The discovery of a chemical compound with antibiotic properties is a helpful case study in the potential — and limits — of using A.I. to develop new treatments
In late February, a paper appeared in the journal Cell with encouraging news regarding one of the world’s most persistent public health problems. Researchers at Massachusetts Institute of Technology and Harvard University had used artificial intelligence to identify a chemical compound with powerful antibiotic properties against some of the world’s most drug-resistant strains of bacteria — a welcome discovery in a world where 700,000 people die every year from drug-resistant infections. It was the first time an antibacterial compound had been identified this way. The researchers named it halicin, in honor of the computer HAL in the film 2001: Space Odyssey.
While the global need for new antibiotics to treat drug-resistant infections is as pressing as it was at the start of the year, the world’s attention has been diverted by the novel coronavirus pandemic, and the hunt for a vaccine that can halt Covid’s spread. Like new antibiotics, new vaccines typically take up to 10 years to deliver. In the case of the Covid vaccine, scientists are working frantically to shorten that timeline. Machine learning has been a vital tool in the fight against Covid, with algorithms assisting in sorting through massive piles of accumulating data and narrowing down potential candidates for vaccines.
The halicin paper is a helpful case study in how machine learning can be a powerful asset in finding elusive treatments. It also highlights artificial intelligence’s limits. Identifying a new antibiotic compound like halicin is only a first small step in the years-long process of evaluating whether that compound can be developed into a safe and effective antibiotic drug. A vaccine is an exponentially more complicated drug than an antibiotic, and while algorithms can point scientists toward promising leads to explore, they can’t do the work of real-world testing.
“At the end of the day, you still have to run the experiments in the laboratory.”
“All of the predictions that any model makes are just predictions,” said Jonathan Stokes, PhD, a postdoctoral researcher in MIT’s James J. Collins Laboratory and the paper’s lead author. “The models can certainly guide your experimentation and point you in the right direction. At the end of the day, you still have to run the experiments in the laboratory.”
Halicin grew from a project at MIT’s Abdul Latif Jameel Clinic for Machine Learning in Health to identify ways that artificial intelligence could make antibiotic discovery faster and more cost-efficient. There are a lot of potentially bactericidal substances out there: Researchers have found antibacterial microbes in everything from soil to Komodo dragon blood to a Dalek parked in the lobby of BBC headquarters to promote the show Doctor Who. But the process of identifying which of those millions of compounds can safely and effectively treat infections in people, and then developing them into actual medications, is time-consuming and expensive.
Fortunately, searching for established patterns in a sea of data is something that intelligent machines do very well. The team built a neural network, an algorithm loosely modeled after the structure of the human brain. They screened a library of 2,335 chemicals to see what happened to the E. coli when exposed to the chemicals, and used those results to train the algorithm to recognize chemical compounds that inhibited bacterial growth.
Then they applied the model to roughly 6,000 compounds in Broad Institute’s Drug Repurposing Hub, a library of U.S. Food and Drug Administration-approved chemicals available to researchers seeking to identify cures for new diseases among drugs originally developed for a different purpose. The machine identified a compound from a diabetes medication that was very good at killing bacteria and very different from the structure of existing antibiotics, making it a promising contender for treating bacteria resistant to existing drugs. With the exception of one highly resistant lung pathogen, halicin worked against every species of bacteria the researchers tested it on, including many highly treatment-resistant strains.
Back in the lab, researchers gave the drug to mice infected with a strain of the pathogen A. baumannii, which has proved impervious to every currently available antibiotic. The bug has plagued U.S. soldiers stationed in Iraq and Afghanistan. Halicin ointment completely cleared the infection in mice within 24 hours.
Encouraged by the results with halicin, the researchers decided to try the model on a dataset that would be virtually impossible for humans to analyze alone. They set the algorithm to work analyzing a set of roughly 107 million chemical compounds with known antibacterial properties, pulled from a larger chemical library.
“I calculated it out: If I were to screen pretty much all day every day, it would take something like 14 years to screen 107 million molecules in the lab,” Stokes said. The machine did it in three days. It identified 23 candidates that met the researchers’ criteria, two of which proved especially promising in early lab tests.
Identifying a compound like halicin is only the first step in the long process of creating an actual antibiotic drug. The chemical compound appears in early preclinical studies to have both strong antibiotic properties and a very different chemical structure from existing antibiotics, which means it could have a better chance against bacterial strains resistant to other drugs.
Yet machine learning has its limits. It takes roughly 10 years and up to $2 billion to turn a preclinical lead into a marketable antibiotic, and those 10 years include a lot of human testing that can’t be cut short or outsourced to a machine. Researchers still don’t know whether halicin can be safely consumed by humans, or what dosage or formulations would be safe and effective. The Cell paper shows how useful A.I. can be for identifying leads in the new treatments we need the most, but a computer can’t tell us whether any chemical can be safely turned into a drug.
“Trying to teach the immune system something, versus just trying to not kill it, is a step change in difficulty.”
And in news that may well sound discouraging to virtually everyone on the planet: As hard and labor-intensive as developing a new antibiotic is, creating a new vaccine is orders of magnitude more complicated.
A vaccine trains the immune system to recognize and attack a virus or bacteria, and it does this by first imitating parts of the pathogen that it’s meant to combat. Creating a substance that’s “trying to teach the immune system something, versus just trying to not kill it, is a step change in difficulty,” said Adam Pah, PhD, a data scientist and a clinical assistant professor at Northwestern University’s Kellogg School of Management. Good data can help build good models, which can be valuable tools in moving the process forward. But there’s no short-cutting the long and laborious process of testing and re-testing the models’ hypotheses.
“Your model may spit out some ideas, but until they’re put to the test — at the very least, an in vitro assay of some sort — they’re not worth much of anything,” said Derek Lowe, PhD, a drug discovery chemist and author of the blog In the Pipeline. “We don’t know enough in any area of drug discovery — vaccines, small molecules, antibodies, what have you — to compute our way to an answer and be sure that it’ll actually work.”
social experiment by Livio Acerbo #greengroundit #thisisnotapost #thisisart