paper✓ VerifiedVC0175
Ni J, Zhang X, Li M, Cai Y, Xie Z, Liu Y. · Computational Biology and Chemistry · 2026-10-09
Ni et al. (Southeast University; Computational Biology and Chemistry 2026) compare STRING/Gene Ontology graph priors with dense, self-edge-only and degree-rewired supports in one fixed graph-attention backbone on four public Perturb-seq datasets (Adamson, Norman, Replogle K562 and RPE1), with GEARS, scGPT and Geneformer evaluated on the same held-out guide conditions. At 200 highly variable genes, STRING-GO improved Pearson r by +0.016 (95% interval +0.010 to +0.022) over dense support; the effect fell to +0.009 at 1000 genes. Of 36 matched contrasts with external models, 32 were within a ±0.030 envelope and four favoured the external model. The authors state the main comparison was formulated during peer-review revision and that the modest gain does not by itself establish biological relevance.
DOIPubMed
preprint✓ VerifiedVC0179
Liang W-L, Yang Y-B, Zhang Q-Y, Feng J-C, Wang Q-F, Zou D, Tang J-L, Liu Y, Wang S-Y, Duan J, Bai Y, Li Y-F, Sun W-Y, Cao Y-F, He R-R. · bioRxiv (preprint, not peer-reviewed) · 2026-10-13
Preprint (not peer-reviewed): Liang et al. (Macau University of Science and Technology; Jinan University) show that conventional evaluation of few-shot adaptation in single-cell perturbation prediction shares one control estimate between calibration responses and held-out targets, which carries control error into predictions and inflates scores. Scoring against an independent control changed the rank of 40 of 52 configuration–dataset entries. They introduce SafeDelta, a split-control criterion that uses calibration data to decide whether to adapt; across six datasets, mean gated gain was non-negative under the original two-control evaluation.
DOI
preprint✓ VerifiedVC0180
Liu Q, Zhang Q, Chen Y, Du J-H, Zhao S, Wang J. · bioRxiv (preprint, not peer-reviewed) · 2026-10-11
Preprint (not peer-reviewed): Liu et al. (Department of Statistics, University of Chicago) revisit the observation that many gene perturbations move cells along a common global transcriptional direction. They report that perturbations contribute differently to this leading trend and that substantial structure remains beyond it, including perturbation-specific effects that transfer across datasets and a common hierarchy of responsive genes; these patterns make simple prediction strategies surprisingly competitive even beyond the leading global component.
DOI
paper✓ VerifiedVC0145
Kernfeld E, Yang Y, Weinstock JS, Battle A, Cahan P. · Genome biology · 2025-11-17
Kernfeld et al. (Genome Biology 2025): PEREGGRN benchmark of 11 perturbation datasets; reports it is uncommon for expression-forecasting methods to beat simple baselines, and discusses a bias-variance explanation and stereotypical perturbation responses for the strength of the training-data mean.
DOIPubMed
paper✓ VerifiedVC0144
Wong DR, Hill AS, Moccia R. · Bioinformatics · 2025-06-01
Wong, Hill, Moccia (Bioinformatics 2025): a 'CRISPR-informed mean' (training-set perturbed-cell mean, with the target gene set by prior knowledge of CRISPRi/CRISPRa) is reported to beat fine-tuned scGPT, GEARS and simple neural baselines on Adamson, Norman and Replogle K562 Essential; also provides a corrected Adamson dataset.
DOIPubMed
paper✓ VerifiedVC0174
Mitra S, Li J, Leslie CS. · Molecular Cell · 2026-10-15
Review (Mitra, Li and Leslie; Memorial Sloan Kettering; Molecular Cell 86:4131–4148, 2026) of AI embedding and foundation models for scRNA-seq and scATAC-seq, machine-learning work on single-cell multiome in regulatory genomics, and strategies to learn "virtual cell" models from Perturb-seq, ending with current limitations of single-cell AI models. The authors note that most early single-cell embedding and foundation models did not gain wide traction among experimental biologists. Declared interest: C.S.L. is on the journal's advisory board.
DOIPubMed