A Framework for Improving the Generalizability of Drug-Target Affinity Prediction Models

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning in drug discovery

Background:

• Accurate prediction of ligand-protein binding affinity is crucial for accelerating drug discovery.
• Existing drug-target affinity (DTA) prediction models often suffer from dataset biases, leading to poor generalization on unseen biomolecules.
• Current methods to improve model generalizability have limitations or may decrease overall performance.

Purpose of the Study:

• To introduce DebiasedDTA, a novel training framework designed to mitigate dataset biases in DTA prediction models.
• To enhance the generalizability of DTA prediction models, enabling more reliable predictions for novel drug-target interactions.
• To provide a broadly applicable solution for improving DTA prediction model robustness.

Main Methods:

• DebiasedDTA employs a sample reweighting strategy within the training framework.
• This reweighting technique aims to counteract dataset-specific biases, promoting learning of generalizable patterns.
• The framework is designed to be compatible with various DTA prediction model architectures and biomolecule representations.

Main Results:

• Extensive experiments demonstrated that DebiasedDTA significantly improves the generalizability of DTA prediction models.
• The framework showed enhanced performance across different biomolecule representations and model architectures.
• Results confirm the effectiveness of DebiasedDTA in predicting drug-target affinities more robustly.

Conclusions:

• DebiasedDTA effectively addresses dataset biases, leading to improved generalizability in DTA prediction.
• The proposed training framework offers a robust and widely applicable method for enhancing drug discovery tools.
• This work contributes to the development of more reliable computational models for predicting drug-target interactions.