Kimia's D2ML paper: ML models that learn directly from unpurified biology screening data

anshulkundaje · x · 2026-11-16

Kimia Tx's first paper D2ML (arXiv:2610.41070, AI4DD workshop) tackles direct-to-biology screening, where unpurified molecules are tested with uncertain concentrations. Assay-aware models predict activity from single-point screens, relative potency from crude dose-response assays, and absolute potency from purified measurements. A joint cross-stage model beat separate networks, lifting crude dose-response selectivity rank correlation from 0.27 to 0.40; model-guided selection of 84 compounds extended the measured Pareto front for ROS1 potency vs TRKB anti-target activity.

Original post →

More from Research

Research channel →