Science
How the method works.
A plain-language walk through the response signal behind Responder Atlas how it works, and where it stops working.
The method
Responder Atlas: a response signal from public pharmacogenomic data, and the method is inspectable.
Responder Atlas is our platform, built on public pharmacogenomic data. For every drug in the DepMap/PRISM training set, we know which cell lines responded and which didn't, and what those cell lines look like at the molecular level. That gives us a map from molecular structure of a cancer cell to how strongly a compound hits it.
When you submit a new molecule, we compute where it sits in that map by chemical structure and read the response signal off its neighbourhood, giving a ranked score across seventeen cancer tissue types plus a confidence tier that reflects how similar your molecule is to what the model has actually seen.
The output is a directional triage signal. Every prediction is accompanied by its provenance: whether the compound was measured in DepMap or inferred from structure, which training-set drug it's most similar to, and how tight that similarity is. Nothing is hidden behind a scalar.
Rigor
The model recovers what's already known.
Tested on four targeted oncology drugs, our unsupervised responder signal ranked each drug's canonical driver first out of hundreds of candidate genes: vemurafenib→BRAF, alpelisib→PIK3CA, selumetinib→KRAS. Categorical recovery on clean-driver cases, manuscript in preparation.
Drug
Vemurafenib
Rank 1 geneBRAFDrug
Alpelisib
Rank 1 genePIK3CADrug
Selumetinib
Rank 1 geneKRAS
Boundaries
Where the method stops working.
Two paths, two different limits. A compound already screened in DepMap returns its own measured cell-line response. Those numbers are observed rather than modelled, and no imputation error applies to them. A compound absent from DepMap is imputed from chemical structure, and that estimate carries real error. Every query states which path produced it.
For the imputed path, cross-tissue selectivity, which means separating the indication where a drug works from the ones where it doesn't, is the hardest axis. It sits at the structural limit of chemistry-only methods. We say so plainly: per-indication scaffold-LDO performance lands around r ≈ 0.39. Useful for triage, not for committing to a single tissue on structure alone. That figure is the limit of the imputed path specifically; it does not describe a measured result.
Both paths share the limits that matter most, and measured data does not soften them. The signal is grounded in cell-line response rather than patient outcomes. A measured cell line is still a cell line, not a patient. It's a hypothesis-generating triage layer, not a clinical or efficacy prediction. Confidence tiers are the honest handle for the imputed path: elevated when your molecule is close to training, exploratory when it isn't, and we surface both.