Development of deep learning models for the discovery of new molecules with antimicrobial activity

• Grant number: 25/07477-4
• Support Opportunities: Scholarships in Brazil - Doctorate
• Start date: January 01, 2026
• End date: April 30, 2029
• Field of knowledge: Biological Sciences - Microbiology - Applied Microbiology
• Principal Investigator: Ricardo Roberto da Silva
• Grantee: Bruno Henrique Dias de Oliva
• Host Institution: Faculdade de Ciências Farmacêuticas de Ribeirão Preto (FCFRP). Universidade de São Paulo (USP). Ribeirão Preto , SP, Brazil

Abstract

The rise of bacterial resistance to antimicrobials poses a challenge to public health, underscoring the need for novel bioactive compounds. Traditional drug discovery methods often lead to rediscovery of known molecules, limiting the development of new therapeutics. Recent advances in artificial intelligence, particularly deep learning models, have proven to be powerful tools for exploring chemical space, enabling targeted design, and optimizing molecular structures. In this context, this project aims to develop deep learning models for generating novel molecules with potential antimicrobial activity against ESKAPE pathogens. To achieve this, we will explore architectures and approaches for both predictive and generative models. Predictive models will be trained to assess the antimicrobial potential of candidate compounds, while generative models will be designed to create optimized molecular structures. Training will be based on a diverse dataset of antimicrobial molecules, including conventional compounds from public databases and structures mined from metagenomic data, inferred through the prediction of biosynthetic gene activity and structure. Additionally, predictive and generative models will be integrated using reinforcement learning, allowing the generative model to be continuously refined based on feedback from the predictor. This approach will allow the efficient exploration of chemical space, promoting the discovery of novel antimicrobial candidates. Finally, by evaluating different data preprocessing strategies, model architectures, and validation techniques, we aim to identify the most promising molecular structures for potential synthesis. (AU)