I have a dream was the Abba-themed title for a panel, at BIO-Europe 2024,hosted by TVM Capital Life Science, which considered the future of artificial intelligence (AI) in drug discovery. The same song also mentions a belief in angels, but the panel participants grounded their views in a more hard-headed consideration of what AI can – and cannot do – to advance drug discovery and development.

Failure is embedded in the process of drug development – and AI is not going to change that, said Stephan Brock, chief technology officer of Heidelberg-based Molecular Health, which develops AI-driven analytic software to improve clinical decision-making. But it may reduce the failure rate – across all stages of discovery and development. And it may also help to accelerate the recognition of failure, saving both time and money. “We want to fail early,” he said. And AI can provide “indicators of failure”.

But it needs to do so in a way that is useful. Brock recalled that a team at Molecular Health won a clinical trial prediction competition with an AI tool it developed that had an accuracy rate of over 80%. But it failed to interest any pharmaceutical companies, he said, because it was a black-box solution that did not provide any insights into why a particular trial had failed. “This is learning from a failure,” he said.

Hartmut Juhl, CEO and founder of Hamburg-based precision oncology firm Individumed Therapeutics cautioned that interpretation of complex biological data will continue to be guided by human rather than machine intelligence. “AI does not help in this question at all,” he said.

Yet, the number of use cases of AI in drug discovery and development is growing rapidly. And the use of AI can enable new ways of working. Boston-based Insilico Medicine, which employs generative AI at every step of the drug discovery and development process, was able to flip the normal high-throughput screening approach, said chief business officer Michelle Chen. Instead of making and screening a library of 10,000 compounds against a given target, it generated 80 ‘synthetic needles’ – compounds with predicted activity, which formed the basis of a more focused screening campaign and a much faster identification of a preclinical drug candidate. But training AI needs to be done thoughtfully. There is, she said, an over-emphasis on positive data. “To train AI properly you need a vast amount of negative data,” she said.

Panel moderator Maria Luisa Pineda, co-founder and CEO of Envisagenics, which has developed AI software for analyzing vast quantities of RNA sequence data, said her firm and its scientific founders have been using AI for two decades, to detect biomarkers to stratify patients for clinical trials, to identify new drug targets for antisense drugs, and to identify novel epitopes generated by alternative RNA splicing that can be targeted with immunotherapies.

Oxford-based Exscientia, a leader in AI-based drug design, has, said its executive vice president of precision medicine Nikolaus Krall, brought three molecules into the clinic and has built its pipeline about five times more quickly and ten times more cost effectively than would be the case with traditional approaches. Its imminent merger with techbio firm Recursion Pharmaceuticals, of Salt Lake City, will bring together two highly complementary AI platforms, he said.