Sci-JEPA v1.0 · production
1.8B-parameter joint-embedding predictive architecture trained on 1.8M disease, protein, and compound triples. The substrate underneath every comparative-oncology query SciRouter answers.
Headline metrics
recall@50 on the cross-species polypharmacology benchmark — i.e. retrieving the right alternative-indication binders for a given target across 12 species.
recall@50 on held-out clinical drug-disease pairs. Higher than contrastive baselines (0.08) by 4.2×.
recall@50 on disease-compound pairs published AFTER the training cutoff — the cleanest test of out-of-distribution generalization we know how to run.
Architecture
We embed disease, protein, and compound modalities into a shared latent space and predict masked tokens of one modality from context in the others. The model learns a richer joint representation than the contrastive baselines that dominate the literature.
Inputs
Backbone
v1.1 roadmap
What you can do with it today
Query "show me human cancers most molecularly similar to canine appendicular osteosarcoma in pediatric populations." Get a ranked list of pediatric OS variants with shared driver genes + accessible models. Co-author the next paper with us.
Stratify your indication-discovery pipeline by cross-species support. A canine target with a strong human-pediatric corollary in Sci-JEPA is a faster path to co-marketing companion-Dx.
Free API access. Cite Sci-JEPA as infrastructure in your manuscripts. Cohort access available pending IRB.
The atlas surfaces inside the Vet Portal (free, credential-gated). Use it to find comparable cases in the literature for whatever shows up on your exam table.