The Covid-19 outbreak has highlighted many aspects of how artificial intelligence is being used in medicine, but drug discovery has inevitably become a major focus.
At least four AI-based proposals have been made towards treating, primarily inhibiting this coronavirus. Given the urgency of the situation, the initial drive has been toward finding existing, approved drugs that might be used.
BenevolentAI, a UK-based start-up, has described baricitinib, which is used to treat rheumatoid arthritis, as one possibility. Deargen, a South Korean AI drug discovery specialist, has suggested atazanavir, used in HIV treatment, as a potentially promising option (NB: The relevant article is ‘pre-print’). Four more have been put forward by a team from China’s Army Medical University based on “high-throughput screening”.
A fourth proposal, from Insilico Medicine of the US, has identified six new molecules that could be synthesised and tested to form the basis of a treatment.
Before going any further, an important point to make is that none of the proposers seeks to represent its work as definitive medical advice. These companies and institutions are not offering nailed-on Covid-19 treatments but rather avenues to explore. They are very clear about that.
Nevertheless, they do illustrate AI’s potential to accelerate the general drug-discovery process and make it more exhaustive.
The basic methodologies behind the work are familiar enough to pharmaceutical professionals.
Here, they all started with China’s release of Covid-19’s genomics in January. Its features were analysed for aspects that could be addressed. In BenevolentAI’s case, its system lighted upon the kinase AAK1. It governs processes that allow materials such as viruses to enter a cell.
Having identified a target, the systems then sought out drugs that already address it. BenevolentAI found 378 drugs that inhibit AAK1, of which 47 were approved.
Finally, the various candidates were assessed based on criteria such as side effects. BenevolentAI found that baricitinib had the least (though this does not mean there are none).
This all sounds very straightforward, and experts say that the same path could be followed by humans. But what humans lack is the ability to process data at the same speed as an AI-based system, draw centrally and simultaneously on all the existing learning built into the algorithms and increasingly granular natural language processing around it, and do all that tirelessly. It took roughly a month for these groups to release their findings.
Another illustration of AI’s importance as an accelerator for drug discovery has also emerged since the Covid-19 outbreak, but here for the treatment of obsessive compulsive disorder (OCD). In late January, it was announced that the first AI-designed drug, currently known as DSP-1181, is to enter clinical trials. The proposed OCD treatment was developed by another UK AI drug-discovery start-up, Exscientia, in partnership with the Japanese pharmaceutical group Sumitomo Dainippon – and the process took just 12 months, compared with an average new drug-discovery timespan of 4.5 years, using traditional methods.
Add to this the broader economic realities that a typical new drug costs an average of $2.6bn (£1.95bn) to develop and that 90 per cent of them fail to complete clinical trials, and the attractions for AI’s increasing use in medicine become self-evident. Pharma companies want faster, cheaper discovery; healthcare providers want to get the most out of the treatments already available. This is not just about one emergency, though Covid-19 may come to illustrate AI’s medical value further.
Certainly, there is a lot of activity around machine learning and drug discovery, much of it having taken off a little under 10 years ago.
According to biomedical AI group BenchSci, there were 203 identifiable start-ups developing these types of processing tool at the start of this month, the majority being in the US but with the UK in second place. Most were founded after 2012.
Nevertheless, Simon Smith, BenchSci’s chief strategy officer, also recently analysed flows of venture capital funding into the sector and detected a trend towards later-stage rounds. That can be seen as an indication of increasing maturity and confidence in the systems companies are offering.
Moreover, all that activity is largely separate from the in-house AI R&D taking place at the existing pharmaceutical giants and within public healthcare. China, for example, may be understated in BenchSci’s data given that much of its research funding is being directed though state bodies (such as the Army Medical University cited earlier) and the country’s triumvirate of national technology champions (Baidu, Alibaba and Tencent). Both Deargen and BenevolentAI worked with major university partners.
AI drug discovery is here, then, to help us defeat Covid-19? Well, up to a point.
First, consider that while at least four candidate treatments have been suggested, all their proposers themselves added heavy caveats (and acknowledge that further peer-review is also necessary).
Then, you immediately notice that the results from each group to go public so far are different. The core steps in the drug-discovery methodology may be consistent, but the algorithms still contain a lot of secret sauce and variation – the AI sector here is maturing, but is not yet fully mature as a whole.
Beyond all that, the same concerns about ‘bias’ present in other branches of AI’s application apply for drug discovery too.
None of this should negate or devalue the work of the groups that have released suggestions or those likely to come forth from others. Rather, it again tells us that things are very fluid and that none of these proposals should be seen as definitive. Rather, as BenevolentAI itself has said, they have been released large to inform inward-facing (or ‘inreach’) science communication, “to assist in the global response”.
That response is vitally needed but – notwithstanding Exscientia’s upcoming OCD drug trial – AI in drug discovery is still probably best seen as a component within a much wider and still human-driven process rather than a standalone solution. And, you suspect, most of the companies in the AI sector prefer it like that for now. Global health emergency or not, hype and medicine make dreadful bedfellows.