Large-scale data-driven pre-trained DNA models enhance performance across diverse genomics tasks

Abstract

Abstract Sequence-based deep learning has advanced genome interpretation, yet most models remain task-specific and rely on retraining, limiting scalability across biological contexts. Here we present SUCCEED, a supervised multi-task DNA foundation model pretrained on 6,389 ENCODE functional genomics tracks to learn transferable regulatory representations. By integrating convolutional layers with a Transformer architecture, SUCCEED captures both local sequence motifs and long-range regulatory dependencies, achieving performance comparable to or exceeding Enformer across benchmark tasks. Through transfer learning, it predicts cell-type-specific epigenomic profiles, denoises sparse chromatin accessibility signals, and predicts three-dimensional chromatin contacts without CTCF input across data scales and cell types. Across diverse genomics tasks, SUCCEED performs comparably to supervised foundation models such as Sei and outperforms self-supervised models trained solely on DNA sequence. Overall, SUCCEED is a transferable and scalable foundation model that provides a unified framework for genome-scale regulatory modeling in complex biological contexts.