We build models that learn how treatment moves a cell’s transcriptomic state — trained on paired pre- and post-treatment patient data, not cell lines.
scDEFT predicts drug treatment effect at the transcriptomic level. It is trained on public paired pre- and post-treatment single-cell data, so it learns how a drug transforms a cell’s state rather than inferring it from cell-line proxies.
The distinctive part is not the forward prediction. Given a patient’s pre-treatment cells, the model can substitute a different drug and predict the post-treatment state that patient would have reached — turning a single cohort into a comparison across therapies.
Every prediction is attributable back to specific gene programs, so the output is a testable hypothesis rather than a score.
Model performance is measured against the reproducibility ceiling of the data itself — how well a measurement agrees with a repeat of the same measurement — rather than against zero. Reporting the fraction of achievable headroom captured, instead of a raw correlation, is what keeps the claims honest about where signal ends and noise begins.
Reported in the scDEFT preprint. Current work extends the approach to oncology cohorts.
Sidney Street AI makes the scDEFT models available as an interactive application for exploring treatment response in single-cell cohorts.
Murthy Devarakonda
murthy@mvdresearch.ai
Open to collaborations with clinical and pharmaceutical groups holding paired pre/post treatment single-cell data.