Discovering molecular features of intrinsically disordered regions by using evolution for contrastive learning

Area of Science:

• Proteomics
• Bioinformatics
• Computational Biology

Background:

• Intrinsically disordered regions (IDRs) are prevalent in proteomes but their functions remain poorly understood.
• Identifying functional molecular features like motifs, repeats, and physicochemical properties in IDRs is a significant challenge.

Purpose of the Study:

• To introduce a proteome-scale feature discovery approach for intrinsically disordered regions (IDRs).
• To leverage evolutionary conservation as a signal for deep learning to identify functional features within IDRs.

Main Methods:

• Developed a deep learning approach called "reverse homology" that uses evolutionary conservation as a contrastive learning signal.
• Trained a neural network on sets of homologous IDRs to distinguish true homologs from random sequences.
• Employed a simple neural network architecture with standard interpretation techniques.

Main Results:

• The neural network successfully learned conserved features within IDRs, interpretable as motifs, repeats, or bulk properties (e.g., charge, amino acid propensities).
• The model generates visualizations highlighting important residues and regions, aiding in hypothesis generation for uncharacterized IDRs.
• Demonstrated the effectiveness of unsupervised neural networks for systematic feature discovery in IDRs.

Conclusions:

• Reverse homology is a promising method for discovering functional features in intrinsically disordered regions.
• Unsupervised neural networks offer a powerful avenue for gaining systematic insights into poorly understood protein sequences.
• This approach can accelerate the functional characterization of IDRs across the proteome.