Testing the Transferability of Perturbational Similarity Maps Across Species, Cell Systems, and Assay Modalities.

Abstract

We evaluated a mouse whole-brain perturbation map in human K562 and RPE1 Perturb-seq, HeLa morphology and nine L1000 cell-line matrices using three endpoints: global pair-rank concordance, source-defined top-k neighbour preservation, and within-validation CORUM retrieval. Strict orthology retained 71, 90, 695 and 273-311 shared targets, respectively. Under the primary zero-anchored RMS representation, Spearman rho was 0.026 and 0.035 in K562 and RPE1, 0.0022 and 0.0055 in the two HeLa conditions, and a median of 0.0055 across L1000 lines. RPE1 depended on representation: across-target standardisation gave rho = 0.102 (95% CI, 0.027-0.201), whereas omitting feature RMS scaling gave a descriptive rho = -0.034. No source-defined top-10 neighbour preservation result survived correction across 13 contexts (minimum q = 0.240). Source pair ranks also changed when response features were restricted. Within-validation CORUM retrieval identified complex structure in several external matrices but did not test preservation of source neighbours. Across 50 cell splits, K562/RPE1 median global rho was 0.600/0.710 and top-10 overlap was 0.497/0.517. Across 50 mouse-level splits of K562/RPE1-matched source profiles, median global rho was 0.731/0.703; a cross-species noise ceiling remains unestimated. The results are conditional on representation and feature selection and do not establish absence of transfer or transferability to neural systems.