Drug Resistance Predictions Based on a Directed Flag Transformer

Area of Science:

• Virology
• Drug Discovery
• Computational Biology

Background:

• SARS-CoV-2 evolution presents a global health challenge, with drug resistance being a major concern.
• PAXLOVID, an antiviral, targets the viral main protease (Mpro), and mutations can confer resistance.

Purpose of the Study:

• To develop a computational framework, CAPTURE, for predicting SARS-CoV-2 drug resistance mutations.
• To analyze the impact of Mpro mutations on nirmatrelvir binding affinity and identify potential resistance-driving variants.

Main Methods:

• Developed CAPTURE, integrating mutation analysis with a resistance prediction module (DFFormer-seq).
• Utilized a Directed Flag Transformer and sequence embeddings from protein and small-molecule-large-language models.
• Analyzed Mpro mutation frequencies and binding affinities to nirmatrelvir.

Main Results:

• Observed increased mutation frequencies near the Mpro binding site, suggesting selective pressure from PAXLOVID.
• Identified experimentally confirmed resistance mutations (H172Y, F140L) and five novel potential mutations.
• Achieved 57% recall and 71% precision in predicting drug-resistant mutations on a test dataset.

Conclusions:

• CAPTURE provides a robust framework for predicting drug resistance and real-time viral surveillance.
• Findings guide the rational design of next-generation antiviral therapeutics against SARS-CoV-2.
• Highlights the impact of selective pressure from antiviral drugs on viral evolution.