Multimodal learning of noncoding variant effects using genome sequence and chromatin structure.

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Noncoding genetic variants, including single-nucleotide polymorphisms, are increasingly linked to complex human traits and diseases.
• Interpreting the functional impact of these variants is challenging, as current computational models often overlook the role of 3D chromatin structure.

Purpose of the Study:

• To develop a computational model that predicts the effects of noncoding variants on epigenetic profiles.
• To incorporate both 1D genome sequence and 3D chromatin structure data for enhanced prediction accuracy.

Main Methods:

• Developed a multimodal deep learning framework integrating convolutional and recurrent neural networks for sequence embedding and graph neural networks for structure embedding.
• Utilized recent DNA language models to bridge the resolution gap between sequence and structure data.
• Employed both unsupervised (zero-shot) and supervised (few-shot) learning approaches.

Main Results:

• The multimodal model significantly outperforms sequence-only models in predicting epigenetic profiles.
• The model effectively captures long-range interactions, complementing sequence-only approaches for regulatory motif identification.
• Demonstrated strong predictive performance for noncoding variant effects on gene expression and pathogenicity.

Conclusions:

• Integrating 3D chromatin structure with 1D sequence data provides a more comprehensive understanding of noncoding variant effects.
• The developed deep learning scheme offers a powerful tool for predicting the functional impact of genetic variants.
• The findings advance the mechanistic interpretation of noncoding variants in human diseases.