)
)
Leveraging Foundation Models and In Silico Perturbation to Accelerate Target Identification in Early Discovery
- How in silico perturbation (ISP) can be used with single-cell foundation models to prioritize driver and rescue genes for target identification in early discovery
- A biologically grounded evaluation of ISP fidelity using embedding-based metrics: cell-state separation, directional shift toward target, baseline testing, and gene recovery ranking
- What we learn from benchmarking zero-shot models against a task-specific perturbation model and why simple baselines still matter
- How fine-tuning, model scale, and domain-matched pretraining change ISP performance, with practical guidance for model selection in different biological contexts
