Early feature extraction drives model performance in high-resolution chromatin accessibility prediction.

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Predicting chromatin accessibility from DNA sequence is crucial for understanding gene expression.
• Current methods often lack the resolution to detect single-nucleotide variant effects.
• The impact of specific deep learning architectural components on high-resolution prediction is not well understood.

Purpose of the Study:

• To systematically evaluate deep learning architectures for fine-grained chromatin accessibility prediction.
• To assess the utility of ConvNeXt V2 blocks, adapted from computer vision, for genomic feature extraction.
• To identify key architectural determinants of prediction accuracy.

Main Methods:

• Integration of ConvNeXt V2 blocks into various deep learning models (CNNs, LSTMs, dilated CNNs, transformers).
• Systematic evaluation of model performance on predicting ATAC-seq signal at 4 bp resolution.
• Comparative analysis of different architectural choices and their impact on prediction accuracy.

Main Results:

• ConvNeXt V2 blocks consistently enhanced performance across diverse architectures, leading to similar prediction accuracies.
• Early feature extraction, facilitated by ConvNeXt V2, was identified as the primary driver of prediction accuracy.
• A ConvNeXt-based dilated CNN model demonstrated superior performance in preserving ATAC-seq signal shape.

Conclusions:

• ConvNeXt V2 blocks are effective high-resolution feature extractors for genomic data.
• The choice of early feature extraction significantly impacts chromatin accessibility prediction accuracy.
• The developed codebase and benchmarks offer valuable tools for high-resolution chromatin modeling.