Toward systematic evaluation of cell–drug representation alignment across biological readouts

Drug perturbation models depend on representations of both the compound and the cellular state, but it remains unclear what is gained by aligning the two rep- resentation spaces using measured biological response. We study this question through three experiments. First, given a post-perturbation expression profile, can the applied drug be identified? Second, given a drug, can its post-perturbation transcriptional response be predicted? Third, given a drug and a cell-line expression state, can drug sensitivity be predicted? The experiments span high-dimensional transcriptional response and the lower- dimensional IC50 endpoint. We compare chemistry-based drug representations, representations pretrained on biological measurements, the biological-response- aware drug embedding (BiRD), and representations obtained by explicit cell–drug alignment. Alignment makes post-perturbation cell representations informative about drug identity. For transcriptional-response prediction on compounds excluded from alignment training, it improves ADMET-AI consistently but marginally, while the other drug representations deteriorate slightly. BiRD gives the strongest results on the transcriptional tasks. For IC50 prediction, aligned representations do not improve ranking on GDSC2 or on an external in-house panel; compound and cell- line main effects account for much of the measured signal. Cell–drug alignment is therefore useful for identifying a perturbation from its transcriptional signature, but the geometry learned at that level does not transfer to a coarser pharmacological endpoint.