This study benchmarks 27 computational methods for predicting single-cell perturbation responses across 29 datasets. It reveals that no single algorithm excels in all tasks. Foundation models improve with large data, but all methods struggle with cellular context generalization.

References:

• Wei, Z., Wang, Y., Gao, Y. et al. Benchmarking algorithms for generalizable single-cell perturbation response prediction. Nat Methods (2025). doi.org