BMS Executives Discuss Company’s AI Approach

While it is common to hear that drug developers are in the early stages of integrating artificial intelligence with their drug research, discovery, and development efforts, one Bristol Myers Squibb (BMS) executive recently offered a more descriptive and no less accurate metaphor stretching back some 45 years.

“The state of artificial intelligence and machine learning is similar to where we were with personal computers in 1980,” Greg Meyers, BMS’ executive vice president and chief digital & technology officer, recently stated on the company’s technologies webpage. “Technology and digital capabilities will transform our ways of working, from how we develop medicines, to how we improve the patient experience, to how we run our core business operations.”

BMS says it is applying AI and machine learning, combining those with its researchers’ expertise in drug targets and mechanisms of action, to conceptualize, create, and evaluate the most promising molecules in its sizeable pipeline (50 compounds, 40+ disease areas) more effectively than ever, through an approach the company sums up as, “predict first.”

The company’s interest in AI stretches to pre-COVID-19 days. In 2019, BMS joined with oncology data-focused Concert HealthAI (now ConcertAI) to apply AI and machine learning to accelerate clinical trials, enable robust protocol design, and generate insights for precision treatment and improved patient outcomes.

That same year, BMS acquired Celgene for $74 billion, and in the process inherited a small molecule oncology collaboration with AI drug development pioneer Exscientia. BMS and Exscientia expanded the partnership to a potentially more than $1.3 billion alliance for developing not only cancer drugs but immunology and inflammation (I&I) treatments as well, led by a potentially first-in-class selective protein kinase C (PKC) theta inhibitor now in Phase I study for immunology indications. The drug is now being developed by Recursion, which combined with Exscientia in November.

Also in 2024, BMS inked a collaboration with VantAI to apply its generative AI platform to design molecular glues as small molecule therapeutics. BMS agreed to pay VantAI up to $674 million tied to achieving discovery, development, clinical, regulatory, and sales milestones, plus tiered royalties. BMS also has an option to expand to additional therapeutic programs.

“AI and machine learning tools aren’t changing what we do at Bristol Myers Squibb, which is discover, develop, and deliver transformational medicines to patients, but they are changing how we do it,” stated Robert Plenge, MD, PhD, executive vice president, chief research officer, and head of research. “These technologies are enabling our scientists to more deeply understand human biology and make more effective use of vast amounts of data. The advancements in predictive molecule invention seen to date have already been immense, informing key aspects of our research strategy as we look to improve the quality and speed of our investigational programs.”

GEN Edge recently discussed BMS’ approach to AI, and successful applications, with Plenge and Meyers. (This interview has been lightly edited for length and clarity.)

GEN Edge: What is BMS’s approach to research? And what is BMS’ research organization?

Robert Plenge, MD, PhD: We like to lay out these five principles because this guides a lot of what we do, not just within research, but also research and development, and how we prioritize opportunities, both internal and external. We start with these five principles. And in some ways, you can think about it as a linear progression from the earliest stages of research all the way to commercialization, but the principles can be applied to any program at any stage in development.

The first principle is how to think about targets, and we think very deeply about the human evidence and the human support for the targets, and we call this causal human biology. We often use genetics to find really good targets, but we can also think about longitudinal profiling, clinical pharmacology, and other things as well.

The second principle is how to match a therapeutic modality to a molecular mechanism of action. That’s a bit of a mouthful, so we shorten that to matching modality to mechanism. And this is basically, how do you create a molecule which recapitulates human biology that has a desired effect on human physiology? Is it a small molecule? Is it a large molecule? Is it a nucleic acid? Is it something that is a living cell therapy? Is it something that engages the immune system?

And the third principle, and this really begins a transition from research into development, how do you take those preclinical ideas and show that they are working in humans the way that you think and predict that they will work? We call that path to clinical proof of concept. And that’s really that moment in clinical development where you say, oh my goodness, we have a drug that we think is going to make a difference in patients, because you really see how the molecule acts.

Once you begin to see that clinical proof of concept, then the fourth principle is to accelerate full development, accelerate clinical development, because you want to get that medicine through all stages of clinical development in as many indications as possible, so you can really get that medicine to as many patients as quickly as possible.

Then finally, the fifth principle is how to maximize market access, and how to show that a medicine really is differentiating from anything else that’s out there, because that’s going to provide the greatest benefit to patients with the greatest need.

Those are the five principles. We apply them in research, we apply them to therapeutic areas, we apply them in the context of any individual asset. Again, we can actually apply those same principles wherever they are, whatever stage they are in the R&D organization.

GEN Edge: How does AI help BMS achieve some of these pillars of research?

Greg Meyers: A lot of what Robert described really explains the complexity of biology. And if you think of physics and chemistry, we have pretty good physical models. It’s one of the reasons why I could launch a projectile in the air at 12 degrees at a certain velocity so that you know exactly where it’s going to land.

Biology doesn’t work that way. If you think about even blood biomarkers, you’ve got thousands of molecules floating around in your blood. And for the most part, only about 5% of them have really been really studied deeply, have good elucidated structural biology behind them, or even are well understood. So, 95% of biology in that situation is really not well characterized. Where AI is helpful is, you start thinking through all the permutations of chemistry and biology and trying to simulate things before you put them into a wet lab. Think of them like a dry lab. It allows you to iterate through things much faster.

A few examples: One is, we look at the amount of time it takes to do something, and we decide to pursue a lead to the time it takes to do tox studies.

And we’ve been able to cut about 25% of that time out by incorporating AI and technology and doing a lot of these dry lab simulations. So that’s an example of where that would be. Research and development, we’re on track to cut almost three years off of our average clinical trial for the use of data, digital, and AI. So, it’s actually pretty pervasive in many parts of our time.

GEN Edge: With those savings, what is that time to tox study and that time for trials with three years?

Meyers: We don’t share that information publicly. We can tell you that we’re seeing meaningful changes in terms of the timelines.

GEN Edge: BMS has its own in-house ChatGPT and Gen AI tools. How are they customized for BMS’ work?

Meyers: Yeah, that’s a good question. In fact, we’ve had about 16,000 employees or active users of this tool. And what we’ve done behind the scenes is we are using OpenAI’s API, which we get through Microsoft Azure. But what we do in addition to that is we’ve also fine-tuned that tool on our own documentation. Think of it as you go to the tool today and if you ask it a question like how tall is the Eiffel Tower, it’s going to default to the base model on that, right?

If you ask a question about something related to one of our products, like what is the ideal storage temperature of an unopened vial of [Opdivo®] nivolumab? There’s actually a document that’s been trained on that will give you an accurate answer, and it won’t hallucinate. It’ll default back to our protected amounts of information. That way, we want to make sure that certain information that is useful to the company, to people that are in important roles, we want to make sure they’re getting accurate information, not whatever happened to be on the internet that that will pick up on.

And then finally, we also are moving toward self-service. All the IT helpdesk knowledge-based documents that our helpdesk people use have all been trained in the model. So, if you ask it a question like, how do I get a new laptop and mouse for an employee who starts next week? It’s going to actually be trained on all of our policies and procedures. It has been customized for BMS to not only take into account the external knowledge, but also the internal knowledge that’s useful to our employees.

GEN Edge: How extensive? Does this take in the drug? Or simply the drug discovery or development process? Does this include operations?

Meyers: We typically focus on prescribing information, stuff that is in the public domain, because any employee can access it. We’re not going to get into clinical trial results; there’s a lot of stuff that’s very confidential. This would be stuff that would be common questions, like is it okay for me to cut a Sotyktu® (deucravacitinib) tablet, or can I crush it? [No and no, according to the prescribing information.] All that information is public, and so we really focus on the public information. We just make sure that it’s highly curated.

GEN Edge: So that would be prescribing info?

Meyers: Exactly.

GEN Edge: BMS has highlighted several applications of AI. What if any additional areas do you see opportunities that would lead to expanding these?

Meyers: There are so many areas that it’s going to be useful in. I think the average employee’s productivity is going to continue to increase as we see agentic AI come online, which we expect to happen soon.

I think we continue to see more and more in the research area, so for all of our small molecules, we’re not going to do an experiment in the lab until a predictive model suggests an experiment is worth it. We’re about 50% in biologics today and expect that to continue to increase as we get more of our biologics programs online.

The average clinical trial takes about 400 documents. We’ve got maybe a dozen or two dozen or so now fully automated, so there’s a lot of room left to go there in terms of helping with document generation. There are just a lot of places that AI can be useful, and I think we’re still in the dial-up days of the internet as far as AI is concerned. There’s a lot of exciting times left to come.

Plenge: One of the things that we did do is, those five principles that I talked about, we wrote an external blog on how we use AI, applied specifically to those first three principles. And then there was a fourth around how we disseminate and democratize information.

Meyers: I would say that you could almost go down every single function of the company: How you do accounts payable, how you do accounts receivable, and how we edit and review legal contracts. You could almost go through every function of the company.

Plenge: Financial modeling.

Meyers: Forecasting, and find some opportunity that we’re either exploring, have explored, fully put in production, something around AI. It’s pretty pervasive.

GEN Edge: BMS has said that as the company refines the tools of AI, it strives to maximize benefits for workforce and patients. How does the company balance those concerns with the desire to maximize the potential benefits of using AI, which could include efficiencies that affect jobs?

Meyers: Our base case assumption is that most employees—certainly it’s my lived experience that I spend a lot of my time doing stuff that I’d rather not be doing. And if I could have an AI co-pilot or a tool that can make it easier for me to find information inside the company, allow me to go through—maybe if you’re in Robert’s organization and you’re trying to figure out, what is all the documentation we have on CD29? You’ve got to find all the SharePoint directories those are in. You’ve got to go read all the documents. You’ve got to go figure out what the summary is. I’d much rather have an AI agent do that for me.

So I think our base case assumption is these tools are really here to help make employees more productive. And what we hope will happen is that they’ll take those productivity gains, reinvest them in being able to bring more medicines to more patients faster. That’s where our focus is at this point.

GEN Edge: How does BMS see the challenge of the biopharma landscape as a big pharma at a time when you have biotech, some that have positioned themselves as all AI drug discoverers and emphasizing their use across their pipelines?

Plenge: I think it’s a big world out there. And I think there’s lots of room for big companies like BMS, and smaller companies that are highly specialized, whether it’s highly specialized in a therapeutic area, a modality, a computational tool like AI.

And I think it’s really about the broader ecosystem and how we all work together. There are many, many steps that are required to deliver a successful medicine to patients. I outlined five principles, but under each of those, there are many, many steps.

And I think as Greg nicely pointed out, we can actually use these computational methods, including AI and ML, to make each of those steps faster and more efficient, which ultimately accomplishes our ultimate goal, which is getting more medicines to more patients faster. I don’t look at it as, it’s one or the other. It’s really a mix of all these things together.

Meyers: You’ll see we have partnerships with insitro; we had a big success with them [$25 million milestone payment in December 2024]. We partner with VantAI. We have a number of relationships with “TechBios” that we’re on good terms with. And the reality of it is there’s no one company that can do it all. I think we see this as an ecosystem of collaboration that we’re more than happy to participate in.

GEN Edge: What roles does AI play in the three platform areas that BMS has articulated: targeted protein degradation, cell therapy, and radiopharmaceutical therapeutics or RPTs?

Plenge: For targeted protein degradation, one of the things that we’ve used AI to do is to expand our chemical libraries. You can imagine you may start with a chemical library that’s of a particular size. And if you wanted to expand that library previously, the way we would expand it is chemists would look at the structure and make their best educated guesses as to what molecules to make next. And you would make those molecules, you would test them. That’s been a good way to expand libraries.

But with the advancement of AI, we can use a machine to make predictions to basically diversify the chemical space and to make, I think, more structures in a more logical way.

And as we’ve done that, we profile the molecules and show that the success rates of AI-expanded libraries are better than those that are just expanded with more conventional methods. That’s one example as it relates to targeted protein degradation.

A second example, and Greg touched upon this with targeted protein degradation, this is true for many of the molecules, whether it’s a targeted protein degrader or some of the other small or large molecules that we have.

We can use AI to optimize the molecule, which is really the late stage of discovery where you’re making sure a molecule has all the properties required to go into humans. And we’ve done that for targeted protein degradation to help with optimization to get all the properties right.

I’ll just give one more example: For cell therapy, we think about how part of what we can do is we can use AI to help us find those binders. I think another really unique way in which we’ve used AI is to help with the manufacturing process, how to make sure that we have the right steps to make the cells outside of a person as we culture and grow those cells to have the right set of characteristics and properties so when we can give them back to the patient, they’re the best possible product, both from a safety perspective and an efficacy perspective. AI has helped us optimize the manufacturing process for cell therapy.

Meyers: And maybe just a quick example. We’re starting to see the first generation of what I would say are AI collaboratively designed molecules sitting in our pipeline. We promoted a drug into Phase II this year for sickle cell disease where AI helped us engineer the properties of the drug.

GEN Edge: Which drug is that?

Meyers: It’s a hemoglobin F (fetal hemoglobin or HbF) inducer. For sickle cell, you’ll have the abnormal cells or the adult hemoglobin molecules that are the ones that sickle. But if you can turn on the fetal hemoglobin HBF, that can actually substitute and effectively functionally correct. We have a molecule that’s in Phase I clinical development that’s an HBF inducer, and a lot of the properties of that molecule were optimized with AI. And again, it helped us get the optimal molecule into patients faster.

GEN Edge: And as for BMS’ pipeline, how many candidates are AI generated besides the HBF inducer?

Plenge: 100% of our small molecules that go from research and get developed have had AI as part of the optimization process. And currently it’s close to 50% for large molecules. AI is basically permuting many, many of the things that we do.