Defesa de Dissertação de Mestrado: Prediction of antimicrobial combinations activity through the use of chemBERTa pre-trained model representation

Orientadores:
Marisa Fabiana Nicolás - Laboratório Nacional de Computação Científica - LNCC
Laurent Emmanuel Dardenne - Laboratório Nacional de Computação Científica - LNCC
Isabella Alvim Guedes - Laboratório Nacional de Computação Científica - LNCC

Banca Examinadora:
Marisa Fabiana Nicolás - Laboratório Nacional de Computação Científica - LNCC (presidente)
Rômulo Gonçalves Agostinho Galvani - Laboratório Nacional de Computação Científica - LNCC
Carlos Alberto Brizuela Rodríguez - Ensenada Center for Scientific Research and Higher Education - CICESE

Suplentes:
Fabio Lima Custodio - Laboratório Nacional de Computação Científica - LNCC
Eduardo Kre mpser da Silva - Fundação Oswaldo Cruz - FIOCRUZ

Resumo:

The combination of two or more drugs is a promising strategy to combat the increasing Antimicrobial Resistance (AMR) crisis. Synergistic combinations are particularly valuable, as they potentially reduce the mutation rate of resistance development. However, the high cost required for experimental screening is challenging, and, up to date, only one research group has made a significant advance. Building upon their study, we developed a two-step Machine Learning approach for the in silico prediction of drug combinations effects.
We first represent drugs as embeddings using the pre-trained chemBERTa model, leveraging transfer learning. In the first classification step, a model distinguishes between additive and non-additive effects. To address class imbalance, we develop a custom undersampling method that maintains the chemical diversity of the dataset. A second model then classifies non-additive effects as either antagonistic or synergistic across six bacterial strains.
The results show that the LightGBM model consistently outperformed other models in both classification tasks. The addictive vs non-additive models demonstrate a high performance (AU-ROC> 0.8), identifying the most common additive effects and providing a potential tool for prioritizing promising drug pairs. Notably, models trained on resistant strains consistently achieved higher performance than those trained on susceptible strains, suggesting that the non-additive effect in resistant bacteria may be restricted to a more predictable space of combinations. While the antagonism vs. synergism models require further refinement, they hold significant potential for predicting synergistic effects.