Novartis CEO proposes measuring AI’s value in drug development by the proportion of programs that win approval
Novartis CEO proposes measuring AI’s value in drug development by the proportion of programs that win approval
On 20 October, Vas Narasimhan published an essay describing three types of delay in drug development. In his model, AI speeds up information processing and operational work; experiments in cells, animals and humans determine the pace of testing.
Bringing a drug to patients involves three types of delay. Scientists find and compare publications; teams prepare documents, coordinate work and make decisions; the drug’s effects are then tested in cells, animals and humans. In his essay of 20 October, Narasimhan calls these parts of the process informational, operational and biological latency.
In Narasimhan’s model, AI can help researchers compare studies faster, compile documents, identify safety risks, improve molecular properties and select patients for trials more precisely. This support is intended to help teams decide earlier whether to continue or stop a program. Biological latency depends on how long an experiment takes to produce an answer.
An industry report from BIO, QLS Advisors and Informa estimates that, for programs in 2011–2020, a drug candidate entering the first phase of human trials had a 7,9% probability of approval in the US; the average time from the first phase to regulatory approval was 10,5 years. Narasimhan proposes that raising this probability to approximately 15% with AI tools would constitute success.
A molecule is designed to act on a biological target: a protein or process that the drug is intended to affect in a disease. Choosing the wrong target can cause a program to fail even when the molecule itself has suitable properties. In Narasimhan’s model, AI helps researchers select hypotheses and identify earlier which experiment is needed to test them.
As generating large numbers of candidates becomes cheaper, decisions about which ideas to fund and which experiments to conduct require greater precision. Narasimhan links these decisions to higher standards of evidence and the early termination of weak programs. He argues that regulators will need to update rules governing AI predictions of risk, trial design and dose modeling. An experiment in cells, animals or humans provides an answer that determines whether a program continues.