The Antibiotics-AI Project at MIT pairs deep learning with experimental biology to discover and design entirely new classes of antibiotics against the world's deadliest bacterial pathogens.
Antibiotics are essential to modern medicine. The continuous evolution of antibiotic-resistant bacteria and a dwindling antibiotic discovery pipeline have resulted in an antimicrobial resistance (AMR) crisis. The United Nations estimates that by 2050, antibiotic resistance will lead to 10 million deaths annually, surpassing cancer. There is a critical need for new antibiotics, and yet pharmaceutical and biotechnology companies have largely abandoned the space in favor of more lucrative markets. Without immediate action to rapidly discover and develop new antibiotics, healthcare systems around the world are at risk of collapse.
To address this challenge, the Antibiotics-AI Project at MIT integrates world-class expertise and decades of experience in artificial intelligence (AI), bioengineering, and life sciences to rapidly discover and design entirely novel classes of antibiotics against the world's deadliest pathogens.
By setting our sights on the superbugs most likely to infect and kill humans over the next 30 years, we are paving the way for improvements in patients' lives and the prevention of an untold number of deaths. Our target pathogens:
A closed loop where the lab and the model teach each other: experiments feed the AI, and the AI points back to the next experiment. Hits are discovered or designed and experimentally validated in vitro and in vivo.
Screen thousands of compounds for growth inhibition against priority pathogens.
Train deep neural networks to predict and explain antibacterial activity from structure.
Search 70B+ molecules, or generate de novo ones, for novel, non-toxic candidates.
Synthesize or procure top hits and confirm efficacy in vitro and in mouse models of bacterial infection.
A deep-learning screen finds readily available compounds active against drug-resistant gonorrhea, including an aminothiazole that inhibits alanine racemase.
Generative models design brand-new antibacterial molecules from scratch; NG1 against gonorrhea and DN1 against MRSA; confirmed in animal models.
Explainable graph neural networks uncover a new structural class of antibiotics active against MRSA and vancomycin-resistant enterococci.
A screen for compounds that kill non-growing, metabolically dormant bacteria identifies semapimod and related outer-membrane disruptors.
Deep learning predicts abaucin, a narrow-spectrum antibiotic that kills the hospital superbug Acinetobacter baumannii while sparing the microbiome.
The first deep-learning screen for antibiotics discovers halicin, a broad-spectrum compound that kills bacteria by collapsing their membrane potential.
An interpretable, white-box machine-learning model links antibiotic-induced metabolic disruption to bacterial killing, illuminating mechanisms of action.
A step-by-step, open-source platform for training explainable deep-learning models that discover and rationalize new antibacterial molecules.
A benchmark of AlphaFold-enabled molecular docking for antibiotic discovery, with datasets mapping where structure-based screening works.
A perspective on how machine learning is reshaping antibiotic discovery, from large-scale screening to generative molecular design.
A review of machine-learning approaches for identifying and designing antimicrobial peptides.
A perspective on applying artificial intelligence across diagnostics, therapeutics, and surveillance to fight infectious diseases.
Empirical growth-inhibition measurements across our panel of priority pathogens; the labelled data that trains our activity models.
Download dataset ↗Viability readouts against 3 human cell lines, used to steer the discovery of safe, selective molecules.
Download dataset ↗Absorption, distribution, metabolism, and excretion profiling to prioritize candidates with drug-like properties.
Code for explainable deep-learning antibiotic discovery, including the structural-class work.
github.com/felixjwong/antibioticsai ↗ github.com/felixjwong/protocol ↗Generative models for de-novo antibiotic design (NG1 and DN1).
github.com/aartikrish/de-novo-antibiotics ↗Message-passing neural networks for molecular property prediction; the D-MPNN framework behind our antibiotic-activity models.
github.com/chemprop/chemprop ↗ Open-source modelA parameter-efficient molecular fingerprinting foundation model, pre-trained on thousands of biological and quantum tasks.
github.com/graphcore-research/minimol ↗The Antibiotics-AI Project is made possible by several philanthropic initiatives and support from our U.S. federal research partners. We are grateful to all of our funders.
An initiative of TED ↗
Flu Lab LLC ↗
Rosamund Zander and Hansjörg Wyss ↗
Mohammed Abdul Latif Jameel ↗
Sea Grape Foundation
An anonymous donor
TARGET project ↗
Defense Threat Reduction Agency ↗
National Institutes of Health (NIH) ↗
Phare Bio ↗
MIT Jameel Clinic ↗