Scholarlink

Tuberculosis (TB) remains a global health threat, intensified by rising drug resistance. We present a pipeline leveraging large language models (LLMs) to extract structured metadata from assay descriptions in ChEMBL, enabling automated curation of biologically relevant datasets for Mycobacterium tuberculosis (Mtb). Using these curated datasets, we train and compare machine learning (ML) models to predict minimum inhibitory concentration (MIC) values, with subsets focused on the H37Rv yielding the best performance ( \(R^2\) up to 0.65, MAE 0.41). We further introduce a multi-task classification model to predict compound activity across resistance profiles—non-resistant, single-drug resistant, and multidrug-resistant (MDR) cases—achieving F1 scores of 0.84, 0.81, and 0.67, respectively. This highlights strong predictive power for susceptible strains while exposing challenges in MDR contexts. Our results demonstrate that LLM-based curation enhances data quality and supports more effective ML-driven drug discovery pipelines, with potential for broader application across resistant pathogens.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Leveraging Large Language Models on Assay Descriptions to Improve the Prediction of Inhibitors for Mycobacterium Tuberculosis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nuno Alves,
  • Nuno S. Osório,
  • Vítor Pereira,
  • Miguel Rocha

摘要

Tuberculosis (TB) remains a global health threat, intensified by rising drug resistance. We present a pipeline leveraging large language models (LLMs) to extract structured metadata from assay descriptions in ChEMBL, enabling automated curation of biologically relevant datasets for Mycobacterium tuberculosis (Mtb). Using these curated datasets, we train and compare machine learning (ML) models to predict minimum inhibitory concentration (MIC) values, with subsets focused on the H37Rv yielding the best performance ( \(R^2\) up to 0.65, MAE 0.41). We further introduce a multi-task classification model to predict compound activity across resistance profiles—non-resistant, single-drug resistant, and multidrug-resistant (MDR) cases—achieving F1 scores of 0.84, 0.81, and 0.67, respectively. This highlights strong predictive power for susceptible strains while exposing challenges in MDR contexts. Our results demonstrate that LLM-based curation enhances data quality and supports more effective ML-driven drug discovery pipelines, with potential for broader application across resistant pathogens.