DIVA

Clinical variant interpretation has a bottleneck that pathogenicity scores alone do not solve. Tools such as AlphaMissense return a single number for how damaging a missense change is, but a diagnostic report needs to say which disease a variant is implicated in — and many proteins are pleiotropic, causing several clinically distinct disorders depending on how they are altered. ClinVar held more than 446,000 variants of uncertain significance as of 2022, and a deleteriousness score alone does little to resolve them.

DIVA (disease-specific variant pathogenicity prediction) addresses that gap by predicting a specific disease alongside a deleteriousness probability for every missense variant. Developed by Yilin Liu, David N. Cooper, and Haiyuan Yu at Cornell University and Cardiff University, it treats variant interpretation as a cross-modal retrieval problem: protein sequence and free-text disease annotations are embedded by two pretrained language models and pulled into a shared space by a contrastive objective, so that a variant's representation sits near the diseases it actually causes.

Because the disease side is an open text vocabulary rather than a fixed label set, DIVA ranks every disease term for an arbitrary new variant without retraining. The work was posted to bioRxiv in September 2025.

#Key Features

• Joint disease and deleteriousness prediction: Each variant receives both a binary deleteriousness probability and a ranked list of candidate diseases, scored by cosine similarity in the learned joint space.
• Contrastive alignment with hard negatives: Training pulls variants toward their annotated diseases and pushes them away from sampled negatives; hard-negative sampling, which draws negatives from diseases associated with the same protein, measurably outperforms random negatives.
• Local and global disease context: Disease knowledge enters through a 128-residue window centered on the variant, which supplies annotations of nearby pathogenic mutations, concatenated with UniProt "function" text describing the protein as a whole.
• Learnable fusion with AlphaMissense: Protein-language-model deleteriousness logits are combined with AlphaMissense scores using protein-specific weights derived from functional-annotation embeddings and optimized end to end.
• Proteome-scale VUS resource: Predictions for 410,032 ClinVar variants of uncertain significance are browsable and downloadable from a public web portal.

#Technical Details

DIVA encodes the alternate protein sequence with ESM-2 at the 150M-parameter scale and encodes disease text with PubMedBERT, a biomedical-domain BERT. Lightweight projection and aggregation modules fuse the sequence representation with local disease features and global functional features into a single variant embedding; the same PubMedBERT encoder embeds each disease name plus its Mondo ontology definition. PubMedBERT is fully frozen and only the last hidden layer of ESM-2 is fine-tuned, restricted to the first two training epochs. The contrastive loss uses a temperature of 0.07.

Training data comprises 151,887 missense variants — 109,053 pathogenic and 42,834 benign across 10,652 proteins — curated from ClinVar, HGMD, and UniProt, mapped to protein residues with Ensembl VEP, and supplemented with rare gnomAD v2 control variants to balance classes; the split is 80/10/10. The disease vocabulary is normalized against the Mondo ontology. On a held-out test set of 11,621 pathogenic and 11,500 benign variants, and on the AlphaMissense cancer-hotspot set (669 pathogenic, 1,693 benign), deleteriousness classification is comparable to or better than AlphaMissense and ESM-2 zero-shot. For disease specificity, DIVA separates true disease labels from random labels with a Cohen's d of 1.962, and from same-protein distractor diseases with 1.875. Larger ESM-2 encoders gave only modest gains at substantially higher training cost.

#Applications

DIVA targets clinical and research variant interpretation, where a curator holding a VUS wants a ranked shortlist of candidate disorders rather than a bare damaging/benign call. Applied across ClinVar, it flags 97,964 VUS as deleterious at an estimated 10% FDR and 52,048 at 5%, and assigns confident disease predictions to 164,294 variants. A relative confidence score — the percentile of a submitted disease's score across the vocabulary — lets curators sanity-check existing annotations, and the ranked output generates hypotheses for pleiotropic genes such as NKX2-5, where different mutations produce distinct cardiac phenotypes. Code, trained weights, and example data are released under GPL-2.0, and the web portal serves predictions without local installation.

#Impact

DIVA pushes a field that has largely reported disease-agnostic damaging scores toward the disease-resolved output clinical genetics actually consumes, and it does so by reusing frozen biomedical foundation models rather than training a specialist from scratch — keeping training cost low enough for a single workstation GPU. Its limits are real: the model is confined to missense variants in protein-coding regions, it depends on AlphaMissense for part of its deleteriousness signal, and its disease vocabulary inherits the annotation biases of ClinVar, HGMD, and UniProt, so poorly curated genes are poorly served. Separation between true and same-protein distractor diseases is also narrower on VUS than on the labeled test set, so the hardest real-world regime remains open. The preprint has not yet been peer-reviewed.