Deep Contrastive Learning for High-Throughput Prediction of Drug Resistance Mutations from Sequences

Area of Science:

• Computational biology
• Drug discovery
• Genomics

Background:

• Mutation-induced drug resistance poses a significant challenge to effective pandemic surveillance and drug discovery.
• Existing computational methods for predicting drug-target interactions are limited by the lack of 3D mutant protein structures and struggle with sequence similarity between wild-type and mutant targets.

Purpose of the Study:

• To develop a novel fine-tuning framework, SimSiam-MuTF, to enhance the prediction of drug resistance variants.
• To improve the accuracy of computational models in predicting drug-target affinity for mutant proteins.

Main Methods:

• Developed SimSiam-MuTF, a fine-tuning framework that aligns latent embedding distances with binding affinity shifts between wild-type and mutant targets.
• Utilized a structure-independent model pre-trained on 1.5 million data points (DeepMutDTA) as a foundation.

Main Results:

• SimSiam-MuTF demonstrated robust performance across varied sequence identities and unseen data, outperforming baseline models.
• Achieved average performance gains of 2.47% (PCC) and 5.10% (SCC) in regression tasks, and 4.00% (AUC) and 4.17% (AUPR) in classification tasks.
• Successfully applied to SARS-CoV-2, HIV-1, and cancer-related targets, showcasing generalization potential.

Conclusions:

• SimSiam-MuTF effectively enhances the detection of drug resistance variants by addressing limitations of sequence-based approaches.
• The framework provides a powerful computational platform to deepen the understanding of mutation-induced resistance and accelerate drug discovery against mutant targets.