Models built around crude direct-to-biology assays pushed a ROS1 series past its training-set potency and selectivity frontier
- Assay-aware models treat single-point screens, crude dose-response potency ratios and purified-compound potency as linked stages, trained jointly in one network.
- The joint model beat separate networks in retrospective tests, with the biggest gain for crude-assay selectivity ranking.
- In a prospective round, models chose dozens of compounds aimed at ROS1 over the TRKB anti-target; after synthesis and testing, several exceeded the training set's best potency-selectivity trade-off.
💡 Points Based on the abstract
The prospective round is the result that counts: some compounds did beat the training set's best trade-off. Still missing is the hit rate among those dozens versus a single-task model or plain medchem picks. The practical test on your own chemistry is how closely crude-assay readouts track purified potency.
D2ML: Models for Direct-to-Biology Small Molecule Screening ↗