Metabolomics
Matched omics layers merge in MOFA2 or DIABLO to surface cross-layer drivers [1], with factor-weight tables, R scripts, and a Methods draft.
| Category | Details |
|---|---|
| Supported platforms / instruments | Metabolomics: Thermo Orbitrap, SCIEX TripleTOF, Agilent, Bruker, Waters LC-MS (raw or peak tables). Transcriptomics: Illumina bulk RNA-seq, microarray. Proteomics: MaxQuant, DIA-NN, or MSstats protein matrices. Metagenomics: shotgun or 16S functional/taxonomic tables on request |
| Input requirements | >15 matched samples with overlapping omics layers for MOFA2 [2]; ≥3 biological replicates per condition recommended for supervised DIABLO contrasts; consistent sample_id across matrices; missing omics per sample documented in manifest |
| Reference builds supported | Human GRCh38 (GENCODE v44 / Ensembl 110); mouse GRCm39 (GENCODE vM33 / Ensembl 110); UniProt Swiss-Prot reviewed proteomes; HMDB 5.0 for metabolite annotation |
| Primary tools (with versions) | MOFA2 1.12; mixOmics (DIABLO) 6.28; limma 3.60; MetaboAnalystR 4.0 [6]; xcms 4.2.3; WGCNA and pairwise correlation modules on request. Versions pinned per project |
| Typical turnaround time | 4-8 weeks (two to three matched omics layers, one cohort); longer for raw preprocessing of all layers or bespoke ML extensions. Confirmed at kickoff |
| Deliverable formats | Factor score and weight tables (.csv, .rds); cross-omics correlation matrices; PDF/SVG factor, heatmap, and network plots; HTML QC report; commented R scripts with renv lock; Methods draft |
| Key cited best-practice reference | Hasin et al. [1], Genome Biology; Argelaguet et al. [3]; Singh et al. [7], Bioinformatics |
| Custom / bespoke analysis | Non-standard omics blocks, integration methods, output formats, or analyses beyond standard MOFA2/DIABLO scoped at kickoff, for example metagenomics functional profiles, spatial layers, or client-specified ML models |
Multi-omics integration models two or more omics feature matrices measured on shared samples. It learns latent factors or sparse components that explain coordinated variation across layers [3]. Single-layer differential testing plus pathway overlay interprets each layer on its own; integration instead models cross-layer covariance so metabolite, transcript, and protein shifts can be read together [4]. Lloyd-Price et al. [5] profiled 132 inflammatory bowel disease subjects with metagenomics, metabolomics, and host transcriptomics, linking microbial and host molecular shifts to disease activity. Pepkio starts from matched matrices or scopes per-layer preprocessing first. Custom block definitions are agreed at kickoff. See the multi-omics integration glossary.
Use multi-omics integration when matched samples cover two or more molecular layers and the question is about cross-layer drivers. A single available layer, or sample IDs that do not align, is a poor fit [1].
| Approach | Best for | Limitations | Approximate cost range |
|---|---|---|---|
| Multi-omics integration (MOFA2 / DIABLO) | Matched samples; cross-layer biomarker panels; unsupervised disease subtyping or supervised outcome prediction | Needs aligned sample IDs; MOFA2 needs >15 samples; per-layer QC and normalization before integration | Quote-based; higher bioinformatics effort than single-omics [1, 8] |
| Single-omics analysis + pathway overlay | Exploratory interpretation when only one layer is measured | No statistical coupling across layers; concordance is post hoc | Lower per-project bioinformatics cost |
| Correlation or WGCNA-only integration | Hypothesis-light screening across two layers | No shared latent factor structure; multiple-testing burden at high dimensionality | Moderate; scoped separately when MOFA/DIABLO is not justified |
Pepkio confirms the biological question, omics blocks in scope, outcome labels (if supervised DIABLO is planned), and that sample_id values match across layers. Samples missing entire omics blocks are flagged. Partial missingness within a block is documented for MOFA2 [2]. Confounded batch-by-condition designs are flagged before preprocessing [1].
Tools used: Custom validation scripts
Output: integration_manifest.csv with sample IDs, omics availability, condition, batch, and design notes
Each layer is audited for orientation (samples × features), missing-value fraction, and metadata completeness. Metabolomics TIC, RNA-seq library complexity, and proteomics detection rates are summarized before normalization [11, 12].
Tools used: MetaboAnalystR 4.0 QC module [6]; custom R diagnostics
Output: sample_qc_summary.csv; per-layer QC flags and exclusion log
When raw LC-MS is in scope, Pepkio aligns peaks, normalizes with QC-RLSC or PQN, and filters low-abundance features [12, 13]. Documented client peak tables are accepted.
Tools used: xcms 4.2.3; MetaboAnalystR 4.0 [6]
Output: metabolite_matrix_normalized.csv; metabolomics PCA and missingness plots
Count matrices undergo size-factor normalization and VST before integration; microarray data is log-transformed and batch-diagnosed [2, 14]. HVG filtering follows MOFA2 guidance when RNA is a primary view [2].
Tools used: limma 3.60
Output: transcript_matrix_vst.csv; hvg_transcript_list.csv
Protein matrices from MaxQuant, DIA-NN, or MSstats exports are log-transformed, filtered by detection rate, and missing-value patterns are profiled before imputation decisions [11]. Low-detection proteins are removed with thresholds logged in the QC report.
Tools used: Perseus-style filtering scripts; limma 3.60
Output: protein_matrix_log2.csv; detection-rate and missingness heatmaps
When batch is not confounded with condition, Pepkio regresses batch per omics block with limma removeBatchEffect before integration, not as a MOFA covariate, which MOFA2 developers advise against [2]. PCA before and after correction is used to check condition-batch separation.
Tools used: limma 3.60
Output: Batch-corrected matrices; PCA plots colored by batch and condition
Feature counts per view are filtered to comparable variance scales so large modalities do not dominate factors [2]. Sample order and column names are synchronized across blocks. Samples that fail per-layer QC are excluded from all views.
Tools used: Custom R harmonization scripts
Output: Aligned multi-block input list (.rds); feature_filter_log.csv
Unsupervised discovery uses MOFA2. Supervised outcome prediction with labeled cohorts uses mixOmics DIABLO block sPLS-DA [3, 7]. Factor count and DIABLO component sparsity are tuned with cross-validation when sample size permits. Concatenation-only baselines are not the primary deliverable [7].
Tools used: MOFA2 1.12; mixOmics (DIABLO) 6.28
Output: MOFA2_model.hdf5 or DIABLO object (.rds); variance-explained and cross-validation metrics
Factor scores are correlated with metadata covariates. Top-weighted features per view are exported for biological interpretation [3]. Cross-omics feature-feature correlations within significant factors are computed and visualized.
Tools used: MOFA2 1.12; mixOmics 6.28; corrplot
Output: factor_scores.csv; factor_weights_<view>.csv; cross-omics correlation heatmaps
Metabolite and gene set enrichment tests pathway overlap across layers. Concordant and discordant pathways are tabulated [4]. Final scripts, figures, QC report, and Methods draft cite exact software versions and preprocessing parameters.
Tools used: MetaboAnalystR 4.0 [6]; clusterProfiler on request
Output: pathway_concordance.csv; final deliverable bundle; Methods draft
MOFA2 is described as needing more than 15 matched samples with overlapping omics measurements to produce stable latent factors [2]. Supervised DIABLO projects benefit from multiple biological replicates per condition, typically ≥3, and larger cohorts improve cross-validation of sparse component selection [7]. Exact minimums are confirmed at kickoff based on your block count and outcome design.
Yes, with documented scope limits. Samples that fail per-layer QC are excluded from all views. MOFA2 handles missing values within a view without imputing them into the likelihood [2]. If an entire omics block is unreliable, Pepkio discusses reducing to two-layer integration or re-measurement before training the model.
Metabolomics: Thermo .raw, SCIEX .wiff, Agilent .d, Bruker .d, Waters .raw, mzML, or MS-DIAL/MZmine peak tables when documented at kickoff. Transcriptomics: Illumina bulk RNA-seq count matrices or .bam-derived counts; microarray when scoped. Proteomics: MaxQuant proteinGroups.txt, DIA-NN reports, or MSstats matrices. Metagenomics: taxonomic or functional abundance tables on request. Preprocessed .csv/.tsv matrices are accepted when preprocessing parameters and sample IDs are supplied.
Matched-sample projects with two to three preprocessed layers often finish in 4-8 weeks from data receipt. Raw preprocessing of all layers, multiple contrasts, or bespoke ML extensions can take longer. Timelines are confirmed at kickoff with milestone check-ins.
Batch is diagnosed per omics layer with PCA and correlation checks. When batch is not confounded with condition, Pepkio regresses batch with limma removeBatchEffect on each layer before MOFA2 or DIABLO training [2, 14]. Fully confounded designs are flagged at kickoff. Integration proceeds only when a defensible contrast remains.
Yes. You retain full ownership of all code, scripts, and results. Pepkio delivers commented R scripts with renv.lock files listing exact package versions, organized by pipeline stage with a README. Objects use standard .csv and .rds formats readable in R. Private Git delivery or agreed file transfer is available.
Yes. Checkpoint reviews occur after per-layer QC, after batch diagnostics, and after factor or component interpretation. You can review contrast definitions, normalization choices, factor count, and feature-selection thresholds within agreed scope. A dedicated scientific contact leads the project.
Support covers clarification of integration methods, preprocessing parameters, factor interpretation, and minor figure or table revisions within agreed scope (typically ≤20% of deliverables). Pepkio drafts Methods and Supplementary text for analyses we performed. Substantial new omics layers or contrasts requested by reviewers are scoped as separate milestones.
No. Pepkio is a fee-for-service provider and does not require co-authorship unless that is discussed in advance. Acknowledgment of bioinformatics support in the Acknowledgments section is the usual practice. Co-authorship is considered only when Pepkio scientists make substantial intellectual contributions beyond routine analysis.
Use MOFA2 for unsupervised shared-variation discovery without a primary outcome label [3]. Use DIABLO when you have a categorical outcome and want a sparse multi-omics biomarker panel. Singh et al. [7] reported a balanced error rate of 17.9 ± 1.9% for a TCGA breast cancer panel under cross-validation. The Methods draft states which method was used and why.
Yes. MOFA2 ignores missing values in the likelihood without hidden imputation [2]. A substantial fraction of samples must share each omics pair, and sample order must align across available blocks. Pepkio documents per-sample omics availability in integration_manifest.csv before training.
Yes. Pepkio accepts client-supplied normalized matrices when preprocessing parameters, feature identifiers, and sample IDs are documented. We audit orientation, missingness, and batch structure before integration. Re-normalization from raw data is scoped separately when QC flags inconsistencies [15].
Tell us:
We'll tell you:
Contact us to start with a free consultation. Need everyday bench calculators? Try our free lab tools.
