Information contracts determine what inverse perturbation benchmarks measure

Abstract

Inverse perturbation benchmarks rank interventions toward a desired cellular state, but high accuracy can answer a different question from biological prioritization. We present InverseBench, which makes candidate information and endpoint construction explicit, and FSTN, a linear pipeline evaluated under it. With identical inputs, candidates and validation, nested ridge reaches Spearman 0.959 for a transcriptional principal-component proxy but 0.445 for independently measured K562 growth fitness across 7,354 genes. Controlled response-entry availability across seven tasks gives heterogeneous gains, 0.442 for CD4 activation and 0.116 for external fitness. Yet predefined shortlist comparisons show no primary-cutoff penalty in K562 or a melanoma co-culture screen, and secondary cutoffs reverse the K562 ordering. FSTN's nonnegative set selector recovers 0.840 of measured two-gene constituents in three picks, against 0.771 for an unconstrained greedy. The framework separates proxy reconstruction, response-informed prioritization and unseen-intervention ranking: an endpoint-dependent accuracy gap does not by itself mean worse biological decisions.