Even a sceptic can see AI will transform the discovery of drugs

A molecule has been found and designed using artificial intelligence with hopes it will lead to a cure for a rare lung disease. This is surely the future.

At every stage in the drugs pipeline there are ways in which AI could, in theory, make it less leaky.

If you haven’t got idiopathic pulmonary fibrosis, ISM001-055 may not seem immediately interesting. In truth the molecule, designed to treat the lung condition, may not even seem that interesting if you do.

A study last month into its effectiveness in about 80 people was written in typically cautious language. The side effects weren’t awful. There were hints it improved lung function. The main conclusion was that there was enough evidence to justify collecting more evidence.

And the reason we should indeed be interested? It was found and designed using AI.

The drug is less an attempt by Insilico, the company behind it, to solve a rare lung disease than to tackle a widespread industry disease.

Today, getting a new drug costs £2 billion ($4.2bn). Of molecules considered promising enough to enter one end of the trial pipeline, 10 per cent emerge successful from the other. Computers have Moore’s law, the idea that performance doubles every couple of years. Pharma has Eroom’s law, the idea that drug discovery does the reverse.

ISM001-055 is a bet Eroom will end. Since the release of ChatGPT, we have used AI to write coursework, write generic PowerPoints and write bad poetry in the style of a pirate. In that time, in parallel, there has been a different, actually useful AI revolution. It seeks not to understand language but biology.

At every stage in the drugs pipeline there are ways in which AI could, in theory, make it less leaky. Currently, we seek drug targets through intuition and experimental data. AI, in theory, allows us to understand what is actually happening at the level of the protein. These are the molecular machines in cells whose structure determines their function but which for so long had structures that were fiendishly hard to determine.

AI, again in theory, allows us to map and parse the fiendishly complex metabolic pathways involving those proteins that, if they go wrong, are what we call “disease”.

Today, to get a drug, scientists look for plausible molecules we know of, screen them, then hope one does something to those proteins that is useful. AI potentially allows us to pick molecules from the 1080-odd possibilities, choosing those most likely to lock into the protein we want.

Will it work? ISM001-055 is the furthest such an AI-made drug has got. Insilico is not alone in betting on others getting further. The UK government announced a project to gather data on how drugs interact with proteins, to train models to identify new drugs and “slash development costs by up to pounds 100 billion”. If that sounds over-ambitious, Demis Hassabis, chief executive of Google DeepMind and Isomorphic Labs, Google’s AI drug company, goes further. He thinks AI will cure all diseases in a decade.

Absurd? It certainly sounds it to me. It also sounded absurd, though, when he suggested that within 10 years AI would be able to make a good guess at the structure of every protein, a grand challenge of biology that would all but guarantee someone got the Nobel. Well, today Hassabis is the someone with the Nobel.

But it doesn’t much matter if you think this has the whiff of hype. Currently, 90 per cent of drugs fail. If we get that to 80 per cent through AI we will have doubled our effectiveness. Humans are sufficiently bad at drug discovery that the bar for AI success is very low indeed.

The Times