Deep Learning-Based Prediction of Variant Effects on Chromatin Accessibility During Dynamic Neuronal Activation

Uncovering the genetic basis of neuropsychiatric diseases (NPDs) is challenging, as most risk variants are noncoding and operate via dynamic, context-dependent mechanisms—many of which remain hidden under static baseline conditions and only emerge during neuronal stimulation. In this work, we evaluate the capacity of ChromBPNet, a sequence-to-function deep learning framework, to predict variant effects across dynamic cellular states. Using iPSC-derived neurons at 0hr (unstimulated), 1hr, and 6hr of KCl stimulation for training, we validated model performance using empirical chromatin accessibility QTLs (caQTLs). ChromBPNet achieves high performance in prioritizing functional variants (average precision $0.51$--$0.60$; $5.6$--$6.6\times$ baseline enrichment) and accurately predicts allelic log fold-changes (Pearson $r=0.61$--$0.69$; 82--86\% directional concordance). Crucially, we observed poor cross-lineage generalization, as models trained on distinct cell types (e.g., microglia) failed to predict neuronal variant effects (Pearson $r \approx 0.2$). These results demonstrate that sequence-based models capture robust regulatory features within a lineage but require context-matched training data to generalize across distinct chromatin landscapes.