COMICAL

A genome-wide association study asks one question at a time: does this variant move this trait? Run it against brain MRI and you get a list of variants per phenotype, never a statement that a particular variant and a particular brain region belong together. Several loci acting on several structures at once — the shape of a polygenic disorder — falls between the tests.

COMICAL (contrastive multiomics association learning) reframes that as a pairing task. Single nucleotide polymorphisms and imaging-derived phenotypes (IDPs) are treated as two modalities describing the same person, each tokenized and pushed through its own transformer encoder, and trained with the symmetric contrastive loss from CLIP to match a genotype against the imaging measurement it belongs with. A disease acts only as a bridge for assembling positive pairs — a GWAS Catalog SNP and an IDP mapped to the same disorder by a phenome-wide association scan — so the training signal is "do these two go together", and the embedding space ends up encoding genotype-phenotype structure rather than a diagnosis.

It was built by IBM Research with a collaborator at Rensselaer Polytechnic Institute and published in Bioinformatics Advances. Where single-modality genomic models such as DNABERT and Enformer learn representations inside one data type, COMICAL's contribution is cross-modal: one pretrained checkpoint whose encoders are frozen and reused, with only a small MLP head trained per downstream target.

#Key Features

• Many-to-many associations: The contrastive objective scores every SNP against every IDP in a batch, so one variant can link to several brain measurements and vice versa — structure a per-marker regression cannot represent.
• Tokenizers built for the substrate: Genotypes are recoded into a 51-token vocabulary of mutation states concatenated with a learned SNP-ID embedding acting as the positional encoding; continuous IDPs use piecewise linear encoding, preserving where a value sits in its range.
• Frozen-encoder transfer: The checkpoint ships with a loader that freezes both encoders and trains only the head, so one representation serves disease prediction, cross-disorder prediction and risk scoring.
• Interpretable latent space: SNP embeddings cluster into eight groups, most of them biologically coherent — APOE and TMEM175 with Alzheimer's disease, LRRK2 and OR4C3 with Parkinson's disease — while IDP embeddings cluster by brain function.
• Multiomics risk score: The same joint embeddings feed a regression head that reproduces UK Biobank enhanced polygenic risk scores from imaging as well as genotype.

#Technical Details

In the released checkpoint, both encoders are two-layer transformers with model dimension 64, four attention heads and a 128-wide feed-forward layer — the paper's text gives 32 and its tuned hyperparameters 64 — totalling roughly 1.2 million parameters in 4.8 MB. Pretraining used 40,426 UK Biobank participants of European ancestry passing genomic quality control, 154 IDPs from T1 structural brain MRI, and 5,603 GWAS Catalog variants tied to eight neurological disorders. Pairing at the top 1% association-strength threshold yields 15,442,732 SNP-IDP pairs, split 70/20/10; trained on the top 0.5% of variants, the model recovered held-out pairs with 97.3% accuracy, rising to 100% after Ray Tune search, and roughly half its significant SNPs had an independent association in ENIGMA consortium summary statistics at P < 5x10^-4.

Downstream, embeddings from the frozen encoders plus age, sex and 40 genetic principal components feed a single-hidden-layer classifier. Against a matched MLP on raw features, COMICAL improves AUC on mood disorder (0.928 vs 0.352), autism spectrum disorder (0.826 vs 0.417), schizophrenia (0.736 vs 0.313) and ADHD (0.683 vs 0.599), and sits at or below baseline on Alzheimer's disease, stroke, multiple sclerosis and Parkinson's disease, whose case counts in this disease-agnostic cohort range from 12 to 489. Schizophrenia was held out of pair-making, so that result is transfer to an unseen outcome. Risk-score regression against enhanced polygenic scores reaches R-squared near 0.70 across five disorders, including when the head is trained on one and evaluated on another.

#Applications

COMICAL suits imaging-genetics groups in large biobanks who want candidate variant-to-region links rather than two separate association tables — an ADHD-associated variant paired with left amygdala grey matter volume, matching an ENIGMA finding no ADHD GWAS had reported. Because the encoders stay frozen, attaching a new clinical outcome costs only a head, and the risk score offers a stratification variable combining genotype with structural imaging. The framework is modality-agnostic by construction: any two omics layers that can be tokenized and paired through a shared mediator fit it.

#Impact

COMICAL demonstrates that CLIP-style contrastive pretraining transfers to paired omics layers, and the Apache-2.0 code with a fixed checkpoint makes that claim checkable rather than rhetorical. Its limits are worth stating plainly. The encoders are small, and the checkpoint's input space is pinned to the UK Biobank feature panel — a fixed SNP index and IDP index — so another cohort must reproduce that indexing rather than simply load weights. Every evaluation is internal to UK Biobank, and pair-making can generate spurious pairs when a SNP and an IDP share a disease label without sharing a mechanism. What it establishes is the shape of the idea: association discovery and downstream prediction from one disease-agnostic multiomics representation.