BIO-AI · SINGLE-CELL GENOMICS · 2026
See which cells went wrong. Faster than ever.
Embed patient cells with foundation models, map them against a healthy reference, and surface the differences that matter — at single-cell resolution.
Powered by
Helical AIFour views, one workflow
Reference Atlas
2,638 healthy PBMC cells embedded by Geneformer, the same model Helical ships in their SDK. UMAP-projected, color-coded by 8 immune cell types.
Disease Projection
5,000 COVID-19 immune cells from Wilk et al. 2020 projected into the healthy reference. Distance-to-manifold quantifies abnormality, per cell.
Distance Analysis
Mean distance-to-healthy bucketed by cell type and disease severity. The heatmap surfaces which immune populations diverge most under viral load.
Model Disagreement
Geneformer and GenePT, two foundation models trained on different objectives, disagree about which cells are normal. Per-cell percentile-rank disagreement maps where they diverge.
How it works
Load
Healthy PBMC reference + COVID query data, both standard .h5ad AnnData files.
Embed
Run Geneformer + GenePT via the Helical SDK. Get 512-dim and 1536-dim cell-level embeddings.
Compare
Project disease into healthy. Measure distance, surface where the models disagree.