AI technology AlphaFold used to discover new cancer drugs in a very fast pace
In order to reduce the time and cost associated with developing promising new medicines, researchers from Insilico Medicine and the University of Toronto have successfully tested the use of artificial intelligence to guide the design of long-acting injectable drug formulations, which shrunk what usually takes years or even decades to less than a month.
The study may be the world’s first to apply a groundbreaking AI technology called AlphaFold to drug discovery research. Although their new drug still needs to go through clinical trials, the paper’s authors say their process demonstrates the “revolutionary” potential of AI in medical research.
According to the study, published last week in journal Chemical Science, the scientists first used AI to scan a type of liver cancer called hepatocellular carcinoma for protein “weak spots.” Once one was detected, another program designed small molecules that could target and take down the specific protein, the paper says. Researchers then tested these molecules on live cells, one of which appeared effective at slowing cancer growth.
The AI-powered drug discovery pipeline starting from detection of the “weak spot” protein, to scanning AlphaFold databases to predict what the protein looks like, to using AI to generate an inhibitor of that protein, to finally testing the inhibitor on live cancer cells.
The entire process, from discovery of the “weak spot” to drug creation and testing, took only 30 days.
The project was the result of a collaboration between Insilico Medicine — a multinational biotechnology company dedicated to using AI to improve health care — and the University of Toronto’s Acceleration Consortium.
Insilico said it’s currently not interested in pursuing clinical trials for the potential drug; it’ll leave that for other researchers to follow now that the molecule has been publicly identified. Instead, the main purpose of the study was to serve as a “proof of concept” of what is now possible with AI, said Alán Aspuru-Guzik, a professor of computer science and chemistry at U of T, director of the Acceleration Consortium and the co-principal investigator who led the study.
“(We’ve had) incredible progress in the technology,” Aspuru-Guzik said. “The fact that we’re already casually talking about discovering a lead for a drug discovery program in only 30 days — it’s incredible.”
AI has been used in biochemistry for years — Aspuru-Guzik is himself one of the pioneers in the field, having first applied AI in his chemistry research over a decade ago. With recent AI advances, the once-obscure space is now exploding, the professor said — and it’s partly thanks to AlphaFold.
Released in 2021, AlphaFold is a revolutionary AI program developed by Google’s DeepMind researchers. It solves one of the greatest puzzles in biochemistry: predicting what proteins look like solely based on their DNA blueprints. In the brief time the program has been out, it’s already identified the structures of over 200 million human proteins — providing an invaluable resource for researchers in life science and biology in general.
It was because of this massive protein database that the team’s project was possible, Aspuru-Guzic said. After Insilico’s in-house AI detected a weak point of liver cancer — a relatively unknown protein called CDK20 — the scientists used AlphaFold’s database to accurately predict what that protein looked like, along with its potential weaknesses.
After that, it was relatively simple process of feeding CDK20’s structure into another of Insilico’s AI programs to design drugs that can take down the protein, Aspuru-Guzik continued.
“People were ridiculing chemists who were working on AI, thinking that we were crazy,” Aspuru-Guzik said. “Now, it’s like a revolution.”
Michael Levitt is a Nobel Prize-winning chemist, a professor of structural biology at Stanford University, a member of Insilico’s scientific advisory board and an author on the study. One of the largest breakthroughs of AI is its ability to sort through an incredible amount of information in seconds, he said.
This allows the program to “scan very broadly for weak spots, not just a single one … It’s AI’s ability to tie (disparate information) together which means that you can now throw a net very broadly and catch things” humans might otherwise miss, he said.
“Basically, it’s a way of getting a much broader sampling of potential drugs.”
It’s better to have multiple good options to try than a single excellent one, Levitt continued, given how easily things go wrong in clinical trials. Just because a drug works on cultured cells doesn’t mean it’ll work as well inside the body: “It may have side effects. It may be too expensive to manufacture. There’s a lot of things that could go wrong,” he said.
Petrina Kamya is the head of AI platforms at Insilico Medicine. She said it’s unlikely the company will pursue further research into the team’s newly discovered drug.
“So now we’ve made the structure public and everything — the whole target is public, the structure is public — it’s difficult to pursue it without somebody else taking the idea and running with it, and maybe optimizing the molecule in some other way,” she said.
“I would have loved to continue to do this, but this was done more or less as a proof of concept to show that it is possible to use a predicted structure for a novel target and come up with some chemical data that is actually usable.”
According to Kamya, Insilico has plans to further automate drug discovery using AI and robotics — for example, they’re now looking at using AI to streamline clinical trials. Kamya also said they’ve recently launched a new lab researching robots capable of creating and testing new drugs as they’re identified by AI.
Speaking generally on the future, however, Kamya admits it’s near impossible to predict how AI would impact health care given how rapidly the field is currently developing.
“Almost every six months something new that comes out that could potentially have an impact on many different aspects of health care,” she said. “I’m afraid I am very wary of making any predictions. The only thing I can say is that there is bound to be an impact and we’re only beginning to scratch the surface regarding what that could potentially be.”
Meanwhile, Levitt believes AI may soon change the face of medicine.
“I’m sure that AI will soon be incredibly important everywhere, from primary health care to preventative health care to pharmaceuticals,” he said. “We’re going to be much smarter with AI than we were without AI.”
“I was in the field from the very beginning and I would say that I didn’t expect to reach this (level of progress) as quickly as we did,” Levitt continued. “We got here in 50 years and I thought it would be 100 years. “This is a massively important step.”
To that end, Allen and the research team have published their datasets and code on the open-source platform Zenodo.
“For this study our goal was to lower the barrier of entry to applying machine learning in pharmaceutical sciences,” Bannigan said. “We’ve made our data sets fully available so others can hopefully build on this work. We want this to be the start of something and not the end of the story for machine learning in drug formulation.”
Part of the article was reported by the Star.
