SubcellPert: Benchmarking Subcellular Perturbation Prediction

Preprint (not peer-reviewed): Wang et al. (Chinese University of Hong Kong) introduce SubcellPert, a benchmark for perturbation-response prediction at single-RNA-molecule resolution: 14 imaging datasets, 13.88 million cells and 3.65 billion RNA molecules; an Oracle task (predict molecule positions given exact target counts) and a Joint task (counts and positions), with fixed perturbation and cellular-context holdouts; eight baselines including classical transport and frozen foundation-model adapters. The authors report that Gaussian transport gives the lowest Oracle spatial error on 8 of 11 cellular-context OOD datasets, and audits show that better centroid accuracy can accompany worse radial distributions. The abstract names no brain-tumor data.

Authors / organisation

Wang J, Cao H, Dang X, Zhou T, Zhao Y, Hsieh C-Y, Gu C, Heng PA.

Venue

bioRxiv (preprint, not peer-reviewed)

Date

2026-10-17

Type

preprint

Year

2026

DOI

10.64898/2026.10.11.758305

Verification

bioRxiv API details (title, 8 authors, posted 2026-10-17, v1, new results, CC BY, corresponding institution CUHK, abstract) + Crossref posted-content record (same title/authors/date) fetched 2026-10-20; full text not read

Benchmark & validationPerturbationSpatial omicsAI virtual cell

Open in the searchable library → · JSON