Workshop report - interdisciplinary metabolomic epidemiology: the pathway to clinical translation.

Metabolomic epidemiology studies are complex and require a broad array of domain expertise. Although many metabolite-phenotype associations have been identified; to date, few findings have been translated to the clinic. Bridging this gap requires understanding of both the underlying biology of these associations and their potential clinical implications, necessitating an interdisciplinary team approach. To address this need in metabolomic epidemiology, a workshop was held at Metabolomics 2023 in Niagara Falls, Ontario, Canada that highlighted the domain expertise needed to effectively conduct these studies -- biochemistry, clinical science, epidemiology, and assay development for biomarker validation -- and emphasized the role of interdisciplinary teams to move findings towards clinical translation.

Metabolite signatures associated with microRNA miR-143-3p serve as drivers of poor lung function trajectories in childhood asthma.

BACKGROUND: Lung function trajectories (LFTs) have been shown to be an important measure of long-term health in asthma. While there is a growing body of metabolomic studies on asthma status and other phenotypes, there are no prospective studies of the relationship between metabolomics and LFTs or their genomic determinants. METHODS: We utilized ordinal logistic regression to identify plasma metabolite principal components associated with four previously-published LFTs in children from the Childhood Asthma Management Program (CAMP) (n = 660). The top significant metabolite principal component (PCLF) was evaluated in an independent cross-sectional child cohort, the Genetic Epidemiology of Asthma in Costa Rica Study (GACRS) (n = 1151) and evaluated for association with spirometric measures. Using meta-analysis of CAMP and GACRS, we identified associations between PCLF and microRNA, and SNPs in their target genes. Statistical significance was determined using an false discovery rate-adjusted Q-value. FINDINGS: The top metabolite principal component, PCLF, was significantly associated with better LFTs after multiple-testing correction (Q-value = 0.03). PCLF is composed of the urea cycle, caffeine, corticosteroid, carnitine, and potential microbial (secondary bile acid, tryptophan, linoleate, histidine metabolism) metabolites. Higher levels of PCLF were also associated with increases in lung function measures and decreased circulating neutrophil percentage in both CAMP and GACRS. PCLF was also significantly associated with microRNA miR-143-3p, and SNPs in three miR-143-3p target genes; CCZ1 (P-value = 2.6 × 10-5), SLC8A1 (P-value = 3.9 × 10-5); and TENM4 (P-value = 4.9 × 10-5). INTERPRETATION: This study reveals associations between metabolites, miR-143-3p and LFTs in children with asthma, offering insights into asthma physiology and possible interventions to enhance lung function and long-term health. FUNDING: Molecular data for CAMP and GACRS via the Trans-Omics in Precision Medicine (TOPMed) program was supported by the National Heart, Lung, and Blood Institute (NHLBI).

Metabolomics reveals mouse plasma metabolite responses to acute exercise and effects of disrupting AMPK-glycogen interactions.

Introduction: The AMP-activated protein kinase (AMPK) is a master regulator of energy homeostasis that becomes activated by exercise and binds glycogen, an important energy store required to meet exercise-induced energy demands. Disruption of AMPK-glycogen interactions in mice reduces exercise capacity and impairs whole-body metabolism. However, the mechanisms underlying these phenotypic effects at rest and following exercise are unknown. Furthermore, the plasma metabolite responses to an acute exercise challenge in mice remain largely uncharacterized. Methods: Plasma samples were collected from wild type (WT) and AMPK double knock-in (DKI) mice with disrupted AMPK-glycogen binding at rest and following 30-min submaximal treadmill running. An untargeted metabolomics approach was utilized to determine the breadth of plasma metabolite changes occurring in response to acute exercise and the effects of disrupting AMPK-glycogen binding. Results: Relative to WT mice, DKI mice had reduced maximal running speed (p < 0.0001) concomitant with increased body mass (p < 0.01) and adiposity (p < 0.001). A total of 83 plasma metabolites were identified/annotated, with 17 metabolites significantly different (p < 0.05; FDR<0.1) in exercised (↑6; ↓11) versus rested mice, including amino acids, acylcarnitines and steroid hormones. Pantothenic acid was reduced in DKI mice versus WT. Distinct plasma metabolite profiles were observed between the rest and exercise conditions and between WT and DKI mice at rest, while metabolite profiles of both genotypes converged following exercise. These differences in metabolite profiles were primarily explained by exercise-associated increases in acylcarnitines and steroid hormones as well as decreases in amino acids and derivatives following exercise. DKI plasma showed greater decreases in amino acids following exercise versus WT. Conclusion: This is the first study to map mouse plasma metabolomic changes following a bout of acute exercise in WT mice and the effects of disrupting AMPK-glycogen interactions in DKI mice. Untargeted metabolomics revealed alterations in metabolite profiles between rested and exercised mice in both genotypes, and between genotypes at rest. This study has uncovered known and previously unreported plasma metabolite responses to acute exercise in WT mice, as well as greater decreases in amino acids following exercise in DKI plasma. Reduced pantothenic acid levels may contribute to differences in fuel utilization in DKI mice.

Lipopolysaccharide-induced interferon response networks at birth are predictive of severe viral lower respiratory infections in the first year of life.

Appropriate innate immune function is essential to limit pathogenesis and severity of severe lower respiratory infections (sLRI) during infancy, a leading cause of hospitalization and risk factor for subsequent asthma in this age group. Employing a systems biology approach to analysis of multi-omic profiles generated from a high-risk cohort (n=50), we found that the intensity of activation of an LPS-induced interferon gene network at birth was predictive of sLRI risk in infancy (AUC=0.724). Connectivity patterns within this network were stronger among susceptible individuals, and a systems biology approach identified IRF1 as a putative master regulator of this response. These findings were specific to the LPS-induced interferon response and were not observed following activation of viral nucleic acid sensing pathways. Comparison of responses at birth versus age 5 demonstrated that LPS-induced interferon responses but not responses triggered by viral nucleic acid sensing pathways may be subject to strong developmental regulation. These data suggest that the risk of sLRI in early life is in part already determined at birth, and additionally that the developmental status of LPS-induced interferon responses may be a key determinant of susceptibility. Our findings provide a rationale for the identification of at-risk infants for early intervention aimed at sLRI prevention and identifies targets which may be relevant for drug development.

Does machine learning have a role in the prediction of asthma in children?

Asthma is the most common chronic lung disease in childhood. There has been a significant worldwide effort to develop tools/methods to identify children's risk for asthma as early as possible for preventative and early management strategies. Unfortunately, most childhood asthma prediction tools using conventional statistical models have modest accuracy, sensitivity, and positive predictive value. Machine learning is an approach that may improve on conventional models by finding patterns and trends from large and complex datasets. Thus far, few studies have utilized machine learning to predict asthma in children. This review aims to critically assess these studies, describe their limitations, and discuss future directions to move from proof-of-concept to clinical application.

Urinary metabotype of severe asthma evidences decreased carnitine metabolism independent of oral corticosteroid treatment in the U-BIOPRED study.

INTRODUCTION: Asthma is a heterogeneous disease with poorly defined phenotypes. Patients with severe asthma often receive multiple treatments including oral corticosteroids (OCS). Treatment may modify the observed metabotype, rendering it challenging to investigate underlying disease mechanisms. Here, we aimed to identify dysregulated metabolic processes in relation to asthma severity and medication. METHODS: Baseline urine was collected prospectively from healthy participants (n=100), patients with mild-to-moderate asthma (n=87) and patients with severe asthma (n=418) in the cross-sectional U-BIOPRED cohort; 12-18-month longitudinal samples were collected from patients with severe asthma (n=305). Metabolomics data were acquired using high-resolution mass spectrometry and analysed using univariate and multivariate methods. RESULTS: A total of 90 metabolites were identified, with 40 significantly altered (p<0.05, false discovery rate <0.05) in severe asthma and 23 by OCS use. Multivariate modelling showed that observed metabotypes in healthy participants and patients with mild-to-moderate asthma differed significantly from those in patients with severe asthma (p=2.6×10-20), OCS-treated asthmatic patients differed significantly from non-treated patients (p=9.5×10-4), and longitudinal metabotypes demonstrated temporal stability. Carnitine levels evidenced the strongest OCS-independent decrease in severe asthma. Reduced carnitine levels were associated with mitochondrial dysfunction via decreases in pathway enrichment scores of fatty acid metabolism and reduced expression of the carnitine transporter SLC22A5 in sputum and bronchial brushings. CONCLUSIONS: This is the first large-scale study to delineate disease- and OCS-associated metabolic differences in asthma. The widespread associations with different therapies upon the observed metabotypes demonstrate the need to evaluate potential modulating effects on a treatment- and metabolite-specific basis. Altered carnitine metabolism is a potentially actionable therapeutic target that is independent of OCS treatment, highlighting the role of mitochondrial dysfunction in severe asthma.

Searching for a technology-driven acute rheumatic fever test: the START study protocol.

INTRODUCTION: The absence of a diagnostic test for acute rheumatic fever (ARF) is a major impediment in managing this serious childhood condition. ARF is an autoimmune condition triggered by infection with group A Streptococcus. It is the precursor to rheumatic heart disease (RHD), a leading cause of health inequity and premature mortality for Indigenous peoples of Australia, New Zealand and internationally. METHODS AND ANALYSIS: 'Searching for a Technology-Driven Acute Rheumatic Fever Test' (START) is a biomarker discovery study that aims to detect and test a biomarker signature that distinguishes ARF cases from non-ARF, and use systems biology and serology to better understand ARF pathogenesis. Eligible participants with ARF diagnosed by an expert clinical panel according to the 2015 Revised Jones Criteria, aged 5-30 years, will be recruited from three hospitals in Australia and New Zealand. Age, sex and ethnicity-matched individuals who are healthy or have non-ARF acute diagnoses or RHD, will be recruited as controls. In the discovery cohort, blood samples collected at baseline, and during convalescence in a subset, will be interrogated by comprehensive profiling to generate possible diagnostic biomarker signatures. A biomarker validation cohort will subsequently be used to test promising combinations of biomarkers. By defining the first biomarker signatures able to discriminate between ARF and other clinical conditions, the START study has the potential to transform the approach to ARF diagnosis and RHD prevention. ETHICS AND DISSEMINATION: The study has approval from the Northern Territory Department of Health and Menzies School of Health Research ethics committee and the New Zealand Health and Disability Ethics Committee. It will be conducted according to ethical standards for research involving Indigenous Australians and New Zealand Maori and Pacific Peoples. Indigenous investigators and governance groups will provide oversight of study processes and advise on cultural matters.

Sensitive and quantitative determination of short-chain fatty acids in human serum using liquid chromatography mass spectrometry.

Short-chain fatty acids (SCFAs) are increasingly being monitored to elucidate the link between gut health and disease. These metabolites are routinely measured in faeces, but their determination in serum is more challenging due to their low concentrations. A method for the determination of eight SCFAs in serum is described here. High-resolution mass spectrometry and gas chromatography were used to identify the presence of isomeric interferences, which were then overcome through a combination of chromatographic separation and judicious choice of MS fragment ion. The SCFAs were derivatised to form 3-nitrophenylhydrazones before being separated on a reversed-phase column and then detected using liquid chromatography tandem mass spectrometry (LC-QQQ-MS). The LODs and LOQs of SCFAs using this method were in the range 1 to 7 ng mL-1 and 3 to 19 ng mL-1, respectively. The recovery of the SCFAs in serum ranged from 94 to 114% over the three concentration ranges tested.

Kynurenic acid may underlie sex-specific immune responses to COVID-19.

Coronavirus disease 2019 (COVID-19) has poorer clinical outcomes in males than in females, and immune responses underlie these sex-related differences. Because immune responses are, in part, regulated by metabolites, we examined the serum metabolomes of COVID-19 patients. In male patients, kynurenic acid (KA) and a high KA-to-kynurenine (K) ratio (KA:K) positively correlated with age and with inflammatory cytokines and chemokines and negatively correlated with T cell responses. Males that clinically deteriorated had a higher KA:K than those that stabilized. KA inhibits glutamate release, and glutamate abundance was lower in patients that clinically deteriorated and correlated with immune responses. Analysis of data from the Genotype-Tissue Expression (GTEx) project revealed that the expression of the gene encoding the enzyme that produces KA, kynurenine aminotransferase, correlated with cytokine abundance and activation of immune responses in older males. This study reveals that KA has a sex-specific link to immune responses and clinical outcomes in COVID-19, suggesting a positive feedback between metabolites and immune responses in males.

A strategy for advancing for population-based scientific discovery using the metabolome: the establishment of the Metabolomics Society Metabolomic Epidemiology Task Group.

Metabolomic Epidemiology is a growing area of research within the metabolomics research community. In response to this, we describe the establishment of the Metabolomics Society Metabolomic Epidemiology Task Group. The overall mission of this group is to promote the growth and understanding of metabolomic epidemiology as an independent research discipline and to drive collaborative efforts that can shape the field. In this article we define metabolomic epidemiology and identify the key challenges that need to be addressed in order to advance population-based scientific discovery in metabolomics.

Data supporting development and validation of liquid chromatography tandem mass spectrometry method for the quantitative determination of bile acids in feces.

Measuring bile acids in feces has an important role in disease prevention, diagnosis, treatment, and can be considered a measure of health status. Therefore, the primary aim was to develop a sensitive, robust, and high throughput liquid chromatography tandem mass spectrometry method with minimal sample preparation for quantitative determination of bile acids in human feces applicable to large cohorts. Due to the chemical diversity of bile acids, their wide concentration range in feces, and the complexity of feces itself, developing a sensitive and selective analytical method for bile acids is challenging. A simple extraction method using methanol suitable for subsequent quantification by liquid chromatography tandem mass spectrometry has been reported in, "Extraction and quantitative determination of bile acids in feces" [1]. The data highlight the importance of optimization of the extraction procedure and the stability of the bile acids in feces post-extraction and prior to analysis and after several freeze-thaw cycles.

Detecting Sex-Related Changes to the Metabolome of a Critically Endangered Freshwater Crayfish During the Mating Season.

Captive breeding is a vital tool in the conservation of highly endangered species, as it is for the Margaret River hairy marron, Cherax tenuimanus, from the south west of Australia. A close relative, Cherax cainii, has almost completely displaced C. tenuimanus in the wild and is a successful aquaculture species, whereas C. tenuimanus has performed poorly in captivity. We used untargeted liquid chromatography-mass spectrometry to obtain metabolomic profiles of female and male C. tenuimanus held in controlled aquarium conditions during their reproductive period. Using repeated haemolymph sampling we tracked the metabolomic profiles of animals just prior to and for a period of up to 34 days after pairing with a similar sized potential mate. We identified 54 reproducible annotated metabolites including amino acids, fatty acids, biogenic amines, purine and pyrimidine metabolites and excretion metabolites. Hierarchical clustering analysis distinguished five metabolite clusters. Principal component-canonical variate analysis clearly distinguished females from males, both unpaired and paired; similar trends in profile changes in both sexes after pairing; and a striking shift in males upon pairing. We discuss three main patterns of metabolomic responses: differentiation between sexes; reactive responses to the disturbance of pairing; and convergent response to the disturbance of pairing for males. Females generally had higher concentrations of metabolites involved in metabolic rate, mobilisation of energy stores and stress. Responses to the disturbance of pairing were also related to elevated stress. Females were mobilising lipid stores to deposit yolk, whereas males had a rapid and strong response to pairing, with shifts in metabolites associated with gonad development and communication, indicating males could complete reproductive readiness only once paired with a female. The metabolomic profiles support a previously proposed potential mechanism for displacement of C. tenuimanus by C. cainii in the wild and identify several biomarkers for testing hypotheses regarding reproductive success using targeted metabolomics.

Early and sustained Lactobacillus plantarum probiotic therapy in critical illness: the randomised, placebo-controlled, restoration of gut microflora in critical illness trial (ROCIT).

PURPOSE: In adults requiring treatment in an intensive care unit, probiotic therapy using Lactobacillus plantarum 299v may reduce nosocomial infection. The aim of this study was to determine whether early and sustained L. plantarum 299v therapy administered to adult ICU patients increased days alive and at home. METHODS: A multicentre, parallel group, placebo-controlled, randomised clinical trial was conducted. Adult patients within 48 h of intensive care admission and expected to require intensive care beyond the day after recruitment were eligible to participate. L plantarum 299v or placebo were administered immediately after enrolment and continued for 60 days. The primary outcome was days alive and out of hospital to Day 60 (DAOH60). Secondary outcomes included nosocomial infections. RESULTS: The median [interquartile range (IQR)] number of DAOH60 in the probiotic (n = 110) and placebo group (n = 108) was 49.5 (IQR 37.0-53.0) and 49.0 (IQR 43.8-53.0) respectively, between-group difference of 0.0 [95% confidence interval (CI) - 6.10 to 7.1, P = 0.55]. Nosocomial infection occurred in 8 (7.3%) and 5 (4.6%) of the probiotic and placebo group participants, respectively, odds ratio 1.62 (95% CI 0.51-5.10), P = 0.57. There were no serious, or probiotic-associated adverse events. CONCLUSION: Early and sustained untargeted administration of probiotic therapy with Lactobacillus plantarum 299v to adult patients admitted to the ICU is safe, but not associated with improved patient outcomes.

Extraction and quantitative determination of bile acids in feces.

With rapid advances in gut microbiome research, fecal bile acids are increasingly being monitored as potential biomarkers of diet related disease susceptibility. As such, rapid, robust and reliable methods for their analysis are of increasing importance. Herein is described a simple extraction method for the analysis of bile acids in feces suitable for subsequent quantification by liquid chromatography and tandem mass spectrometry. A C18 column separated the analytes with excellent peak shape and retention time repeatability maintained across 800 injections. The intra-day and inter-day precision and accuracy was greater than 80%. Recoveries ranged from 83.58 to 122.41%. The limit of detection and limit of quantification were in the range 2.5-15 nM, respectively. The optimized method involved extracting bile acids from wet feces with minimal clean up. A second aliquot of fecal material was dried and weighed to correct for water content. Extracting from dried feces showed reduced recovery that could be corrected for by spiking the feces with deuterated standards prior to drying. Storage of the extracts and standards in a refrigerated autosampler prior to analysis on the LC-MS is necessary. Multiple freeze-thaws of both extracts and standards lead to poor recoveries for some bile acids. The method was successfully applied to 100 human fecal samples.

Improvement in the Prediction of Neonatal Hypoxic-Ischemic Encephalopathy with the Integration of Umbilical Cord Metabolites and Current Clinical Makers.

OBJECTIVE: To validate our previously identified candidate metabolites, and to assess the ability of these metabolites to predict hypoxic-ischemic encephalopathy (HIE) both individually and combined with clinical data. STUDY DESIGN: Term neonates with signs of perinatal asphyxia, with and without HIE, and matched controls were recruited prospectively at birth from 2 large maternity units. Umbilical cord blood was collected for later batch metabolomic analysis by mass spectroscopy along with clinical details. The optimum selection of clinical and metabolites features with the ability to predict the development of HIE was determined using logistic regression modelling and machine learning techniques. Outcome of HIE was determined by clinical Sarnat grading and confirmed by electroencephalogram grade at 24 hours. RESULTS: Fifteen of 27 candidate metabolites showed significant alteration in infants with perinatal asphyxia or HIE when compared with matched controls. Metabolomic data predicted the development of HIE with an area under the curve of 0.67 (95% CI, 0.62-0.71). Lactic acid and alanine were the primary metabolite predictors for the development of HIE, and when combined with clinical data, gave an area under the curve of 0.96 (95% CI, 0.92-0.95). CONCLUSIONS: By combining clinical and metabolic data, accurate identification of infants who will develop HIE is possible shortly after birth, allowing early initiation of therapeutic hypothermia.

Dissemination and analysis of the quality assurance (QA) and quality control (QC) practices of LC-MS based untargeted metabolomics practitioners.

INTRODUCTION: The metabolomics quality assurance and quality control consortium (mQACC) evolved from the recognized need for a community-wide consensus on improving and systematizing quality assurance (QA) and quality control (QC) practices for untargeted metabolomics. OBJECTIVES: In this work, we sought to identify and share the common and divergent QA and QC practices amongst mQACC members and collaborators who use liquid chromatography-mass spectrometry (LC-MS) in untargeted metabolomics. METHODS: All authors voluntarily participated in this collaborative research project by providing the details of and insights into the QA and QC practices used in their laboratories. This sharing was enabled via a six-page questionnaire composed of over 120 questions and comment fields which was developed as part of this work and has proved the basis for ongoing mQACC outreach. RESULTS: For QA, many laboratories reported documenting maintenance, calibration and tuning (82%); having established data storage and archival processes (71%); depositing data in public repositories (55%); having standard operating procedures (SOPs) in place for all laboratory processes (68%) and training staff on laboratory processes (55%). For QC, universal practices included using system suitability procedures (100%) and using a robust system of identification (Metabolomics Standards Initiative level 1 identification standards) for at least some of the detected compounds. Most laboratories used QC samples (>86%); used internal standards (91%); used a designated analytical acquisition template with randomized experimental samples (91%); and manually reviewed peak integration following data acquisition (86%). A minority of laboratories included technical replicates of experimental samples in their workflows (36%). CONCLUSIONS: Although the 23 contributors were researchers with diverse and international backgrounds from academia, industry and government, they are not necessarily representative of the worldwide pool of practitioners due to the recruitment method for participants and its voluntary nature. However, both questionnaire and the findings presented here have already informed and led other data gathering efforts by mQACC at conferences and other outreach activities and will continue to evolve in order to guide discussions for recommendations of best practices within the community and to establish internationally agreed upon reporting standards. We very much welcome further feedback from readers of this article.

Kynurenic acid underlies sex-specific immune responses to COVID-19.

Coronavirus disease-2019 (COVID-19) has poorer clinical outcomes in males compared to females, and immune responses underlie these sex-related differences in disease trajectory. As immune responses are in part regulated by metabolites, we examined whether the serum metabolome has sex-specificity for immune responses in COVID-19. In males with COVID- 19, kynurenic acid (KA) and a high KA to kynurenine (K) ratio was positively correlated with age, inflammatory cytokines, and chemokines and was negatively correlated with T cell responses, revealing that KA production is linked to immune responses in males. Males that clinically deteriorated had a higher KA:K ratio than those that stabilized. In females with COVID-19, this ratio positively correlated with T cell responses and did not correlate with age or clinical severity. KA is known to inhibit glutamate release, and we observed that serum glutamate is lower in patients that deteriorate from COVID-19 compared to those that stabilize, and correlates with immune responses. Analysis of Genotype-Tissue Expression (GTEx) data revealed that expression of kynurenine aminotransferase, which regulates KA production, correlates most strongly with cytokine levels and aryl hydrocarbon receptor activation in older males. This study reveals that KA has a sex-specific link to immune responses and clinical outcomes, in COVID-19 infection.

Study protocol for the safety and efficacy of probiotic therapy on days alive and out of hospital in adult ICU patients: the multicentre, randomised, placebo-controlled Restoration Of gut microflora in Critical Illness Trial (ROCIT).

INTRODUCTION: The effect of early and sustained administration of daily probiotic therapy on patients admitted to the intensive care unit (ICU) remains uncertain. METHODS AND ANALYSIS: The Restoration Of gut microflora in Critical Illness Trial (ROCIT) study is a multicentre, randomised, placebo-controlled, parallel-group, two-sided superiority trial that will enrol 220 patients in five ICUs. Adult patients who are within 48 hours of admission to an ICU and are expected to require intensive care beyond the next calendar day will be randomised in a 1:1 ratio to receive early and sustained Lactobacillus plantarum 299v probiotic therapy in addition to usual care or placebo in addition to usual care. The primary endpoint is days alive and out of hospital to day 60. ETHICS AND DISSEMINATION: ROCIT has been approved by the South Metropolitan Health Service Human Research Ethics Committee (ref: RGS00000004) and the St John of God Health Care Human Research Ethics Committee (ref: 1183). The trial results will be submitted for publication in a peer-reviewed journal. TRIAL REGISTRATION NUMBER: Australian and New Zealand Clinical Trials Registry (ANZCTR12617000783325); Pre-results.

"notame": Workflow for Non-Targeted LC-MS Metabolic Profiling.

Metabolomics analysis generates vast arrays of data, necessitating comprehensive workflows involving expertise in analytics, biochemistry and bioinformatics in order to provide coherent and high-quality data that enable discovery of robust and biologically significant metabolic findings. In this protocol article, we introduce notame, an analytical workflow for non-targeted metabolic profiling approaches, utilizing liquid chromatography-mass spectrometry analysis. We provide an overview of lab protocols and statistical methods that we commonly practice for the analysis of nutritional metabolomics data. The paper is divided into three main sections: the first and second sections introducing the background and the study designs available for metabolomics research and the third section describing in detail the steps of the main methods and protocols used to produce, preprocess and statistically analyze metabolomics data and, finally, to identify and interpret the compounds that have emerged as interesting.

Migrating from partial least squares discriminant analysis to artificial neural networks: a comparison of functionally equivalent visualisation and feature contribution tools using jupyter notebooks.

INTRODUCTION: Metabolomics data is commonly modelled multivariately using partial least squares discriminant analysis (PLS-DA). Its success is primarily due to ease of interpretation, through projection to latent structures, and transparent assessment of feature importance using regression coefficients and Variable Importance in Projection scores. In recent years several non-linear machine learning (ML) methods have grown in popularity but with limited uptake essentially due to convoluted optimisation and interpretation. Artificial neural networks (ANNs) are a non-linear projection-based ML method that share a structural equivalence with PLS, and as such should be amenable to equivalent optimisation and interpretation methods. OBJECTIVES: We hypothesise that standardised optimisation, visualisation, evaluation and statistical inference techniques commonly used by metabolomics researchers for PLS-DA can be migrated to a non-linear, single hidden layer, ANN. METHODS: We compared a standardised optimisation, visualisation, evaluation and statistical inference techniques workflow for PLS with the proposed ANN workflow. Both workflows were implemented in the Python programming language. All code and results have been made publicly available as Jupyter notebooks on GitHub. RESULTS: The migration of the PLS workflow to a non-linear, single hidden layer, ANN was successful. There was a similarity in significant metabolites determined using PLS model coefficients and ANN Connection Weight Approach. CONCLUSION: We have shown that it is possible to migrate the standardised PLS-DA workflow to simple non-linear ANNs. This result opens the door for more widespread use and to the investigation of transparent interpretation of more complex ANN architectures.