Enformer Recognizes cis-Regulated Genes by Prediction Magnitude but Fails to Predict Effect Direction: A Magnitude–Direction Dissociation in Individual-Level Expression Prediction

We cross-reference Enformer's individual-level per-gene Pearson R (Sasse et al. 2023, Nature Genetics, ROSMAP cohort n=839) with GTEx v8 Brain Cortex eGenes (n=9,082) across 6,808 genes, and report a fundamental dissociation: prediction magnitude (|R|) monotonically predicts cis-eQTL gene identity (eGene rate 31.6% → 82.5%; |R| model AIC=8274 vs R+R² AIC=8378, ΔAIC=104), but prediction direction for eGenes is near-random (54.0% correct; max D10=62.5%). Direction errors are irreversible by fine-tuning on ROSMAP (Mann–Whitney p≈10⁻¹⁰⁰), yet PrediXcan recovers correct direction for R<0 eGenes (PrediXcanR=0.295), consistent with the hypothesis that Enformer's non-brain ENCODE training corpus lacks the brain-specific regulatory programs needed to correctly orient directional effects.