Generative AI Designs New Antibiotics From Scratch

​​​​​​Biotechnology and pharmaceutical industries are being disrupted by artificial intelligence (AI) machine learning. Pioneering scientists are applying AI to shorten drug discovery and design. A new study by researchers at McMaster University and Stanford University shows how an AI generative model can rapidly design novel small-molecule antibiotic candidates that are effective against Acinetobacter baumannii (A. baumannii), a harmful, antibiotic-resistant bacteria.

Antimicrobial resistance (AMR) is a worldwide public health threat that occurs when harmful bacteria, viruses, fungi, and parasites develop, over time, a resistance to the drugs designed to destroy them or inhibit their growth. Globally, antimicrobial resistance played a role in 4.95 million deaths, of which 1.27 million deaths were directly caused by bacterial AMR in 2019, according to the World Health Organization (WHO). In the U.S., the Centers for Disease Control and Prevention (CDC) estimates that there are over 2.8 million antimicrobial infections that cause 35,000 deaths per annum. Although antimicrobial resistance is a natural process that involves genetic mutations that occur with the passing of time, the WHO attributes the inappropriate usage and overutilization of antimicrobials for humans, animals, and agriculture as the main culprit of the AMR global health crisis. To put this in context, if the AMR crisis is left unaddressed, a wide range of healthcare outcomes is impacted, ranging from simple dental and medical procedures to more complex surgeries, transplants, and even chemotherapy for cancer treatment.

Discovering new drugs to combat antimicrobial resistance is challenging. The traditional drug development process is a long, costly, and risky endeavor that is ripe for disruption. On average, drug development costs $2.6 billion and takes 10 to 15 years from discovery to regulatory approval, according to PhRMA. Statistically, an overwhelming majority of drug candidates will never be approved by the U.S. Food and Drug Administration (FDA). There was only a 7.9 percent overall likelihood of approval by the FDA from Phase I for all developmental drug candidates between 2011 and 2020, according to a report by Biotechnology Innovation Organization, PharmaIntelligence Informa, and Quantitative Life Sciences.

Artificial intelligence machine learning algorithms can help accelerate drug discovery and design. Generative artificial intelligence is a deep learning model that can produce entirely new content, such as images, text, audio, video, or simulations that resemble the data used for training the algorithm. ChatGPT, DALL-E, Midjourney, Stable Diffusion, and Gemini (formerly Bard) are examples of generative AI.

“Our entire pipeline—including training set curation, model training, molecule generation, chemical synthesis, and experimental validation—could be performed in about three months, demonstrating that generative AI is a powerful tool for rapidly exploring vast chemical spaces for new drug candidates that are easy to acquire in the laboratory,” wrote corresponding authors Jonathan Stokes at McMaster University and James Zou at Stanford University along with researchers Kyle Swanson, Gary Liu, Denise Catacutan, and Autumn Arnold.

For this study, a team of scientists created a generative AI model called SyntheMol that uses Monte Carlo tree search (MCTS) to construct novel compounds from roughly 132,000 building blocks consisting of molecular fragments and 13 validated chemical reactions.

“These building blocks allow for the exploration of a chemical space of nearly 30 billion molecules that are easy to synthesize, with synthesis success rates of over 80 percent within three to four weeks,” the researchers wrote.

SyntheMol was tasked to design molecules that have antibacterial activity by inhibiting the growth of A. baumannii, a Gram-negative bacterium that is resistant to multiple drugs and available antibiotics.

Gram-negative bacteria are highly resistant to most classes of antibiotics. It is named after their reaction to a microbiological method to identify and classify bacteria called the Gram stain, which was developed in 1884 by Danish physician Hans Christian Gram. Gram-negative bacteria stain pink due to its cell walls with thinner layers than Gram-positive bacteria, which stain purple. Infections caused by Gram-negative bacteria cause urinary tract infections, E. coli infections, Salmonella infections, sepsis, pneumonia, meningitis, the plague, cholera, typhoid fever, cat-scratch disease (Bordetella), Legionnaire’s disease, clostridium botulinum, gonorrhea, peptic ulcers, and more.

To create a training database for their AI model, the scientists screened three chemical libraries consisting of bioactive compounds for over 9,000 molecules and more than 5,300 commercially available synthetic molecules for growth inhibition against A. baumannii.

The scientists directed SyntheMol to produce molecules that can be synthesized with two or three building blocks and one chemical reaction to facilitate quick and cost-effective development. The team then screened the high-scoring generated compounds and narrowed the pool down to 150 molecules, of which 58 were successfully synthesized and experimentally validated. After screening for toxicity, the pool of molecules narrowed down to six non-toxic drug candidates with antibacterial properties against A. baumannii. Furthermore, the AI model provided guidance on how to create the generated novel small molecules.

Pioneering researchers are using state-of-the-art generative AI to help shorten the time frame for drug development. This proof-of-concept illustrates the revolutionary potential of AI to directly address the growing threat of antimicrobial resistance, improve health care worldwide, and accelerate scientific innovation for a better future ahead.