CellVELA: Functional Adaptation of Cell Foundation Models for Cancer Vulnerability Discovery

Single-cell atlases capture the heterogeneous states within a tumour, whereas genome-scale CRISPR screens measure which genes a cancer model depends on for fitness. We introduce CellVELA (Cell Vulnerability Estimation via Latent Adaptation), a framework for functional adaptation of cell foundation models using CRISPR loss-of-fitness supervision available once per cancer model. CellVELA improves held-out vulnerability prediction over the unadapted foundation-model representation and outperforms a published target-wise dependency mapper. Matched controls, however, change how that gain should be interpreted. Replicating model-level phenotypes across individual cells more than doubles the apparent benefit of adaptation, while under correct model-level supervision both direct expression and a dimension-matched random projection outperform the pretrained representation. Adaptation utility and the value of pretraining are therefore separable. In matched post-training comparisons, representation-level adaptation remains effective, whereas the tested in-backbone methods add little. The same overall pattern remains largely unchanged under reduced supervision and leave-one-lineage-out evaluation, and the advantage over the published mapper persists on an independent CRISPR platform. An osteosarcoma case study further shows how patient-derived predictions and held-out functional evidence can narrow a genome-wide vulnerability landscape to a concrete prospective perturbation hypothesis.