A Critical Re-evaluation of "BRAIN-MAGNET: A functional genomics atlas for interpretation of non-coding variants" by Deng et al., Cell 2025; doi: 10.1016/j.cell.2025.10.029

This work provides a comprehensive, figure-by-figure, methodologically detailed critique of the Cell 2025 article “BRAIN-MAGNET: A functional genomics atlas for interpretation of non-coding variants” by Deng et al. Using the full published figures, Extended Data, and Supplementary Figures, this commentary examines the reproducibility, statistical validity, experimental design, donor variability, enhancer calling methodology, CRISPRi perturbation robustness, machine-learning model integrity, and disease-variant interpretation.

The analysis identifies major concerns in five categories:

(1) Reproducibility and donor effects: Enhancer maps show substantial heterogeneity across donors, with inconsistent ATAC-seq and CUT&Tag quality metrics and insufficient IDR-based reproducibility analyses.

(2) Overinterpretation of co-accessibility and enhancer–promoter pairing: Correlation-based co-accessibility is repeatedly treated as causality, despite the lack of orthogonal chromatin interaction data (Hi-C, PLAC-seq, HiChIP).

(3) Underpowered CRISPRi validation: Low knockdown efficiency, minimal effect sizes, single-donor cellular context, and visual exaggeration of effect sizes limit interpretability.

(4) Machine-learning model concerns: MAGNET-ML is at high risk of training/testing leakage, lacks benchmarking against state-of-the-art regulatory deep-learning models (Enformer, Basenji2, DeepSEA, ExPecto), and shows insufficient external validation.

(5) Overstated disease applications: GWAS enrichments depend on circular logic, insufficient LD control, and limited variant-level perturbation.

Extended Data and Supplementary materials are evaluated in detail, revealing threshold instability, correlated feature structures, unrealistic null models, donor-level inconsistencies, and limited robustness analyses.

The commentary concludes that BRAIN-MAGNET is a promising but preliminary atlas and that its claims regarding enhancer accuracy, regulatory causality, machine-learning predictions, and neuropsychiatric disease relevance require substantial refinement and validation.