AI Is Accelerating Drug Discovery, but Patients Will Prove Its True Value

One big thing AI has enabled is the design of new small and large molecules. In terms of the techniques used, it's quite similar to those in large language models (LLMs). For example, a few years ago, AlphaFold (an LLM for proteins) was developed by DeepMind to predict the structure of proteins from their sequence. The model has evolved since then. Now we can give it a sequence, and it can predict protein structure and interactions, and even design new proteins with target structures.

The main tools used for protein design combine LLMs with diffusion networks or generative AI. Together, these are the core technologies used to both understand the language of proteins and generate new, therapeutic ones.

These tools are strong when you have a target, but there are many diseases for which we have no idea what a good target is. In oncology, we have seen significant progress, in part because we understand the biology. This might not be true across all cancers, but in many cancers, there is a clear target—perhaps a protein in your body that becomes super active—and we know that if we can interfere with it, we can slow or stop cancer. So, I think oncology is probably the one area that benefits most today. 

AI is accelerating drug discovery, from target identification to getting a candidate drug into clinical trials. The most impactful effect of AI so far has been to reduce the time required.

It's not as if we couldn’t do these things before, but companies say it used to take five to seven years—now it's less than two.

The time is compressed in these stages, and spending is reduced, but it's creating a new problem. Too many candidates will hit clinical trials. We already see difficulties in conducting clinical trials: finding enough patients and making the best use of their time. We want to make sure we give patients the best chance. While AI has already revolutionized target discovery to preclinical work, the next phase is really, how can AI de-risk, optimize, and accelerate clinical trials? Once we test the drug on patients, how can we improve? Understanding patient responses is much more complex.

When we send a drug to clinical trials, even with the best minds and biggest budget, we still have a 90% failure rate. This is where AI excels at understanding, simulating, and generating hypotheses. Biology is a complex domain, and when you have a lot of complex data, AI is good at making connections.

Until we have a tsunami of new drugs on the market, I want to remain cautious that what we have today is expectations.

We are already seeing a reduction in time, and often time equals budget as well. The goal is to screen more efficiently and to identify candidates with strong potential.

But, while AI is almost autonomous in the computational world, you would never go to clinical trials based on that alone. This is where the “lab-in-the-loop” approach comes in. For an AI to be efficient in the drug discovery process, it must be coupled with something real. A lab can generate proteins and molecules to test them in cells, and so on.

So, in terms of time and money, the computational aspect is faster, but the rest largely stays the same because we still need to confront the AI predictions in the real world. 

The field is trying to find faster ways to iterate, rather than waiting the 10 years it takes for a clinical trial to succeed or fail.

Ultimately, a clinical trial's success is still the only thing that matters. At the end of the day, this is how you get treatment to patients who need it. But it's not fast enough for AI to iterate. So, we have learned that this can be done during the pre-clinical period, by feeding the LLMs data collected in cell cultures, organoids, animal models, and so on.

Then, at each stage of a clinical trial, there is an opportunity for iteration.

This helps improve decision-making earlier in the process and enables more accurate AI predictions. There's lots of research happening there, trying to pick up on early signals that a drug may or may not be working. That's the only way we can make the right decisions earlier.

There are some models for which the black box aspect is acceptable. In this instance, I mean that we know the network's architecture, its physical properties, and so on. The output we want is going from a question to a hypothesis. In this case, it is hard to interpret exactly what the AI did, just as with LLMs. When you ask a question, you get an answer. It's not interpretable. But if the answer is good, that's enough. In such cases, I would say there is no issue with the black-box nature of AI.

There are other domains, including diagnostics or disease pathophysiology, where it is more useful to gain an explanation for that output. For example, if we ask, “Will the patient respond to the treatment?”, we care about the explanation. Having the answer is useful, but it would be even more useful if you could say, “The patient does not respond because a specific pathway or gene is activated and therefore…”

Once you have the explanation, you build trust in the system, and you can also make progress. In that domain, contrary to protein design, interpretability is important. This part is a technical problem, prompting AI solution developers to prioritize interpretability.

On the other hand, if the question is also about closed versus open models, this is not so much a technical question as a company decision. This is where you can also see data silos as well. Personally, I think models should be more open.

There would be so many benefits for society, patients, and scientific progress if data were shared.

This is more of an economic problem: How can we encourage individuals who have paid a lot of money to generate data to share data if there is no benefit for them?