#25 · Structural bio & protein design

Perturbation models and benchmarks reshaping therapeutic target nomination

The dominant thread is the evaluation and prioritization of perturbation-to-target logic. Two preprints argue that current perturbation-prediction benchmarks for transcriptomic and virtual cell models do not measure scientific value, and the SBB principles report that deep models, including foundation models, often fail under rigorous baselines. A published drug-blind study reaches a parallel conclusion: cancer response models generalize only through shared mechanisms, so drug-blind benchmarks reflect the dataset more than the model. On the nomination side, a GNN-Transformer screen over LINCS L1000 signatures nominates Class I HDAC inhibitors for pan-cancer use, and a master-regulator method designs combination therapies for coexisting diffuse midline glioma cell states, validated in vivo. A proteome-calibrated glioblastoma digital twin simulates EGFRvIII and PTEN perturbations. What is new is coupling prediction to cell-state resolution during target nomination, most explicitly in the master-regulator design, while the evaluation work questions whether such predictions carry mechanistic content.