paper✓ VerifiedVC0146

Deep learning perturbation models can outperform baselines on calibrated metrics.

Miller et al. (Nature Biotechnology 2026; authors from Shift Bioscience, competing-interest statement in paper): positive-control baselines and a 'dynamic range fraction' calibration framework over 14 datasets and 18 metrics; argues MSE and Pearson delta are often miscalibrated and that deep-learning models can beat uninformative baselines on calibrated metrics. Counter-position to VC0048/VC0049/VC0052/VC0144/VC0145.

Authors / organisation: Miller HE, Mejia GM, Leblanc FJA, Swain B, Wang B, de Lima Camillo LP.
Venue: Nature biotechnology
Date: 2026-10-01
Type: paper
Year: 2026
DOI: 10.1038/s41587-026-03307-w
PMID: 42823510
Verification: Europe PMC (metadata; preprint 10.1101/2025.10.20.683304 indexed as PPR1104931) + Nature article page via doi.org read 2026-10-16 (full text, incl. competing interests); PMID 42823510

Benchmark & validation Critique

Cited by GlioState topics

Linear baselines and systematic variation in perturbation prediction: what 'a strong baseline' means, and where the papers disagree (Validation)

Entities mentioned

Open in the searchable library → · JSON

Deep learning perturbation models can outperform baselines on calibrated metrics.

Cited by GlioState topics

Entities mentioned

Related papers (shared tags)

A comparison of computational methods for expression forecasting.

Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines.

Simple controls exceed best deep learning algorithms and reveal foundation model effectiveness for predicting genetic perturbations.

Benchmarking foundation cell models for post-perturbation RNA-seq prediction.

Zero-shot evaluation reveals limitations of single-cell foundation models.

Biological domain shift and statistical nesting concerns in generative AI-based spatial tumor growth prediction for pediatric diffuse midline glioma.