Cancer Cell Line Heterogeneity Imposes a Primary Bottleneck for Virtual Perturbation Screening at Scale

UniPert-G2CP (Li et al., Cell, 2026) bridges genetic and chemical screens from molecular representation to phenotype modeling across five cancer cell lines [1]. Here we extend this architecture to 162 cell lines, 32,039 compounds, and genome-wide output (12,328 genes), and find that the resulting prediction platform reveals a striking performance divide: normal and primary cell lines achieve substantially higher prediction fidelity (mean per-cell-line Pearson Correlation Coefficient (PCC) = 0.480, median 0.502) than cancer lines (0.296, median 0.324; Mann-Whitney p = 3.93e-11), identifying cancer cell line heterogeneity as a primary bottleneck for virtual perturbation screening at scale. To understand the sources of this divide, we analyzed per-cell-line performance, genome-wide directional accuracy, mechanism-clustering (SMD), and compound-protein interaction (CPI) enrichment. Directional accuracy on top-5% effect-size genes reaches 73.8% (genome-wide 60.8%), with pathway-dependent recovery: of four literature-supported perturbation-gene pairs queried across three cell lines, two were recapitulated (dexamethasone-TSC22D3/NFKBIA/FKBP5; bortezomib-BAG3/DNAJB1/HSPA1A), one was absent (CD36 depletion-PPARG/CEBPA in ASC), and one was partially recapitulated (metformin-SLC7A5 in HEPG2). Mechanism-clustering SMD of the learned embedding reached 1.636 (vs. original 1.85; 88.5% retention at 32x cell-line coverage), exceeding the ECFP4 fingerprint baseline (1.613), while self-consistency Mantel rho=0.852 confirmed the model retains compound mechanism structure internally. Overall held-out performance: genetic perturbation PCC=0.442 (978-gene subset); novel drug PCC=0.3047 (genome-wide). Analysis of CPI enrichment reveals that training-data overlap inflates apparent performance: 54.9% of Touchstone evaluation pairs overlap with our ChEMBL-derived training CPI pairs, reducing effective EF from 139 to 109 at top 0.5% yet remaining far above random (1.0). These findings establish the first large-scale characterization of cell-type-dependent generalization in perturbation-to-phenotype prediction. The observed performance stratification between normal and cancer lines generates testable hypotheses for why virtual cell models degrade on heterogeneous cancer contexts, and provides a diagnostic framework for identifying where and why such models fail--informing future architecture improvements targeting the CPI vocabulary gap, protein encoder design, and cell-type-aware training strategies.