Multimodal learning decodes the global binding landscape of chromatin-associated proteins

Area of Science:

• Genomics
• Molecular Biology
• Computational Biology

Background:

• Chromatin-associated proteins (CAPs) regulate gene expression and cell differentiation.
• Genome-wide binding profiles of CAPs are crucial for understanding gene regulation.
• Current techniques limit obtaining CAP binding data, especially in primary cells.

Purpose of the Study:

• To develop a scalable method for predicting CAP binding landscapes.
• To understand the principles governing CAP binding.
• To explore CAPs' roles in cell fate transitions and development.

Main Methods:

• Development of Chromnitron, a multimodal foundation model.
• Utilizing DNA sequence motifs, chromatin accessibility, and protein functional domains as features.
• In silico perturbation experiments to validate model learning.

Main Results:

• Chromnitron accurately predicts CAP binding landscapes across hundreds of proteins in unseen cell types.
• The model learned underlying principles of CAP binding from multimodal features.
• Discovery of novel CAPs involved in T cell exhaustion.
• Prediction of dynamic CAP binding landscapes during neurogenesis.

Conclusions:

• Chromnitron enables accurate prediction of CAP binding, overcoming technical limitations.
• The model accelerates discovery in regulatory genomics, particularly in human primary cells.
• This work opens avenues for engineering regulatory genomics and therapeutic opportunities.