Leveraging large language models to predict antibiotic resistance in Mycobacterium tuberculosis

Area of Science:

• Genomic Medicine
• Computational Biology
• Infectious Diseases

Background:

• Antibiotic resistance in Mycobacterium tuberculosis (MTB) presents a major global health threat.
• Accurate prediction of resistance is crucial for effective treatment and controlling resistant strain spread.
• Current methods require extensive data curation and struggle with emerging drug resistance.

Purpose of the Study:

• To develop and evaluate a novel approach using large language models (LLMs) for predicting antibiotic resistance in MTB.
• To leverage natural language processing techniques on genomic data for resistance pattern identification.
• To explore the potential of LLMs in adapting to new antibiotics and uncovering novel resistance mechanisms.

Main Methods:

• Training and evaluation of a LLM-based model (LLMTB) on genomic data from 12,185 CRyPTIC isolates.
• Utilizing transformer-based LLMs and natural language processing to analyze complex genomic sequences.
• Assessing model performance in predicting resistance to various antibiotics.

Main Results:

• The LLMTB model achieved high performance in predicting antibiotic resistance in MTB strains.
• LLMs demonstrated adaptability to new drugs through fine-tuning or few-shot learning, reducing data curation needs.
• The model identified critical genes, intergenic regions, and novel resistance mechanisms, providing deeper biological insights.

Conclusions:

• LLMTB offers a transformative shift in antibiotic resistance prediction for MTB.
• The approach enhances diagnostic capabilities and supports personalized treatment plans for tuberculosis.
• This method contributes significantly to combating tuberculosis and antimicrobial resistance globally.