Transcriptomic similarity guides source selection beyond tissue lineage

Perturbational and pharmacogenomic maps remain incomplete across cellular contexts, leaving a practical source-selection question before response prediction: which measured context should inform an unmeasured target? We tested whether baseline transcriptomic similarity ranks candidate sources for held-out targets, separating source ranking from aggregation and transfer-uncertainty decisions. Similarity tracked matched-response concordance in LINCS and GSE306429; selected-source utility exceeded random expectation in GSE306429 (ΔU = 0.034; P = 0.002) and in exact-time-matched MIX-Seq tasks (ΔU = 0.046). In the full GDSC2 analysis, the drug-balanced ranking effect was 0.359 (95% CI, 0.337-0.383). In a lineage-restricted analysis of the same eligible subset, within-tissue ranking remained positive at 0.130 (95% CI, 0.119-0.141), retaining 35.8% of the same-subset global effect (0.363; 95% CI, 0.342-0.386). Global transcriptomic top-1 selection exceeded tissue-matched random selection by 0.154 (95% CI, 0.143-0.164). Thus, baseline similarity adds source-ranking information after tissue matching in the evaluated setting, while ranking alone does not resolve whether to aggregate sources or how to quantify transfer uncertainty.