Drug Resistance Predictions Based on a Directed Flag Transformer

Abstract

The evolving SARS‐CoV‐2 virus threatens global public health, particularly with potential resistance to PAXLOVID, whose active ingredient, nirmatrelvir, targets the viral main protease (Mpro). CAPTURE (direCted flAg laPlacian Transformer for drUg Resistance prEdictions) is developed to assess Mpro mutations' impact on nirmatrelvir binding and identify drug‐resistant variants. CAPTURE integrates a mutation analysis with a resistance prediction module using DFFormer‐seq, a novel ensemble model combining a Directed Flag Transformer and sequence embeddings. Analysis of Mpro mutations from May to December 2022 revealed increasing mutation frequencies near the binding site, suggesting PAXLOVID's widespread use accelerated drug‐resistant evolution. CAPTURE identified potential resistance mutations, including experimentally confirmed H172Y and F140L, and five others awaiting validation. Evaluated on Mpro mutant data, CAPTURE achieved 57% recall and 71% precision in predicting drug‐resistant mutations. This work establishes a robust framework for predicting resistance and enabling real‐time viral surveillance, guiding the design of next‐generation therapeutics.