Cross-modality and self-supervised protein embedding for compound-protein affinity and contact prediction.

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Compound-protein affinity and contact (CPAC) prediction is crucial for rational drug discovery.
• Existing structure-free methods struggle with limited protein structures and scarce labeled data.
• Accurate CPAC prediction requires understanding both interaction strength and contact patterns.

Purpose of the Study:

• To address challenges in CPAC prediction, including structure naivety and data scarcity.
• To develop structure-aware and task-relevant protein embeddings.
• To improve the generalizability of CPAC models for unseen proteins.

Main Methods:

• Introduced cross-modality learning to integrate 1D amino-acid sequences and 2D contact maps.
• Employed self-supervised learning strategies on massive unlabeled protein data for pre-training.
• Utilized recurrent neural networks (RNNs) for sequences and graph neural networks (GNNs) for contact maps.

Main Results:

• Individual protein modalities showed varying strengths in predicting affinities versus contacts.
• Cross-modality embedding combined with self-supervised learning significantly improved model generalizability.
• The integrated approach enhanced predictions for both affinity and contact for novel proteins.

Conclusions:

• Cross-modality and self-supervised learning effectively overcome limitations in CPAC prediction.
• The developed methods provide a more robust framework for structure-aware protein embedding.
• This approach advances rational drug discovery by improving compound-protein interaction predictions.