PIE: Generalizing perturbation effects across unseen perturbations, contexts and datasets

Preprint (not peer-reviewed): Verma et al. (Arc Institute) introduce PIE, which reformulates perturbation-response learning around population-level effects and directly predicts differentially expressed genes and expression changes for each (context, perturbation) pair from external biological knowledge, baseline expression and observed responses, including genes unseen in training. The authors report state-of-the-art performance on most metrics on the Replogle–Nadig dataset across all tested generalization settings, and 1.2 to 3.2 times the DEG AUPRC of the strongest baseline in each setting (unseen contexts, perturbations, datasets, and jointly unseen contexts and perturbations). Self-reported; no independent evaluation yet.

Authors / organisation

Verma R, Adduri A, Bevilacqua B, Eraslan B, Burke D, Goodarzi H, Roohani Y.

Venue

bioRxiv (preprint, not peer-reviewed)

Date

2026-10-05

Type

preprint

Year

2026

DOI

10.64898/2026.10.02.756297

Verification

bioRxiv API details (title, 7 authors, posted 2026-10-05, v1, new results, CC BY, corresponding institution Arc Institute, abstract) fetched 2026-10-20; full text not read; KB id VC0166 (renumbered 2026-10-20)

PerturbationBenchmark & validationZero-shotArc Institute

Open in the searchable library → · JSON