DeepREAL: a deep learning powered multi-scale modeling framework for predicting out-of-distribution ligand-induced...

Area of Science:

• Computational biology
• Pharmacology
• Machine learning

Background:

• Predictive modeling of drug-target interactions is crucial for drug discovery.
• A gap exists in correlating these interactions with clinical outcomes, specifically predicting genome-wide receptor activities and function selectivity (e.g., agonist vs. antagonist) for novel chemicals.
• Key challenges include scarce receptor activity data for genome-scale modeling and deploying models on data with shifted distributions.

Purpose of the Study:

• To develop an end-to-end deep learning framework for multi-scale modeling of genome-wide ligand-induced receptor activities.
• To address data scarcity and distribution shift issues in predicting receptor activity and function selectivity.

Main Methods:

• Developed DeepREAL, a deep learning framework utilizing self-supervised learning on millions of protein sequences.
• Employed pre-trained binary interaction classification to handle data distribution shifts and scarcity.
• Validated performance on G-protein coupled receptors (GPCRs) in simulated real-world, out-of-distribution scenarios.

Main Results:

• DeepREAL achieves state-of-the-art performance in out-of-distribution settings for predicting receptor activities.
• The framework effectively addresses challenges of data scarcity and distribution shift.
• Demonstrated applicability beyond GPCRs, with potential for extension to other gene families.

Conclusions:

• DeepREAL provides a robust solution for predicting genome-wide receptor activities and function selectivity.
• The framework enhances drug discovery by improving the correlation between drug-target interactions and clinical outcomes.
• DeepREAL's approach offers a scalable and adaptable method for future drug development research.