Perhaps in common with many people who don’t work in pharmaceuticals, I vaguely imagined that the industry found much of its inspiration from nature. Scientists might start with an old wives’ tale that, say, cowslips cure genital warts, and then work towards medicalising it — searching for an active ingredient which can then be synthesised, tested and ultimately marketed as a sterile pill in a little cardboard box, to be taken three times a day.
That was, indeed, how some common drugs were developed — and how some still are, most notably with efforts to make medical use of the constituents of cannabis plants. Yet this ignores the fact that the great bulk of modern drugs are increasingly being identified via theoretical work. There are an almost infinite number of molecules which could exist, but don’t necessarily exist in nature. Many of these theoretical molecules will have potential therapeutic properties. If those properties can be discovered and the molecules synthesised, they might form the active ingredient of an effective medicine.
But how to look for the few magic needles in the vast haystack of possible chemicals? Remarkably, it can be done without actually creating the substances themselves. The chemical structures of theoretical molecules can be modelled and analysed for their likely effect on the body. If they are ever to be turned into medicines, they will of course have to be exhaustively tested in real-life situations. But the initial discovery work can now be carried out entirely within a virtual laboratory.
Inevitably, it is an area in which Artificial Intelligence (AI) is becoming involved. Oxford-based company Exscientia announced this month that, in work for drugs company GSK, it has identified a molecule which it thinks has potential to treat chronic obstructive pulmonary disease (COPD). The molecule was identified using an AI platform called Centaur Chemist — and, while it may still be a long way from being available as a drug, the deployment of AI, says Exscientia’s CEO Andrew Hopkins, should help cut the time it takes to discover new drugs by three quarters.
‘Previously, we would have had humans integrating data,’ Hopkins says. ‘Now that is all done by AI. The number of experiments which will need to be conducted on the drug is much reduced.’ The company is also working with several other drugs companies, including Roche. Last month it did a £19 million deal with Celgene to develop cancer drugs.
If there ever was an industry which could do with a little help from AI to speed things up, it is pharmaceuticals. The industry, says Hopkins, is ‘incredibly unproductive’ and has been getting more so as it runs out of low-hanging fruit — such as drugs identified through observation of nature — and must look much harder for new products. ‘In 2010 the industry had an internal rate of return on investment of 10 per cent. That is now down to 1.9 per cent.’
It takes more than ten years to take a new treatment from inception to being marketed. Each human trial, according to Cambridge Consultants, another UK company involved in using AI in pharmaceutical research, costs between £23 million and £30 million. Four in every five per medical trials of new drugs are not completed. Even when a drug does become a success, pharmaceutical companies are endlessly accused of profiteering — their critics tend to overlook the fact that the revenue from one successful product has to cover the development costs not just of that drug but of nine others which failed along the way.
Which drug company wouldn’t, then, jump at the opportunity to use AI to speed up elements of this process? So much is being made of the promise of AI in pharma that it now has its own conference and trade show. It is also being pushed as a means of identifying the ideal participants for human trials of new drugs. At present, 30 per cent of participants in drugs trials are failing to last the course — with each drop-out costing companies around £20,000.
There is, however, reason for caution. Like the drugs themselves, not every application of AI is a magic cure-all. After a boom in initiatives and start-ups, pharmaceutical companies are beginning to grow a bit shy of the whole area. As one senior vice president of a pharma giant was quoted as saying at the ‘AI in Pharma’ summit in Boston in October: ‘Last year I was paying seven figures for anything AI. Now I am asking all these AI companies to prove it. I don’t trust that these technologies can deliver. A lot of companies claim to be AI when just last year they were selling me data reports and market research.’
The case for AI in medicine took a blow last year when questions started to be asked about IBM’s AI system Watson for Oncology, which had been billed as the future for cancer diagnosis when it was launched in 2013. The system was supposed to accumulate the wisdom of oncologists across the world and share it in hospitals which didn’t have access to the best specialists, effectively automating the process of diagnosis.
But an internal IBM document obtained last summer by the medical website STAT revealed that some patients had been recommended for treatments which were ‘unsafe and incorrect’. In one case, a patient with lung cancer who was already suffering from severe bleeding was recommended a drug which can cause haemorrhaging. No patients had actually been given inappropriate drugs.
None of this means, of course, that AI isn’t going to make a huge difference in the development of new drugs. But it will be some time before it becomes clear just how much AI is going to transform the pharma industry.
Comments