A self-supervised DNA foundation model with collapse-resistant multimodal fusion

Genomic foundation models pretrained on DNA sequence have achieved strong performance across a range of tasks, but sequence-only representations cannot fully capture regulatory information reflected by additional DNA-centric modalities. Existing multimodal genomic models are often optimized for specific prediction tasks rather than for learning reusable embeddings shared across downstream analyses. However, directly fusing heterogeneous genomic modalities is challenging because sparse, peak-shaped regulatory signals and dense sequence representations have markedly different statistical structures, making naive multimodal alignment prone to degenerate near-zero solutions. We present a self-supervised DNA-centric multimodal foundation model that addresses this gap, integrating DNA sequence embeddings with local and global chromatin accessibility in a shared multimodal encoder to produce reusable window-level embeddings that support both masked reconstruction during pre-training and downstream prediction tasks. We diagnose this heterogeneous-modality alignment failure and show that global normalization substantially alleviates collapse, enabling effective joint learning across modalities. The resulting embeddings improve multiple downstream evaluations of regulatory function, including regulatory activity prediction, regulatory signal ranking and chromatin accessibility peak detection, achieving a 4.6-fold AUPRC improvement over the DNA-only baseline in peak detection, and further improving external validation on ClinVar, GTEx eQTL and PBMC caQTL datasets. The framework is extensible to additional regulatory modalities, providing a methodological basis for multimodal DNA foundation models.