For the past sixty years, researchers have made a groundbreaking discovery in combating antibiotic-resistant Staphylococcus infections using a novel antibiotic identified through the application of artificial intelligence (AI) and machine learning. This significant advancement in global healthcare was achieved by a collaboration between researchers from the Massachusetts Institute of Technology (MIT), Harvard University, and the Broad Institute of MIT and Harvard located in Cambridge, Massachusetts.
Antibiotic resistance is a major cause of death worldwide and poses a significant public health threat. According to the Antimicrobial Resistance Review commissioned by the UK government, it is estimated that by 2050, 10 million people will die annually due to antibiotic resistance. A study published in The Lancet in 2019 revealed that globally, 4.95 million deaths were attributed to bacterial antibiotic resistance, with 1.27 million deaths directly caused by antibiotic resistance. The Centers for Disease Control and Prevention (CDC) reported in their 2019 Antibiotic Resistance Threats in the United States that there are over 2.8 million cases of antibiotic-resistant infections in the US each year, resulting in 35,000 deaths.
Antimicrobials are substances that disrupt or inhibit the growth of microorganisms, including antiviral agents, antiparasitic agents, antibiotics, and antifungal agents. Antibiotic resistance occurs when microorganisms mutate or adapt, rendering antibiotics ineffective. Improper or excessive use of antibiotics accelerates this natural process, leading to harmful viruses, bacteria, parasites, and fungi developing resistance to antimicrobial agents and antibiotics.
For instance, the overuse of antibiotics by humans and their inclusion in livestock feed is contributing to the emergence of drug-resistant strains. According to a report by Landers et al. titled “Review of Antimicrobial Use in Food Animals: Perspective, Policy, and Potential,” approximately 88% of pigs are fed antibiotics such as tetracyclines or tylosin, 42% of beef cattle are fed tylosin, and virtually all dairy cows receive post-milking prophylactic doses of antibiotics such as penicillin, β-lactams, or cephalosporins.
According to the World Health Organization’s Global Antimicrobial Resistance and Use Surveillance System (GLASS) report in 2022, the median methicillin-resistant Staphylococcus aureus (MRSA) rate in 76 countries was 35%, with over 40% inhibition of third-generation cephalosporin-resistant Escherichia coli (E. coli).
Staphylococcus aureus, also known as golden staph, is a Gram-positive bacterium that can cause various infections in humans, such as skin infections, sepsis, and life-threatening pneumonia. According to a report by the Institute for Health Metrics and Evaluation in January 2022, MRSA caused over 120,000 deaths globally in 2019.
Dr. James Collins, a professor at MIT and a co-author of the study, stated, “Our research demonstrates that a graph-based chemical substructure search can provide better understanding and interpretation of graph neural networks, thereby summarizing the model’s predictions.”
The scientists utilized a graph neural network (GNN) artificial intelligence platform called Chemprop. The AI graph neural network utilizes information from molecular bonds and atoms to make predictions for each molecule. A graph neural network is an artificial neural network capable of processing graph data structures for prediction, analysis, and classification tasks.
The researchers noted, “An enticing implication of this study is that deep learning models in drug discovery can be made interpretable.”
The scientists screened over 39,300 compounds for their growth-inhibitory activity against methicillin-sensitive Staphylococcus aureus strain RN4220, resulting in the identification of 512 active candidate compounds. The screening data were used to train a collection of artificial intelligence graph neural networks to predict whether new compounds inhibit bacterial growth based on their molecular chemistry, atoms, and bonds.
An important aspect of drug discovery is eliminating compounds that may harm human cells or be toxic to them. Therefore, the researchers performed a reverse screening of the training database of over 39,300 compounds to predict cytotoxicity. To assess general cytotoxicity and hepatotoxicity, they reverse screened for cytotoxicity in human liver cancer cells (HepG2). To gain further insights into in vivo cytotoxicity, the researchers also reverse screened human lung fibroblasts (IMR-90) and human primary skeletal muscle cells (HSkMCs).
According to the orthogonal model for predicting cytotoxicity, 40% of the 512 active antimicrobial candidate compounds exhibited cytotoxicity, resulting in 306 compounds that showed no cytotoxicity against the three cell types used for screening.
Subsequently, the scientists retrained four sets of 20 AI models using each complete training database to predict the antibiotic capabilities and cytotoxicity against the three cell types (HepG2, HSkMCs, and IMR-90). These four AI ensembles collectively processed input data from over 12 million compounds, with over 11.2 million compounds sourced from the Mcule purchasable database and over 799,000 compounds from the Broad Institute database.
After further screening, the researchers obtained a set of 283 compounds, which were subjected to experimental growth inhibition testing against MRSA in the laboratory. This led to the discovery of two antibiotic candidate compounds that were subsequently tested in mice for local and systemic treatment against MRSA.
“Our approach revealed a diverse set of compounds with antibiotic activity against Staphylococcus aureus. Among them, we identified a structurally distinct class that exhibited high selectivity, overcame resistance, possessed favorable toxicological and chemical properties, and demonstrated efficacy in both local and systemic MRSA infection models in mice.”