Reflector Bio

Reflector Bio develops machine learning models that predict the functional, phenotypic response of small molecules from chemical structure rather than predicting binding affinity. Its "bulk foundation model" scores compounds across readouts including viability, gene expression and morphology, and is described as pre-trained on more than 10 million functional experiments spanning over 1,400 cell lines. The stated use cases are compound prioritization and triage before follow-up experiments, early-stage safety profiling and off-target/selectivity risk flagging ahead of in vivo studies, hit-to-lead decision support, and target-agnostic classification of mechanism of action.

The company positions the approach as complementary to structure-based and target-based computational methods: structural tools model molecular binding, while Reflector models how cells respond to treatment, capturing polypharmacology, pathway crosstalk, protein expression dynamics and cell-context-dependent effects. A second strand of the offering is the reuse of legacy or archived screening data, where customer datasets — including small screens of roughly 100 to 400 compounds and proprietary assays with no public equivalent — are used to fine-tune the pre-trained model, re-score existing libraries against new endpoints and virtually screen unsynthesized analogs. MassBio describes the company as using in silico phenotypic screening to identify small-molecule hits by integrating different assays and readouts across the screening cascade.