Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
The artificial intelligence hype cycle depends on flashy announcements that break records. In April, San Francisco-based start-up Xaira delivered just that, announcing that it had raised $1bn in one of biotech’s biggest ever launches.
Xaira claims that drug development is poised for an AI revolution. It is not alone. Demis Hassabis, co-founder of Google DeepMind, famous for solving the 50-year-old scientific challenge of protein shape prediction, argues that biology could be “perfect” for AI, as it is fundamentally an information processing system. He heads Isomorphic Labs, Alphabet’s AI drugs offshoot which has agreed partnerships worth up to $3bn with Eli Lilly and Novartis. It aims to halve the drug discovery stage to just two years.
Growing numbers of AI-derived compounds are under development. The World Health Organisation has identified at least 73, though none are yet approved for use in humans. Some companies are getting close. Insilico Medicine, which recently filed for a Hong Kong IPO, was the first to get an AI-designed drug into Phase II clinical trials.
But AI is still no substitute for the experimentation that underpins understanding of a disease. The sector has already experienced trouble. On the day of Xaira’s launch, BenevolentAI announced major lay-offs. The London-based company set out to unite human and machine intelligence but its shares have lost 94 per cent of their value since it went public with a €1.5bn valuation in December 2021 via a merger with a special purpose acquisition company.
Developing innovative new drugs is expensive and inefficient. The pharmaceutical industry has no shortage of funds or motivation when it comes to improving drug discovery success rates using AI. Around 200 “AI-first” biotechs have secured more than $18bn in the decade to 2023, according to consultants BCG. Both AI use and success rates vary.
Using computing in drug design is far from new, originating as far back as the 1970s. Insights are only as good as the data used to train models. Predicting the toxicity of drug candidates is held back by the paucity or relevance of publicly available information. There is a lot of data on profitable and hotly pursued research fields like cancer, for example. There is less on relatively neglected areas such as mental health or infectious diseases.
AI is not a magical solution to these problems. Data gaps can be filled through experimentation, but it takes time and deep pockets.
Read the full article here