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

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Epigenetics

Background:

• Current sequence-based deep learning models for genome interpretation are often task-specific, requiring extensive retraining and limiting their scalability.
• A need exists for versatile, transferable models that can generalize across diverse biological contexts and data scales.

Purpose of the Study:

• To introduce SUCCEED, a supervised multi-task DNA foundation model designed for transferable regulatory representation learning.
• To evaluate SUCCEED's performance and scalability across various genomics tasks, including epigenomic profile prediction and chromatin contact prediction.

Main Methods:

• SUCCEED integrates convolutional layers with a Transformer architecture for capturing both local sequence motifs and long-range regulatory dependencies.
• The model was pretrained on 6,389 ENCODE functional genomics tracks.
• Transfer learning was employed to adapt the model for specific downstream tasks such as epigenomic profiling and chromatin contact prediction.

Main Results:

• SUCCEED achieves performance comparable to or exceeding Enformer on benchmark tasks.
• The model successfully predicts cell-type-specific epigenomic profiles, denoises sparse chromatin accessibility signals, and predicts 3D chromatin contacts without CTCF input.
• SUCCEED demonstrates comparable performance to supervised foundation models like Sei and outperforms self-supervised models trained solely on DNA sequence.

Conclusions:

• SUCCEED offers a transferable and scalable foundation model for genome-scale regulatory modeling.
• The model provides a unified framework for analyzing complex biological contexts and diverse genomics tasks.
• SUCCEED advances the field of genome interpretation by enabling efficient and accurate predictions across multiple functional genomics applications.