Predicting gene essentiality and drug response from preclinical perturbation screens with layered ensemble of autoencoders and predictors

Description

Bodinier and colleagues introduce LEAP, a tool meant to forecast how lab disease models will respond to gene knockouts and drugs before the experiments are run. Such screens, which disturb cells systematically and measure the fallout, generate huge datasets, but models trained on them often fail on new cell lines or resist interpretation. LEAP, described in 2026, layers together many predictors, each trained on a different compressed view of gene activity.

Across tasks, the layered ensemble consistently beat single predictors on cell lines the models had never seen, lifting correlation scores by a few points. One lean version combined LEAP with simple per-drug models and nearly matched the best performance while staying fast. The team also distilled the ensemble to highlight which biological pathways drive its predictions. They argue the method can prioritise which lab experiments are worth doing first.

Catalog record

Source: https://europepmc.org/articles/PMC13493823