This month’s announcement that Alphabet’s Isomorphic Labs, led by the talented and charismatic Deepmind co-founder Demis Hassabis, had inked discovery collaboration deals with both Lilly and Novartis was just the latest indication that AI-enabled pharma R&D was ready for the big time. Billions of dollars have been invested in companies promising to “revolutionise pharma R&D” through the application of artificial intelligence – but it is now clear that global pharma companies, and not just tech bro’s, are buying into the claims.
And there is no doubt that there is plenty of room for improvement. Both the cost and the time required for each new drug approval have increased exponentially for decades (aptly tagged ‘Eroom’s Law’, being the inverse of price drops and speed gains in the semiconductor industry). Much of the eye-watering cost in getting new drugs approved results from the high failure rates (since their cost must be recouped from the winners to sustain economically viable R&D), which also negatively impact the speed of progress. Ergo anything that can improve success rates would be hugely valuable.
Step forward artificial intelligence. After decades slowly developing in the shadows, featuring more often in science fiction novels than television news programs, the arrival of large language models, and ChatGPT in particular, thrust artificial intelligence into the national consciousness. What is more it provided real-world evidence of its amazing powers accessible to anyone with a browser. It seems like a no-brainer that improved intelligence of the artificial sort might do better than humans and unlock rapid advancement and untold riches. It came as no surprise, therefore, that billions have been invested (or at least spent) on the application of these new approaches to drug discovery and development.
There is, of course, no evidence of whether the claims made for AI in discovery and development are hype or reality. With close to a decade between concept and approval of a new medicine, it is still too soon to know whether there has been any impact at all from introducing AI into various parts of the process. Even when time has elapsed, such evidence will be hard to obtain: many approvals are for more me-too drugs against a validated target, or simple reformulations and line-extensions, making anything but a step-change in R&D efficiency all but impossible to identify.
So what about the theoretical basis for the claim that AI can “revolutionise” drug discovery and development? What happens if we look beyond the most superficial truism that AI can solve lots of problems so why not this one?
To address that, we have to look more carefully at what AI is – and of course it is not one thing, but a collection of tools designed to improve decision-making. The big step forward that large language models, such as ChatGPT, represent comes from a surprisingly simple mathematical innovation. For decades, we have modelled large datasets by creating latent variables (which are combinations of the source data) so as to reduce the degrees of freedom, or complexity, of the underlying system being modelled. A simple example is called projection, where simple linear combinations of the data are used to extract the internal patterns within very large datasets – tools that were rightly called pattern recognition methods. And pattern recognition is not intelligence.
But relatively recently, it became clear that you could logically invert the process and instead of reducing the dimensionality of the problem, you could increase it. If you have millions of datapoints, instead of trying to distil them down to a handful in a conventional pattern recognition approach, you could combine them to create billions of latent variables. Instead of simply describing the data, you can now extrapolate from it. And that simple inversion converts a pattern recognition tool into something that can talk to you, like ChatGPT.
Whereas pattern recognition helps you understand what is literally inside the box, these new approaches can go beyond what was already known. They have moved outside the box, and we typically consider out-of-the-box thinking as intelligent or creative. And hey presto, ChatGPT can create genuinely new responses based on, but never present in, the underlying data.
So if you want creative, you will be impressed – ask ChatGPT for a script for an advert for a box of chocolates and you will get a very passable effort, as good as all but the most highly-paid humans might generate.
Intelligence (artificial or organic) makes the best use of the available information to guide future decisions. But that presumes that the necessary information to make a better decision exists and is available. In short, whether AI of any kind will materially improve outcomes depends on what is causing poor decision-making in the first place.
Consider target selection, which is probably the most critical decision for any drug R&D project (after all, if the target is wrong, then everything that follows will by definition be wasted). Can AI help make better decisions here? Only if the necessary information to select valid targets already exists in the totality of the data available. Undoubtedly, there is a lot still to be discovered in biology and it is plausible that our decision-making on target selection is limited principally by the absence of data about large swathes of relevant biology. But, being generous to the potential for AI, its not unreasonable to assume that our decision-making based on what is known is sub-optimal.
Yet standing in our way, there is not just one elephant in the room but two, who together represent a massive, unmovable obstacle in the way of AI “revolutionising drug R&D”
The first is the length of the development process. This matters because AI, like organic intelligence, depends upon learning to improve. Whatever the specific mathematical algorithm used, and particularly if you need it to extrapolate beyond the data (as we surely need if it is to aid decision making in R&D), iterative improvement of the models is key. And you cannot learn beyond the data without making new predictions and then seeing whether they are correct. The success, or otherwise, can then guide refinement of the way the existing data is used to build the model.
Unfortunately, it takes a decade or more to know whether your predictions in the drug discovery and development game were correct. No surrogate marker (selecting a development candidate, starting human trials, or (arguably) even launching a drug are reliable indicators of the only things that matter: real-world clinical utility and commercial success.
The same limitation that stops us, today, knowing which of the myriad “AI-enabled” biotech platforms, each with different algorithms and models powering their approach, might be the best also stops them from iteratively improving their toolkit.
All the places (outside drug R&D) where today AI is delivering value lack this issue. Want to know if your large language model is providing convincing human-like output? Just ask a human and you know within seconds. Want to know if your AI-identified tweaks to the productivity of your chemical plant do indeed improve output? Turn the dials and watch what happens.
Does AI improve decision making in drug R&D? Ask me in 2050
The second elephant is even more insidious and difficult to circumvent. The subservience of “the science” to other factors in the decision-making process. Whether “the science” supporting a decision has been developed by a human or a computer, we have a frightening tendency to find ways to ignore it.
DrugBaron has written extensively (for example here and here) about these systemic biases that befuddle decision-making in pharma and biotech: the implicit bias towards continuing a project rather than killing it; the “shackled” thinking that leads investors to prefer stories over data; the imperative to find a drug that fits the need of the commercial organisation in preference to the one that has the greatest chance of working.
The ways in which the financing framework for drug discovery and development works creates, in effect, a straightjacket that constrains and diminishes decision-making quality throughout the R&D process. Why do companies move assets into expensive phase 3 trials on the basis of shaky post hoc sub-group analyses, when even schoolboy statistics tells them they will likely fail? Because of the asymmetric returns – if you succeed you make billions, if you lose the worst that can happen is you lose your table stake.
Even if, and it is still a big if, AI improved the recommendations coming from “the science” to a small degree, that benefit will be substantially diminished by the global factors that dominate the decision-making.
Do you believe your investor directors will listen to the computer after decades of ignoring the CSO?
It is, therefore, a certainty that AI will not “revolutionise drug R&D” any time soon. Not because of the (very real) limitations of AI, which will no doubt be gradually improved upon, but because of the peculiar nature of long, complex processes and real-world decision-making.
But, of course, “revolutionising drug R&D” was always too big an ask; hype created by founders looking for any angle to win investment dollars. Just because AI will not solve every problem doesn’t mean it cannot help with some. And it already is. Where the problem itself is well-defined and the feedback loop short, AI-enabled tools can deliver what look like miracles. Take RxNfinity, the first “dynamic chemical space” that leverages AI strategies to design small molecule ligands for “undruggable targets” or with complex pharmacologies: it has delivered hit rates above 50% with sub-micromolar novel compounds on multiple client projects, far exceeding the capabilities of competitor discovery technologies. Here, you know what you want and can determine quickly if you have it or not, driving optimisation of the underlying algorithm. And there are plenty of other examples such as optimising CDRs for high-affinity specific binding of antibodies.
AI in drug discovery and development is here to stay. But providing an incremental boost to productivity, finding molecules (small or large) that conventional approaches couldn’t find, for example. Yet if you misunderstood the biology, or decision-making isn’t as closely tied to the science as you would like, then these AI-enabled advances will not save you. Strip away the hype, and AI platform biotechs are only doing the same things as the rest of us – warts and all.
RxCelerate Ltd is an outsourced drug development platform based near Cambridge, UK. We specialize in delivering an entire road map of drug development services from discovery and medicinal chemistry through to formal preclinical development and clinical up to Phase IIa. In the last five years, we have witnessed dramatic changes in the drug development …
