and Verrucomicrobiota across samples. Interestingly, Campylobacteriota was not detected using the One Codex Targeted Loci pipeline but reported by DADA2. Alternative faceting to highlight difference on a per-sample basis is provided in (Fig. S4 ). Figure 5: Relative abundance of phyla for wastewater samples (1-7) faceted by sequencing approach and pipeline combination. Metagenomic approaches consistently reported significantly higher richness (Chao1) than 16S rRNA gene sequencing approaches at all taxonomic ranks tested (p<0.002; mean increase of 1354 genera), with the One Codex Metagenomic Classifier identifying significantly more unique taxa than ResPipe (p<0.001; mean increase of 787 genera) (Fig. 6A ). Conversely, 16S rRNA gene sequencing approaches produced profiles with significantly more even representation of taxa (Pielou's evenness; p<0.001) than metagenomic approaches (Fig. 6B ). Despite higher richness in metagenomic results, overall community complexity (Shannon index) was significantly higher for profiles generated by 16S rRNA gene sequencing than metagenomic approaches (p<0.007); except for ResPipe at the genus level. (Fig. 6C ). Interestingly, One Codex Targeted Loci classifier (16S rRNA gene data) produced taxonomic profiles with higher variability across the wastewater samples than all other sequencing-pipeline combinations (Fig. 6ABC ). AMR profiles Across all wastewater samples, AmpliSeq One Codex Report identified a total of 3220 AmpliSeq targets (2889 called as present, 331 called as probable) while AmpliSeek identified a total of 2157 (1970 called as present, 187 called as potentially present), and the shotgun metagenomics-ResPipe pipeline identified a total of 526 CARD sequences with lateral coverage of 1 (i.e. 100% nucleotide identity over the full length of the reference sequence). When deduplicated to account for targets shared across wastewater samples and redundant AmpliSeq amplicon design (Fig. S1 ), the AmpliSeq One Codex Report detected the most unique AMR genes (n=367), followed by AmpliSeek (n=300) and shotgun metagenomics-ResPipe (n=132), respectively. The distributions of AmpliSeq targets identified by the One Codex Report and AmpliSeek showed the same ratio of increased present calls as the evaluation of the defined mock (i.e. 1.4-fold more AmpliSeq targets called as present by One Codex than AmpliSeek). Similarly consistent with the methodological comparisons on the mock sample, almost all AMR genes called as present by One Codex but absent by AmpliSeek were hits identified at <100% nucleotide identity (817/826; 99%; median nucleotide identity: 98% [IQR:97-99%]). The nine targets identified as present and at 100% identity by the One Codex pipeline but called absent by AmpliSeek possessed low numbers of mapped reads by One Codex (median of one read mapping only). The single AmpliSeq target not targeted by the One Codex AmpliSeq Report (aph2prime-Ia_NT) was called as present by AmpliSeek across all wastewater samples. Discussion In this study, we have compared the performance of metagenomic and targeted sequencing and different bioinformatics approaches in accurately reconstructing taxonomic and AMR gene composition of a defined mock microbial community, and a set of polymicrobial "real-world" wastewater samples containing complex microbial communities. We highlight that results are significantly influenced by both the sequencing and bioinformatics approach used, and this is an important consideration for studies using these methods. Mock community evaluation Metagenomics outperformed targeted 16S rRNA sequencing and more accurately estimated true relative abundance of mock genera regardless of the bioinformatic pipeline utilised. While 16S rRNA gene sequencing successfully captured all mock genera, relative abundances differed significantly from the known, true, composition. This was not unexpected as targeted approaches are subject to primer and amplification biases already known to potentially distort taxonomic profiles , however, the degree of bias has not been previously quantified. This was most apparent for One Codex Targeted Loci results where Klebsiella and Bacillus abundances were respectively under-and overestimated by up to three-fold. Conversely, shotgun metagenomics is less susceptible to these biases, and when combined with statistical methods for normalising abundances (as used by both ResPipe and One Codex Metagenomic Classifier), accurately estimated mock genus distributions. For profiling taxonomy, ResPipe and DADA2 slightly outperformed the One Codex Metagenomic and Targeted Loci classifiers. Since classification methods were similar within each sequencing approach (i.e. k-mer based for shotgun metagenomics and alignment-based for 16S rRNA gene profiling), differences in the reference taxonomic databases are most likely responsible for differences in performance 16 , consistent with previous studies demonstrating classification is dependent on database choice 20, 24 . Almost all metagenomics and AmpliSeq approaches detected all relevant unique AMR genes known to be present in the mock microbial community. However, the DNA Amplicon App performed poorly, reporting no reads mapping to 21 AmpliSeq target sequences known to be present in the mock and reported as present by the One Codex AmpliSeq Report and AmpliSeek. This appears attributable to differences in the read mapping workflow rather than alignment algorithm per se since both One Codex AmpliSeq Report and the DNA Amplicon App use the BWA algorithm to align reads. However, instead of directly aligning reads to the reference AmpliSeq target sequences as performed by One Codex and AmpliSeek, the DNA Amplicon App first aligns reads to a reference database of whole genomes containing the reference AMR genes 27 . We noted that this initial genome-based screen appeared to erroneously filter out reads which would otherwise map directly to sequences present in the AmpliSeq AMR panel. However, full details of the DNA Amplicon App workflow are unclear and detailed evaluation of this pipeline was therefore beyond the scope of this study. For AmpliSeq based methods, when focussing on AMR genes known to be present in the mock with 100% nucleotide sequence identity, AmpliSeek slightly outperformed One Codex AmpliSeq Report as the latter reported more false positives. False positives have been previously described with the AmpliSeq panels and attributed to mispriming events associated with primers erroneously binding to non-target sequences 28 . However, mispriming did not appear to be the main cause in this study since most false positives shared high sequence identity to the reference AMR genes (>98%), indicating that these are more likely related to pipeline-specific differences in calling genes present/absent. In fact, the highly related nature of most false positive calls made by the One Codex AmpliSeq Report would be consistent with the default 2% error BWA mapping threshold used in the pipeline to account for sequencing error. The difficulty in calling presence/absence of related AMR genes where the reference sequences used as markers are very similar is highlighted in our analyses: For example, the amplicon reference sequence for blaFEC (blaFEC_T) shares high nucleotide identity with amplicon reference sequences denoting blaCTX-M-3, blaCTX-M-15 and blaCTX-M-32 (blaCTX-M-3_T.2 [99%], blaCTX-M-15_T.3 [99%] and blaCTX-M-32_T.1 [98%]). Modifying the mapping threshold may avoid these issues since recent work has suggested the median error rate of Illumina MiniSeq sequencing reads may be as low as 0.6% 29 . For AmpliSeek we retained stringent mapping of reads as the default to enable differentiation between highly similar reference markers, guard against mispriming, and call specific alleles with high confidence, but this threshold can be relaxed depending on the users' requirements. "Real-world" wastewater samples On real-world samples, where the true diversity was much higher than in the mock community, differences in the relative abundance of taxa and the number of unique AMR genes observed between approaches when profiling the mock community were greatly amplified. In these samples, metagenomic approaches identified more unique taxa than 16S rRNA gene profiling at all taxonomic ranks tested. This represents a commonly cited key advantage of shotgun sequencing, where higher richness of bacterial taxa can be reported than for 16S rRNA-based methods , and is particularly important in the characterisation of complex microbial communities 31 . Taxonomic classification is also heavily dependent on the reference database used 17 . In this respect, the One Codex Metagenomic Classifier yielded significantly higher richness than ResPipe likely owing to respective reference species database sizes of ~115,000 and ~20,000 complete genomes. Taxonomic richness reported by ResPipe was also seemingly capped across most ranks with minimal variation between wastewater samples which may be a result of saturating the reference database as previously suggested 10 . Metagenomic abundance estimates were largely skewed towards a small number of taxa; in contrast, 16S rRNA gene-based estimates were consistently more even across all ranks regardless of pipeline, and more sensitive to the detection of taxa missed by shotgun metagenomic outputs. This likely partly reflects the sensitivity trade-off of shotgun sequencing where sequencing effort can be swamped by high abundance taxa and/or taxa with large genomes; especially when sequencing depth is low and community complexity is high 1 , which can result in missing the presence of lowabundance, but significant, taxa 33 . However, users of 16S rRNA gene sequencing for taxonomic profiling need to be aware of the likely greater divergence of profiles from the truth. In our hands, different 16S rRNA gene pipelines also notably impacted results; for example, Campylobacteriota were absent from One Codex outputs but classified as present using DADA2, further highlighting differences across reference databases. Interestingly, despite being more sensitive in profiling AMR targets in the mock, metagenomics detected less than half the number of AMR genes reported by targeted AmpliSeq approaches. This is potentially driven by low abundance AMR genes which are missed by an agnostic approach, but directly targeted and amplified by AmpliSeq. However, we used the strictest definition of AMR gene presence requiring read matches covering all of the reference sequence (lateral coverage=1), so our metagenomic profiling was a conservative estimate. As discussed, sequencing depth impacts the sensitivity of target detection in complex samples, with the relative simplicity of the mock composition (n=11 bacterial genomes, 77 CARD sequences, 43 AmpliSeq sequences) masking this limitation. As seen in the evaluation on the mock, the less stringent mapping thresholds of the One Codex Report produced more present calls than AmpliSeek due to calling target variants as present. In fact, the proportion of increased present calls made by One Codex Report compared to AmpliSeek was consistent between mock results and wastewater results. These findings support similar performance of both AmpliSeq pipelines in wastewater as in the mock despite significantly increased complexity, where AmpliSeek results represent increased confidence in the presence of the exact marker sequence. Limitations We have not evaluated all methods for profiling taxonomic and resistome composition and therefore potentially missed approaches with increased performance. However, our pragmatic evaluation focussed on accessible approaches (i.e. kit-based and all-in-one pipelines) which are commonly used, attractive to researchers and have not undergone comparative evaluation. The relative simplicity of the mock DNA community may mask sensitivity issues owing to sequencing depth but use of in silico simulated datasets would otherwise omit assessment of library preparation amplification reactions. To mitigate this, we enriched the mock with additional AMR determinants and presented performance differences between the mock and "real-world" samples. Our in-house pipeline AmpliSeek was internally validated on the same mock dataset used for performance metrics which inherently optimises performance, however, the "real-world" samples represented an external validation confirming similar performance with and without scoring threshold optimisation. We also guarded against wastewater metagenomic sensitivity issues by using optimised depth as determined by previous ultradeep pilot sequencing performed on wastewater form the same sampling site 10 . Our resistome characterisation focusses on presence/absence but a major advantage of shotgun metagenomics is quantitative measurement of genes which we have not explored here. Additional factors such as read length and approaches to normalization may also optimise each approach but exploring these were beyond the scope of this study. Conclusion A complementary sequencing approach encompassing 16S rRNA or shallow-metagenomic sequencing and the AmpliSeq AMR panel represents a potential lower-cost alternative to deep shotgun sequencing for taxonomic and resistome profiling of highly diverse, complex, polymicrobial samples. Accuracy of profiling was also to some extent dependent on the bioinformatics pipeline used; for AmpliSeq AMR gene profiling we have optimised an in-house pipeline (AmpliSeek) which outperformed the other methods tested on this dataset and is open source. However, further validation of this in-house pipeline is needed, since it was also developed on the same dataset. We recommend careful consideration and further validation of sequencing approaches and bioinformatic pipelines in the context of study goals and sample type to be analysed. Validating sequencing and bioinformatic approaches remains challenging due to difficulties in representing the true complexity of "real-world" samples; any conclusions drawn should always be considered in the context of the methods used. Figure 1 : 1 Figure 1: Schematic overview of evaluated sequencing approaches and respective bioinformatic pipelines tested for taxonomic and resistome profiling of the DNA mock. Figure 2 : 2 Figure 2: (A) Theoretical taxonomic composition of DNA mock as compared to actual generated pipeline outputs faceted by sequencing approach. Genus "Other" represents genera present at <1% relative abundance in pipeline outputs. For the theoretical facet, "Other" represents DNA composition from yeasts included in the commercial DNA standard. (B) Taxonomic profiling error scores represented by 1-BC and 1-MAPE where 1.00 is perfect reconstitution of theoretical genera abundances. Figure 3 : 3 Figure 3: (A) Detection sensitivity of known AMR genes present in the mock at 100% identity. NB the reference AMR gene database used to annotate the mock was specific to, and hence different for, each approach. Error bars represent 95% confidence intervals (CIs), and facets divide sequencing approach and bioinformatic pipelines. (B) Specificity of AMR gene detection presented as for (A) and with magnified y-axis in (C). Figure 4 : 4 Figure 4: Heatmap of AmpliSeq pipeline calls labelled by reference marker ID and sequence identity.Each column represents a bioinformatic pipeline output where black and grey fill reflect detected and potentially detected calls respectively. Labels (green, blue, orange) denote the sequence identity of reference markers when annotated in the mock DNA community. Figure 6 : 6 Figure 6: Distribution of diversity metrics across all wastewater samples coloured by sequencing approach and pipeline combination. Panels: 6A reports richness via log10 Chao1, 6B reports abundance distribution via Pielou's evenness, 6C reports overall alpha diversity via Shannon index.
10.35470/2226-4116-2022-11-4-217-226	cc-by	null	null	openalex	We consider one of the cybernetic methods in biology related to the study of DNA chains. Namely, we are considering the problem of reconstructing the distance matrix for DNA chains. Such a matrix is formed on the basis of any of the possible algorithms for determining the distances between DNA chains. The objects of research of these algorithms (for mammals), as a rule, are one of the following 3 variants: the main histocompatibility complex, the mitochondrial DNA, and "the tail" of the Y chromosome. In the paper we give an improved algorithm for restoring the distance matrix for DNA chains. Compared to our previous publications, the following changes have been made to the algorithm. We abandoned the use of the branches and bounds method, but at the same time significantly improved the greedy auxiliary algorithm used in it. In this paper, we apply only this greedy algorithm to the general solution of the distance matrix reconstruction problem. As a result of the conducted computational experiments carried out on one of the two considered criteria for the quality of the algorithms, significant improvements were obtained compared to the results given in our previous publications. At the same time, the total running time of the algorithm remained almost the same as in the previous version.Introduction In this paper, we continue to consider one of the cybernetic methods in biology related to the study of DNA chains. Namely, we are considering one of the important tasks of this topic, i.e., the problem of reconstructing the distance matrix for DNA chains. In this case, the distance matrix is formed on the basis of any of the possible algorithms for determining the distances between DNA chains of monkeys, as well as any specific object of study. "Plants, animals and bacteria all contain the essential biological molecule known as DNA or deoxyribonucleic acid. DNA contains all the information required to build and maintain living organisms. You can think of it as nature's very own top-secret instruction manual . . . " "This manual is written in multiple combinations, but limited to just 4 letters: A, T, G and C. Each letter denotes a nitrogenous base: A for adenine, T for thymine, G for guanine and C for cytosine. Every living being has a huge supply of these 4 bases, each of which is attached to a pentose sugar and a phosphate molecule. Together, they are known as a nucleotide. These nucleotides are arranged in two long coiled strands like a hair braid." "Every single cell which builds up a living organism carries information for various functions necessary for the survival of the cell. This genetic information in each cell is stored in molecules called nucleic acids. The most stable form of nucleic acids is called deoxyribonucleic acid (DNA). Each of the DNA strands forms helical structures that are long polymers of millions of linked nucleotides. These nucleotides consist of one of four nitrogen bases, a five-carbon sugar, and a phosphate group. The nitrogen bases -A (Adenine), T (Thymine), G (Guanine), C (Cytosine) encodes the genetic information . . . " ( https://www.scienceabc.com/pure-sciences and https://whatisdna.net/ ) However let us remark, that the total length of the human genome exceeds 3 10 9 characters, which is about 200 000 times longer than mt DNA (see below). This fact indirectly confirms the need to apply heuristics when considering DNA algorithms. It is important to note that currently it is easy to find only a few similar algorithms on the Internet, [Needleman and Wunsch, 1970; Winkler, 1990; van der Loo, 2014] etc. (the authors' usage of Internet searches give about 10 similar algorithms only); see also the description of our algorithm in [Melnikov, Pivneva, and Trifonov, 2017 ] and some of our other papers cited there. The objects of research of these algorithms (for mammals) are, as a rule, one of the following 3 variants: mt DNA, the mitochondrial DNA, inheritance in the "direct female line", see [Maloy and Kelly (Eds), 2013; Cibelli et al. (Eds) , 2014] etc.; for human, the the length of its sequence exceeds 16 000 characters; "the tail" of the Y chromosome, inheritance in the "direct male line", see [Sykes, 2003, p. 290] etc.; for human, the the length of its sequence exceeds 50 000 characters; MHC, the main (major) histocompatibility complex, see [Lennarz and Lane (Eds) , 2013] etc.; usually, we cannot say about its length. "The structure of MHC allows to bind peptides of varying lengths because both ends of the peptide are free . . . ", see ibid. "The MHC complex encodes the α-chains of the MHC class I molecules human leukocyte antigen (HLA)-A, HLA-B, and HLA-C and the αand β-chains of the MHC class II molecules HLA-DR, HLA-DP, and HLA-DQ, all of which are expressed in a co-dominant fashion." "MHC class I is expressed by all nucleated cells and platelets in jawed vertebrates, although the amount on the cell surface varies among cell types and under different inflammatory conditions." "The folded MHC class II molecule consists of two transmembrane proteins, an α-chain and a β-chain, which together form a protein . . . Peptides bound by MHC class II molecules typically are longer than 10 amino acids and occasionally more than 20 amino acids." However, such a small number of variants (less than 10 algorithms and 3 objects of research) does not negate the need to create effective algorithms for processing DNA chains, in particular, constructing (for one of these variants) a matrix of distances between such chains. At the same time, the distances between DNA sequences are often used in scientific and popular science literature. However, as we already said, there are several different algorithms for calculating them, and for each pair of sequences, the operation of any of these algorithms takes a lot of time. For example, the practical programming results show that on an average modern computer, it takes about a day to build such a 30 × 30 matrix for mtDNAs using the Needleman -Wunsch algorithm [Needleman and Wunsch, 1970] ; therefore, for such a 300 × 300 matrix, about 3 months of continuous computer operation is expected. Such dimensions come from real problems: for example, in the class of mammals there are about 30 orders, in the order of primates there are about 20 families, more than 80 genera and more than 500 species. At the same time, the exact values differ in different classi-fication options, but they are not interesting to us: we are interested in approximate values only. Thus, even for a relatively small number of species (smaller than the total number of primate species), calculating the distance matrix on conventional computers is hardly feasible; the use of supercomputers, firstly, is not always possible, and, secondly, it often requires significant revision of existing software. In this regard, the task of restoring such a partially filled matrix arises. We started publishing our variants of similar algorithms for restoring partially filled matrices in [Melnikov, Pivneva, and Trifonov, 2017] (the simplest algorithm was described very briefly there), after which we returned to this problem in [Melnikov and Trenina, 2018a; Melnikov and Trenina, 2018b] , where a variant of the algorithm using the method of branches and boundaries was described in detail. Remark. At the same time, it is worth noting that in the last two papers we made computational mistakes, which, however, did not affect the overall assessment of the results of calculations given in these papers at all. For example, it was said that we leave about 30-35 % of the elements in the matrices of dimension about 30 × 30, obtained for processing, while we left a significantly smaller number of elements, i.e., about 10 %; thus, we solved a more difficult task. Certainly, the latter fact indicates much greater possibilities in the application of the algorithms we are considering. In addition, some computational errors were made when obtaining obtaining the values of σ and δ, which, again, did not affect the evaluation of the calculation results. Despite the previously successful results of calculations, we return in this paper to the algorithm variants that do not use the branches and boundaries method: as our recent work has shown (primarily in the subject areas related to graph theory and the development of ultra-large communication networks, [Melnikov and Terentyeva, 2021] etc.), even greater improvement in the quality of the algorithm can often be achieved without improving the auxiliary heuristics of the branches and boundaries method. We are improving the algorithms that formulate the greedy function of this method only; however, these algorithms can also be called auxiliary to the method of branches and boundaries. In this paper, we describe a similar improvement of the greedy algorithm, now for the task of reconstructing the matrix of distances between DNA chains. As the obtained results of computational experiments show (see Section 5), they are better than ones obtained using simple variants of the branches and boundaries method. Let us repeat that for restoring partially filled matrices, i.e., for the inverse problem of matrix processing, we used the method of branches and boundaries before, but in this paper, we do not use it. In connection with the above, the question arises about the concept of "partially filled matrix": how to determine this partial filling. It is clear that the greater the percentage of values will not be calculated, the less time will be spent on these calculations: after all, as follows from the above estimates, the calculation of one value for two considered mt DNA s requires about 3 minutes of computer operation, which is approximately equal to the total time required for matrix recovery even using the long-running method of branches and boundaries. On the other hand, a too small percentage of the values left in the matrix (i.e., calculated by the special previous algorithms), of course, cannot give adequate results; in this regard, we have been using in this work the percentage of values calculated by the algorithm of about 10-12 %. The second important question that cannot but arise on the basis of the above text is how exactly we can analyze the quality of the solution obtained using the recovery algorithm(s). For more information, see Section 5 below. The paper has the following structure. In Section 2, we consider a brief description of the greedy algorithms of restoring the distance matrix. In Section 3, we give the theoretical substantiation of the possibility of improvement of the greedy algorithm (without the variants of the method of branches and boundaries). In Section 4, we formulate two possible quality criteria for the numerical solution of such restoring problems: the first criterion compares the matrix reconstructed by the simplified algorithm under consideration with the matrix obtained as a result of applying a general formation algorithm for each of its elements; and the second criterion considers the discrepancy in a special way, it applies the same algorithms that are used as auxiliary ones of the general recovery algorithm considered in this paper. In both cases, the goal is to reduce the values obtained by the applied criterion. In Section 5, we give some results of computational experiments; we evaluate these results well. In Section 6, we formulate some problems for the future solution. And in Conclusion (Section 7), we briefly repeat the content of this paper. 2 A brief description of the greedy algorithms of restoring the distance matrix It is clear that with smaller dimensions of the matrix, a larger percentage of non-deleted elements is required. Thus, for small dimensions (of the order of 10), algorithms often do not work with a small number of nonremovable elements. Of course, in principle, there are options when, under the conditions we have given (i.e., about 10 % of non-removable elements with matrix dimensions of about 30), it is impossible to restore the matrix: for example, when an empty line is obtained. When calculating using the elementary formulas of probability theory, we find that the probability of this event is about 25 %. However, we simply do not consider such examples as source data. We considered the simplest heuristic for filling in the distance matrix without using the method of branches and boundaries in [Melnikov and Trenina, 2018a] . (We note in advance that the results of the computational experiments given in that paper will be compared below with newer results using other heuristic algorithms.) Further, as we have already noted, our publications were devoted to the application of the branches and boundaries method; but in this paper, we again abandon it. At the same time, we complicate the greedy heuristics of [Melnikov and Trenina, 2018a] . Let us briefly describe the greedy matrix filling algorithm used in this paper. First of all, we choose an element that, if filled in, forms the largest number of newly formed triangles; i.e., for n×n-dimensional distance matrix (m i,j ), we choose the pair (i, j) such that: m i,j < 0 (in our natural notation, this means that there is no corresponding matrix element in the input data); and the following formula is achieved: max 1⩽i,j⩽n i̸ =j 1⩽k⩽n k̸ =i, k̸ =j sgn(m k,i ) + sgn(m k,j ) . If there are several such elements, we choose any of them. Next, we consider all the resulting triangles, and minimize the total value of the badness. One of the variants of such an assessment of badness for one triangle is the formula α -β γ , where α, β and γ are the angles of the derived triangle, and α ⩾ β ⩾ γ. (Note that earlier we sometimes used another formula, a-b c , where a, b and c are the sides of the derived triangle, and a ⩾ b ⩾ c. At the same time, in both cases, if the three sides do not satisfy the triangle inequality, we assumed a large value as the value of badness, usually from 1.0 for the case a=b+c to 2.0 for "absolutely impossible" triangles.) The total value of badness is always (i.e., both for choosing a value in the described algorithm and for a posteriori evaluation of the quality of the algorithm) considered simply as the sum of the values of badness of all triangles. In the described algorithm, we are trying to minimize this badness value for all newly formed triangles. The minimization method is given in the next section, where the justification for the possibility of piecemeal filling of the matrix is given, to obtain a value of badness close to optimal (i.e., in terminology of [Melnikov et al., 2018c] , "to obtain a pseudo-optimal solution"). Simplifying it, we can say that that, taking the average values of the maximum sides of the formed triangles (i.e. α in previous formulas; note that in [Melnikov and Trenina, 2018a] , this value was counted final) as the beginning of the iterative process, we get a pseudo-optimal value in a few iterations. (We usually limited the number of iterations to 10, this gives an acceptable value of the calculation time.) 3 The theoretical substantiation of the possibility of improvement of the greedy algorithm Thus, in this paper we reconstruct the matrix of distances between DNA sequences of different species of organisms. We shall restore the distance function of the matrix and find its derivative. The main problem is that it is an ill-posed problems, which means that a small error in the source data can lead to a large error in the calculated derivatives [Tikhonov, 1963; Groetsch, 1984; Hanke and Scherzer, 2001; Chaikovskii and Zhang, 2022] , etc. If we introduce the coordinate axes x and y, located in the horizontal and vertical axes, respectively, and consider the matrix as a two-dimensional array with noisy data defined on the domain Ω ∈ R 2 , then the matrix elements will represent the noisy values of the function of two variables u δ i,j . It is natural to assume that the domain Ω is divided into N ⩽ n 2 parts {Ω i } N i=1 , and there is the only one value u δ i,j in each parts. Denote d i as the diameter of Ω i and let d = max{d i }. In this case, we obtain the deterministic model max\|u(x i , y j ) -u δ i,j \| ⩽ δ between the noisy data {u δ i,j } and the corresponding exact values {u(x i , y i )} at grid points X n := {x 1 < x 2 < < x n } and Y n := {y 1 < y 2 < < y n }. Let us suppose that the reconstructed function u ε (x, y) is formed according to the following optimization problem: u ε = arg min v∈C 1 (Ω)   1 n 2 n i=1 n j=1 v(x i , y j ) -u δ i,j 2 + ε ∂ 2 v(x, y) ∂x 2 2 L 2 (Ω) + ∂ 2 v(x, y) ∂y 2 2 L 2 (Ω) , (1) where v(x, y) is a cubic spline, and regularization parameter ε satisfies the expression 1 n 2 n i=1 n j=1 u ε (x i , y j ) -u δ i,j 2 = δ 4 . Then by [Wang and Wei, 2005, Theorem 3 .3], the following proposition holds. Proposition 1. Let u(, ) ∈ H 2 (Ω). Let u ε (x, y) be the minimizer of the problem (1). Then for ε = δ 2 , we have ∥u ε (, ) -u(, )∥ H 1 (Ω) ⩽ C 1 d 1/4 + C 2 √ δ, where C 1 and C 2 are some constants depending on the area Ω and on ∥∆u(x, y)∥ L 2 (Ω) . Based on Proposition 1, we obtain the following fact. For sufficiently small values d and δ, after solving the optimization problem (1) values ∂u ∂x , ∂u ∂y can also be found with sufficiently high accuracy. We can do it taking derivatives ∂u ε ∂x , ∂u ε ∂y . Due to the fact that with an increase in the value of the norm ∥∆u(x, y)∥ L 2 (Ω) the value of constants C 1 and C 2 will increase, we conclude that the smoother the function u(x, y) is, the smaller this norm will be, and the more accurate the regularization result will be. If the function is not smooth enough, we shall need more noisy data to obtain the necessary accuracy. Another smoothing technique can be used is the convolution, see [Gulliksson et al., 2016, Theorem 3 ] and [Lin, Cheng, and Zhang, 2018, Section 4] for details. It also follows from the proposition 1 that if we set δ → 0, the error of restoring the function will depend mainly on the diameter d, which is the larger, the more missing data {u δ i,j } at the grid points. And due to the fact that 65 % of data is missing in the task we have set, we need to introduce additional conditions to restore the function. One of such conditions is the regularity found in the paper [Melnikov, Pivneva, and Trifonov, 2017] for distance matrices, which consists in the fact that the three elements of the matrix (m i,j , m k,j , m k,i ) form the sides of an isosceles triangle. Thus, the formulas below reduce such a metric to a function of several variables, and a "triangular" norm for determining the quality of the distance metric can be introduced, which can be represented in the following way. For the matrix M = (m i,j ) and its elements m i,j , we always 1 suppose that i, j, k ∈ { 1, 2, . . . , n } and do not consider diagonal elements (i.e., elements m i,i and the arithmetical expressions with these elements are ignored in formulas). The total error σ is defined as follows: n-1 i=1 n j=i+1 σ i,j , where one of the calculation variants is the sequential usage of the following formulas: r (1) i,j,k = max m i,j , m k,j , m k,i , r (2) i,j,k = min (m i,j , m k,j , m k,i ), and σ i,j is as follows: max 1⩽k⩽n k̸ =i,k̸ =j 2r (1) i,j,k + r (2) i,j,k -m i,j -m k,j -m k,i r (2) i,j,k . Then the original problem can be reformulated into the problem of minimizing the error value σ (as we already said, it often was called "badness" in our previous papers) by piecemeal filling in the missing elements. It is very important that we fill missing elements in the table sequentially, piecemeal, "step by step"; thereby we greatly simplify the implementation of the corresponding algorithm. Filling the table in this way, we obtain a matrix with noisy data {u δ i,j } and then we restore the u ε function by solving (1). The level of noise δ generated by this algorithm for restoring missing values can be estimated by analyzing the results of violations of the "isosceles triangle" regularity in [Melnikov, Pivneva, and Trifonov, 2017] . Thus, by sequentially filling in the missing elements of the matrix, we can guarantee a consistent improvement of the resulting solution, which theoretically justifies the possibility of abandoning the branch and boundary method, which works much more longer than the greedy algorithm for obtaining the value of one element considered here. 4 Quality criteria for the numerical solution of the problem As we said before, an important question arises, is how exactly can we analyze the quality of the solution obtained using the recovery algorithm(s). However, the above model of calculations does not give a complete answer to the question of the quality of matrix restoring. Therefore, the simplest quality criterion would be a comparison (by some natural metric) of the reconstructed matrix and the actually obtained distance matrix, which we can obtain for some examples of a small dimension; however, it is obvious that such a comparison can be made only a limited number of times, probably during the initial debugging of the algorithms. Therefore, in this section we formulate two possible quality criteria for the numerical solution of such restoring problems: (1) the first criterion compares the matrix reconstructed by the simplified algorithm under consideration with the matrix obtained as a result of applying a general algorithm for the formation of each of its elements, like [Melnikov and Trenina, 2018a; Melnikov and Trenina, 2018b] ; we shall denote by σ the value of this criterion; (2) and the second criterion considers the discrepancy in a special way, it applies the same algorithms that are used as auxiliary ones of the general recovery algorithm considered in this paper; we shall denote by δ the value of this criterion (or by d in some previous papers). In both cases, the goal is to reduce the values obtained by the applied criterion. The exact formulas are as follows. (1) For σ, we usually set σ = 2 n (n-1) n-1 i=1 n j=i+1 (m i,j -m i,j ) , where all the elements m i,j are obtained by applying the original algorithm (for instance, already cited Needleman -Wunsch algorithm), i.e., without restoring any elements. Note that for obvious reasons, we cannot often use this method, and also we cannot apply it for large matrices obtained by some distance determination algorithms; therefore the following criterion δ can be called more universal. (2) For δ, we usually set δ = n-2 i=1 n-1 j=i+1 n k=j+1 δ i,j,k (we specifically note once again, that the values m i,j are not used here). Each value δ i,j,k (where (2b) If a ⩾ b + c (i.e., the triangle inequality is violated 2 ), we choose in advance a constant ω (usually, ω = 2) and set 1 ⩽ i, j, k ⩽ n, i ̸ = j, i ̸ = k, j ̸ = k) is δ i,j,k = min a b + c , ω . (2) (2c) Otherwise, for usual triangle, we count its angles; let they be α, β and γ, where α ⩾ β ⩾ γ. (2d) Then we set δ i,j,k = α -β γ . Let us especially note that δ, unlike σ, is calculated quickly, despite we need to consider ∼ n 3 triangles. Table 1 . The result of the recovery program using a greedy algorithm, which is some more complicated than the algorithm of [Melnikov and Trenina, 2018a] . In both cases, the branches and boundaries method is not used. Like [Melnikov and Trenina, 2018a] σ = 0.038, δ = 0.044 Let us also note the relationship of both of these criteria with the task we are considering: for example, for a "random" matrix, we obtained significantly worse results of calculation by the criterion δ, even for small dimensions; to say, for such a matrix of dimension 13 × 13 we obtained δ in limits about 0.4 -0.5, this is several times higher than the corresponding values for the "correct" matrices of dimension 28 × 28 and significantly lower percentage of initial fullness, see the next section. Some results of computational experiments First of all, let us give a few comments on the given large tables of results. Both are given for the reader's possible verification of these results; at the same time, anyone can either simply recognize the table as a picture, or request from the authors, after which we shall send the same tables in the form of text. Having these tables, anyone can simply check their characteristics (badness, etc., according to the formulas given in the paper). It is possible to say, simplifying a little, that the topic of the article is how to obtain the missing values in these tables with minimum badness. As we have already noted, we evaluate the results of computational experiments well. Namely, an improvement in the performance of the algorithm was obtained (according to both criteria given in the previous section), compared with the simplest variant of the branch and boundary method, see [Melnikov and Trenina, 2018a; Melnikov and Trenina, 2018b] etc. More precisely, by the words "simplest variant", we mean that the simplest greedy heuristic used to select the next separating element. Here we apply a more complex greedy heuristic, while abandoning the method of branches and boundaries. It is clear that even more successful results (from the point of view of the quantitative criteria formulated above) we would have obtained by using both the branch and boundary method and a more complex greedy heuristic at the same time; however, it seems that we shall not satisfy acceptable time constraints. Though, we did not conduct detailed computational experiments for this case. To perform all the computational experiments described in the article, we used a computer with the following characteristics: Intel(R) Core(TM) i7-8700 CPU @ 3.20GHz In all our computational experiments, the total time of the computer was extremely short, and we did not record it, since it is significantly less than the time required to output the results of the algorithm (especially in comparison with the time necessary for the initial filling of only one cell of the matrix). For comparison, we repeat once again: such calculations by the method of branches and boundaries for dimensions of the order of 30 × 30 take about 1 second; the calculation of the distance by the Needleman -Wunsch algorithm between two sequences describing mtDNA takes about 3 minutes; and the filling of the entire matrix by the Needleman -Wunsch algorithm of the order of 30 takes about 1 day. Thus, let us briefly describe the computational experiments already carried out. In this paper, only one variant of the input data is used (28 monkeys, mt DNA, we briefly talked about this variant above); but we note in advance that similar results were obtained on all variants of the input data used. The initial matrix filled in as a result of the Needleman -Wunsch algorithm is given in [Melnikov and Trenina, 2018a, Tab. 8 ]. The initial matrix with about 10 % of remaining elements is given in [Melnikov and Trenina, 2018a, Tab. 9 ]. Let us remind once again that the calculations of the matrices in that paper were correct, but the calculations of the values σ and δ were erroneous. However, the mistakes did not affect the relative quality indicators of the algorithms. In the current paper, we present completely correct results, they are easy to check. The column designations in Table 3 are clear, and the row designations have the following meaning: (A) corresponds to the matrix, obtained by the best algorithm of [Melnikov and Trenina, 2018a] (that matrix was given on [Melnikov and Trenina, 2018a, Tab. 13] ); let us remind once again that the algorithm does not use branches and boundaries method; (B) corresponds to the matrix, obtained by the best algorithm of [Melnikov and Trenina, 2018b] (that matrix was given on [Melnikov and Trenina, 2018b, Tab. 17] ); the algorithm uses branches and boundaries method; (C) corresponds to the matrix, obtained by the simplest greedy algorithm of the current paper; the algorithm does not use branches and boundaries method; see its results (i.e., the obtained matrix and its characteristics) in The most successful values of σ and δ are highlighted in bold. A brief discussion of the obtained results is given in Conclusion, Section 7. Some possible directions for further work on this subject Let us formulate some problems for the future solution, i.e., consider a brief description of the nearest directions for further work related to the modification and improvement of the described algorithms. For the first possible direction, we shall temporarily assume that the original processed algorithm under consideration (the Needleman -Wunsch algorithm is one of the possible examples only) is obviously not optimal, and requires improvement. Note that this fact, of course, is always true, regardless of everything else: for example, almost all the original algorithms (i.e., determining the distance between two given DNA sequences) are based on the long-known algorithm for constructing the Levenshtein metric (or Levenshtein distance), [Levenshtein, 1966] , to which the "penalties" are additionally added. At the same time, the numerical values of such penalties are always selected based on preliminary expert assessments (or even simply assumed to be equal to 1), [Christen, 2012a; Christen, 2012b; Yu et al., 2016; Suganthan et al., 2018] . However, it is clear that any selfstudy procedure should give an improvement of such values (penalties, etc.). Therefore, another inverse problem arises: to achieve such an improvement, where the minimization of the value δ is used as the criterion. In contrast to the first possible direction briefly described before, in other variants for further work we consider the given algorithm (the Needleman -Wunsch algorithm etc.) as something "God-given", i.e., not as the subject to change; we try, as above in this paper, to tune in to it. The second direction. We introduce a special "reliability coefficient" (let it be R < 1, to say, R = 0.9 in the following description), which we use as follows. We consider that the initial values of the matrix (in the example considered in the paper, the remaining 10 % of the elements after the removal) have a weight of 1.0. The elements derived from only the initial ones (i.e., in the beginning of filling) have a weight of R. And in the general case (i.e., after filling in some elements) we proceed as follows. As in the greedy algorithm already discussed in this paper, we form all possible triangles obtained together with the element selected for filling, i.e. if the considered unfilled element of the matrix is m i,j , then, as before, we consider all such k that m i,k and m j,k are already filled. However, we calculate the obtained values with the reliability coefficients already assigned to these values, i.e., we minimize the general function, which includes values with these coefficients; for the reliability coefficients R i,k and R j,k , we assume that the reliability coefficient of the considered triangle is R ∆ = R i,k + R j,k 2 . (3) The resulting value obtained as a result of minimization is placed in a matrix with its new reliability coefficient equal to the a priori value of R multiplied by the average reliability coefficient of all considered triangles that form the element m i,j : using (3) and assuming we are considering m triangles, this new coefficient can be written as follows: R (1) ∆ + R (2) ∆ + + R (m) ∆ R m . And, of course, the best value of the reliability coefficient R should be obtained as a result of some selflearning process. The third direction. Here we propose to continue the simultaneous application of increasingly complex greedy heuristics and the method of branches and boundaries. We hope that the results will surpass those obtained in this paper and in [Melnikov and Trenina, 2018a; Melnikov and Trenina, 2018b] . At the same time, we note that almost no time is spent on the method of branches and boundaries, at least in comparison with the time that needs to be spent on filling in only one initial element of the matrix. The fourth direction. This is a detailed study of the quality of algorithms depending on the dimension of the matrix and the percentage of its initial fullness. (Note that we have not actually started solving this problem yet.) The fifth direction. We believe that it is possible to choose which elements of the matrix should be initially filled, of course, within a predetermined total number of them. This possibility sometimes reflects the subject area under consideration: after all, the total time of such filling will practically not depend on specific elements, but will depend on their number only. Is it true that in this case, the elements for the initial filling should be selected so that there would be approximately the same number of them in all rows of the matrix? (Answering this question is the fifth possible direction of work.) The sixth direction represents a new approach to comparing different heuristics for distances between DNA chains, i.e., an alternative approach to the one discussed in [Melnikov, Pivneva, and Trifonov, 2017] and some of our other papers cited there. Namely, after filling in several distance matrices with various algorithms (i.e., algorithms for obtaining distances between pairs of DNA chains), we consider all possible triangles in these matrices; note again that their number is quite large, of the order ∼ n 3 . Next, for each initial filling algorithm, we consider a list of these triangles, ordered, for example, by non-increasing values of the badness (considered, for example, as δ i,j,k , see ( 2 )). The main idea of this heuristics is that we assume that all the algorithms described in the literature and on the Internet for obtaining distances between pairs of DNA chains are logically correct. Therefore, considering some "natural" metric on such ordered sequences of triangles, for the "best" algorithm for the initial filling of matrices (the best algorithm for calculating the distance between pairs of DNA chains), we obtain the minimum value of the sum of the distances to other ordered sequences of triangles. As such a natural metric on ordered sequences of triangles, we can choose some natural function from the pairwise correlation between the sequences. In our preliminary calculations, we choose a linear function as such one: having a value 0 in the case of matching sequences; having a value 1 (the maximum possible value) in the case of the maximum possible number of the minimum number of exchanges required to convert from the first sequence to the second one; note that for a matrix of the order n × n, the number of its triangles is ∼ n 3 , therefore the number of possible exchanges is ∼ n 6 ; and intermediate values otherwise; these values are calculated, as we already noted, using the simplest linear function. Note that our version of the pairwise correlation is obtained here with another version of the linear function, i.e. when simultaneously replacing 0 with 1 and 1 with -1 in the items above. After that, we propose to consider "pairwise correlation between pairwise correlations": for this, we need to arrange the heuristic algorithms for the initial placement of DNA chains in two ways (i.e. according to [Melnikov, Pivneva, and Trifonov, 2017] and according to the above). The seventh direction. And of course, as a possible direction for further work, it is necessary to consider new objects of application of the described algorithms: other species (besides monkeys), Y-chromosomes instead of mt DNA s, other initial filling algorithms (instead of Needleman -Wunsch) . . . Besides, for monkeys, we propose to consider a very strong increase in dimension (it is optimal to consider all types, 500 +) with a simultaneous decrease in the percentage of initially filled matrix cells (to say, to 5 % instead of 10 %). Of course, these seven directions do not limit further possible work on the topics described here . . . Conclusion In this article, we propose an improved algorithm for recovering missing data in distance matrices. An unusual feature of the results is that the reduction of the quality criteria σ and δ used in several of our articles by different algorithms occurs independently of each other, i.e., there is a binary relationship between the algorithms, which consist in the fact that the first of them gives the best results for both criteria, forms a partial order only. Similar relative results are obtained for other objects of study (not for monkeys etc.). Of course, the best option would be the simultaneous minimization of σ and δ, which we shall achieve in the subsequent work. However, the results presented in this paper are of big interest. Acknowledgement This work was supported by the Natural Science Foundation of Guangdong Province (No. 2019A1515110971) and Shenzhen National Science Foundation (No. 20200827173701001). the "badness" of corresponding triangle; it is usually counted in the following way.(2a) Firstly, we rename m i,j , m i,k and m j,k into a, b and c, where a ⩾ b ⩾ c. Table 2 . 2 , the value of the table 315 corresponds to the obtained value 0.315, etc. The result of the recovery program using a complicated greedy algorithm. The branches and boundaries method is not used. 0 313 258 334 328 341 334 326 505 324 344 421 315 334 335 335 262 334 332 337 334 997 506 329 338 325 325 344 313 0 330 313 303 342 323 222 505 320 340 420 312 313 334 334 297 328 331 332 292 997 504 269 336 316 320 342 258 330 0 335 343 339 333 320 505 327 347 422 319 335 334 334 296 332 331 343 341 997 505 335 337 273 309 344 334 313 335 0 329 340 292 330 505 327 344 420 328 311 334 334 332 322 333 333 257 997 503 333 335 333 330 344 328 303 343 329 0 343 332 236 505 330 341 417 328 329 336 336 325 333 334 304 327 997 505 329 338 320 328 343 341 342 339 340 343 0 341 341 505 341 351 421 341 340 341 341 341 341 341 344 337 997 506 343 341 341 341 345 334 323 333 292 332 341 0 332 505 318 344 421 326 329 335 335 334 279 334 337 331 997 499 334 331 334 332 344 326 222 320 330 236 341 332 0 505 329 344 420 326 330 336 336 322 333 333 333 330 996 504 325 337 298 324 344 505 505 505 505 505 505 505 505 0 505 505 505 506 505 505 505 505 505 505 505 505 996 523 505 505 505 505 505 324 320 327 327 330 341 318 329 505 0 344 421 302 348 354 354 328 350 351 337 333 996 508 333 354 332 347 360 344 340 347 344 341 351 344 344 505 344 0 419 344 359 361 361 344 361 360 326 341 996 509 332 360 344 359 268 421 420 422 420 417 421 421 420 505 421 419 0 420 427 428 428 420 428 427 315 419 996 514 420 427 420 427 428 315 312 319 328 328 341 326 326 506 302 344 420 0 355 361 361 297 358 357 337 332 996 509 332 360 330 353 366 334 313 335 311 329 340 329 330 505 348 359 427 355 0 361 361 355 358 359 356 264 999 678 355 361 356 357 367 335 334 334 334 336 341 335 336 505 354 361 428 361 361 0 301 358 357 264 362 361 997 534 401 300 401 364 367 335 334 334 334 336 341 335 336 505 354 361 428 361 361 301 0 358 355 321 362 361 997 534 401 256 401 364 367 262 297 296 332 325 341 334 322 505 328 344 420 297 355 358 358 0 359 326 337 331 996 509 330 342 327 354 366 334 328 332 322 333 341 279 333 505 350 361 428 358 358 357 355 359 0 356 361 359 996 287 360 265 359 359 367 332 331 331 333 334 341 334 333 505 351 360 427 357 359 264 321 326 356 0 360 358 997 534 399 310 398 361 367 337 332 343 333 304 344 337 333 505 337 326 315 337 356 362 362 337 361 360 0 303 996 509 336 361 336 359 365 334 292 341 257 327 337 331 330 505 333 341 419 332 264 361 361 331 359 358 303 0 997 509 330 360 333 357 366 997 997 997 997 997 997 997 996 996 996 996 996 996 999 997 997 996 996 997 996 997 0 995 997 997 997 997 997 506 504 505 503 505 506 499 504 523 508 509 514 509 678 534 534 509 287 534 509 509 995 0 509 531 509 533 535 329 269 335 333 329 343 334 325 505 333 332 420 332 355 401 401 330 360 399 336 330 997 509 0 401 332 395 404 338 336 337 335 338 341 331 337 505 354 360 427 360 361 300 256 342 265 310 361 360 997 531 401 0 401 397 401 325 316 273 333 320 341 334 298 505 332 344 420 330 356 401 401 327 359 398 336 333 997 509 332 401 0 244 405 325 320 309 330 328 341 332 324 505 347 359 427 353 357 364 364 354 359 361 359 357 997 533 395 397 244 0 402 344 342 344 344 343 345 344 344 505 360 268 428 366 367 367 367 366 367 367 365 366 997 535 404 401 405 402 0 σ = 0.079, δ = 0.103 0 313 258 332 328 341 332 326 505 317 344 317 286 332 330 330 262 332 329 333 332 997 332 329 332 270 270 344 313 0 330 284 245 342 284 222 505 315 340 316 311 284 317 285 297 280 321 302 292 997 273 269 286 316 315 342 258 330 0 334 343 339 332 320 505 320 347 340 283 334 330 330 269 332 329 343 341 997 332 335 331 273 264 344 332 284 334 0 285 340 292 283 505 325 344 329 325 253 322 285 328 282 325 293 257 997 276 283 285 327 327 344 328 245 343 285 0 343 284 236 505 326 341 323 325 280 323 282 322 277 326 304 295 997 273 267 281 320 326 343 341 342 339 340 343 0 341 341 505 341 351 332 341 340 341 341 341 341 341 344 337 997 341 343 341 341 341 345 332 284 332 292 284 341 0 282 505 318 344 318 322 280 324 283 327 279 327 292 287 997 276 281 281 327 329 344 326 222 320 283 236 341 282 0 505 324 344 314 324 276 324 278 321 276 326 297 290 996 271 262 277 298 323 344 505 505 505 505 505 505 505 505 0 505 505 505 506 505 505 505 505 505 505 505 505 996 505 505 505 505 505 505 317 315 320 325 326 341 318 324 505 0 344 313 302 314 309 307 324 315 311 330 329 996 310 329 311 325 312 327 344 340 347 344 341 351 344 344 505 344 0 325 344 326 360 360 344 328 359 326 341 996 355 332 356 344 359 268 317 316 340 329 323 332 318 314 505 313 325 0 311 308 310 307 310 311 311 315 309 996 308 306 310 306 311 323 286 311 283 325 325 341 322 324 506 302 344 311 0 309 306 305 297 310 308 330 330 996 306 329 308 291 276 354 332 284 334 253 280 340 280 276 505 314 326 308 309 0 304 266 337 268 308 300 264 999 282 292 268 336 309 353 330 317 330 322 323 341 324 324 505 309 360 310 306 304 0 301 308 303 264 304 301 997 300 299 300 331 307 355 330 285 330 285 282 341 283 278 505 307 360 307 305 266 301 0 335 264 321 297 291 997 289 288 256 330 335 355 262 297 269 328 322 341 327 321 505 324 344 310 297 337 308 335 0 338 326 330 329 996 338 328 332 264 272 354 332 280 332 282 277 341 279 276 505 315 328 311 310 268 303 264 338 0 306 300 294 996 287 291 265 336 336 353 329 321 329 325 326 341 327 326 505 311 359 311 308 308 264 321 326 306 0 307 305 997 304 303 310 334 305 355 333 302 343 293 304 344 292 297 505 330 326 315 330 300 304 297 330 300 307 0 303 996 300 293 300 327 336 353 332 292 341 257 295 337 287 290 505 329 341 309 330 264 301 291 329 294 305 303 0 997 294 288 293 327 336 354 997 997 997 997 997 997 997 996 996 996 996 996 996 999 997 997 996 996 997 996 997 0 995 997 997 997 997 997 332 273 332 276 273 341 276 271 505 310 355 308 306 282 300 289 338 287 304 300 294 995 0 265 294 336 330 349 329 269 335 283 267 343 281 262 505 329 332 306 329 292 299 288 328 291 303 293 288 997 265 0 294 327 330 347 332 286 331 285 281 341 281 277 505 311 356 310 308 268 300 256 332 265 310 300 293 997 294 294 0 334 328 346 270 316 273 327 320 341 327 298 505 325 344 306 291 336 331 330 264 336 334 327 327 997 336 327 334 0 244 344 270 315 264 327 326 341 329 323 505 312 359 311 276 309 307 335 272 336 305 336 336 997 330 330 328 244 0 344 344 342 344 344 343 345 344 344 505 327 268 323 354 353 355 355 354 353 355 353 354 997 349 347 346 344 344 0 Table 1 1 of this paper; (D) corresponds to the matrix, obtained by the compli- cated greedy algorithm of the current paper; the algorithm does not use branches and boundaries method; see its results in Table 2 of this paper; Certainly, all the calculation results shown in this table can be quickly checked using a simple supportive com- puter program. Table 3 . 3 General results of some computational experiments T σ δ (A) 0.091 0.110 (B) 0.029 0.133 (C) 0.079 0.103 (D) 0.038 0.044 I.e., when considering summation, as well as when taking minima and maxima. In some our previous papers, we wrote that the number of such violations ranged from 2 (the minimum value in previous calculations) to several dozen (for matrices about 30 × 30); it depended on the subject area, as well as on the specific algorithm. In the case under consideration here (28 species of monkeys of different genera, mt DNA, Needleman -Wunsch algorithm), we obtained no such violations at all, then the item (2b) was not used.
10.5888/pcd15.180026	cc-by\|public-domain	null	null	openalex	Background A growing body of literature shows that health promotion and disease prevention strategies and messages may not be effective in reaching racially and ethnically diverse communities, unless those strategies are culturally and linguistically adapted for target communities (1) . The Racial and Ethnic Approaches to Community Health for Asian Americans (REACH FAR) project is a multilevel, evidence-based program of health promotion and disease prevention for Asian American communities in New York and New Jersey. Guided by a socio-ecological framework, social marketing principles, and a community based participatory approach, the project implemented multilevel, evidence-based strategies culturally adapted to address hypertension and improve access to healthy food options for Asian Americans in various community settings (1) . The strategies were delivered through a multisector coalition made up of a lead academic agency, New York University Center for the Study of Asian American Health; the NYC Department of Health and Mental Hygiene (NYC DOHMH); 4 community-based organizations: UNITED SIKHS, serving the Asian Indian population; Diabetes Research, Education, and Action for Minorities, serving the Bangladeshi community; Kalusugan Coalition, Inc., serving the Filipino community; Korean Community Services of Metropolitan New York, Inc., serving the Korean community; and other organizations and groups serving these communities. Coalition partners implemented strategies at various community sites, including faith-based organizations, ethnic restaurants, grocery stores, pharmacies, and primary care practices. REACH FAR has 2 program arms to deliver culturally and linguistically adapted resources. The first arm is focused on improving access to healthy foods and beverages and includes implementation of policies adapted for communal meals at faith-based organizations, healthy food options and labeling at restaurants, and strategic placement and discounting of healthy food products at grocery stores (2) . The second arm is focused on improving access to hypertension management and cardiovascular disease prevention programs by offering the NYC DOHMH's Keep on Track (KOT) blood-pressure monitoring program at faith-based organizations and increasing access to Million Hearts (https://millionhearts.hhs.gov/) blood pressure medication adherence resources at faith-based sites, community-based pharmacies, and health care providers' offices (3). The REACH FAR coalition leveraged existing partnerships and fostered new relationships with implementation sites to tailor and disseminate strategies for their respective communities. Geographic areas with high concentrations of targeted Asian American communities were selected at project conception. Hence, a primary motivation for creating our map was to assess the population and geographic reach of the REACH FAR project by comparing Asian American population clusters (ie, geographic areas with high concentrations of Asian Indian, Bangladeshi, Filipino, and Korean populations) with the locations of implementation sites and participants. A desire lines (or spider diagram) approach was used to visualize KOT program participation. Mapping provides a visual means of evaluating program reach and is a concise way of communicating the role of implementation sites in delivering culturally, linguistically, and geographically targeted health prevention strategies. Methods We geocoded the addresses of 12 implementation sites active as of 2016 and residential addresses of KOT program participants reported in baseline surveys from 2015 and 2016 (4-6). All research conducted with human participants was reviewed by the NYU School of Medicine Institutional Review Board and approved as an expedited study. Data processing was performed with version R version 3.4.1 (R Foundation for Statistical Computing). Straight lines were created between successfully geocoded participants' residences (n = 587; 86% of participants) and faith-based organizations (n = 12); each line represents a KOT program participant (7) . Population estimates from Census 2010 Summary File 1 data and household language from ACS 2015 5-Year estimates were obtained for Census tracts in NY and NJ (8, 9) . Processed data were exported as Esri shape files (10) . The map construction and geo-processing related to population clusters were completed with ArcGIS 10.4.1 (Environmental Systems Research Institute [Esri]). The Asian American subgroup population clusters are significant hot spots, positive z scores, determined by the ArcGIS Hot Spot Analysis (Getis-Ord Gi*) tool (Esri) with the Contiguity Edges Corner and Apply False Discovery Rate correction parameters; the shading indicates intensity of clustering with darker shading indicating more clustering of high values (larger z scores). Overlaps in population clusters were found by using the ArcMap Intersect tool, version 10.3 (Esri). Main Findings An overarching objective of REACH FAR is delivery of targeted strategies for 4 Asian American communities. The map demonstrates that program implementation sites are located in population clusters of targeted Asian American subgroup populations, as intended at program inception: Asian Indian in Middlesex and Queens counties, Bangladeshi in Kings and Queens counties, Filipino in Queens county, and Korean in Queens and Bergen counties. By mapping KOT program participants, we found that large concentrations of congregants residing in areas neighboring each faith-based site are reached by the KOT program. However, the maps also demonstrate that the programs also reached community members residing outside of immediate or neighboring ethnic enclaves. We suspect that people attending these events prefer or need culturally and linguistically adapted resources that are unavailable in the neighborhoods in which they reside. Culturally and linguistically adapted materials regarding healthy eating and blood pressure control developed for limited English-proficient Asian Americans were disseminated at implementation sites (11), reaching 1,353,201 people as of September 30, 2017. The map reveals opportunities for collaboration, the areas where population clusters overlap, and gaps in coverage in areas with population clusters where the project did not have implementation sites. PREVENTING CHRONIC DISEASE Action Our map provides a visual illustration of the network of faithbased organizations and community-based organizations coordinating to promote healthy eating and heart health and the progress of implementation as illustrated with the KOT program. To this end, mapping products are being used in the following ways to enhance future coordination and collaboration between partners: The map is being presented to the NYC DOHMH to demonstrate the need for reaching Asian American populations. Before our coalition efforts, the KOT program had not been implemented in any faith-based organizations serving Asian Americans. Our results demonstrate both success and potential for future NYC DOHMH engagement efforts with Asian American communities. 1. The map is being reviewed at coalition meetings to discuss opportunities for scaling the program to expand program reach. For example, the map suggests that partnering with community-based organizations serving the Asian Indian and Filipino communities in Hudson County or the Bangladeshi community in Bronx County may expand reach. 2. opinions expressed by authors contributing to this journal do not necessarily reflect the opinions of the U.S. Department of Health and Human Services, the Public Health Service, the Centers for Disease Control and Prevention, or the authors' affiliated institutions. Centers for Disease Control and Prevention www.cdc.gov/pcd/issues/2018/18_0026.htm
10.7759/cureus.8736	cc-by	null	null	openalex	Background Shared governance is considered a model for mounting autonomous decision making in nursing profession and practice. This study aimed to assess how registered nurses in an outpatient department in a tertiary care hospital perceive shared governance. Methods We conducted a cross-sectional study among a convenient sample of registered nurses in an outpatient department. A self-administered, Index of Professional Nursing Governance (IPNG) questionnaire was used to measure the study outcome. A descriptive analysis was used to describe nurses' characteristics and study outcomes. Results A total of 186 nurses completed the questionnaire. Of whom, 151 (92.1%) were female, and 78 (47.3%) were aged between 20 and 30 years. Only 54 (29.3%) and 59 (31.7%) had indicated a shared decision in terms of controls and influence scales, respectively. The majority of the nurses indicated traditional shared across shared governance scales except in the access information scale. Conclusion The findings showed a prevalent traditional nursing management style in the study setting. Supportive strategies and education must be provided for both managers and staff nurses to develop and implement shared governance in their practice.Introduction In a complex health care system, nurses are a vital component in providing optimal patient quality care and promote outcomes. Nursing nowadays facing multifaceted challenges, like shortages in the workforce, increased workloads, and patient acuities that have reshaped the attention to the quality of nursing care while fulfilling increased patient needs and demands. In addition to these challenges, nurses are astounded with increased rules that inflate their work profile and load, increase job dissatisfaction, and decrease bedside time spent with the patient. Although these challenges increased nurses' responsibility and accountability to accomplish professional practice, it was not escorted with increasing power or authority to figure out required changes to impact nursing practice . Therefore, healthcare organizations have developed several professional practice models to direct disciplines' clinical practice, empowering and authorizing health staff, and improve provided quality of care [2, 3] . The nursing profession has encouraged and supported nursing involvement in nursing practice models such as shared governance . Nurse participation in shared governance guarantees accountability for the safety and quality of care and the autonomy of the nurses . Shared governance has been defined as "a decentralized approach which gives nurses a greater authority and control over their practice and work environment; engenders a sense of responsibility and accountability; and allow active participation in the decision-making process, particularly in the administrative area from which they were excluded previously" . Governance generally comprises structure and process where a group of individuals direct, control, and regulate their goal-oriented efforts . Shared governance as an official program should include nurses in control and authority decisions by attaining the right to control their practice and extending their governance beyond that to higher and different tasks like scheduling, and evaluating personnel, budgeting, that were historically controlled solely by managers . Moreover, shared governance is a model that includes shared decision-making between the healthcare workforce members and is centered on the principles of partnership, equity, accountability, and ownership [7, 8] . Shared governance structure supports patient care directly, promotes nurses' control over their practice and accountability for quality patient care . For nurses, it results in feelings of empowerment that allow for professional autonomy. The application of shared governance structure leads to enhance the provision of quality of care [4, 9] , cultivate collaboration between healthcare professionals, advance the quality of care and clinical effectiveness; upsurge employees confidence; assist in developing individual and professional skills; growth of professional profile; which in turn lead to improve personal communication; advance knowledge and skills; intensification of professionalism and accountability; reduce duplication of effort [4, 10, 11] . Nurses' satisfaction is a vital end-point of shared governance. Investigating shared governance among nurses at King Fahad Medical City is important because of the influence of attention on human factors in nursing and maintain workforce retention. Therefore, this study aimed to assess the perception of shared governance among nurses in a tertiary care hospital, Riyadh, Saudi Arabia. Materials And Methods Study design and settings A cross-sectional study was conducted to assess perceptions of shared governance levels among nurses. A convenient sampling method was used for recruiting the nurses working in ambulatory care in a tertiary care hospital, Riyadh, Saudi Arabia. All ambulatory care nurses were invited to participate in the study by an invitation letter attached to a copy of the questionnaire and cover sheet describing the study's objective and voluntary participation. Participants' identity was kept anonymous. Sample characteristics A total of 186 nurses included in the study. The majority of participating nurses 151 (92.1%) were females, and 148 (92.8%) were staff nurses. About 78 (47.3%) of the nurses were aged between 20 and 30 years. One-hundred and thirty (79.9%) had a bachelor degree, while 25 (15.2%) had a nursing diploma. Nearly half of the nurses had six-ten years of experience, and 119 (72.5%) had one-five years of experience in the current institution (Table 1 ). TABLE 1: Participants' demographics Table 2 displays the descriptive statistics for the six scales of the IPNG scale. Regarding the "Control" scale, the results showed that slightly more than two-thirds (68.5%) of the nurses (n=126) perceived traditional decision making and 54 (29.3%) had indicated a shared decision making. Furthermore, 122 (65.6%) of the nurses perceived traditional decision making in the "Influence" scale compared to 59 (31.7%) who perceived a shared decision making. The results of "Participation" scale indicated that more than half of the nurses (54.1%) perceived having a traditional decision making. Moreover, the "Ability" scale showed a traditional decision making among 110 (59.5%) of the nurses. Nevertheless, 72 (39.9%) of the nurses revealed the "Ability" scale is governed by the shared decision making. ( Discussion Overall, the study showed a traditional management decision making as indicated by the nurses. The distributions of the percentages of nurses' responses were all more or less asymmetrical and concentrated to the left side of the chart, indicating a traditional model of governance. The most proportioned distribution that makes relatively somewhat balance between nurses' shared governance model, and the traditional model of governance was for the participation scale. Overall, the findings reflect the administrative driven model of governance at the study site, which means that the governance and management type had no effect on the perception of staff nurses and that nurses managers are working in an environment that is not equally sharing decisions and not enabling staff nurses to control over their practice. In the control scale, the results indicated that nurses had limited control over their practice. The findings were inconsistence with the similar studies reported in the literature which revealed that nurses' perceptions of their work setting are more related to the shared governance model [12, 13] . These studies indicated that nurses and administration were equally involved in decision-making activities concerning their control over professional practice [12, 13] . Relating to the influence over practice and resources, participating nurses perceived they lack influence or formal authority in several daily procedures, including patient care locations, obtaining and monitoring supplies, patients' admissions, and discharges, creating new clinical and administrative positions, and generating schedules. The results disagree with Hashish et al., who reported that nurses indicated high influence over practice and resources . Along with the findings of the current study align with Seada and Etway who showed that nurses perceived various parts of their influence over different activities are being done only with an administrator decision with limited staff inputs . The findings of our study showed that nurses perceived a lack of shared ability with nursing management to engage in nursing profession committees mostly concerning their clinical practices, staff scheduling, and strategic planning. Moreover, nurses perceived they have limited ability to play a part in committees that relating multidisciplinary professionalism, organizational expenses, and budgets. George et al. revealed that nurses from a non-shared governance hospital had less engagement in decision making compared to nurses in a shared governance model, which comes in line with our results . However, several studies have reported results supporting nurses' ability to participates in nursing profession committees [12, 13] . The findings of this study were consistent with results of a study conducted by Tourangeau et al. ; they stated that nurses perceived they have the least extent of control over professional practice; as well as, they perceived slight contribution or control in several areas that affect directly the patients' bedside care from nursing, quality care standards, educational progress, and determining the structure of nursing care for their work. Also, Seada and Etway revealed that nurses had lowermost scores regarding their perception of shared governance which showed that they did not have control over their professional work setting . Shared governance is characterized by collaborative decision-making between management and staff working, together at the organizational level and unit level . The values of shared governance have been united in nursing structures to providing a transformational structure for staff nursing care and enhancing an organization's productivity and performance . The three essential values related to shared governance are the responsibility of delivering for nursing care should belong to clinical staff, nurses' authority for being renowned by the organization, and quality patient care accountability and nursing professionalism must be acknowledged by the clinical staff . Nursing shared governance denotes to the played nurses' role in decision-making and liability for patient care . Shared governance affords the context for a cooperative milieu of nursing leaders and nurses. Both, they can frame a partnership of shared decision for operational and clinical practices . Worldwide, nurses are the prime profession in healthcare settings. Therefore, participation and influence in shared governance are necessary to enable constant transformation, advance and progress of the nursing profession. Bringing this to reality, fitting models, and processes that deliver and inspire participation in decision-making are indispensable . Limitations Limitations of this study include the employment of basic descriptive statistics. A further bivariate and multivariate analysis could be considered in the future. Other acknowledged limitations are the relatively modest sample size and limited study site; therefore, the findings may not be representative of the entire. A prospective study with a bigger sample size and multi-settings will be powerful to generalize the data. Conclusions The results showed that the studied setting lacks a shared governance model in place to engage nurses in decision-making which can enable them to control their professional practice. The findings of our study showed a traditional management governance type in the study setting. The study findings can be utilized to enhance nurses' work milieu and improve shared governance. These results would be an added value for nursing top management to enhance nurses' perception of shared governance by emerging or implementing shared governance model. In the light of our findings, nurse managers must adopt and apply strategies to empower nurses like shared governance that provide nurses the chance to control their nursing practice and improve quality nursing care. Table 2 ) 2 Factor Traditional (management) decision Shared decision making Nurses are the decision maker scales making (86-172) n(%) (173-344) n(%) (345-430) n(%) Controls 126 (68.5) 54 (29.3) 4 (2.2) Influence 122 (65.6) 59 (31.7) 5 (2.7) Official authority 124 (67) 56 (30.3) 5 (2.7) Participation 100 (54.1) 81 (43.8) 4 (2.1) Access information 74 (40.2) 105 (57.1) 5 (2.7) Ability 110 (59.5) 72 (39.9) 3 (1.6) TABLE 2 : Sub-scales of nursing shared governance 2 Kaddourah et al. Cureus 12(6): e8736. DOI 10.7759/cureus.8736
10.5194/isprs-annals-v-3-2022-131-2022	cc-by	null	null	openalex	Calculating solar-sensor zenith and azimuth angles for hyperspectral images collected by UAVs are important in terms of conducting bi-directional reflectance function (BRDF) correction or radiative transfer modeling-based applications in remote sensing. These applications are even more necessary to perform high-throughput phenotyping and precision agriculture tasks. This study demonstrates an automated Python framework that can calculate the solar-sensor zenith and azimuth angles for a push-broom hyperspectral camera equipped in a UAV. First, the hyperspectral images were radiometrically and geometrically corrected. Second, the high-precision Global Navigation Satellite System (GNSS) and Inertial Measurement Unit (IMU) data for the flight path was extracted and corresponding UAV points for each pixel were identified. Finally, the angles were calculated using spherical trigonometry and linear algebra. The results show that the solar zenith angle (SZA) and solar azimuth angle (SAA) calculated by our method provided higher precision angular values compared to other available tools. The viewing zenith angle (VZA) was lower near the flight path and higher near the edge of the images. The viewing azimuth angle (VAA) pattern showed higher values to the left and lower values to the right side of the flight line. The methods described in this study is easily reproducible to other study areas and applications.INTRODUCTION Remote sensing has proved to be highly effective and efficient in studying a diverse variety of natural and ecological resources. Other than satellite and aerial remote sensing, recent advances in Unmanned Aerial Vehicles (UAV) and sensor technology has opened more opportunities to study vegetation dynamics, specifically in agricultural applications (Maddikunta et al., 2021) . Since UAVs can be flown at lower altitudes than satellites or aircrafts, the resulting products offer higher spatial resolution and with more accurate canopy spectra (Tao et al., 2020) . The canopy spectra can be used to model or represent different plant traits. For instance, different vegetation indices (e.g., normalized difference vegetation index, NDVI) can indicate overall crop health that improves precision agriculture practices (Radoglou-Grammatikis et al., 2020) . Additionally, UAV sensors offer highthroughput plant phenotyping that accelerates current crop breeding operations (Song et al., 2021) . Moreover, UAV-based imageries can be used to train advanced machine learning models, which predict various crop traits, disease, yield, and seed quality at plot-level (Bhadra et al., 2020; Maimaitijiang et al., 2020; Nguyen et al., 2021) . Hyperspectral sensors can collect reflected spectra from crop canopies with higher spectral resolution. A typical hyperspectral image (HSI) often contains hundreds or even thousands of bands for a wide range of wavelengths. Generally, the wavelengths can vary from Very Near Infrared (VNIR, 400-1000 nm) to Shortwave Infrared (SWIR, 900-2500 nm) with different spectral resolution (1 to 10 nm). Plants reflect electromagnetic radiation, which contains information about their biophysical composition and physiological status (Segarra et al., 2020) . Numerous studies * Corresponding author have utilized the broader range of HSI products to study different characteristics of plants and vegetation (Mariotto et al., 2013; Fernandes et al., 2015; Banerjee et al., 2020; Wang et al., 2021) . The quality of HSI-based inference heavily depends on the accuracy of HSI post-processing techniques. Generally, the HSI sensor provides the raw Digital Number (DN) or radiance (in Wsr -2 m -2 ), which is then converted to unitless top-ofatmosphere (TOA) reflectance and surface reflectance (SR). Empirical Line Method (ELM) is the widely used calibration technique to convert radiance into SR using different calibration targets on the ground (Markelin et al., 2008; Wang and Myint 2015; Ortiz et al., 2017) . The principal assumption behind this technique is that the objects on the ground represent a Lambertian surface, which appears uniformly bright from all directions of view and reflects the entire incident light (Mao et al., 2020) . However, the crop canopy architecture is far from being a Lambertian surface and exhibits anisotropic effects (Jiao et al., 2014) . Therefore, several studies have identified that multiple viewing angles or viewing geometry of sensors play an important role in the pixel-level SR (Vermote et al., 2009; Zhang et al., 2014) . For example, Galvao et al., (2009) retrieved highly accurate Vegetation Indices (VIs) from Hyperion and MODIS satellite data when using backward observations. Similarly, Gu et al., (2015) found improved Leaf Area Index (LAI) estimation accuracy from backward observations compared to forward observations in CHRIS/PROBA data. Huang et al., (2011) demonstrated that multi-angular hyperspectral observations could retrieve the vertical distribution of chlorophyll content in winter wheat. In terms of UAV-based observations, several studies have utilized snapshot (or frame) hyperspectral cameras to derive multi-angular spectral information. For instance, Roosjen et al., (2018) achieved improved results in estimating LAI and chlorophyll content of potato by using multi-angular data. They introduced a goniometer-based simulation method for HSI footprints with high overlaps. The multiple viewing angles were converted to zenith and azimuth angles, which were used to simulate PROSAIL spectra and derive better LAI and chlorophyll retrieval accuracy. Similarly, Mao et al., (2020) found that the effect of multi-angular observations was significant in deriving VIs for soybean and maize. They also extrapolated different viewing angles from a snapshot hyperspectral camera mounted with a UAV and corrected for the Bi-directional Reflectance Function (BRDF) effect. Therefore, the availability of multiple solar-sensor zenith and azimuth angles is highly important to accurately study different plan characteristics. Alternative to snapshot cameras, push-broom hyperspectral sensors (or line-scanner sensors) are now widely used with UAVs. The push-broom sensor captures one line per exposure that forms one image line after the other (Barreto et al., 2019) . Therefore, push-broom sensors can outperform snapshot cameras, as the latter systems require a compromise between spatial coverage, spatial resolution, and spectral resolution (Aasen et al., 2015; Yi et al., 2021) . However, extracting the solar-sensor zenith and azimuth angles from a push-broom sensor is not as straightforward as snapshot cameras. While the snapshot camera provides 2D scenes captured across overlapping flight lines in relatively higher time interval, the push-broom sensor captures line by line 1D spectra across its flight path. Due to the line-by-line scanning mechanism, push-broom sensors suffer from wind-related motions during data acquisition (Jaud et al., 2018) . As a result, push-broom sensors require high accuracy Global Navigation Satellite System (GNSS) and Inertial Measurement Unit (IMU) onboard the UAV to ortho-rectify the lines and generate a geometrically accurate hyperspectral cube (Yuan and Zhang 2008) . Due to the availability of GNSS/IMU system onboard the platform, the solar-sensor geometry can be directly calculated using linear algebra and spherical trigonometry. Therefore, the objective of this study is to develop an automated framework that can calculate the solar-sensor zenith and azimuth angles for each pixel in a hyperspectral cube collected by a push-broom UAV scanner with cross-grid flight pattern. STUDY AREA AND DATASETS Experimental Setup The experiment was setup in the Planthaven Farms at OFallon, Missouri, United States (Figure 1 ). The site was located slightly northeast from Saint Louis city close to the Mississippi River to the north. The field was planted with 220 rows of maize on May 25, 2021, where 2 rows were marked as one plot. Total 55 different genotypes or cultivars of maize were planted with 2 replicas. The field was approximately 75 m long and 20 m wide. During the growing season, average temperature was between 23-24C and average annual precipitation was 1092.2 mm for the study area. UAV Flight A DJI M600 Pro UAV was used to collect the hyperspectral data for the study area (Figure 1c ). The UAV was equipped with a Headwall Nano-Hyperspec VNIR push broom camera (Headwall Photonics, Massachusetts, United States), a FLIR Vue Pro thermal camera (FLIR Systems, Oregon, United States), and an APX-15 GNSS/IMU (Applanix Corporation, Ontario, Canada) unit all attached to a DJI Gimbal (Figure 1d ). The APX-15 UAV GNSS/IMU records the precise time, position, and orientation of the sensor at 200 Hz interval. Full specifications of the sensor and the GNSS/IMU unit is provided in Table 1 . Two UAV flights were conducted on July 20th and August 4th of 2021. Each flight was planned in a cross-grid pattern (Figure 1f ) in UgCS mission planning software (v4.0.187, SPH Engineering, Latvia) with 4 length wise and 9 width wise lines, resulting in total 13 hyperspectral cubes. The altitude and velocity for both flights were 50 m and 3 m/s. The ground sampling distance (GSD) was found 3.01 cm from both flights. METHODS The extraction of solar-sensor geomtery contain three major parts (Figure 2 ): 1) hyperspectral cube processing, 2) locating viewing point for each pixel, and 3) calculating solar-sensor geometry. The process was automated using Python libraries which are available in a public repository with test datasets (https://github.com/remotesensinglab/uav-solar-sensor-angle). Locating Viewing Point for Each Pixel The viewing point in terms of each pixel was required to calculate both sensor zenith and azimuth angles (Figure 2b ). First, the GNSS data was extracted from APX-15 device and converted to an ASCII text file which contained latitude, longitude, and timestamp information. Additionally, the coordinates of each pixel was calculated by converting the raster data into a geospatial text file using GDAL v3.3.1 (GDAL/OGR 2020). The raster image also had longitude, latitude, and timestamp information. Therefore, only the corresponding GNSS observations for a HSI cube were filtered by matching the timestamp from the cube and GNSS points. Finally, the GNSS points which had the shortest distance from each pixel location were identified by representing the point pairs in a matrix form. It was done by calculating Euclidean distance from each pair of pixels and GNSS coordinates. The information was preserved in a text file as comma-separated value (CSV) format, which contained the unique ID of the closest GNSS point for every pixel in the HSI cube. Solar-Sensor Angle Calculation Overview of solar zenith angle or viewing zenith angle (VZA, θ V ), solar zenith angle (SZA, θ S ), sensor or viewing azimuth angle (VAA, φ V ), and solar azimuth angle (SAA, φ S ) calculations in terms of cross-grid UAV with push-broom hyperspectral sensor are illustrated in Figure 2c . Solar Zenith Angle (SZA): The solar zenith angle (SZA) is similar to the VZA, but instead of the sensor as the moving vector, the position of the sun becomes the point of interest. SZA is a function of the raster location coordinates (longitude and latitude) and time of the day. SZA (θ S ) can be calculated from Solar Elevation Angle (α S ) using Equation 3 . θ S = 90 o -α S ( The pixel coordinates were attached with corresponding sensor points and each sensor point included the time information in UTC format. Also, the coordinates were converted from UTM to a geographic coordinate system (World Geographic System 1984) , so the values were available as latitude and longitude. A python package called PVLIB (v0.9.0) was used to calculate δ, h and eventually θ S for all pixel coordinates. Solar Azimuth Angle (SAA): Solar azimuth angle (SAA) is a function of time and coordinate for each pixel location and can be calculated using Equation 8 . φ S = cos -1 [ sin(δ) cos(φ)-cos(δ) sin(φ) cos(h) cos(α S ) ] (8) where φ, δ, h and α S are the latitude, solar declination angle, hour angle and solar elevation angle, respectively. Viewing Zenith Angle (VZA): The viewing zenith angle (VZA) is the angle between the vector from sensor and raster point (VR ⃗⃗⃗⃗⃗⃗ ), and the surface normal (Z ⃗ ⃗ ) from the raster point (also known as zenith), which can be defined as θ V . The UAV was flown at a 50 m altitude for the whole mission. Therefore, a perpendicular vector from the sensor point to the XY surface can be drawn as VV ⃗⃗⃗⃗⃗⃗ , where VV ⃗⃗⃗⃗⃗⃗ is 50 m. The angle between VR ⃗⃗⃗⃗⃗⃗ and RV ⃗⃗⃗⃗⃗⃗ is known as Viewing Elevation Angle (α V ) and can be calculated using Equation 1 . α V = tan -1 VV ⃗⃗⃗⃗⃗ RV ⃗⃗⃗⃗⃗ = tan -1 50 √(x v -x r ) 2 +(y v -y r ) 2 (1) where the coordinates of R and V are (x r ,y r ) and (x v ,y v ), respectively, calculated in a Universal Transverse Mercator (UTM) projection system. Therefore, RV ⃗⃗⃗⃗⃗⃗ can be calculated as the Euclidean distance between R and V ́. If α V is known, then θ V can be calculated using Equation 2 . θ V = 90 o -α V (2) For every pixel coordinate, corresponding θ V values were calculated and converted to degrees. Viewing Azimuth Angle (VAA): Azimuth angle can be calculated in the XY plane, where it is the clockwise angle between a point of interest and the true north (Y ⃗⃗ ). For calculating the sensor or viewing azimuth angle (VAA, φ V ), an arbitrary north vector for any pixel coordinate (x r ,y r ) was created by adding 100 m to y r (Figure 2c2 ). The VAA can be calculated using Equation 7 . φ V = cos -1 [ a ⃗ b ⃗ \|a ⃗ \|\|b ⃗ \| ] ( 7 ) where a is the true north vector and b ⃗ is the vector between the raster point and corresponding sensor point. RESULTS AND DISCUSSIONS The SZA (θ S ), SAA (φ S ), VZA (θ V ), and VAA (φ V ) were calculated for all 13 HSI cubes, but we will discuss only the 1 st HSI cube (Figure 3 ). Additionally, the descriptive statistics of the angles are provided in Table 2 . Table 2 . Descriptive statistics of the angles for the 1 st HSI cube. Solar Angles The SZA and SAA shows different angular pattern in the resulting rasters (Figure 3b and 3c ). The SZA started decreasing along with the flight direction, whereas the SAA started to increase along with the flight direction. The total duration for capturing this HSI cube was 42.32 seconds, which resulted in low standard deviation for the angular values of SZA and SAA (Table 2 ). The verification of our calculation of solar angles was done by calculating the solar position based on National Oceanic and Atmospheric Administration (NOAA) Solar Calculator (NOAA 2021), which is an online tool to calculate approximate solar position in terms of coordinates and local time. To verify our result with the NOAA Solar Calculator, 5 randomly selected points were selected, and corresponding solar angles were extracted. Table 3 shows the SZA and SAA calculated based on NOAA Solar Calculator and our method, and the absolute differences observed. Calculator and our method. Abs. Diff. indicates the absolute differences between two methods and T is the order of points mentioned in Figure 3 . The SZA and SAA values calculated by NOAA Solar Calculator and our method showed slight differences at the decimal level. Since we used highly accurate PVLIB Python library to calculate the solar angles, we could provide coordinates and time information up to any decimal level possible. For instance, the time information in our method had 6 decimal places for second values. On the other hand, the NOAA calculator could only take the second values as integer. Moreover, NOAA (2021) indicates that due to the variations in the atmospheric conditions and uncertainty in the algorithms, there could be slight differences in the solar position calculations. These could be attributed to the slight differences in the solar angle values. However, having precise coordinate and time information in the angle calculation is highly preferable for remote sensing applications, specifically in HSI-based processing. Sensor Angles The pattern of VZA (Figure 3d ) can be explainable in terms of the flight path. Since the flight path runs through the middle of the cube, the VZA values are close to zero near the flight path and starts increasing at the edge of the image. Since this is zenith angle from the sensor, there should be higher angles at the edge rather than the middle. The VAA shows comparatively larger range of angular values (Figure 3e ). Since azimuth angle is calculated as the clockwise angle from the north vector of each raster pixel, the right side of the cube resulted with smaller angular values, whereas the left side comprised of higher values. However, the pattern in the VAA raster may seem binary, but the inset map on Figure 3e shows an enlarged portion of the right side. The inset map shows that there exists angular variation along the flight path and the variation can be seen perpendicular to the flight line. Therefore, the standard deviation is the highest for VAA with larger range of angular values (Table 2 ). Limitations The major issue encountered in this study was the lack of camera calibration. We used the vendor provided software (Headwall Spectral View v3.1.4) to ortho-rectify the HSI cubes. However, when the cubes were plotted in a GIS environment, it was noted that the overlapping regions from two consecutive HSI cubes did not exactly match. Therefore, the HSI cubes were georeferenced with a Light Detection and Ranging (LiDAR)-derived RGB point cloud using 6 control points. The LiDAR mission was also flown on the same days the HSI missions were performed. The LiDAR UAV point cloud was corrected using a GNSS base station established during the data collection time and the vendor provided software named, Phoenix LiDARMill (v2.0, Phonix LiDAR Systems, Texas, United States). After correction, the position accuracy was around ±0.1 cm. When the HSI cubes were georeferenced with the RGB point cloud, corresponding cubes matched properly with each other. However, the problem can be solved by performing a camera calibration. Probably the internal operating parameters (IOPs), boresight angles or the lever-arm offsets were changed from the initially approximated values by the vendor. LaForest et al., (2019) performed similar camera calibration technique to perform time-delay adjustment to similar type of HSI UAV platform that had accurate GNSS/IMU information. The same methodology can be applied to our study to improve upon the ortho-rectification of the HSI cubes as well. Therefore, careful considerations should be made when working with push-broom sensors and calibration flights should be conducted using randomly placed ground control points (GCPs) on the ground. CONCLUSIONS The study demonstrates a simple methodology for calculating solar-sensor zenith and azimuth angles for a push-broom hyperspectral sensor equipped in a UAV. The results show that the method can deliver all the angles in raster format, which can be very helpful to perform BRDF corrections or radiative transfer model-based applications in remote sensing of vegetation. If this work is needed to be reproduced for other study areas using similar sensor and platform, then it will be easy to do so by utilizing the automatic Python workflow developed from this work. In future, we will improve the camera calibration issue incurred in this study and apply these angle rasters to perform radiative transfer modeling-based applications for plant phenotyping. ACKNOWLEDGEMENT This work has been supported in part by NSF/USDA (2020-67021-31530), NASA (80NSSC20M0100), USGS AmericaView Grant (G18AP00077) and NSF Plant Genome Research Program (1733606). Figure 1 . 1 Figure 1. Study area and data collection instruments, (a) a RGB image of the maize field, (b) a close-up view of the field marked with a yellow box in (a), (c) the DJI M600 Pro UAV equipped with sensors, (d) a close up view of the sensor package which includes a Headwall Nano-Hyperspec VNIR camera, A FLIR Vue Pro thermal camera and APX-15 GNSS/IMU unit all attached in a DJI Gimbal, (e) the location of the study area, and (f) the cross-grid flight pattern created in UGCS flight planning software (v4.0.187) which was used for data collection. Figure 2 . 2 Figure 2. Overview of methods for the extraction of solar-sensor zenith and azimuth angles, i.e., (a) hyperspectral cube processing involves converting digital number (DN) to radiance, reflectance, and ortho-rectified images, (b) locating corresponding sensor location for each pixel in a hyperspectral cube, and (c) calculating three solar-sensor zenith and azimuth angles. 3) sin α S = sin φ sin δ + cos φ cos δ cos h (4)where, φ is the latitude of the location, δ is the solar declination angle and h is hour angle. Declination angle (δ) is the angle between the line joining the centers of the Sun and the Earth and its projection on the equatorial plane. The value of δ can range from -23.44 to 23.44 and calculated using Equation5, where d is the number of days since the beginning of the year. Hour angle (h) is the position of the sun relative to solar noon and can be calculated using Equation6, where LST is the local solar time. Solar hour angle is 0 at solar noon and it increases by 15 after each hour. Figure 3 . 3 Figure 3. Resulting angles for the 1 st HSI cube, where (a) RGB true color composite, (b) solar zenith angle (θ S ), (c) solar azimuth angle (φ S ), (d) viewing zenith angle (θ V ), and (e) viewing azimuth angle (φ V ). The black dots (in b and c) are test points that were verified with alternative source. The inset map in the VAA (e) shows a small zoomed-up portion, which indicates the changing angular pattern in VAA. Table 1 . 1 Sensor specifications Sensor Specifications Headwall Wavelength (nm) 400 -1000 Nano Spatial bands 640 Hyperspec Spectral bands 269 VNIR Field of View () 50.684 Focal length (mm) 12 Dimension (mm) 76(L)×76(W)×119(H) Weight (g) 680 APX-15 Channels 336 GNSS/IMU Dimension (mm) 67(L)×60(W)×15(H) Weight (g) 60 Table 3 . 3 Comparison of solar angles between NOAA Solar This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper.https://doi.org/10.5194/isprs-annals-V-3-2022-131-2022 \| Author(s) 2022. CC BY 4.0 License. This contribution has been The double-blind peer-review was conducted on the basis of the full paper.https://doi.org/10.5194/isprs-annals-V-3-2022-131-2022 \| Author(s) 2022. CC BY 4.0 License.
10.1056/nejm188512241132607	public-domain	null	null	openalex	later a cyclitis developed which soon subsided. In December, Sn. 1 was read, though there were opacities in the vitreous. During the next six months Miss X.. was kept indoors by sciatica. She read and sewed at will, and later was able to exercise in her garden. September 14th. Feeling very well, she went out to drive at 11 a.m., (eleven months after the extraction.) The day was bright, and she said she looked pretty steadily at the various objects that came into view and with absorbing interest and enjoyment till her return at 1 p.m. On leaving her carriage everything around her suddenly took on a dark-pink color, and this redenvironment continued after she had entered her house, it persisted after the lamps had been lighted though of a different shade and this distressing red vision re- mained unaltered for six days. Although living but a few blocks away, she delayed consulting me till the 21st, (feeling doubtful, very likely, if she should get any relief?), when the light was less annoying, though all objects still seemed pink. The eye was not inflamed and the sight was better than in December. At the end of another week the erythropsia had entirely disappeared, having faded grad- ually away. As the patient has again been confined to her house, it is uncertain whether her trouble will return when she is again subjected to the full light of day. SOME FACTS BEARING ON THE \l=AE\TIOLOGY OF CARCINOMATOUS DISEASES. BY G. A. WHEELER, M.D., CASTINE, MAINE. The predisposing causes of cancer are stated by nearly all authors who have written about this disease, as unknown. It is usually admitted, however, that all causes which tend to lower the vitality of the system are among the number, and by many there is thought to be a kinship between tuberculous and carcinoma- tous affections. As phthisis undoubtedly bears a cer- tain relation to the humidity or dryness of the soil, the query is forced upon us whether carcinoma may not also have a similar relation either to the humidity or to the composition of the soil. I am led to these reflections by a consideration of the cases of malignant disease which were presumably cancerous, though not actually demonstrated to be such by the microscope, which have occurred within the last eighteen years, in the small village of Castine, Maine, and in a particular part of the town. In fact, nearly all on one particular street, or quite near to it. The few other cases mentioned, having at some previous time lived on or near this street. The soil of this portion of the village is superficially a sandy loam, beneath which is a layer of coarse gravel and still deeper a blue clay. All the drinking water upon the street is extremely hard but no analy- sis of it has ever been made. The street commences near the .summit of a hill about two hundred feet above the level of the ocean and slopes down to the sea. The drainage is excellent. In the accompanying plan of the village the houses where the cases about to be mentioned occurred, are numbered to correspond with those cases. 1. N. G., male, age about fifty. Lived in the house for many years. No hereditary predisposition known. Occupation, a ship carpenter. Autopsy showed seirrlius of the cardiac orifice of the stomach. Death from starvation. 2. J. N., male, age about fifty. Occupation, ship carpenter. No hereditary predisposition known. Had a tumor of the thigh for which he was admitted to the Massachusetts General Hospital, but for which noth- ing was done there. He afterwards went to the Bos- ton Homoeopathic Hospital, where he was operated upon, but the tumor returned, and he subsequently died, and the death was reported as from cancer. Had lived in this house for years. 3. Wife of the above. Age about forty-five. She was attended by the writer a few weeks before her death, which was due to malignant disease of the womb. 4. C. G, male. Age about fifty. Merchant. No hereditary tendency. Had occupied the house for years. He was operated upon unsuccessfully by the late Dr. Greene, of Portland, for scirrhus of the rectum. 5. S. H., female, single. Age sixty-nine. No he- reditary tendency. Had occupied the house for years. Had a scirrhus of the breast removed at the Maine General Hospital. Died from constitutional infection. Had also a large abdominal tumor. 6. J. II. N., male. Carpenter. Age about sixty. Had lived, in the house from childhood. His mother is said to have had a "sore" cancer of the face. Autopsy showed scirrhus of the pylorus. 7. B. K., female, widow. Age about fifty-five. Had lived in this house and in town only two years. No hereditary tendency known. Died from carcinoma uteri. Is supposed to have had the disease when she came to town, but this is a matter of doubt. Her sister died in the same house two years later of phthisis. 8. B., female, married. Age about sixty-live. No hereditary tendency. Had an abdominal tumor which from its hard, knobby feel I diagnosticated as scirrhus. Had marked cachexia when she died. No autopsy. Had lived in this house for years. 9. D. L., female, married. Age about sixty-five. No hereditary tendency known. Had resided in this house for years. Diagnosis, carcinoma uteri. Was under the charge of Dr. Stevens, now deceased. 10. M. G, female, widow. Age about eighty. No hereditary tendency. Had carcinoma mamma'. Had resided in this house for years. 11. F. G, female, married. Age forty-four. Cauli- flower excrescence of cervix uteri. Removed by Or. F. F. Sauger and myself. Disease extended and fin- ally destroyed recto-vaginal septum. No hereditary tendency. Had lived in this house twenty years. 12. G. T., male. Age fifty-five. Trader. Died from a malignant disease affecting the stomach, and probably involving other organs. No autopsy. No hereditary tendency, Had occupied this house for ten or twelve years but had lived in the one opposite 4 and 5, for a still longer period. 13. S. D., female, married. Age about forty-five. No hereditary tendency known. Died from what was diagnosticated by her attending physician as carcinoma uteri. She had lived in this house for several years. but had also lived many years in the one marked 5. Had a child die of phthisis. il. L. W., female, widow. Age thirty-seven. Mother died of heart disease, but hada suspicious tu- mor of the abdomen. This patient was operated upon by Dr. E. F. Sänger and myself for cauliflower excres- cence of cervix uteri. Is still living, but is likely, to succumb to the disease eventually. This patient has lived in this house for many years, but formerly lived two houses from the one marked G. According to common report there have been one or two other cases of cancer in this same locality, but I have been unable to verify the fact. I hayre one pa- tient on hand, however, who resides just above 4, who has had a polypus uteri removed, and who now has a large uterine tumor of some kind, the nature of which is still in doubt, but which I am fearful may prove to be carcinomatous. The surprising thing about these cases is that they should have been confined to such a limited territory. After a residence of fifteen years in this town I have not known and am unable to learn of any other cases of presumed carcinoma occurring in any other portion of this village, and but few cases have occurred in my practice elsewhere. If there be not some local cause for these cases it is remarkable that they should have occurred where they did. I am aware that so limited a number of cases can have no great weight towards establishing a local cause, but it seems to me they should have some, and they are given in the hopes of inducing other physicians to pay some attention to the localities in which their cases occur. Report s of Societies. Dr. Abbott, in opening the debate, said that some years since, he had a case of hydatid mole similar to that reported by Dr. Dunn. When he reached the woman, he found that the mass had come away entire, and consisted of a great number of cysts, varying from the size of a pea to a plum, and in quantity, he should judge, more than a pint. There was no trace of a foetus, and no haemorrhage followed the delivery. The mass was passed, he believed, early in pregnancy, and there were no bad symptoms before or after the delivery. This yvas the only case of uterine hydatids he had ever seen. Dr. Doe remarked that he had a case similar to that mentioned by Dr. Abbott. A large mass of cysts, suf- ficient to fill a six-quart pail, came away at about the seventh month. No foetus was discovered. De. A. E. McDonald mentioned the following case : Mrs. IL, aged twenty-five years, multípara ; family history^good ; no syphilis. She came under my care about ten years ago, and yvas treated with sat- isfactory results for uterine displacement and chronic en- do-mitritis. About a year afterward, she again consulted me, stating that she had not menstruated for a period of four months previously, but was then having slight sanguineous discharge, with occasional pain in the region of the uterus. Her abdomen appeared unusually large for that of a woman in the fourth month of ges- tation, but I thought she might have made a mistake as to the date of her conception. In view of her con- dition, I did not deem a critical examination advisable. Restin the recumbent position was advised, with medi- cine to relieve pain. I was summoned on the following day, and found her having quite strong expulsive pains at short intervals. My attention was called to the chamber-vessel which she had used a short time before my arrival, and 1 found in it about a gill of saccules attached to each other, and which had come away during one of her expulsive efforts while on the vessel. These saccules were about the size of small grapes, and semi-transparent, corresponding somewhat to Gooch's simile of " White currants floating in red cur- rant juice." My finger, on introducing it into the vagina, came in contact with a large, soft mass, which, to the feel, resembled a recent blood-clot, excepting that it was more spongy. I succeeded in removing the mass in its entirety, and found it to be a true mole of the vesicular or hydatidiform variety. I discovered no trace of a foetus in the mass, that having disappeared in the earlier stages of gestation. The whole quantity which came away would more than lili a pint measure. There was -no haemorrhage to contend with in the case, and the patient made a speedy and perfect recovery. Since then, she has given birth to two healthy children who are now living, aged respectively five and three The Boston Medical and Surgical Journal as published by The New England Journal of Medicine. Downloaded from nejm.org at NYU WASHINGTON SQUARE CAMPUS on June 24, 2016. For personal use only. No other uses without permission. From the NEJM Archive. Copyright 2010 Massachusetts Medical Society. SUFFOLK DISTRICT MEDICAL SOCIETY. SECTION OF OBSTETRICS AND GYNAECOLOGY. ROBERT B. DIXON, M.D., SECRETARY. November 18, 1885, Dr. James R. Ciiadwick in the Chair. Dr. W. A. Dünn read a paper entitled IIYDATIDIFORM MOLE OF THE UTERUS.1 . W. Driver remarked that during a large 1 See page 612. The Boston Medical and Surgical Journal as published by The New England Journal of Medicine. Downloaded from nejm.org at NYU WASHINGTON SQUARE CAMPUS on June 24, 2016. For personal use only. No other uses without permission. From the NEJM Archive. Copyright 2010 Massachusetts Medical Society.
10.1371/journal.pone.0230645	cc-by	null	null	openalex	Houttuynia cordata Thunb. has been used as a traditional medicine to treat a variety of ailments in Asian countries such as China, Japan, South Korea, and Thailand. In Thailand, H. cordata fermentation products (HCFPs) are commercially produced and popularly consumed throughout the country without experimental validation. Anti-inflammatory activity of H. cordata fresh leaves or aerial parts has previously been reported, however, the antiinflammatory activity of the commercially available HCFPs produced by the industrialized process has not yet been investigated. The aim of this study was to evaluate in vitro and in vivo anti-inflammatory potential of the selected industrialized HCFP. LPS-induced RAW264.7 and carrageenan-induced paw edema models were used to evaluate the antiinflammatory activity of HCFP. The phenolic acid components of HCFP aqueous and methanolic extracts were investigated using HPLC analysis. In RAW264.7 cells, the HCFP aqueous and methanolic extracts reduced NO production and suppressed LPS-stimulated expression of PGE 2 , iNOS, IL-1β, TNF-α and IL-6 levels in a concentration-dependent manner, however, less effect on COX-2 level was observed. In Wistar rats, 3.08 and 6.16 mL/kg HCFP reduced paw edema after 2 h carrageenan stimulation, suggesting the second phase anti-edematous effect similar to diclofenac (150 mg/kg). Whereas, 6.16 mL/kg HCFP also reduced paw edema after 1 h carrageenan stimulation, suggesting the first phase antiedematous effect. Quantitative HPLC revealed the active phenolic compounds including syringic, vanillic, p-hydroxybenzoic and ferulic acids, which possess anti-inflammatory activity. Our results demonstrated for the first time the anti-inflammatory activity of the industrialized HCFP both in vitro and in vivo, thus validating its promising anti-inflammation potential.Introduction Inflammation is a host response against infection, foreign stimulant and tissue injury. Although inflammation is a process of the immune response in our body, it can damage the body when it is out of control. While acute inflammation is a normal part of the defense response, chronic inflammation is a complex process stimulated by activating inflammation or immune cells. During the inflammatory process, many types of cells are activated, and these cells secrete various pro-inflammatory mediators, including cytokines (IL-1β, TNF-α, IL-6), nitric oxide (NO) and prostaglandin E2 (PGE 2 ) . Overproduction of inflammatory mediators leads to chronic inflammation, which can cause many diseases such as rheumatoid arthritis, cancer and allergies . During the inflammatory response, immune cells are also activated by adhesion molecules-activated signals to increase the migration capacity to inflamed tissue and finally to form heterotypic cell clustering between the immune cells, endothelial cells, and inflamed cells. Indeed, various inflammatory stimuli such as LPS and proinflammatory cytokines activate immune cells to up-regulate such inflammatory states [2, 3] . Hence, these cells are useful targets for developing new anti-inflammatory drugs and exploring the molecular anti-inflammatory mechanisms of a potential drug. Many drugs have been developed to treat inflammation and nociceptive symptoms however, undesired adverse effects of these clinical anti-inflammatory drugs have consistently evidenced . Non-steroidal anti-inflammatory drugs (NSAIDs) are normally used for the treatment of pain and inflammatory conditions, however, many NSAIDs are associated with undesired side effects including congestive heart failure, bleeding of the gastrointestinal tract and chronic kidney disease . Therefore, the search for alternative substitutes such as plantderived anti-inflammatory agents with the ease of availability and fewer side effects is urgently required to develop safe drugs for clinical use. Houttuynia cordata Thunb. is a perennial herbaceous plant mostly distributed in East Asia, and generally grown for local vegetable consumption in the North and Northeast of Thailand. H. cordata has been used as a medicinal plant possessing many biological properties including antioxidant, anticancer, and anti-inflammatory activities . As a traditional medicine in China, H. cordata has been used to treat ulcers, sores, heatstroke, diarrhea, and dysentery . In Korea, it has been used for the treatment of pneumonia, bronchitis, dysentery, dropsy, uteritis, eczema, herpes simplex, chronic sinusitis and nasal polyps . In Thailand, it has been used as an immunostimulant herb and anticancer agent . Nowadays, H. cordata is considered for a high-value industrial crop in Thailand and it has been fermented with probiotic bacteria to yield a H. cordata fermentation product (HCFP) commercially available. The microbial fermented herbal plant is a promising alternative source for many flavonoid molecules including anthocyanins, flavones and flavanones . Probiotics are microorganism exerting health-promoting functions in humans and animals , improving the nutraceutical value of the herbal plant products by breaking down undesirable phytochemicals, and producing certain desirable compounds . The fermentation process has increased the flavonoid content of H. cordata fermentation products conferring excellent anti-inflammatory effects in LPS-stimulated cells . Accordingly, many HCFPs have been commercially distributed and popularly consumed throughout Thailand. Previous studies reported that the industrial process caused a reduction in phenolic content of natural products [14, 15] , which may affect their biological properties. Anti-inflammatory activity of H. cordata fresh leaves or aerial parts has previously been studied [16, 17, 18] , however, the anti-inflammatory activity of HCFPs produced by industrialized process has not yet been investigated. Therefore, we aimed at investigating the anti-inflammatory activity of the industrial HCFP in LPS-stimulated RAW264.7 cells as well as its phenolic acid content to provide information for the general public or consumers. Here, we demonstrated the phenolic acid profiles and antiinflammatory activities of aqueous and methanolic extracts of the industrial HCFP Dokudami manifested by inhibiting the production of NO, PGE 2 and inflammatory cytokines such as TNF-α, IL-1β, and IL-6. Furthermore, the anti-inflammatory activity of this industrialized product was also confirmed using the rat paw assay. Materials and methods Materials The dietary supplement H. cordata fermentation product (HCFP), Dokudami, was obtained from the Prolac (Thailand) Co., Ltd., Lamphun, Thailand. The information on plant ingredient and serving suggestion of the HCFP were obtained from the label on its container. The major ingredients of this HCFP are composed of 99.3% (w/w) aerial parts of H. cordata and 0.7% (w/w) sugar cane powder. Serving suggestion is as follows: 5-15 ml twice a day in the morning before bedtime and before meal. H. cordata was cultivated by the Prolac (Thailand) Co., Ltd. in an organic farm in Chai Badan district, Lopburi province, Thailand. The fermentation product Lot no. 14/5/2015 was used throughout the study. RAW264.7 cells were obtained from Dr. Pramote Mahakunakorn, Faculty of Pharmaceutical Science, Khon Kaen University, Thailand. Male Wistar rats (250-300 g) were obtained from registered animal breeders, Nomura Siam International Co., Ltd., Bangkok, Thailand. LPS (E. coli 0111: B4) and diclofenac sodium were purchased from Sigma-Aldrich (St. Louis, MO, USA). Griess reagent for nitrite determination was purchased from Molecular Probes (Invitrogen, USA). All antibodies used in this study were purchased from Cell Signaling (USA). PGE 2 EIA was purchased from ANOVA (Taiwan). The ELISA kits for measuring cytokines (IL-1β, TNF-α, IL-6) were purchased from BioLegend (California). RPMI 1640 medium, fetal bovine serum (FBS), trypsin-EDTA and penicillin/streptomycin were obtained from Gibco/Invitrogen Crop. (Grand Island, NY, USA). Cell culture and animals RAW264.7 macrophage cells were cultured in RPMI 1640 medium with 10% fetal bovine serum (FBS), 1% penicillin and streptomycin and incubated at 37 ̊C in an atmosphere containing 5% CO 2 . Wistar rats were recovered from transportation for 1 week before the study. The rats were maintained at Northeast Laboratory Center, Khon Kaen University, Thailand. Details of animal welfare are as follows: Shelter: case size 37.5 x 48 x 18.5 cm (wide x length x high), with sterilized-wood shavings for bedding, Food: sterilized commercial food, Water: reverse osmosis (OR) with choline 3-4 ppm, Environment enrichment: social housing, free excess food, and water, Environment: temperature: 23±2 ̊C, humidity: 30-60% RH, dark: light cycle: 12:12 h, illumination: 350-400 Lux, ventilation: 10-15 ACH, noise: no exceed 85 Decibels. The experimental procedure was approved by the Institutional Animal Care and Use Committee of Khon Kaen University, based on the Ethic of Animal Experimentation of National Research Council of Thailand. The approval number was IACUC-KKU-101/60. Preparation of the lyophilized powder of HCFP aqueous extract To obtain polar phytochemical compounds, 50 mL of HCFP was centrifuged at 2,815 x g for 15 min and the supernatant (aqueous fraction) was filtered through Whatman grade No. 4 filter paper. The filtrate containing water-soluble constituents was lyophilized to obtain a lyophilized powder (aqueous extract). The extraction yield was 13.80 ± 0.57 mg/mL. The HCFP lyophilized powder was re-dissolved in double distilled water to obtain desired concentrations. Preparation of HCFP methanolic (phenolic-rich) extract To obtain both polar and nonpolar phenolic compounds in the free forms, 140 mL of methanol was added to 60 mL of HCFP and then the mixture was stirred for 2 h at room temperature. The filtrate was evaporated to 60 mL by rotary evaporator, then added with 60 mL of 2 M NaOH and stirred continuously for 12 h at room temperature. The mixture was centrifuged at 1,700 x g for 20 min and then filtered through Whatman grade No. 4 filter paper. The supernatant was repeatedly extracted three times with 80 mL of diethyl ether and the aqueous phase was collected and the diethyl ether phase was discarded. The aqueous phase was adjusted to pH 1.5 by 10 M HCl and filtered through Whatman grade No. 4 filter paper. The filtrate was extracted further with 80 mL of diethyl ether for three times, in which the portion of diethyl ether was collected. Sodium sulphate (Na 2 SO 4 ) anhydrous was used to dehydrate the diethyl ether phase, which was then filtered through the filter paper. A rotary evaporator was used to evaporate the filtrate to 5 mL, which was then finally evaporated to dryness under a gentle stream of nitrogen gas. The extraction yield was 5.36 ± 0.96 mg/mL. Cell viability assay The viability of RAW264.7 cells was determined colorimetrically using 3-(4,5-dimethylthiazolyl)-2-2,5-diphenyltetrazolium bromide (MTT) reagent (Invitrogen, USA). The cells at a density of 8 x 10 3 cells/well were seeded in 96 well plates. After 24 h, various concentrations of HCFP aqueous (5-1,500 μg/mL) and methanolic (4-18 μg/mL) extracts were added to the cells and incubated for 24 h. The MTT solution was added to each well and incubated for 2 h at 37 ̊C. After removing the solution, each well was added with DMSO to dissolve the formazan dye. The absorbance of formazan was measured using the microplate reader (Bio-Rad, USA) at 550 nm and 655 nm as a reference wavelength for subtraction of optical density caused by cell debris. Nitrite determination The nitrite concentration in the culture medium of treated and untreated RAW264.7 cells was measured as an indicator of NO production according to Griess reaction . Briefly, the cells (1x10 5 cells/well) were seeded into 24-well plates for 24 h, and then pre-treated cells with various concentrations of HCFP aqueous (25-750 μg/mL) and methanolic (4-12 μg/mL) extracts for 2 h. After 2 h incubation, the cells were incubated with LPS (1 μg/mL) for 24 h. The treatment with diclofenac (DCF; 25 μg/mL) was used as a positive control. The cultured medium was then collected and mixed with an equal volume (1:1) of Griess reagent (Invitrogen, USA). After 10 min incubation at room temperature, the absorbance at 550 nm was measured using a microplate reader. Prostaglandin E2 (PGE 2 ) determination The PGE 2 metabolite is measured by using an enzyme immunoassay (EIA) kit (Abnova, Taiwan) based on the conversion of all major PGE 2 metabolite into a single stable derivative. The cells (1 x 10 5 cells/well) were seeded into 24-well plates and cultured for 24 h. The cells were pre-treated with various concentrations of HCFP aqueous (25-750 μg/mL) and methanolic (4-12 μg/mL) extracts for 2 h and thereafter incubated with LPS (1 μg/mL) for 24 h. The treatment with diclofenac (DCF; 25 μg/mL) was used as a positive control. Subsequently, PGE 2 concentration in a culture medium was determined with PGE 2 EIA kit according to the manufacturer's instructions. Western blot analysis RAW264.7 cells (1 x 10 6 cells) were seeded into a 5.5-cm culture dish and cultured for 24 h. Cells were pre-treated with various concentrations of HCFP aqueous (25-750 μg/mL) and methanolic (4-12 μg/mL) extracts for 2 h and then incubated with LPS (1 μg/mL) for 24 h. The treatment with diclofenac (DCF; 25 μg/mL) was used as a positive control. The treated cells were harvested and lysed with lysis buffer (25 mM Tris-HCl pH 7.6, 150 mM NaCl, 5mM EDTA, 1% NP-40, 1% sodium deoxycholate, 0.1% SDS) for 1 h on ice. The protein concentration was determined by Bradford protein assay (Bio-Rad, USA). Equal amounts of protein (30 μg) were loaded and separated on 12% SDS-polyacrylamide gel and afterward the proteins were transferred to the PVDF membrane. The membrane was blocked with a blocking solution, 5% skim milk in phosphate-buffered saline containing Tween-20 (PBST), for 1 h at room temperature, and then incubated with monoclonal anti-iNOS, anti-COX-2, anti-β-Actin (1:1000 dilutions, Cell signaling, Germany) for overnight at 4 ̊C. The blots were washed twice with PBST and then incubated with horseradish peroxidase (HRP)-conjugated secondary antibody (1:1000 dilutions, Cell signaling, Germany) for 2 h at room temperature. Blots were washed again twice with PBST and PBS, respectively. The protein bands were visualized using ECL detection reagent (GE healthcare, UK). RNA isolation and RT-PCR analysis Cells (1 x 10 6 cells) were seeded into a 5.5-cm culture dish and cultured for 24 h. Cells were pre-treated with various concentrations of HCFP aqueous (25-750 μg/mL) and methanolic (4-12 μg/mL) extracts for 2 h and then incubated with LPS (1 μg/mL) for 6 h. Total RNA from treated cells was isolated using Trizol reagent (Invitrogen, USA) according to the manufacturer's protocol and the RNA was kept at -70 ̊C until used. Total RNA (1 μg) was used for reverse transcription reaction using M-MuLV reverse transcriptase (NEB, UK), 0.5 μM specific reverse primer, deoxyribonucleotide triphosphate (dNTP, 0.2 mM) and 1 U RNase inhibitor. The reaction was incubated at 42 ̊C for 1 h and the M-MuLV reverse transcriptase was then inactivated by heating at 65 ̊C for 20 min. The PCR reactions were carried out in a total volume of 25 μl containing 2.5 U of Taq DNA polymerases, 0.2 mM dNTP, 1X reaction buffer, and 0.5 μM of forward and reverse primers as listed in S1 Table . After initial denaturation for 30 sec at 95 ̊C, the amplification by 30 cycles of 94 ̊C for 45 sec (denaturing), 50-55 ̊C for 45 sec (annealing), 72 ̊C for 45 sec (extension), was carried out. The PCR products were analyzed by 1.5% agarose gel electrophoresis. The level of mRNA expression was quantitated by Quantity One software 4.4.1 (Bio-Rad) using β-actin band intensity as the internal control. Determination of pro-inflammatory cytokines (IL-1β, IL-6 and TNF-α) Cells (1 x 10 5 cells/well) were seeded into 24 well plates and cultured for 24 h, and then pretreated with various concentrations of HCFP aqueous (25-750 μg/ml) and methanolic (4-12 μg/mL) extracts for 2 h. Thereafter, the cells were incubated with LPS (1 μg/mL) for 24 h. The levels of these cytokines in the cultured medium of treated RAW264.7 cells were quantified using ELISA kits according to the manufacturer's instructions. The absorbance at 450 nm was measured using the fluorescence microplate reader (SpectraMax M5, Molecular Devices, USA), and 570 nm was used as a reference wavelength. In vivo experiment Animal studies were performed in obligation with the Institutional Animal Care and Use Committee at Khon Kaen University, Khon Kaen, Thailand (Approval ID: AE101/60) and were performed according to guidelines established by the Ethical Principles and Guidelines for the Use of Animals for scientific purposes, National Research Council of Thailand. All experiments were carried out with six animals in each group. In this study, carrageenan-induced inflammation in the rat paw was used as a model system for in vivo anti-inflammatory study. Male Wistar rats (250-300 g) were divided into four different groups, (1) negative control (carrageenan-treated), (2) HCFP 1 (concentration 1-treated), (3) HCFP 2 (concentration 2-treated), and (4) positive control (Diclofenac-treated). HCFP (3.08 and 6.16 mL/kg) and Diclofenac (150 mg/kg) were administered by orally 1 h before carrageenan induction. The rats received a sub-plantar injection of 100 μL of 1% (w/v) suspension of carrageenan lambda in the right hind paw. The volume of rat paw in all animals was measured at 1, 2 and 3 h after carrageenan injection by using Plethysmometer (Ugo Basile Model 7140, Italy). After λ-carrageenan injection and measurement the volume of rat paw at three hours, the rats were sacrificed by injecting pentobarbital sodium anesthetic. The results were expressed as the changes in paw volume from the baseline value. The percentage of paw edema was calculated using the following equation: %Paw edema 1⁄4 ðV À ViÞ 100=Vi Where V = Paw thickness after carrageenan injection and Vi = Paw thickness at 0 time. HPLC analysis Phenolic acid compositions in HCFP aqueous and methanolic extracts were analyzed by using reverse-phase HPLC as previously described , with some modifications. The columns used to identify phenolic acids in HCFP aqueous and phenolic extracts were Inertsil 1 -ODS-4 C18 column (4.6 mm i.d. x 250 mm, 5 μm particle size) and Waters system C18 column (3.9 mm i. d. x 150 mm, 5 μm particle size), respectively, due to availability of the columns at Facilities Service Center, Faculty of Science, Khon Kaen University, Thailand. The linear gradient of solvents A (100% acetonitrile) and B (1% acetic acid in deionized water) for Inertsil 1 -ODS-4 C18 column was as follows: 0 min, 3% A: 97% B; 5 min, 8% A: 92% B; 15 min, 8% A: 92% B; 25 min, 10% A: 90% B; 55 min, 10% A: 90% B. The linear gradient of solvents for Waters system C18 column was as previously described . The internal standard (m-hydroxybenzaldehyde; 1 μg) was used to ensure the accuracy of phenolic acid identification. Statistical analysis Data are expressed as mean ± S.D. form two or three independent experiments. The data analysis was performed by one-way ANOVA with Duncan's post hoc test. Differences were considered to be significant at p < 0.05. Results Effect of HCFP aqueous and methanolic extracts on cell viability in RAW264.7 cells To study the effect of HCFP aqueous and methanolic extracts on inflammatory responses in vitro, RAW264.7 macrophage cells, which play an important role in the maintenance of tissue homeostasis, were used as a model system. The concentrations of both extracts that had no adverse effects on the growth of RAW264.7 cells were determined using MTT assay. Both HCFP aqueous (Fig 1A ) and methanolic (Fig 1B ) extracts showed no toxicity against RAW264.7 cells at the concentration ranges of 5-750 and 4-12 μg/mL, respectively. The cell viability of more than 90% as compared with a control group was considered non-toxic. Thus, these concentration ranges of HCFP aqueous and methanolic extracts were selected for further study on the anti-inflammatory effect. Effect of HCFP aqueous and methanolic extracts on nitric oxide (NO) production of LPS-stimulated RAW264.7 cells NO is a pro-inflammatory mediator produced by activated macrophages that induce inflammation under pathological conditions . To investigate the effect of HCFP aqueous and C, D ) and PGE 2 levels (E, F) in LPS-stimulated RAW264.7 cells. RAW264.7 cells were incubated with aqueous (5-1,500 μg/mL) and methanolic (4-18 μg/mL) extracts for 24 h. Cell viability was assessed by MTT assay. The results were reported as a percentage of cell viability compared with untreated controls and expressed as mean ± S.D. of three independent experiments. For determinations of NO production and PGE 2 levels in LPS-stimulated RAW264.7 cells, the cells were pre-treated with indicated concentrations of aqueous (C, E) and methanolic (D, F) extracts for 2 h and then stimulated with LPS (1 μg/mL) for 24 h. The nitrite production and PGE 2 levels in cultured medium were determined by using Griess reagent and PGE 2 EIA kit, respectively. Statistically significant inhibition of NO production and reduction of PGE 2 levels ( p < 0.05) were found as compared with the LPS group. Data were obtained from three and two independent experiments, respectively. https://doi.org/10.1371/journal.pone.0230645.g001 methanolic extracts on NO production, RAW264.7 cells were pre-treated with aqueous (25-750 μg/mL) and methanolic (4-12 μg/mL) extracts and thereafter stimulated with LPS (1 μg/ mL). NO production was determined by the measurement of nitrite released into the cultured medium using the Griess reagent. The NSAID drug diclofenac (DCF; 25 μg/mL), a positive control for comparing the activity of HCFP extracts, inhibited NO release by 75.40% in LPSstimulated RAW264.7 macrophages (Fig 1C and 1D ). The maximum (750 μg/mL) and minimum (25 μg/mL) concentrations of the HCFP aqueous extract reduced NO production by 74.40% and 35.22%, respectively (Fig 1C ). Notably, the HCFP methanolic extract at maximum (12 μg/mL) and minimum (4 μg/mL) concentrations reduced NO production by 62.26% and 12.66%, respectively (Fig 1D ). Accordingly, our results showed that both HCFP aqueous and methanolic extracts inhibited NO production in a concentration-dependent manner in LPSstimulated RAW264.7 cells. Effect of HCFP aqueous and methanolic extracts on PGE 2 production in LPS-stimulated RAW264.7 cells PGE 2 produced from arachidonic acid through the function of cyclooxygenase (COX) enzymes during inflammatory responses exacerbates the inflammatory process through several signaling modules . We sought to investigate the inhibitory effect of HCFP extracts on PGE 2 levels in LPS-stimulated macrophages, which may be an effective strategy for treating inflammatory disorders. Similar to the effect on NO production, both HCFP aqueous and methanolic extracts caused a dose-dependent inhibition of PGE 2 production in LPS-stimulated RAW264.7 cells (Fig 1E and 1F ). PGE 2 level was increased to 3,503.11 pg/mL in LPS treatment, whereas in the absence of LPS, PGE 2 level was reduced to 179.42 pg/mL. PGE 2 levels were significantly reduced in the cells treated with aqueous (25-750 μg/mL) (Fig 1E ) and methanolic (4-12 μg/mL) (Fig 1F ) extracts, especially at the highest concentrations tested (58.59% and 51.00% reduction by aqueous (750 μg/mL) and methanolic (12 μg/mL) extracts, respectively). However, diclofenac (25 μg/mL) inhibited PGE 2 production by 93.04% in LPSstimulated RAW264.7 macrophages (Fig 1E and 1F ). Effect of HCFP aqueous and methanolic extracts on expressions of inducible nitric oxide synthase (iNOS) and cyclooxygenase-2 (COX-2) at both mRNA and protein expression levels in LPS-stimulated RAW264.7 cells NO is produced from the conversion of L-arginine to L-citrulline by iNOS , whereas PGE 2 production is mediated by COX-2 . We sought to investigate whether the observed inhibition of HCFP aqueous and methanolic extracts on LPS-induced NO and PGE 2 production (Fig 1C -1F ) was related to the modulation of iNOS and COX-2 using RT-PCR and Western blot analysis. The mRNA and protein expression levels of iNOS and COX-2 were minimally detected and undetectable, respectively, whereas their levels were significantly increased by LPS treatments (Figs 2 and 3 , respectively). iNOS mRNA induction was significantly suppressed by pre-incubation with the NSAID drug diclofenac (25 μg/mL), however, the greater suppression was observed for pre-incubation with both HCFP aqueous (25-750 μg/mL) (Fig 2A and 2B ) and methanolic (6-12 μg/mL) extracts (Fig 2C and 2D ). iNOS protein level was significantly decreased for pre-incubation with both HCFP aqueous (25-750 μg/mL) (Fig 3A and 3C ) and methanolic (4-12 μg/mL) (Fig 3B and 3D ) extracts. Pre-incubation with diclofenac did not cause a significant decrease in either mRNA ( Effect of HCFP aqueous and methanolic extracts on expressions of IL-1β, TNF-α and IL-6 at both mRNA and protein expression levels in LPSstimulated RAW264.7 cells Interaction between LPS and the membrane receptor CD14 of macrophages caused the induction of pro-inflammatory cytokines including IL-1β, TNF-α and IL-6 . These pro-inflammatory cytokines have been considered as targets for anti-inflammatory therapies . To investigate the anti-inflammatory action of aqueous and methanolic extracts of H. cordata fermentation product, the production of the pro-inflammatory cytokines was evaluated by both RT-PCR and ELISA. The mRNA levels of IL-1β, TNF-α, and IL-6 were up-regulated in LPS treated cells compared with untreated controls (Fig 2A -2D ). Diclofenac treatment caused a significant decrease in mRNA levels of IL-1β and TNF-α but not IL-6 (Fig 2A -2D ), whereas treatments with both HCFP aqueous (25-750 μg/mL) (Fig 2A and 2B ) and methanolic (4-12 μg/mL) (Fig 2C and 2D ) extracts caused a significant decrease in IL-1β, TNF-α, and IL-6 mRNA levels. Based on ELISA results, induction of IL-1β, TNF-α, and IL-6 protein levels was dose-dependently suppressed by pre-incubation with both HCFP aqueous (25-750 μg/mL) (Fig 4A , 4C and 4E ) and methanolic (4-12 μg/mL) (Fig 4B , 4D and 4F ) extracts. LPSinduced IL-1β production was inhibited by 68.78 or 62.09% when treated with the highest Effect of HCFP on carrageenan-induced paw edema in Wistar rats To evaluate the in vivo anti-inflammatory activity of the HCFP Dokudami, the carrageenaninduced paw edema model was chosen as it is sensitive and reproducible in vivo test for NSAID drugs and has long been established as a valid model for studying new anti-inflammatory drugs . The formation of paw edema was gradually increased within the first hour after carrageenan injection (Fig 5 ). A common clinical NSAID drug diclofenac (DCF) was used as a positive control pre-treated at 150 mg/kg. DCF significantly (p < 0.05) reduced paw edema after 2 h carrageenan stimulation. Similarly, the HCFP (3.08 and 6.16 mL/kg) also significantly (p < 0.05) reduced paw edema after 2 h carrageenan stimulation. However, pretreatment of HCFP at a concentration of 6.16 mL/kg caused a significant (p < 0.05) reduction of paw edema after 1 h carrageenan stimulation. 1 ). Whereas, seven phenolic acids were identified in methanolic extract of HCFP including protocatechuic, p-hydroxybenzoic, vanillic, syringic, p-coumaric, ferulic and sinapinic acids (Fig 6D and Table 1 ). Among the identified phenolic acids of HCFP, the most abundant phenolic acid in both aqueous and methanolic extracts was syringic acid (Table 1 ). Quantification of phenolic composition in HCFP by HPLC Discussion This study was based on the extensive use of industrial HCFPs as a dietary supplement in Thailand without scientific testing on their biological properties. Therefore, we aimed at investigating the anti-inflammatory activity of a commercialized fermented broth of H. cordata both in vitro (LPS-induced RAW264.7 model) and in vivo (carrageenan-induced paw edema model). To study the anti-inflammatory potential of the industrial HCFP, the inherent cytotoxic effects of the HCFP extracts on the cellular model used in this study were predetermined using assay. All concentrations of the HCFP extracts used in this study induced negligible cytotoxic effects on RAW264.7 macrophages (cell viability > 90%) (Fig 1A and 1B ), indicating that the inhibitory effect of both extracts on production of inflammatory mediators is not attributed to cytotoxicity. LPS from gram-negative bacteria has been shown to possess a dose-dependent cytotoxic activity in RAW264.7 cells , therefore, the non-toxic concentration of LPS was predetermined (Data not shown). LPS at concentration of 1 μg/ml was used in the present study, and not toxic to RAW264.7 cells, consistent with the result from previous study . In the present study, we reported for the first time that the industrial HCFP possessed antiinflammatory activity. Both HCFP aqueous and methanolic extracts successfully inhibited the production of inflammatory mediators (NO and PGE 2 ). During the inflammatory process, large amounts of the pro-inflammatory mediators like nitric oxide and prostaglandins E 2 are generated by the inducible nitric oxide synthase (iNOS) and cyclooxygenase (COX-2), respectively. Nitric oxide is one of pro-inflammatory mediator that responds to pathogenic infections. During the inflammatory process, NO is generated by macrophages to eliminate foreign pathogens, recruiting other cells to the infected area and subsequently resolving the inflammation. However, the excessive amount of NO is also harmful to normal tissue surrounding the infected area because it binds with other superoxides radical and acts as a reactive radical to damage normal cell function. Gram negative bacterial LPS is well known to increase iNOS expression and NO production, leading to the initiation of an inflammatory response . Therefore, the inhibition of NO production is a key therapeutic consideration in both searching for anti-inflammatory agents and developing a novel treatment for inflammatory disorders. In the present study, both aqueous and methanolic extracts of H. cordata fermentation product reduced NO production in LPS-stimulated RAW264.7 cells in a concentration dependent manner (Fig 1C and 1D ). The decreased NO production was correlated well with the dose-dependent decrease of iNOS mRNA ( TNF-α, IL-1β, and IL-6 are the main pro-inflammatory cytokines that are primarily produced by macrophages and have various pro-inflammatory effects on many cell types [31, 32] . Over-production of TNF-α caused the release of various inflammatory mediators including NO, PGE 2 , IL-1β and IL-6. Excessive production of cytokines (TNF-α, IL-1β and IL-6) has linked in several physiological effects, including septic shock, inflammation and cytotoxicity . Thus, the inhibition of cytokine production or function is a key mechanism in the control of inflammation . In the present study, aqueous and methanolic extracts of H. cordata fermentation product reduced production of cytokines in LPS-stimulated RAW264.7 cells in a concentration-dependent manner both at mRNA ( Carrageenan-induced paw edema is an animal model suitable for evaluating inhibition of edema. Biphasic edema induced by carrageenan , includes the first phase (1 h) involving the release of serotonin and histamine and the second phase (over 1 h) mediating by prostaglandins, cyclooxygenase products. In the present study, both doses of the industrial HCFP significantly reduced paw edema at 2 and 3 h after carrageenan injection. This finding suggests that HCFP produces an anti-edematous effect during the second phase which is similar to DCF (Fig 5 ). Interestingly, the highest dose of HCFP (6.16 mL/kg) showed a significant reduction of paw edema at 1 h, suggesting an anti-edematous effect during the first phase. Further animal study on the mechanism underlying inhibition of the first/second phase edema is of interest. The identification of active components in HCFP extracts is an important pharmacological goal. Our HPLC results demonstrated that the amount of all identified phenolic acids in HCFP methanolic extract were much greater than those in HCFP aqueous extract (Table 1 ). Syringic acid was present in the greatest amounts in both HCFP aqueous (88.23 μg/g of extract) and methanolic (2,268.34 μg/g of extract) extracts, followed by vanillic, p-hydroxybenzoic, and ferulic acids, respectively (Fig 6 and Table 1 ). Yoo et al. demonstrated that syringic, vanillic, p-hydroxybenzoic and ferulic acids. In addition, p-coumaric acid has been shown to possess anti-inflammatory activity both in vitro and in vivo [37, 38] . Accordingly, these phenolic acids may contribute to HCFP-mediated inhibition of the production of inflammatory cytokines and mediators. Phenolic acids in the water-soluble constituents of H. cordata fermentation product were previously identified and quantified , but their amounts were greater than those found in the present HCFP aqueous extract (Table 1 ). The discrepancy may be due to a lot-to-lot variation of the industrial HCFP. Further study on synergistic anti-inflammatory effects of HCFP individual phenolic acids both in vitro and in vivo is of interest. Conclusions Our results demonstrated that the aqueous and methanolic extracts of H. cordata fermentation product possessed anti-inflammatory activity by inhibiting the production of NO, PGE 2 and inflammatory cytokines (TNF-α, IL-1β, IL-6) in LPS-stimulated RAW264.7 cells. The antiinflammatory activity of the industrial HCFP was confirmed by the inhibition of inflammation in the carrageenan-induced rat paw edema model. Our results suggest that this industrial HCFP may be considered as an anti-inflammatory dietary supplement. The health benefits of this industrial HCFP warrant further clinical studies. Fig 1 . 1 Fig 1. Effect of aqueous and methanolic extracts of HCFP on RAW264.7 cell viability (A, B), NO production (C, D) and PGE 2 levels (E, F) in LPS-stimulated RAW264.7 cells. RAW264.7 cells were incubated with aqueous (5-1,500 μg/mL) and methanolic (4-18 μg/mL) extracts for 24 h. Cell viability was assessed by MTT assay. The results were reported as a percentage of cell viability compared with untreated controls and expressed as mean ± S.D. of three independent experiments. For determinations of NO production and PGE 2 levels in LPS-stimulated RAW264.7 cells, the cells were pre-treated with indicated concentrations of aqueous (C, E) and methanolic (D, F) extracts for 2 h and then stimulated with LPS (1 μg/mL) for 24 h. The nitrite production and PGE 2 levels in cultured medium were determined by using Griess reagent and PGE 2 EIA kit, respectively. Statistically significant inhibition of NO production and reduction of PGE 2 levels ( p < 0.05) were found as compared with the LPS group. Data were obtained from three and two independent experiments, respectively. Fig 2) or protein (Fig 3) levels of COX-2. Similarly, COX-2 mRNA induction was not significantly suppressed by pre-incubation with both HCFP aqueous (50-750 μg/mL) (Fig 2A and 2B) and methanolic (8-12 μg/mL) (Fig 2C and 2D) extracts. COX-2 protein level was not significantly decreased for pre-incubation with both HCFP aqueous (50-750 μg/mL) (Fig 3A and 3C) and methanolic (4-12 μg/mL) (Fig 3B and 3D) extracts. The above results suggest that inhibition of NO and PGE 2 production Fig 2 . 2 Fig 2. Effect of aqueous and methanolic extracts of HCFP on mRNA expression of iNOS, COX-2, IL-1β, TNF-α, and IL-6 in LPS-stimulated RAW264.7 cells. The cells were pre-treated with indicated concentrations of aqueous and methanolic extracts for 2 h and then stimulated with LPS (1 μg/mL) for 6 h. The mRNA expression levels of aqueous extract-treated (A) and methanolic extract-treated (C) cells were determined by reverse transcription-PCR. Bar graphs showed the relative fold of mRNA expression of aqueous extract-treated (B) and methanolic extract-treated (D) cells, p < 0.05 compared with the LPS group. https://doi.org/10.1371/journal.pone.0230645.g002 Fig 3 . 3 Fig 3. Effect of aqueous (A) and methanolic (B) extracts of HCFP on protein levels of iNOS and COX-2 in LPS-stimulated RAW264.7 cells. The cells were pretreated with the indicated concentration of aqueous extract or phenolic extract for 2 h and then stimulated with LPS (1 μg/mL) for 24 h. The protein expression levels were analyzed by western blot. Bar graphs showed the relative fold of protein expression in aqueous-(C) and methanolic-(D) treated cells. Data were expressed as mean ± S.D. p < 0.05 compared with the LPS group. https://doi.org/10.1371/journal.pone.0230645.g003 Fig 4 . 4 Fig 4. Effect of HCFP on production of IL-1β (A, B), IL-6 (C, D) and TNF-α (E, F) in LPS-stimulated RAW264.7 cells. The cells were pretreated with the indicated concentration of aqueous extract or phenolic extract for 2 h then stimulated with LPS (1 μg/mL) for 24 h. The IL-1β, IL-6, and TNF-α in cultured medium were determined by ELISA kits. Data were expressed as the mean±S.D. of two independent experiments. Statistically significant inhibitions of IL-1β, IL-6 and TNF-α production ( p < 0.05) were found as compared with the LPS group. https://doi.org/10.1371/journal.pone.0230645.g004 The component profiles of HCFP aqueous and methanolic extracts were analyzed by HPLC. The representative chromatograms were shown in Fig 6. Six phenolic acids were identified in HCFP aqueous extract including p-hydroxybenzoic, vanillic, syringic, p-coumaric, ferulic and gallic acids (Fig 6B and Table Fig 5 . 5 Fig 5. Effect of HCFP and DCF on carrageenan-induced paw edema in Wistar rats. Bar graphs show percentages of changes in paw edema. Data are expressed as mean ± SD of n = 6 rats/group. Asterisk " " indicates a significant difference at p < 0.05 as compared with the control group. https://doi.org/10.1371/journal.pone.0230645.g005 Fig 6 . 6 Fig 6. HPLC chromatograms of phenolic acid standards (A, C) and base hydrolyzed HCFP aqueous (B) and methanolic (D) extracts, where 1 = gallic acid, 2 = protocatechuic acid, 3 = p-hydroxybenzoic acid, 4 = vanillic acid, 5 = caffeic acid, 6 = m-hydroxybenzaldehyde, 7 = syringic acid, 8 = p-coumaric acid, 9 = ferulic acid and 10 = sinapinic acid. The m-hydroxybenzaldehyde was used as an internal standard (I.S.). https://doi.org/10.1371/journal.pone.0230645.g006 Fig 2 ) 2 and protein (Fig 3) levels. PGE 2 , an inflammatory mediator, is produced by the metabolism of arachidonic acid by COX enzymes at inflammatory sites . PGE 2 production increased following LPS treatment. After the RAW264.7 cells were pre-treated with HCFP aqueous and methanolic extracts, the PGE 2 levels in LPS-stimulated RAW264.7 cells were decreased in a dose-dependent fashion (Fig 1E and 1F). However, the NSAID diclofenac (25 μg/mL) exhibited more potent inhibitory activity against PGE 2 production than HCFP extracts at all concentrations tested. The decreased PGE 2 production (Fig 1E and 1F) in both DCF and HCFP treatments was not correlated well with the COX-2 mRNA (Fig 2) and protein (Fig 3) levels. The reduced PGE 2 levels may be due to the inhibition of COX-2 activity by DCF and HCFP treatments. Indeed, DCF has been shown to selectively inhibit COX-2 activity . Table 1 . 1 Phenolic acid compositions of aqueous and methanolic extracts of HCFP. a Results are expressed as means ± SD of three determinations. n.d., not detected. https://doi.org/10.1371/journal.pone.0230645.t001 Fig 2 ) 2 and protein (Fig 4) levels. PLOS ONE \| https://doi.org/10.1371/journal.pone.0230645March 25, 2020 PLOS ONE \| https://doi.org/10.1371/journal.pone.0230645 March 25, 2020 / 18
10.2196/43376	cc-by	null	null	openalex	"Background: Chronic musculoskeletal pain (CMSP) affects between 13% and 47% of the population, with(...TRUNCATED)
10.21496/ams.2020.002	cc-by	null	null	openalex	"The article introduces an innovative concept of a screw of an extruder used for granulation of poly(...TRUNCATED)
10.48550/arxiv.1705.03274	cc-by	null	null	openalex	"After Run 1 of the LHC, global fits to b → s observables show a deviation from the Standard Model(...TRUNCATED)

End of preview. Expand in Data Studio

HC4 (Healthcare Comprehensive Commons Corpus)

HC4 is a large-scale pretraining dataset containing over 65 billion tokens from diverse healthcare-related sources. The corpus was curated to enable systematic investigation of how data composition influences language model behavior, including potential demographic biases.

Dataset Overview

Dataset Name: HC4 (Healthcare Comprehensive Commons Corpus)

Size: 153GB (around 65 billion tokens)

Number of samples: 9.7+ million documents from diverse sources including peer-reviewed scientific literature collected from PubMed Central, Semantic Scholar, OpenAlex repositories

Repository: m42-health/HC4

Purpose: Pretraining large language models for healthcare applications

Format: .parquet files

License: Open licenses for each data sample permitting commercial use and redistribution

Organization: M42 (Abu Dhabi)

How to load the dataset

from datasets import load_dataset

dataset = load_dataset("m42-health/HC4")

Details

This dataset is accompanied by a peer-reviewed research paper accepted at EMNLP 2025 Conference, which presents comprehensive bias analysis methodology for clinical LLMs and provides transparency in dataset composition and curation.

Reference Paper: "Building Trust in Clinical LLMs: Bias Analysis and Dataset Transparency" (EMNLP 2025)

Link to arxiv page: https://arxiv.org/pdf/2510.18556

Downloads last month: 629

Size of downloaded dataset files:

164 GB

Size of the auto-converted Parquet files:

164 GB

Number of rows:

9,769,240