Multi-Omic Joining

Multi-omic joining is the act of putting measurements from different molecular layers into one table or one model, aligned on shared keys — usually a gene or protein identifier, a genomic coordinate, and a time point. “Omics” means a layer measured comprehensively rather than one assay at a time: DNA (genomics), RNA (transcriptomics), protein (proteomics), small molecules (metabolomics). “Integration” is what you do after you have more than one of them on the same subject.

How it works

There are three levels of joining, and they fail differently.

Early integration concatenates features into one matrix. 20,000 RNA columns, 3,000 protein columns, 40 chemistry values. The problem is scale: RNA in log2 CPM has a variance structure nothing like NPX or mg/dL, and any distance-based or penalized method will be dominated by whichever block has the most columns. If you do this, standardize per block and down-weight by sqrt(n_features) in that block.

Intermediate integration learns a shared latent space and keeps the blocks separate. MOFA and similar factor models decompose variance into factors, then report how much variance each factor explains in each layer, which tells you whether a signal is transcriptional, proteomic, or both. Kernel methods compute a similarity matrix per layer and combine kernels; deep generative approaches (autoencoders, GANs) learn a joint embedding and can impute one layer from another. A 2025 technical review walks the full progression from classical statistical methods to deep generative models and is the single best map of the method space ¹. Generative adversarial approaches have been used specifically to fill in a missing omics layer for samples where only one was measured ².

Late integration analyzes each layer separately and merges conclusions: rank genes by RNA, rank proteins by abundance, intersect. It is the weakest statistically and the easiest to interpret. For n=1 longitudinal data it is often the right starting point.

Most published integration assumes many subjects and few time points. Your data is the reverse.

In your own data

The files you will be joining:

• WGS: a VCF or gVCF (GRCh38), variants keyed by CHROM:POS:REF:ALT. Annotate with VEP or bcftools csq to get Ensembl gene IDs.
• RNA-seq: a counts matrix from salmon or featureCounts, rows as ENSG00000141510.17 with version suffixes, plus a quant.sf with TPMs.
• Proteomics: Olink NPX (log2, relative, plate-normalized) or mass-spec intensities, keyed by UniProt accession or gene symbol.
• Blood panels: LOINC codes if you are lucky, free-text analyte names if you are not.
• CGM: 5-minute interval readings, an epoch column and a mg/dL column.

Concrete steps we would take. Strip Ensembl version suffixes before joining (sed 's/\..*//') or you will silently lose every gene where the annotation release differs between your quantification and your reference. Map protein IDs to genes through UniProt’s idmapping file, not through gene symbols: symbols are ambiguous and get auto-corrected by spreadsheets. Check the join yield explicitly. If you merge 3,000 Olink proteins against a 20,000-gene RNA matrix and get 1,900 rows, ask where the other 1,100 went before you interpret anything.

Then resample onto a shared time axis. CGM is every 5 minutes, blood draws are every few months, RNA is whenever you sequenced. Pick the coarsest layer as your grid and summarize the finer ones into windows (mean glucose, glucose CV, time above 140 mg/dL for the 14 days preceding each draw). Do not interpolate proteins across a six-month gap.

Expect RNA and protein to disagree. Transcript and protein abundance correlate around r = 0.4 across genes, and joining them is a way to see where post-transcriptional regulation is doing something, not a validity check on either measurement ³.

Common mistakes: joining on gene symbol; ignoring that one layer’s batch is confounded with your time axis (if every autumn draw ran on one plate, plate and season are the same variable); treating zeros in proteomics as absence rather than below-LOD ⁴.

Limitations

Sample size is the binding constraint. Most integration methods derive their power from variation across subjects. With one person you have variation across time, which is fewer effective degrees of freedom than the feature count by orders of magnitude, so any latent-factor model will overfit unless you restrict it hard. Unsupervised methods are also sensitive to normalization and to the number of latent factors chosen, and results shift with those choices ⁵.

Integration produces associations, not mechanism. Reviews of post-integration target discovery find reproducibility and causal evidence to be the weak links, not the modeling ⁶. A joined table that shows a protein moving with a glucose window is a hypothesis for the next sampling round. Anything that looks clinically relevant belongs in front of a physician, who can order the confirmatory test on a validated assay.

Woolf Software builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.