Introduction

The year 1944 witnessed the simultaneous emergence of artificial neural networks, laying the foundation for deep learning, and the discovery of streptomycin, the first aminoglycoside antibiotic.
This historical synchrony ultimately connects deep learning and antibiotics.

Why in news?

In December 2023, scientists introduced a groundbreaking alliance between deep learning and antibiotics by leveraging deep learning techniques to discover a new class of antibiotics, addressing a multi-decade gap in antibiotic development.

Different Approach: Unlike previous applications of deep learning in drug discovery, this study focused on identifying chemical motifs or substructures used by the deep learning model to evaluate compounds for antibiotic potential, rendering the model “explainable”.
Proven Efficacy: The research successfully demonstrated the effectiveness of two compounds from the newfound antibiotic class against methicillin-resistant Staphylococcus aureus (MRSA) infections, a major cause of human fatalities in 2019.
Recognition and Expansion: Experts praised the study for its contributions to antibiotic research and its potential to enhance drug development strategies.

Neural Networks: Deep learning relies on artificial neural networks, comprising layers of artificial “neurons” that process inputs and yield outputs through training and testing phases.
Training and Testing: Deep learning networks are trained on large datasets with annotated inputs to learn specific tasks. During testing, they classify novel inputs based on their learned knowledge.
The Black Box Issue: Most deep learning models lack transparency in explaining how they arrive at their conclusions, remaining “black boxes.”
Explainable Deep Learning: In contrast, the study’s model was designed to be explainable, allowing it to not only predict antibiotic potential but also elucidate the substructures contributing to this property.

Experimental Screening: The research began by screening over 39,000 compounds to inhibit S. aureus growth, shortlisting 512 active compounds.
Graph Neural Network (GNN): A GNN was trained on the dataset, representing atoms as nodes and bonds as edges on a mathematical graph.
Selecting Non-Toxic Compounds: To ensure safety, 306 compounds were identified that didn’t harm human cells, and other GNNs were trained to identify cytotoxic compounds.
Identifying Potential Antibiotics: The GNNs evaluated a database of over 1.2 crore compounds, identifying 3,646 potential antibiotics based on substructures.
Substructure Rationales: The study introduced “rationales” to explain the substructures that conferred antibiotic properties to molecules.
Efficacy Against MRSA and VRE: Certain compounds, including N-[2-(2-chlorophenoxy)ethyl]aniline, exhibited inhibition of MRSA and vancomycin-resistant enterococci (VRE).
Mouse Models: One compound effectively reduced MRSA-related skin and thigh infections in mouse models.

Transparency in Drug Discovery: The study’s significance lies in rendering deep learning approaches to drug discovery more transparent and reproducible across drug categories.
Future Exploration: Researchers are applying substructure rationales to design new antibiotics and explore applications in drugs targeting age-related disorders.
Addressing a Lacuna: An identified shortcoming is that explainability analysis occurred after predicting antibiotic properties. Implicitly incorporating explainability in deep learning models is proposed as a more robust approach.