Gene expression prediction from epigenomic data links regulatory elements to transcriptional output. Multilayer perceptron (MLP) models rely primarily on histone modification signals at candidate cis-regulatory elements (CREs) and often ignore genomic and 3D chromatin context. These models are often overparameterized and may interpret correlated signals as causal regulatory effects. These limitations motivate the incorporation of biological priors.

We introduce a biologically guided pruning framework to systematically evaluate sparsification strategies in gene expression prediction models. Using H3K27ac signals at CREs, we train MLPs on bulk RNA-seq data and impose sparsity via three approaches: data-driven pruning, pruning guided by linear genomic distance between CREs and transcription start sites (TSSs) of target genes, and pruning informed by Hi-C chromatin interaction frequencies between CREs and TSSs.

On a benchmark of 200 genes, Hi-C-guided pruning achieves the strongest predictive performance, improving median correlation and reducing mean squared error. Stratifying genes by predictability (threshold r = 0.7) reveals heterogeneous effects, with larger gains in low-correlation genes and smaller improvements in high-correlation genes.

We then apply this approach to 1,574 genes from an enhancer-gene interaction map and evaluate biological validity. Hi-C-pruned models improve recovery of functional CRE-gene regulatory relationships, increasing AUC from 0.159 to 0.180 for CRISPRi-validated links and showing stronger enrichment of eQTL-supported associations across multiple tissues from the GTEx Consortium, relative to the unpruned model.

These results demonstrate that chromatin-informed sparsification improves predictive performance and recovery of experimentally supported regulatory interactions.

Co-authors: Dennis Hecker, Shamim Ashrafiyan, Marcel Schulz,