Benchmarking biological knowledge-graph priors against single-cell foundation models for Perturb-seq prediction.

Abstract

Perturb-seq prediction increasingly relies on either biological graph priors or single-cell foundation models, but their relative value is difficult to interpret when architectures, split definitions and graph sources vary simultaneously. We benchmarked STRING functional associations and Gene Ontology co-annotation against dense, self-edge-only and exact-degree-rewired supports in one fixed graph-attention backbone, with a separate 200-HVG STRING-GO-control-co-expression sensitivity. Four public Perturb-seq datasets, three nested gene-space scales and three optimisation seeds were evaluated on held-out guide conditions; GEARS, scGPT and Geneformer were evaluated on the same held-out conditions, output-gene panels and Pearson-r metric after model-specific validity checks. At 200 highly variable genes, STRING-GO improved the equal-dataset Pearson r by +0.016 (95% interval [+0.010, +0.022]) relative to dense support. Dataset effects were positive in Adamson (+0.017), Norman (+0.015), Replogle K562 (+0.017) and Replogle RPE1 (+0.014), and 122/200 guide conditions improved (0.61 [0.54,0.68]). The top-20-gene Jaccard difference was +0.009 [+0.006,+0.012]. STRING-GO exceeded self-edge-only support by +0.024 [+0.018,+0.029] and the mean of ten exact-degree rewires by +0.011 [+0.005,+0.017]. The native-panel effect declined from +0.016 at 200 genes to +0.009 at 1000 genes; the 1000-200 interaction was -0.006 [-0.011,-0.002]. The primary graph-support comparison was formulated and conducted during peer-review revision rather than planned as an a priori confirmatory study. All three external models passed implementation-specific checks. Across 36 protocol-matched dataset-scale contrasts, 32 were within the descriptive ±0.030 reference envelope and four were outside, all favouring the external model. These results demonstrate a modest empirical predictive improvement under the evaluated matched protocol; they do not by themselves establish practical biological relevance.