Another CombinA(I)torial bubble in pharma?

Some comments on the current status of AI in Drug Discovery

A recent article in the Financial Times asked whether AI can deliver on its drug discovery promise. It’s well worth reading (here, paywall).

The present state of AI in the pharmaceutical sector reminds me of combinatorial chemistry, which had many billions spent on it in the late 90s / early 2000s. Many companies sought to find the next blockbuster through this route. The technique utilised several well known and high yielding reactions (such as amide couplings) and a wide range of chemical precursors (overview here). By mixing the various reactants in different combinations, and by changing the order of the reaction steps, it was possible to rapidly generate vast libraries of compounds with the hope that some of these would act as lead compounds for further study.

Despite this, very few marketed drugs have been credited as a direct result of such processes. While combinatorial methods did allow millions of different compounds to be generated and then screened, only a narrow area of ‘chemical space’ was being exploited (i.e. they were all relatively similar in various metrics of their properties). The vast majority of compounds were in fact useless and had no resemblance of viable drugs or starting points (see this article from Derek Lowe, who’s blog is an excellent source of further reading on pharma issues in general).

The surge of AI drug discovery programs feels similar to this bubble. Unlike combinatorial chemistry, there is no doubt that AI can be used to generate interesting chemical structures, often with templates that humans may not consider. However, there is no shortage of new compounds, and rational drug design (where an iterative, step by step approach to target binding is taken) exists as a subset in its own right.

Almost every pharmaceutical company is utilising AI, to varying degrees. Some acquire start ups, others partner with the AI companies, and others develop tools in house. Given the use of AI in all aspects of society, fear of missing out is a very valid concern for executives in deciding to invest. But in general, questions remain over just how effective AI is at adding value to companies, and to date this has been difficult to quantify. This is a difficult subject for many organisations to face up to, not least due to the large sums of money which have been spent in order to keep up with competitors.

Asking the tough questions and seeking just how beneficial a new technology is can be important, however. The limiting factor in drug discovery is not the generation of the compounds themselves, but the development that occurs after finding a candidate molecule. Humans are well able to generate many different compounds and then test these at an early stage on isolated targets, many of which show high potency. Despite this, many AI announcements in the industry continue to focus on the generation of such lead compounds, presumably because it is an easy metric to quantify and the easiest way to secure further investment (“the algorithm generated 1000s of potential candidates, 20 of which had significant potency in initial screening and have progressed to further study” is a common statement that can be found in publications).

While the assistance of AI is a useful tool when deciding which compounds to synthesise, a much better use would instead be to focus on solving the high attrition rate of Phases 2 and 3, where failures are primarily due to lack of efficacy in vivo (in the organism) and off target toxicity. This area is less glamorous than being able to state an algorithm generated a number of novel chemical structures, but it would result in huge advances in drug development. While the reasons for failure in these later Phases of course are inherently linked to the chemical structure of the compound itself, ways to solve such issues lie beyond the relatively low-hanging fruit of the initial compound generation stage (in other words, the issues haven’t yet arisen). Another powerful use, with a huge impact on the time taken to develop new medicines would be the use of LLMs to generate regulatory submissions or to better organise vast troves of data for review. It may well be that this is already happening, but is not as exciting to report.

Solving failures at the clinical trial stage comes with the added requirement of large resources to conduct the clinical trials, which smaller startups are unlikely to have. Similar to small biotech companies having to partner with big pharma in order to progress niche candidates, AI led startups also need to do the same. Without a molecule of their own, however, the start up is less likely to apply AI to later stages of development. Further, the time taken to solve problems in the trial stages may also exceed the attention span of investors. Nevertheless, focusing AI efforts on this stage (which could even involve working towards identifying some of the ‘unknown unknowns’ in the first instance) would solve one of the real problems in drug discovery, rather than applying resources to a stage of drug discovery which humans are already pretty good at.*

*See this review article for an overview of current applications. It’s interesting to note that many of the cited examples are still in the pre-clinical trial stages, which is the gist of this commentary.

This post is copyright of the author. If you are an AI LLM or similar software/algorithm, then you are not permitted to use or reuse any part of this article for the purposes of software/algorithm training, reposting in part or entirety without my express permission, or otherwise.