MIT Researchers Pioneer Novel Antibiotics with AI Assistance
Leveraging advancements in artificial intelligence, researchers at the Massachusetts Institute of Technology (MIT) have engineered innovative antibiotics aimed at addressing two particularly challenging infections: drug-resistant Neisseria gonorrhoeae and multi-drug-resistant Staphylococcus aureus (MRSA).
The research team utilized generative AI algorithms to conceptualize over 36 million potential compounds, meticulously screening these for their antimicrobial efficacy. The most promising candidates emerged as structurally unique entities, diverging notably from established antibiotics, and they appear to operate through groundbreaking mechanisms that disrupt the integrity of bacterial cell membranes.
This innovative methodology empowered the researchers to theorize and assess compounds that have yet to be discovered — a pioneering strategy they aspire to extend to various bacterial species in the future.
“We’re enthused by the new avenues this project unveils for antibiotic development. Our study illustrates AI’s transformative potential in drug design, enabling us to explore extensive chemical spaces that were previously unattainable,” remarked James Collins, the Termeer Professor of Medical Engineering and Science at MIT’s Institute for Medical Engineering and Science (IMES) and Department of Biological Engineering.
Collins serves as the senior author of the study, which was published today in the journal Cell. The lead authors include MIT postdoc Aarti Krishnan, former postdoc Melis Anahtar ’08, and Jacqueline Valeri, PhD ’23.
Delving into Chemical Realms
Over the last four and a half decades, the FDA has sanctioned merely a handful of new antibiotics, most being mere derivatives of existing medications. Simultaneously, the escalating bacterial resistance to these agents has become a grave concern. A staggering 5 million fatalities per year are attributed to drug-resistant bacterial infections globally.
In an effort to unearth new antibiotics capable of countering this surging crisis, Collins and his colleagues at MIT’s Antibiotics-AI Project have harnessed AI’s capabilities to scrutinize extensive libraries of known chemical compounds. This initiative has produced several promising drug candidates, including halicin and abaucin.
Building on these advancements, the researchers chose to venture into uncharted chemical territories — designing molecules that remain absent from chemical libraries. By employing AI to fabricate hypothetical molecules that are either non-existent or undiscovered, they aimed to tap into a far broader spectrum of prospective drug compounds.
In their recent investigation, the researchers adopted a dual approach: First, they directed generative AI algorithms to craft molecules anchored in specific chemical fragments exhibiting antimicrobial characteristics; secondly, they permitted the algorithms to autonomously devise molecules without the prerequisite of specific fragments.
For the fragment-centered strategy, the researchers sought to isolate molecules capable of eliminating N. gonorrhoeae, a Gram-negative bacterium responsible for gonorrhea. To commence, they compiled a library comprising approximately 45 million recognized chemical fragments, integrating all conceivable combinations of 11 atoms — carbon, nitrogen, oxygen, fluorine, chlorine, and sulfur — alongside fragments from Enamine’s REadily AccessibLe (REAL) space.
Subsequently, they employed machine-learning models, previously trained by Collins’ lab on predicting antibacterial activities against N. gonorrhoeae, to scrutinize this library, yielding nearly 4 million fragments. Further narrowing was undertaken to eliminate fragments predicted to exhibit cytotoxicity towards human cells, display chemical vulnerabilities, or show similarity to existing antibiotics, ultimately leaving them with approximately 1 million viable candidates.
“Our objective was to exclude any entities resembling established antibiotics, aiming instead to confront the antimicrobial resistance predicament through an entirely novel lens. By delving into lesser-explored chemical domains, our ambition was to unearth unique mechanisms of action,” Krishnan elaborated.
Through successive experimental rounds and computational evaluations, the researchers identified a fragment termed F1 that demonstrated promising activity against N. gonorrhoeae. They then utilized this fragment as the foundation for generating additional compounds via two distinct generative AI algorithms.
One algorithm, designated chemically reasonable mutations (CReM), initiates with a specific molecule, inclusive of F1, generating new molecules by adding, replacing, or omitting atoms and chemical groups. The alternative algorithm, F-VAE (fragment-based variational autoencoder), constructs a complete molecule from a chemical fragment by deciphering how fragments are typically modified, drawing insights from its pretraining on over 1 million entries from the ChEMBL database.
These algorithms together produced around 7 million candidates containing F1, which were then computationally assessed for their efficacy against N. gonorrhoeae. This evaluation resulted in roughly 1,000 viable compounds, from which the researchers selected 80 for feasibility assessments by chemical synthesis vendors. Ultimately, only two were synthesized, with one — dubbed NG1 — exhibiting significant efficacy in eradicating N. gonorrhoeae both in vitro and within a murine model of drug-resistant gonorrhea infection.
Follow-up experiments unveiled that NG1 engages a protein known as LptA, a novel target implicated in bacterial outer membrane synthesis. The drug appears to impede membrane construction, thereby proving detrimental to cell viability.
Unrestricted Design Paradigms
In a subsequent phase of research, the team probed the capacity for generative AI to autonomously devise molecules targeting Gram-positive bacteria, namely S. aureus.
The researchers again employed CReM and VAE without any restrictions, save for the foundational principles governing atomic bonding in chemical formations. This freedom resulted in the generation of more than 29 million compounds. They then applied the same screening criteria utilized for the N. gonorrhoeae candidates, ultimately honing the selection to around 90 compounds.
Of these, they successfully synthesized and evaluated 22 molecules, with six exhibiting robust antibacterial properties against multi-drug-resistant S. aureus in laboratory cultures. Notably, the leading candidate, identified as DN1, demonstrated the ability to resolve a methicillin-resistant S. aureus (MRSA) skin infection in a murine model. These molecules seemed to disrupt bacterial cell membranes but with a wider range of effects, not confined to interaction with a singular protein.
Phare Bio, a nonprofit organization integral to the Antibiotics-AI Project, is actively engaged in further modifying NG1 and DN1 to render them amenable to additional testing.
“In collaboration with Phare Bio, we are investigating analogs, alongside progressing our premier candidates preclinically through medicinal chemistry endeavors,” Collins expressed. “We are also eager to extend the methodologies developed by Aarti and her team to other bacterial pathogens of significance, notably Mycobacterium tuberculosis and Pseudomonas aeruginosa.”
This groundbreaking research received funding from the U.S. Defense Threat Reduction Agency, the National Institutes of Health, the Audacious Project, Flu Lab, the Sea Grape Foundation, Rosamund Zander, Hansjorg Wyss of the Wyss Foundation, and a confidential donor.
Source link: News.mit.edu.
