This repository provides a framework for fine-grained prediction of chromatin accessibility from DNA sequence, using ConvNeXt V2 blocks as powerful feature extractors. By integrating these blocks into diverse model architectures—including CNNs, LSTMs, dilated CNNs, and transformers—we demonstrate consistent performance gains, with the ConvNeXt-based dilated CNN achieving the most robust and shape-preserving predictions of ATAC-seq signal at 4 bp resolution. Our codebase includes benchmarks and tools for cell-type-specific chromatin modeling at high resolution.

Input to the model: DNA sequence (.fa)

Output of the model: ATAC-seq signal (.npz or .bigwig)

Use the links below to get started:

• Installation
• Model Training
• Model Evaluation
• Predicting SNV Effects on ATAC-seq
• Exporting Predictions