VB-Atlas Training

Embeddings

VB-Atlas

The map of causal human biology.

VB-Atlas learns embeddings from tens of thousands of association studies to create a map of relationships between genes and phenotypes. It finds connections in the data itself, not by mining literature or what's already been published.

Query any gene and instantly find others with similar association profiles. Explore relationships that span disease areas and molecular mechanisms. Navigate biological space in ways that were never possible with isolated studies.

Dense embeddings make massive datasets immediately accessible to language models, enabling powerful AI-driven discovery.

Applications

  • Identify novel genes with profiles similar to known disease drivers
  • Find druggable targets analogous to undruggable ones
  • Hypothesis-free navigation through association networks

Example: Novel Target Discovery

Entyvio (vedolizumab) is an α4β7 integrin antagonist for IBD. We mapped the ITGA4 subunit in VB-Atlas to identify new IBD targets: 5 of the top 10 similar gene clusters hit known IBD GWAS regions, including CEBPB, a novel target with clear QTL-based variant-to-gene mapping and existing oncology inhibitors that could be repurposed.

VB-Atlas visualization

PCSK9 in VB-Atlas

Querying PCSK9 reveals genes with similar association profiles across thousands of studies. Each cluster represents genes that behave similarly across phenotypes, potential alternative targets or pathway members.

Predictive Models

VB-Predict

Molecular models that see what clinicians can't.

VB-Predict creates molecular models of disease using proprietary transcriptomic, proteomic, and metabolomic data combined with neural networks.

These models predict clinically relevant traits from multi-omic data, dramatically expanding the therapeutic area scope of every study. Apply a model trained in one cohort across entirely different populations. Identify the molecular features most predictive of disease outcomes. Discover novel biomarkers and targets.

Applications

  • Predict clinically relevant traits from multi-omic data
  • Identify molecular features most predictive of disease outcomes
  • Discover novel biomarkers and therapeutic targets

Example: Fibrosis Prediction

The ELF score is a clinically validated marker of fibrosis, but expensive to measure. We built a model to predict ELF from proprietary proteomic data, applied it to 38,000 UK Biobank participants, ran a GWAS, and replicated known fibrosis loci while discovering new ones.

VB-Predict visualization in Locus Explorer

Replicating Known Fibrosis Loci

Locus Explorer comparing a liver cirrhosis GWAS from FinnGen (top) with a GWAS of predicted ELF score from UK Biobank proteomics (bottom). The well-known PNPLA3 missense variant is highlighted, validating that VB-Predict captures real biology.

Ready to get started?

Learn how Inference can accelerate your research and transform your drug discovery pipeline.

Contact Us →