Nature Biotechnology · 1 October 2026 · From issue 45
Deep learning perturbation models can outperform baselines on calibrated metrics
Gene-editing prediction programs beat simple guessing once scientists fixed broken scoring tests.
Henry E. Miller, Gabriel M. Mejia et al.
Many tests used to judge gene-editing prediction programs were broken. These tests often could not tell a good program from a random guess. New, fairer tests were built and checked against 14 datasets and 18 scoring methods. With these better tests, the smart programs actually did better than simple guessing.
Integrity screen: passed (3 checks) Checked 15 November 2026. Read the source
The plain-language reading of this paper is in the full issue, for subscribers.
Read the full issue