Addressing the Goldman Sachs Healthcare C-Suite Unscripted Conference last year, Merck & Co. Chairman and CEO Robert A. Davis told investors the pharma giant was spending undisclosed “significant” capital toward automating processes, analyzing data, and supporting decision-making through AI.
“We’re making meaningful investments in artificial intelligence [and] machine learning across what we’re doing in the labs, starting to think differently about how we approach customers,” Davis said, according to a transcript
posted on Merck’s website.
Less than a year earlier in an interview with Francesca Cornelli, PhD, Dean of Davis’ alma mater Northwestern University’s Kellogg School of Management (class of 1993), the Merck CEO recalled how his company’s use of AI and ML stretched back to the early 2000s, primarily focused on accelerating drug discovery. Back then, Merck found the benefits limited by the extent of computing power, a lack of structured data and the inability to augment internal data through time with external data.
“We used these tools, but I would say we didn’t see a huge benefit from them. It’s probably been only in the past few years, as computing power has changed markedly with the advent of the cloud and the ability to amass large amounts of data, that we’ve started to really reap the benefits of it,” Davis said, adding: “It’s now one of the core strategic priorities for our company.”
Iya Khalil, PhD, Merck’s Vice President and Head of Data, AI and Genome Sciences, shared a recent example of an AI-developed drug behind one of the company’s announcements at the recent American Society of Clinical Oncology (ASCO) Annual Meeting 2025, held May 30-June 3 at Chicago’s McCormick Place Convention Center. Merck presented safety and efficacy results from the open-label Phase I KANDLELIT-001 trial (NCT05067283) evaluating MK-1084, a next-generation KRAS G12C inhibitor, alone and in combination with other therapies in patients with KRAS G12C-mutant solid tumors, including advanced colorectal cancer and non-small cell lung cancer.
In patients with advanced KRAS G12C-mutated CRC and NSCLC, antitumor activity and a manageable safety profile were seen with MK-1084, either as a monotherapy or in combinations with Merck’s multi-indication blockbuster cancer immunotherapy Keytruda® (pembrolizumab), with or without chemotherapy (carboplatin and pemetrexed).
“What we were able to do is use AI and machine learning to optimize for the properties of that drug—make it safer, better, and more effective. And that’s a real example where the AI machine learning models and methods that we built at Merck took in our historical data, new data that we collected on that compound, and got us a better molecule,” Khalil said. “The proof is in the trial itself: We ran the trial and we have a better drug, MK-1084.”
Earlier this year, Merck unveiled its new TEDDY (Transformers for Enabling Drug DiscoverY) family of artificial intelligence foundation models. TEDDY is designed to overcome the limits of existing gene regulatory network (GRN) models by incorporating biological annotations, scaling to a larger and more diverse dataset, improving the ability to make inferences across diseases and cell types (especially with unseen data), and leveraging model size (ranging from 10M to 400M parameters) and biological knowledge.
During last week’s Biotechnology Innovation Organization (BIO) International Convention, Khalil discussed Merck’s approach to AI and successful applications of the technology with GEN Edge. (This interview has been lightly edited for length and clarity.)
GEN Edge: Companies of all sizes talk about using AI in drug discovery. What makes Merck stand out in that area?
Iya Khalil, PhD: We are using AI, but we’re also building our own AI. And we are incorporating vast amounts of data, both data from our historical datasets on our compounds, historical datasets from our patient datasets, preclinical datasets, and generating new data to make the AI better. A key thing is that we really believe that active learning—creating what we call a flywheel engine where you feed the AI data—and then the AI makes predictions, and then it makes it better—is a core component, a differentiating component.
We have generated and created this AI, and we’ve embedded it in our functions, whether it’s the function for finding new disease biology and we make AI that’s specific to that. It’s not like a one-size-fits-all solution, but AI that will help us learn disease biology. We have built AI solutions that help us with discovering our molecules and optimizing our molecules so that they have the best potency, selectivity, less off-target effects, best safety profiles. And we’re leveraging AI and machine learning in our early trials to do patient selection and make sure the right patients get the right drug at the right time.
GEN Edge: How and where has Merck integrated AI in the drug discovery process?
Khalil: We have integrated it in the target selection phase. We have integrated it in the lead discovery and lead optimization phase. We have integrated it into our safety—how we enable and generate the insights needed to understand safety signals for our drugs and optimize our drugs there. And we have leveraged it and integrated it into our clinical trials around precision medicine, and identifying who are the patients from their genetic, genomic background that would best respond to a drug. And then we’re using it also in cases for our trials just helping our clinicians and trial designers understand the already existing knowledge base that’s out there, whether it’s how we design past trials, how we think about designing new trials and more of an operational way, just to make it faster for us and more efficient for us.
GEN Edge: We often think about AI as replacing human involvement in drug discovery, but Merck says it sees AI as augmenting human ability. Can you explain the human role in AI driven processes and what can AI help humans do?
Khalil: We’ll take an example around target discovery and disease biology discovery. We’re using AI right now where we have brought in all the patient data we could find, single cell patient data, and looking at their genomics. How does each cell for each disease express itself across many, many different patients?
We brought in data on 116 million cells across 24,000 donors, 413 different tissue types, 860 different cell types, and across 122 different diseases. We used AI to build what we call a foundation model. This model aims across all of that disease biology and all of that variability in patients. Down to the molecular level, it aims to learn a representation of the biology. That’s all encoded in the model.
Then what we’re able to do is we’re able to take that model and feed it very specific data sets on specific disease biology, and have the biologists learn what are the pathways and start to use their expert knowledge and scientific knowledge—start to go, ‘Okay, I am better off targeting this mechanism versus that mechanism.’ But the model brings it all together in one place. It brings together what would maybe take many decades of trying to learn from human intuition all in one place and then enables that scientist to just make much better decisions even faster.
GEN Edge: Why 116 million and not more?
Khalil: That was as much as we could get in April. And more is coming. We’re generating more data, and we’re also involved in a number of partnerships where we have access to data and we’ll be getting more data over time.
GEN Edge: What type of cells make up that 116 million?
Khalil: Everything from your liver cell, heart cells, colon cells, immune cells and different types of immune cells: T cells, B cells. Every cellular type potentially imaginable in the human body that we could get data on.
GEN Edge: Earlier this year, Merck introduced TEDDY, a new family of AI foundation models. You wrote in a commentary how TEDDY overcomes the limits of gene regulatory network models, improving the model’s ability to make inferences across diseases and cell types. What inferences are researchers looking for and able to find with this?
Khalil: We built the TEDDY model because we wanted to really understand the variability and biology across all the genes in the human genome, variability and biology across all diseases, as well as down to the patient level. So, we can take these models and after training them, we can bring in a patient’s profile. If you have a tissue and look at your gene expression, or if it’s from a tumor patient and feed it into the model, then it will predict for me how all the genes relate to each other. I can say, okay, in that patient, this pathway is the most active. This gene or these sets of genes are the ones that are actually causing the disease. And that becomes a basis for new therapeutics that we would develop at Merck.
It also becomes the basis for making a precision drug. Because we can say that for that patient, this is the actual pathway mechanism you want to target. It’s possible that for patients, even in the same indication of lung cancer, that they might have certain pathways that are active and in a different patient, it will be a different set of pathways. And we want to know what those are so that we can get the best treatment for that patient.
GEN Edge: Another feature of interest was the integration of key biological data—disease type, tissue type, cell type—as supervisory signals. Which of these would take priority?
Khalil: We need to learn from all of them. This is the beauty of advanced AI and machine learning methods. It’s that you can feed it data from cell different cell types, tissue types, disease types, other features as well and the molecular changes that are happening. And it can learn across all of them. You don’t have to tell it which one is most important.
The analogy for this, because we’re using from a computer science perspective, AI perspective, the technology that was developed for large scale LLM’s like ChatGPT. We’re using transformer-based models instead of feeding the models letters, and words, and essays, and books, we’re feeding at the biological data. And just like those LLM’s are able to learn across letters, words, concepts, essays, books and give you that answer that allows it to act where you ask it that specific query and that question, same thing here.
We learned across all of the data and in the TEDDY publication that we announced, our goal is to make these models better. We plan to continue to improve them and enable them to really capture human disease biology at scale. So we will be feeding these models other data types eventually. We will be feeding at imaging data pathology data, clinical data, multimodal data coming from proteins and enzymes, etc.
GEN Edge: You mentioned constant improvement. Any particular tweaks to TEDDY since April?
Khalil: Since April, the main thing that we have been adding to the models are disease specific data types. We are pushing through the models data on perturbing these cells and perturbing them in different disease contexts. It might be for eye diseases. It might be for immunology diseases. It might be for cancer.
Now that we’ve built the model to learn a representation of the biology, that’s allowing us to make predictions that are more accurate than what we’ve seen. So, we’re able to predict a disease indication state with 72% accuracy, 10% greater accuracy than current state-of-the-art models. Now we’re feeding them very specific disease datasets. And we hope to get to [learn]—in macular degeneration, what’s the best way to go after a mechanism that would be new, that the models are revealing are new, and potentially would have higher efficacy than current drugs and treatments?
This is a model that trained on all data. We’re not telling it what a cancer cell is or what a liver cell is or what liver disease is or what cancer is. We’re just giving it all the information, looking at the labels, and then from that it infers that.
GEN Edge: Are disease tissue and cell type the only or main signals, or are there others?
Khalil: We used RNAseq from the molecular profiles in each cell type, disease indications, the cell types, the tissue types.
GEN Edge: TEDDY was trained on 116 million cells, a larger data set than those existing models used.
Khalil: That’s right. That was our goal: To bring together the biggest tasks that we could, what was out there, and also scale it to learn that how many parameters does it need to train, right? We started with 10 million and we ramped up the compute and the training to 400 million parameters. And we were able to learn what we call a scaling law for biology.
Again, the analogy of using GPT-4 [OpenAI’s multimodal large language model): There were certain scaling laws that they knew—the computer scientists who worked on GPU for the more data from languages, and what language could do, and all the information that’s on the Internet. Eventually it would get your biology AP test correctly. It would scale to get that. We’re doing that here with cell biology feeding it as much data as possible, playing around the parameter it is and to see how accurately can we protect what disease that patient has? How accurately can we predict what pathways and mechanisms are dysregulated in that?
GEN Edge: TEDDY has two variant models, TEDDY G and TEDDY X, described in your commentary. When would you use G? When would you use X instead?
Khalil: The G we think about as the foundation model that’s able to read the recipe of the cell. It gives you, for which gene, what cell type or disease type it’s more important for. That’s the TEDDY G. And it’s sort of like, okay, you’re trying to imagine how to move one cell state or one disease state to the other. What genes do I need to perturb it to move it? The TEDDY X is for, if you’re a laboratory scientist and you’re working in the lab and you want to know exactly what is the precise level of that gene if I move another gene.
GEN Edge: Merck says that it has applied AI for more than a decade. How has that application evolved over time?
Khalil: Our foray into, let’s call it advanced data science methods started over maybe even 20 years ago where we were trying to understand how does Keytruda® work, and who does it work for? And through aggregating and leveraging the same types of large-scale datasets that we used to build TEDDY, the teams figured out what are some key biomarkers that you could measure in patients, starting from your PD-L1 levels to the status of how many mutations you’ve accumulated. And it does a fairly decent job of helping us select patients for Keytruda itself.
That gave us the idea that, you know what? There’s a lot of power in these human genetic genomic datasets. It’s something we should continue to invest in. And we did. And we had a number of partnerships over the years to get more of that cancer data. But we also realized we needed to invest in our own datasets. So, we routinely collect genetic and genomic data from our clinical trials, and we built out wet lab capabilities to generate that genetic/genomic dataset at large scale where we can profile many cells, many tissues, and even go in and perturb them, down to like perturbing a specific gene with CRISPR, so we can learn how that works. We made those investments.
Then, the key thing is that now we want to go from the most advanced data science methods to the most AI methods. So, I built out a team that is focused solely on AI/ML. And it has deep computer scientists and deep experts coming from some of the best computer science labs in the country…we felt that we had to bring in that deep expertise in-house and then enable that expertise, working with our IT and our own technology team to build off of the tech stacks that they built that allow us to build those models ourselves and scale them those models. We’re not just borrowing from the outside world. We’re actually innovating internally with the AI and building it internally.
GEN Edge: How committed is Merck in AI?
Khalil: We are committed, I can say that. And our commitment is that we have built and embedded AI and ML teams for target discovery, for composition of matter, for safety, and for understanding disease biology, and biology for clinical trials. It’s not just doing it in one area in one box; it’s across all of that. And then, it’s not a sort-of, let’s build a one-size-fits-all solution that works for everything. No, each of those boxes has its own teams of innovators that are innovating with AI and using it deeply within that workstream, and for that use case within their own teams.
GEN Edge: Like many companies, Merck says it strives to maximize AI benefits for its workforce and for patients. How does Merck balance those concerns with the desire to maximize the benefits of using AI, including efficiencies which over time could even include the loss of jobs?
Khalil: I think of AI as solving two related problems: How do you get to just better efficiencies? Can you run your trial faster? Can you launch your drug at speed, and potentially with a more optimized workforce to do it? The other problem is on the scientific side—the deep research that’s basically humbling us to unravel human biology that we didn’t know, discover molecules that we couldn’t before, and target those molecules in the right places.
As any big company, you have to do both of these things. That’s where the trend is going. But here’s what we’re learning: As the AI gets better, it’s actually helping the scientists and researchers do their jobs better. So, this is not just about efficiency. It’s not just about less. It’s actually making more with the workforce that you do have.
GEN Edge: In terms of a faster process? Or more candidates looking?
Khalil: Insights you couldn’t get to it before. Last year, we started a target ID project in a therapeutic area, and a mechanism that we had not explored before. What would have taken years for us to get to before, we should have our first best-in-class target this year. We have molecules that are in the lead optimization phase and the project teams have to figure out how to optimize that model. And instead of it taking months and months, it takes weeks. But also, the result from that is better.
GEN Edge: In what therapeutic area?
Khalil: The first case is in immunology. And then the drug optimization, I gave the KRAS example—that’s a great one, but we’re applying it across many therapeutic areas for drug optimization [Merck’s therapeutic areas of focus are oncology, vaccines, infectious disease, cardiometabolic disorders, immunology, neuroscience, and ophthalmology].
GEN Edge: Looking ahead, what additional areas does Merck see opportunities for AI that would lead to maybe expanding its use?
Khalil: The opportunities for me are being able to go deeper. We have been very strategic about putting the AI to help R&D across better targets, molecules, safety, and our first-in-human trials. For me, it’s about going deeper. I’ll give you an example. With many traditional chemo agents, we still don’t know who they’re going to work for initially. So, it’s like going and figuring out, what are all the questions that we can now tackle that we now know that will help me design a better trial?
And we might take certain diseases like SLE [systemic lupus erythematosus], where we know very little still about really what drives it molecularly, and genetically, and genomically. How can we tackle something like that? We have this great toolkit. We can look at targets where we may not even yet have great hits for them. How can we use AI to get us those initial hits faster, and drug things that maybe previously were very difficult to drug if not undruggable? It’s now being able to take these technologies, the AI along with the wet lab, along with the data—dry lab and wet lab together—to drive deeper answers to our most important questions for human health and biology.
