Advertisement

If you have an ACS member number, please enter it here so we can link this account to your membership. (optional)

ACS values your privacy. By submitting your information, you are gaining access to C&EN and subscribing to our weekly newsletter. We use the information you provide to make your reading experience better, and we will never sell your data to third party members.

ENJOY UNLIMITED ACCES TO C&EN

Drug Development

Q&A: What Novartis’s biomedical research head sees in AI

The pharma giant is partnering with software firms to imagine new drugs, find targets, and design clinical trials

by Rowan Walrath
January 15, 2025

 

Fiona Marshall.
Credit: Novartis
Fionna Marshall

One year ago this week, Novartis biomedical research president Fiona Marshall called artificial intelligence a technology with “the potential to transform how we discover new drugs.” Marshall was adding some color to the pharma giant’s then-new collaboration with Isomorphic Labs, the DeepMind spin-off launched in 2021 by Demis Hassabis, coinventor of the protein-folding software AlphaFold and a winner of the 2024 Nobel Prize in Chemistry.

Now, Marshall says, Novartis’s partnerships with Isomorphic and other AI firms are beginning to bear fruit. Novartis has leaned into these relationships to bolster its computational abilities and help unlock the secrets of targets that its own chemists haven’t quite cracked. Novartis is also working with Microsoft, Deciphex, Generate:Biomedicines, and, as of November, Schrödinger, one of the original computational chemistry firms. Novartis has doled out hundreds of millions of dollars on these collaborations with the promise of billions in return.

“We’re using AI end-to-end across all the things that we’re doing,” Marshall says.

C&EN’s Rowan Walrath spoke to Marshall during the J.P. Morgan Healthcare Conference this week about the possibilities of AI. This interview was edited for length and clarity.

When we last spoke, it was right after the Nobel Prize announcement. We talked about the Isomorphic partnership in that context. I thought this would be a good opportunity to catch up on where these AI partnerships stand.

The AI partnerships are fully up and running. We’re already starting to see some ways that AI can change how we do things. For example, some of the programs that we’re working on with Isomorphic; we’ve already found chemical material for some of these targets we’ve worked on for a long time that is different from what we found by traditional means. It can definitely lead you into new chemical space. This is AI-enabled virtual screening, essentially. We tried to pick targets that we were really struggling with.

We’re also seeing people who were skeptics about the use of AI—which is obviously one of the barriers to entry, how do you bring everybody along with you?—starting to say, “OK, actually, this looks quite good. I’d like to have it in my program.”

With Schrödinger, we’re doing a multitarget collaboration. They have a very good platform called LiveDesign where everyone can access it and see how compounds can be designed through computational chemistry. Again, it’s about bringing the tools to people so that they can engage in this. That even applies to basic things like [Microsoft] Copilot.

Generate:Biomedicines is taking AI into the biologics design space. It’s all about reducing the cycle time, improving the productivity, reducing attrition. We can really now use AI to understand safety profiling of compounds. We have a collaboration with the company Decipherex about digitizing all our historical safety data and then seeing if you can see safety signals earlier on, which allows you to stop one molecule and then move to another.

You mentioned that you are trying to focus these AI efforts on targets that you have struggled with in the past. Can you shed any light on what those targets are or what properties have been challenging?

I can’t name the specific targets, but they cover totally different target classes. Some are intracellular, some are cell-surface receptors. There are some targets where the issue is safety and selectivity, where the protein may be closely related to other proteins. Sometimes certain targets, you get stuck in a particular chemical space that always tends to have some tox[icological] effects, and you sort of go deeper and deeper. No matter how many compounds you bring through, you get stuck on the tox. Trying to go to a totally other chemical series is a way of getting out of that hole.

If you were to run a million-compound screen, that’s a tiny, tiny fraction of total chemical space. Whereas when you use AI and virtual screening, you can actually explore an infinitely large chemical diversity space that you couldn’t if you were doing lab screening. If you now have a really good understanding of the 3D structure of the protein—which you can get from AlphaFold—that allows you to go into new chemical space.

When it comes to using software to help with clinical trial design and to speed up recruitment, what does that look like in concrete terms?

Let’s say you’re going to design a clinical study. You have to set all the parameters of that study in terms of what type of patients would be enrolled, what stage of the disease would they be at, what other drugs are they already taking. Can you add your drug onto standard of care, or would you be taking them off the drug? How long does it take to collect data? What are the endpoints of the trial? How would you power it from a statistical point of view? Think of all these different parameters that have to go into the trial design.

Then, what you want to understand is: How long would it take to recruit 200 patients? Then, by changing the parameters—let’s say a simple thing like age, or we only want patients who’ve been on this drug but not this drug; we don’t want patients with these comorbidities—you can start to alter all the parameters, and then you can say, “OK, if I alter this, how is that going to change the recruitment rate?”

AI can really help bring all this data together.

How do you balance that need for speed with the need to reach patients who have been underserved by the medical establishment, who haven’t been targeted to be enrolled in clinical trials because of their race or gender or because they live far from a hospital system or don’t have proper transportation?

This is a very important thing. We do want the trials to represent the total patient population. Of course, that’s very important. This can actually help with that because, for example, we could put in the parameters that we want to make sure we include this percentage of people by gender or ethnic diversity, and that can be set up as a prior part of the protocol design. It would then help you to see, OK, if you want to recruit this number of Black African Americans, these are the hospitals where you would be most likely to be able to access those patients. That’s definitely something we’re very committed to, making sure we have that clinical diversity.

From a change management perspective, for chemists and other scientists within Novartis who are not accustomed to working with software tools, how do you onboard them?

It is through training, basic training but also sharing best practices. We have virtual communities where people share, like “this is a thing you should try out,” and sharing ideas and best practices and really trying to upskill everybody to the same sort of standard, basic understanding of how you can use AI. We’re finding people are very enthusiastic about learning.

Also, it’s not just about doing the new things. It’s also letting go of the old ways of working. Take a chemistry program, for example. You could use AI to design your molecules, or you could design them by making hundreds and hundreds of compounds like you always used to make. Now, what you quite often find is that the chemist will now do both—the new approach and the old approach at the same time—and see what comes out.

What we’re doing at the moment is collecting the data on—OK, these were the compounds that were designed by AI and generative chemistry; these are the ones where you were enabled by AI. These are ones where you weren’t enabled by AI. Can you see a more rapid improvement in overall properties in the AI-enabled design? Scientists like to be convinced by data. So if you can then take them and say, “Look in this program how the AI-enabled compounds rose to the top more quickly.” That’s a way of bringing people on board and getting them to let go of the old. You need to let go of those things and then commit to the new. Otherwise, you don’t speed up.

Article:

This article has been sent to the following recipient:

0 /1 FREE ARTICLES LEFT THIS MONTH Remaining
Chemistry matters. Join us to get the news you need.