Adversarial validation of in silico perturbation profiles for intelligence-associated genes in human prefrontal cortex

BACKGROUND: Genome-wide association studies (GWAS) have identified numerous intelligence-associated loci, but the cell-type-specific mechanisms through which these genes influence cognition remain largely unknown. Standardized adversarial-validation frameworks for assessing in silico perturbation claims from foundation models are currently lacking. METHODS: We developed a reproducible adversarial-validation framework and applied it to 21 intelligence-linked genes using Geneformer-based in silico gene-deletion perturbations in human dorsolateral prefrontal cortex (DLPFC) single-cell RNA-seq data (500 cells, 3 donors). The framework incorporates expression-matched empirical null testing (61 control genes), control gene-pair combinatorial nulls (20 random pairs), cross-model comparison with scGPT, donor-aware pseudobulk analysis, and donor-aware combinatorial permutation testing. RESULTS: Individual gene perturbation effects did not survive multiple-testing correction against expression-matched nulls (0/21 FDR-significant), demonstrating that raw embedding shifts are substantially confounded by expression level. A gene-set-level rank-sum test showed a modest collective signal (Wilcoxon two-sided p=0.051), which we interpret as suggestive but not confirmatory. Combinatorial analysis revealed directionally consistent but statistically fragile super-additivity (Mann-Whitney p=0.045; donor-aware permutation p=0.200; bootstrap 95% CI including zero). Cross-model concordance with scGPT was limited by tokenization incompatibilities (6/21 overlapping genes; ρ=0.66, p=0.16). Cell-type stratified analysis generated illustrative patterns (Friedman p=0.70), retained as hypotheses for future validation. The study is substantially underpowered at the donor level (n=3; estimated power ≈ 27%; n≥20 recommended). CONCLUSIONS: This study provides a reproducible adversarial-validation framework for stress-testing foundation-model perturbation findings. The framework reveals that expression-level confounding substantially inflates raw perturbation signals and recommends expression-matched nulls as a minimum standard. All biological findings warrant replication in independent multi-donor cohorts (n≥20) and experimental validation via CRISPRi or Perturb-seq.