Mycobacterium Tuberculosis Drug Resistance Prediction from Whole Genome Sequences Using Hierarchical CNN-Transformer Mutation Pattern Encoding
TL;DR
A Hierarchical CNN-Transformer Mutation Pattern Encoding (HCT-MPE) framework that hierarchically encodes nucleotide-, gene-, and regionlevel mutations, uses CNN layers to extract local mutation patterns, and employs Transformer self-attention to learn long-range genomic dependencies before feature fusion and drug-specific classification is proposed.
Abstract
Drug-resistant Mycobacterium tuberculosis (Mtb) remains a major challenge, creating a need for rapid and accurate whole-genome sequencing (WGS)-based resistance prediction. Existing CNN-based approaches have achieved strong performance, but their ability to jointly capture hierarchical mutation patterns and long-range dependencies among genomic variants remains limited. This study proposes a Hierarchical CNN-Transformer Mutation Pattern Encoding (HCT-MPE) framework that hierarchically encodes nucleotide-, gene-, and regionlevel mutations, uses CNN layers to extract local mutation patterns, and employs Transformer self-attention to learn long-range genomic dependencies before feature fusion and drug-specific classification. The model will be implemented in Python using PyTorch and evaluated using the open-access CRyPTIC Consortium WGS–pDST dataset from Zenodo, containing 44,405 WGS samples and 36,738 samples with both WGS and pDST information. The proposed framework is targeted to achieve 97.6% accuracy, 97.2% F1-score, and 98.1% ROCAUC, representing an expected approximately 8.6% improvement in accuracy over a representative 89.0% baseline performance. The framework is expected to provide robust and scalable genomic drug-resistance prediction. Keywords: Mycobacterium tuberculosis; Drug-Resistance Prediction; Whole-Genome Sequencing; Hierarchical CNN-Transformer; Mutation Pattern Encoding.