Identification of Heterogeneous Erlotinib Response Gene Sets Using Sample-Specific Counterfactual Causal Attribution

Identifying cell-line-specific gene sets associated with drug response is difficult because pharmacogenomic data are high-dimensional, continuous-valued, and rarely paired with validated sample-level mechanism labels. We present a mechanism-guided framework for model-based counterfactual attribution of Erlotinib response from transcriptomic data. The framework first uses a BIRD-inspired evidence-integration step, motivated by Bayesian inference from abduction and deduction (BIRD), to combine expression--response association, mutation support, KEGG topology, and curated EGFR tyrosine kinase inhibitor biology in an unlabeled omics setting. This produces a BIRD-inspired panel, referred to as BIRD-20, designed to preserve Erlotinib-relevant mechanism diversity rather than to directly implement the original BIRD framework. Given a fixed panel and biologically adjusted ordering, we approximate conditional counterfactual causal effect (CCCE) scores with causal normalizing flows (CNFs), enabling one-, two-, and three-gene counterfactual interventions in continuous expression and Erlotinib logIC50 space. Conditional on full-cohort-selected panels, five-fold held-out evaluation showed that BIRD-20 had higher CNF-CCCE hit rates than the KEGG-26 pathway-union comparator and similar or modestly higher rates than the text-derived GPT-20 comparator, especially for two- and three-gene interventions. The resulting profiles should be interpreted as sample-level counterfactual hypotheses for downstream biological validation, not as experimentally verified causal mechanisms.