Phenotypic Screening Meets Machine Learning in Small-Molecule Discovery Through Cellular Signatures, Representation Learning, Mechanism Inference, and Compound Prioritization

Download PDF

Phenotypic screening measures compound‑induced biological change without requiring a predefined molecular target, yet modern assays increasingly generate images, transcriptional states, and other high‑dimensional readouts whose interpretation depends on computational representation. This integrative review synthesizes peer‑reviewed evidence published from 2017 through 2025 across image‑based profiling, perturbational transcriptomics, representation learning, mechanism inference, chemical–phenotypic modeling, compound prioritization, and experimental confounding, interpreting studies according to task definition, validation design, mechanistic evidence, and transferability rather than pooling across non‑equivalent benchmarks. The literature supports a progression from measured cellular responses to reusable signatures and learned representations that can support perturbation retrieval, activity prediction, response modeling, mechanism‑oriented hypothesis generation, and prioritization of compounds for testing, with morphological, transcriptional, and chemical representations providing overlapping but non‑equivalent information. This review proposes an integrated phenotype‑to‑decision framework in which mechanism inference and compound prioritization branch from learned biological‑response spaces while remaining conditioned by cell context, technical structure, applicability domain, and experimental validation. Strong internal prediction does not itself establish target identity, causal mechanism, external generalization, or prospective discovery value. Phenotypic machine learning is most defensible as an evidence‑organizing and hypothesis‑prioritizing strategy, and its discovery value depends on separating representational usefulness, mechanistic attribution, and experimentally validated decision use while treating discordance, confounding, and domain shift as scientific information rather than preprocessing nuisances.