We build models that learn how drugs act on human biology — trained on patient data, benchmarked carefully, and built to produce hypotheses you can test.
Most models of drug response are trained on cell lines and evaluated against zero. We do neither.
Our models learn from paired pre- and post-treatment patient data, so they capture how a treatment actually moves human cells rather than a proxy for it. And we measure performance against the reproducibility ceiling of the data itself — the agreement between a measurement and a repeat of that same measurement. It is the only way to know where signal ends and noise begins.
scDEFT, below, is the first model in this line. More are in development.
scDEFT predicts drug treatment effect at the transcriptomic level, treating a drug as a conditioning operator on cell representations. It predicts forward — how a treatment moves a cell’s state — and counterfactually, substituting a different drug to predict the state a patient would have reached instead.
The distinctive part is neither. It is the backward attribution. scDEFT ranks the dimensions of its learned representation by how strongly they separate eventual responders from non-responders, then names each one by correlating it against measured gene expression — real counts, not model output — and runs a variance decomposition to establish whether it is a genuine program (genes changing within cells) or merely a shift in cell-type composition.
What comes out is a named set of biological programs: the ones a therapy engages in patients it helps, and the ones that remain unaddressed in patients it does not. On the IBD atlas this independently recovered published cell-state remission signatures — without being directed to look for them.
A backward stage ranks the latent dimensions separating responders from non-responders and maps them to genes, under a cell-composition control — surfacing candidate targets and the partners worth hitting alongside them.
Responders can be separated from non-responders before treatment begins, from baseline cells alone — the question a trial design actually has to answer.
A compound that failed in an unselected population may still work in a definable subgroup. The counterfactual lets you ask which patients would have responded, and why, from cohorts that already exist.
Reported in the scDEFT preprint. Current work extends the approach to oncology cohorts.
Sidney Street AI makes our models available as an interactive application for exploring treatment response in single-cell cohorts.
Murthy Devarakonda, PhD
murthy@mvdresearch.ai
Open to collaborations with clinical and pharmaceutical groups pursuing target and co-target nomination, patient stratification, or drug rescue.