Reproduce “GeneDisco: A Benchmark for Experimental Design in Drug Discovery” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Arash Mehrjou, Ashkan Soleymani, Andrew Jesson, Pascal Notin, Yarin Gal, Stefan Bauer, Patrick Schwab
GeneDisco introduces an open benchmark for evaluating batch active-learning methods that choose genetic perturbation experiments in drug discovery. Using four real genome-wide CRISPR screens, the study shows that acquisition performance depends on the assay, gene representation, model, batch size, and objective: diversity-based methods often improve overall prediction, while uncertainty-based policies are more effective at finding genes with extreme phenotypic effects.
Key insights
- Acquisition policies should be judged jointly with the predictive model and treatment representation; no method consistently performs best across assays and experimental settings.
- Diversity-oriented sampling generally supports broad function estimation and lower prediction error, whereas uncertainty-oriented sampling more often identifies genes with strong positive or negative effects early.
- The objective of the experimental campaign determines the preferred policy: global predictive accuracy and high-impact target discovery can favor different sampling behaviors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.