In recent years, artificial intelligence programs for drug design have steadily shown they can help explore far beyond just small-molecule drugs—by illuminating how proteins fold and interact, all the way up to large antibodies. Get ready to add another category to the list.
Researchers at the University of Washington’s Institute for Protein Design have put forward a paper showing its AI algorithms can help create millions of never-before-seen drug-like peptides—chains of amino acids that are much smaller than your typical protein, but still capable of delivering potent effects within the human body.
Peptides have formed the basis of dozens of drugs and treatments—including insulin, for example—and serve as the “P” in the hormone GLP-1, the target of today’s blockbusters Ozempic and Wegovy.
According to researchers in UW’s Baker Lab, their AI programs were able to construct ring-shaped peptides known as macrocycles—a group they say has the potential to cross tricky cell membranes to interrupt the body’s pain signals, disrupt viral infections or interfere with growing tumors, while possibly being delivered as an oral pill.
“Certain snails, sponges and other marine animals produce macrocycles with potent activities, and scientists have had some success turning these natural products into medicines,” said Patrick Salveson, who helped lead the study while at the Baker Lab, which was published this week in the journal Science.
“But until now, there hasn’t been a way to systematically create new macrocycles that might treat specific diseases. Our work shows that this promising class of chemicals can be systematically explored using computational design,” said Salveson, who currently serves as co-founder and chief technology officer of Vilya Therapeutics, a UW spinout that made its debut in 2022 and has licensed the tech for development.
The work to develop deep-learning programs for generating peptides has been slow in part because there haven’t been enough molecule models in the relevant size range to help train the algorithms.
“The challenge here was to come up with an efficient way to model these chemicals on the computer,” said study author Adam Moyer, formerly of the Baker Lab, and now director of molecular design and co-founder of Vilya.
“We found a solution that combines the accuracy of AIMNet, which is an AI-powered tool for simulating quantum mechanics, with the speed of a more traditional software approach. We optimized these steps to create a new way to quickly build ring-shaped compounds,” Moyer said.
Previously, members of the Baker Lab, Vilya and others also showed that they could build on the work of DeepMind’s AlphaFold protein-computing network and adapt it for smaller cyclic peptide chains, to help custom design new potential drugs for a variety of diseases.
Now, the researchers have demonstrated that they can systematically generate and manufacture small macrocycles with four or fewer amino acids—with subsequent X-ray and nuclear magnetic resonance tests showing that they come close to aligning with the original AI blueprints.
To start, they picked one that could potentially inhibit a protein from the coronavirus behind COVID-19, as well as additional macrocycles capable of blocking a protein essential to the survival of cancer cells without affecting healthier pathways.
Almost all were able to cross artificial cell membranes in preclinical experiments and could survive enzymatic degradation processes for more than 24 hours, according to the researchers.
Another Baker Lab spinout made headlines earlier this week, with the antibody-focused Xaira Therapeutics posting a $1 billion debut fundraising—led by Arch Venture Partners, which also backed Vilya.
Meanwhile, drugmakers such as Merck & Co. have called macrocyclic peptides “the next wave of drug discovery,” and earlier this year inked a $220 million deal with the AI designer Unnatural Products. And last year, Roche’s Genentech signed a $1 billion pact with the Japanese biotech PeptiDream, which is developing a peptide discovery platform.