(263g) Predicting Anti-Microbial Activity of Natural Products through Machine Learning and Large Language Models

Antibiotics have been the foundation of modern medicine. However, due to the global antibiotic-resistance, the efficacy of these essential antibiotics becomes threatened. Unfortunately, the discovery of new antibiotics is a rising challenge¹. Natural products (NPs) have been indispensable in various fields, including medicine, yet exploring their bioactivity remains challenging due to their limited availability in small quantities². Machine learning (ML) has emerged as a powerful tool for predicting bioactivity, offering the potential to revolutionize NP research. This project aims to leverage ML models and large language models (LLM) to enhance our understanding of NPs, specifically focusing on predicting anti-microbial bioactivity. In this work, leveraging LLMs, we focused on 12 gram-negative bacteria and trained ML models tasked to predict the minimal inhibition concentration (MIC) of each bacterium. We trained a transformer-based regression model using labeled database and used the trained model to guide the experimental validation of a small focused chemical library (Figure 1). Using LLM, our model has outperformed previous methods based on cheminformatics³. Our model has achieved high precision and recall across all 12 different gram-negative bacteria (Figure 2). After the model has been trained, we curated a large NP database by combining NPAtlas and Coconut and gathered >440,000 NPs for the model to predict. Based on the model’s prediction 19 NPs are selected for in vitro experimental validation, out of which 4 compounds showed growth inhibition bioactivity against Acinetobacter baumannii and 1 for Klebsiella pneumoniae, leading to a hit-rate of 26.3%.

Reference:

1. Brown, Eric D., and Gerard D. Wright. "Antibacterial drug discovery in the resistance era." Nature 529.7586 (2016): 336-343.
2. Cox, Georgina, et al. "A common platform for antibiotic dereplication and adjuvant discovery." Cell chemical biology 24.1 (2017): 98-109.
3. Stokes, Jonathan M., et al. "A deep learning approach to antibiotic discovery." Cell 180.4 (2020): 688-702.