Chapter 148
Defeating Infectious Agents — 12
While that fund is a start, we face the prospect of ten million deaths from antibiotic resistance by 2050, according to the World Health Organization. Thanks to academic labs, using AI and structure-based discovery, we are seeing a jump forward in promising candidates that are going into clinical trials. The table below summarizes recent discovery of potent molecules against multiple species of concern, using different AI models and structure-based drug discovery.
Overcoming Antibiotic Resistant Species
Method
Number of Compounds Screened
Citation
C. difficile; M. tuberculosis; A. baumannii; Enterobacteriacae
Deep neural network
107M – >23 – >8
Stokes J, Cell, February 20, 2020
Burkholderia cenocepacia
Deep neural network
225K – >43 – >5
Rahman A, PLoS Comp Biol, October 13, 2022
Staph aureus; E. coli; Pseudomonas aeruginosa
Structure-based drug discovery
NA
Wu K, Science, February 16, 2024
Staph aureus
Graph neural network
12M – >283 – >4
Wong F, Nature, December 20, 2023
A. baumannii
Generative AI
13K – >58 – >6
Swanson K, Nature, March 22, 2024
E. coli, K. pneumonaie
Structure-based drug discovery
NA
Huseby D, PNAS, April 5, 2024
Source: compiled by author (original)
MIT researchers discovered Halicin, a drug initially developed to treat diabetes but later found to have antibiotic properties, and related molecules using deep-learning AI. They assessed over 107 million molecules, whittling this down to 23, and identified 8 with potent activity against several pathogens, some verified with experimental models. A similar approach was applied to find an effective antibiotic against Burkholderia cenocepacia from over 225,000 molecules, with five compounds having very strong activity. The first new structural class of antibiotics in over three decades came from a graph neural network, a type of AI. How it worked, starting with over twelve million compounds to get to four nontoxic effective ones, was clearly not simple (these compounds were nontoxic in animal models against methicillin-resistant Staphylococcus aureus). A hunt for an antibiotic to A. baumannii was the first to use generative AI, ingesting data from over thirteen thousand molecules to get to six with considerable activity. The beauty of this work is that it focused on easy-to-synthesize small molecules, which would have practical advantages for mass production at low cost. AI has also enabled new antibiotic peptide discovery at scale by mining the proteomes of extinct organisms (the “extinctome”!) and from the global microbiome public database. Both efforts predicted tens of thousands of candidate antimicrobial peptides, for which initial assessment of more than one hundred validated them to have functional activity against bacteria. Large-scale screening of drugs that are not antibiotics has also been undertaken, yielding many candidate molecules with potent activity. A large language model like ChatGPT was used to reengineer an existing antibiotic to kill a multidrug-resistant, toxic bacterium.