Montreal AI startup with a touted advisor licenses its 'few-shot' platform to Repare for synthetic lethality molecules
The rise of AI in drug discovery has presented a buffet of tasty options for drugmakers to identify better molecules for hard-to-hit targets. That promise is mixed, to say the least, but startups like Montreal’s Valence Discovery think they’ve cracked the code — and at least one small biotech is convinced.
Valence, a recently uncloaked AI startup specializing in deep learning in drug discovery, has licensed its tech to Canada’s Repare Therapeutics to identify effective synthetic lethality molecules, the companies said Wednesday.
Valence’s platform seeks to overcome a crunch on available training data sets by using what it calls a “few-shot learning” approach that needs less starting data to turn out valuable molecules. According to CEO Daniel Cohen, that approach stems from research at the Montreal Institute for Learning Algorithms (Mila), the Montreal tech “ecosystem” spawned by Yoshua Bengio, a computer scientist with a reputation as a deep learning pioneer.
Bengio has joined Valence as an advisor, and Cohen thinks the team’s proprietary platform offers something differentiated over the many deep learning players on the market.
“The goal here is to help our partners very rapidly design high-quality drug candidates that have been optimized for whatever potency, selectivity, safety properties are relevant to that particular drug discovery program,” Cohen said.
Valence didn’t divulge where Repare was looking to target its molecules, Cohen said.
While its work with Repare will target synthetic lethality drugs, Valence is taking a broad approach to potential partnerships — but the end goal is to offer its services to drug discovery programs where available clinical data are slim and targets are hard to reach. Also, the company is working to build a better “AI-generated” molecule that has a high chance of success in the lab.
“AI-generated molecules are of no value to anyone if they can’t be easily made in the lab,” Cohen said. “Most AI systems for drug design today yield low-quality molecules that are very difficult to make. If you give an AI-generated molecule to a medicinal chemist, maybe seven out of 10 times they’ll laugh at you. What we’ve done … is focused on developing new classes of design technologies that allows us to enforce a very high degree of quality into our molecules.”
Valence is one of a number of early players that have skated on the promise of an “AI-discovered” molecule — a questionable claim that has whetted the industry’s appetite for a revolution in the discovery process.
AI has emerged as a crucial tool in both discovery and preclinical R&D to mitigate the biopharma industry’s abysmal attrition rate for new drugs, effectively giving molecules a higher chance of success, those companies hope, before they ever hit in vitro.
While some of those AI discovery firms, like the UK’s Exscientia, have eventually morphed into full-on biotechs with their own in-house molecules, Valence is focusing at the moment on working with its partners, Cohen said.