Deep learning unlocks new antibiotic classes, revolutionizing the fight against resistance

A recent Nature journal study uses a deep learning approach to explore chemical sub-structures that helped discover structural classes of antibiotics.

Study: Discovery of a structural class of antibiotics with explainable deep learning. Image Credit: kasarp studio / Shutterstock.com

Developing novel antibiotics

Previous studies have indicated an interval of 38 years between the introduction of fluoroquinolone and oxazolidinones, thus demonstrating the extensive time needed to discover a new structural class of antibiotics. The ongoing antibiotic resistance crisis has emphasized the importance of developing new antibiotics. 

Antibiotic resistance and a lack of new antibiotics have increased morbidity from bacterial infections. Typically, antibiotics are discovered based on structure-guided and rational design, natural product mining, evolution and phylogeny analyses, high-throughput screening, and in silico screens using machine learning.

Discovering novel antibiotic agents with a large structural diversity of chemical space is extremely difficult. Deep learning methods have been used to identify potential antibiotics from large chemical libraries to overcome this challenge.

For example, halicin and abaucin were identified from the Drug Repurposing Hub, which comprises over 6,000 molecules. This approach was also used to identify antibacterial agents from the ZINC15 library, which contains approximately 107 million molecules. The ZINC15 library used Chemprop, a platform for graph neural networks based on typical black box models or models that cannot be interpreted easily. 

The current study applied graph neural network models trained with large datasets linked to antibiotic activity measurements and human cell cytotoxicity. It was hypothesized that model predictions could be explained by chemical sub-structures determined using graph search algorithms.

Since antibiotic classes are categorized based on shared sub-structures, the current study proposed that sub-structure identification could be used to explain model predictions.

Models for antibiotic activity and human cytotoxicity

The current study aimed to discover antibiotic classes effective against Staphylococcus aureus, a Gram-positive pathogenic bacterium. This bacterium was selected as it has shown resistance against many first-line antibiotics and causes many difficult-to-treat nosocomial infections.

A total of 39,312 structurally diverse antibiotic agents were screened for antibiotic activity against a methicillin-susceptible strain of S. aureus. About 1.3% of all compounds exhibited antibacterial activity.

Chemprop was used to train ensembles of graph neural networks. The screening data was used for binary classification predictions, which indicated whether a new compound can inhibit bacterial growth based on its chemical structure. 

The graph neural network operates by implementing complex steps based on atoms and chemical bonds of each molecule. Through convoluted steps, each model generated a prediction score between zero and one, which indicated the probability of the molecule’s antibacterial activity.

Chemprop models with RDKit-computed molecular features that exhibit molecular features can be used to predict antibiotic activity. The model outperformed other deep learning models, such as random forest.

Orthogonal models were used to predict cytotoxicity in human cells. The outcome was used to identify compounds that could be effective against S. aureus.

Some compounds were found to be cytotoxic for human liver carcinoma cells (HepG2), human lung fibroblast cells (IMR-90), and human primary skeletal muscle cells (HSkMCs). Compared to HepG2 and HSkMCs, cytotoxicity models were more predictive for IMR-90 cells.

Discovery of novel structural classes of antibiotics

The current study identified putative structural classes of antibiotics through graph-based explanations of deep-learning model predictions. These models were trained on the antibiotic activity and cytotoxicity of 12,076,365 compounds.

Multiple compounds that showed antibacterial activity against S. aureus were identified. However, one structural class exhibited superior selectivity and the ability to overcome resistance. Importantly, this class of antibiotics also indicated favorable chemical and toxicological properties.

A mouse model revealed that the new structural class of antibiotics was effective against both topical and systemic treatment of methicillin-resistant S. aureus (MRSA) infection. Furthermore, structure-activity relationship analyses indicated that this structural class could be optimized for higher sensitivity and selectivity against Gram-positive pathogens and improved permeability against Gram-negative pathogens.

The current study highlighted the deep learning approach’s effectiveness in discovering new antibiotic classes. A new structural class of antibiotics can be identified based on predictions of single compound hits and analyzing their chemical substructures. In addition to the down-sampling of chemical space, another advantage of this approach is the ability to automate the identification of unprecedented structural motifs.

A better understanding of graph-based rationale predictions could enable the discovery of new antibiotic classes. The current study approach could be used as the foundation for developing future predictions using deep learning models.