henhaobio
Sign in

bioRxivBioinformaticsPreprint

Illuminating dark genes and dark metabolites: calibrated functional association across kingdoms with autoencoders and correlation

Yannick MahlichiD, Dylan H. RossiD, Lummy Maria Oliveira MonteiroiD, Sutanay ChoudhuryiD, William C Nelson, Jason McDermottiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Motivation Despite continuing advances in sequencing and computational function prediction, large parts of gene, protein, and metabolite space remain functionally undetermined. Most function assignment is based on homology searches and annotation transfer from well-studied proteins, but this approach fails to leverage available experimental omics data generated via technologies like mass-spectrometry. Results To address this gap, we have developed the VaLPAS (Variation-Leveraged Phenomic Association Screen) framework, a Python package that provides a user-friendly platform to discover associations between expression patterns of genes, proteins, and metabolites in multi-omic datasets. VaLPAS uses various statistical and learning methods, from simple statistical methods (correlation) to a sophisticated autoencoder, to shed light on the functional dark matter of protein space using 'guilt by association' with molecules of known function. We have used VaLPAS to examine transcript, protein, and metabolite data from seven organisms across two kingdoms using two different benchmarks. We present results demonstrating the ability of VaLPAS to make high-confidence predictions of the functional associations of a subset of genes and proteins. Furthermore, we show that VaLPAS can provide insight into the functional relevance of detected but unidentified metabolites by cross-omics association, a unique feature of this approach. This ability of cross-omics association enables VaLPAS to serve as a tool for generating hypotheses of functional interaction that can be validated with targeted experiments, representing a significant step forward in elucidating dark functional matter. Availability and implementation VaLPAS is hosted on github (https://github.com/PNNL-Predictive-Phenomics/valpas/).

In-depth summary

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 2 (8 Oct 2026) · CC BY 4.0

VaLPAS pairs a bi-directional autoencoder with correlation for omics guilt-by-association with calibrated confidence; in L. starkeyi its autoencoder linked 176 unidentified metabolites to known-function proteins at ≥50% confidence.

Function assignment relies mostly on homology transfer; even in E. coli and S. cerevisiae 20% to 35% of protein-coding genes remain poorly characterized, and an estimated over 95% of metabolomics peaks cannot be given a structure. The paper says no prior work had systematically compared correlation and autoencoder guilt-by-association across organisms, omics types and sample sizes with interpretable confidence, or tested linking unknown metabolites to known protein pathways.

Key results

  1. ShownAcross four bacterial transcriptomics compendia, AUROC against STRING relationships rose with sample number and the autoencoder generally outperformed Pearson correlation, though accuracy was modest and differed by species.

    AUROC was above random (0.5) but modest; exact values appear only in Figure 2. Against KEGG module co-membership the trend was weaker, and Pearson matched or slightly beat the autoencoder for E. coli at 200 samples. Neighborhood-only STRING evidence generally scored best. In P. putida both methods kept over 50% performance with 75% of samples scrambled. · 3.1 Benchmarking Pearson and autoencoder for…; Figure legends

  2. ShownScore rank tracked confidence: top-ranked pairs had the highest positive predictive value (PPV) against STRING for both methods. With 200 samples per species, both gave over 1000 predictions at ≥50% confidence in every bacterium except S. elongatus.

    PPV was plotted against prediction rank because autoencoder scores were compressed for E. coli. Pearson led for E. coli; the autoencoder was equal for S. typhimurium and better for P. putida and S. elongatus. With 20 P. putida samples, confidence was above 70% for 100 predictions and above 30% for up to 1000. · 3.1 Benchmarking Pearson and autoencoder for…; Figure legends

  3. ShownIn S. cerevisiae transcriptomics the authors report the autoencoder was significantly better than Pearson at small sample sizes (down to 11), with Pearson matching or overtaking it only at 200 samples.

    AUROC against STRING across subsample sizes (Figure 4A); a KEGG benchmark gave similar results (Figure S5). At 200 samples both methods gave high-confidence associations at the top score ranks (Figure 4B). The text gives no numeric AUROC values and names no statistical test. · 3.1 Benchmarking Pearson and autoencoder for…; Figure legends

  4. SuggestedUsing 321 P. putida transcriptomes, the VaLPAS autoencoder linked 139 of 1,510 dark genes (no KEGG orthology, no TIGRFAMs hit by HMMER) to annotated partners at confidence above 70%; Pearson correlation yielded more such annotations.

    Confidence came from STRING-calibrated score thresholds (Figure 3). At the 90% level, VaLPAS (method not specified) associated 26 dark genes vs 9 from the guilt-by-association results of Tavis and Hettich, and the two sets did not overlap. The authors note this comparison is indirect because the data and methods differ. · 3.2 Characterization of dark genes in…; 4. Discussion

  5. SuggestedAcross omics types, on matched L. starkeyi proteomics and metabolomics the autoencoder linked 176 unidentified metabolite peaks to proteins of known function at ≥50% confidence, whereas Pearson gave no predictions above 20% confidence.

    Confidence was calibrated against KEGG pathway relationships. The data were 36 samples from nine carbon and nitrogen culture conditions in a single study, and the predictions are pathway-level hypotheses rather than metabolite identities. · 3.3 Characterization of dark metabolites in…; 4. Discussion; 5. Conclusion

Methods, data and reproducibility

VaLPAS (Python) scores feature pairs in sample-by-feature omics matrices by correlation, B-spline mutual information, cosine or Jaccard similarity, or by embedding cosine similarity from a bi-directional autoencoder trained to reconstruct 15%-masked matrices. Predictions are scored by AUROC and rank-wise positive predictive value (PPV) against STRING v12.0 (co-expression excluded) and KEGG modules.

  • E. coli transcriptomics compendium from the iModulon database, analysed at several sample sizes including 200 · Escherichia coli
  • P. putida transcriptomics compendium from the iModulon database; the full set was used for the dark-gene analysis · Pseudomonas putida · n = 321 samples
  • S. enterica transcriptomics from the iModulon database (called S. typhimurium in Results) · Salmonella enterica
  • S. elongatus transcriptomics from the iModulon database · S. elongatus
  • S. cerevisiae transcriptomics from GEO, analysed at sample sizes from 11 up to 200 · Saccharomyces cerevisiae · GSE42527
  • Matched transcriptomics, proteomics and metabolomics (source citation missing in the text) · Rhodotorula toruloides
  • Matched proteomics and metabolomics under different carbon and nitrogen culture conditions (Monteiro et al. 2026) · Lipomyces starkeyi · n = 36 samples from nine experimental conditions
Methods & tools
VaLPAS, STRING v12.0, KEGG REST API, TIGRFAMs v15.0, HMMER v3.3, iModulon database, GEO, scikit-learn, bspline-mutual-information, numpy, pandas, networkx
Software & environment
Python package run from the command line or through an API, with a Jupyter notebook example; it depends on bspline-mutual-information, numpy, networkx, pandas and scikit-learn (no versions given). HMMER v3.3, TIGRFAMs v15.0 and STRING v12.0 were used for benchmarks.
Reproducibility gaps
No compute requirements or dependency versions are stated. The text omits the source citation for the R. toruloides data. Part of the code was generated with Claude Opus 4.7, which the authors say they validated.
Limitations (6)
  • A trained autoencoder applies only to its own input matrix and cannot easily be transferred to another dataset, a restriction the statistical methods share. · stated by the authors
  • Full performance needs experiments spanning many conditions, which favours large compendia. Cross-omics use also needs sample-matched proteomics and metabolomics across multiple conditions, and such datasets are still scarce. · stated by the authors
  • The dark-metabolite predictions come from a limited dataset (36 samples, nine conditions, one study), reach only modest confidence at the top thresholds, and suggest identity only at the pathway level. · stated by the authors
  • Performance varies from dataset to dataset, so the authors advise caution when applying VaLPAS to less well-studied organisms. · stated by the authors
  • No targeted experiment tests any dark-gene or dark-metabolite association. Confidence for dark features is extrapolated from benchmarks built on already-annotated genes and on STRING and KEGG relationships. · noted in this summary
  • The text is inconsistent in places. The Discussion cites A. baumanni results in Figures 2 and 3, but Methods lists no A. baumanni dataset. R. toruloides data are listed, but no R. toruloides result appears in the main text. · noted in this summary

Read this if you want to rank uncharacterized genes or unidentified metabolomics peaks by co-variation with annotated proteins using benchmark-calibrated confidence, or are choosing between correlation and an autoencoder for that.

Citations
0

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISDickinson '22Wiegand '26Belcour '22Huang '25Alabed '26Liu '24
53% semantic matchBioinformatics
Quinn Dickinson, Andreas Aufschnaiter, Martin Ott+1 more·Bioinformatics·Sep 15, 2022· 23 citations