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

Single-cell transcriptomes resolve heterogeneous cancer states, whereas genome-scale CRISPR screens measure genetic dependencies at the cancer-model level. We introduce CellVELA (Cell Vulnerability Estimation via Latent Adaptation), which adapts cell foundation models for selective vulnerability prediction using loss-of-fitness supervision. Across held-out cancer models, CellVELA improves over the unadapted representation and a published dependency mapper. Matched controls substantially narrow this interpretation: assigning model-level phenotypes to individual cells inflates the apparent gain, while direct expression and a dimension-matched random projection outperform the pretrained representation under correct supervision. The advantage over the published mapper nevertheless persists on an independent screening platform. In osteosarcoma, patient-derived predictions nominate a small candidate set, and disease-specific screening changes which candidate advances to prospective validation. Together, these results show how matched controls and independent evidence refine both what model improvements establish and which predictions merit experimental testing.