Artificial intelligence & machine learning

Artificial intelligence is here to stay at the FDA, but its integration and evolution may not be a straight line now that one of its biggest proponents, former commissioner Dr. Marty Makary, has stepped down.

Makary was pushing for a more centralized approach to AI, rolling out an agency-wide tool called Elsa, a large language model that could accelerate reviews and evaluations by taking on support tasks such as writing and summarizing reports. The agency also announced plans in December to expand the use of agentic AI to support premarket reviews, inspections and administrative tasks.

Between 2024 and 2025, the number of AI use-cases reported by the FDA skyrocketed by 148% amid a broader government push to leverage the tech, according to a Bipartisan Policy Center study.

Acting commissioner Kyle Diamantas said AI remains a top priority for the FDA. Still, recent high-level departures, including Makary, Jeremy Walsh, chief artificial intelligence officer, and Sridhar Mantha, acting chief information officer, have raised questions about the future of AI implementation.

“It's unclear now what the leadership and governance structure are around FDA-wide efforts,” said Tala Fakhouri, chief artificial intelligence and regulatory strategy officer at Parexel, and a former FDA AI policy official.

FDA officials have spoken about AI at recent conferences, but they were all representing individual divisions, Fakhouri said, noting that this could portend a return to the more fragmented, department-specific AI approach that existed prior. At the same time, efforts to increase transparency around AI use and enact AI-related policy-making could also slow, she said.

These potential setbacks primarily concern the agency’s internal use of AI, and not those governing how drug companies apply the technology in their own work. So far, there hasn’t been a change in the FDA's policies as they relate to the sponsor use of AI. 

“And I don't expect to see a change there, which is good,” Fakhouri said.

Evolving use of AI

FDA’s internal use of AI to streamline staff workloads has evolved steadily in recent years. Elsa had its origins in a CDER-developed program, CDER GPT, said Fakhouri. The agency expanded on that original program to use a retrieval-augmented generation system to reduce AI hallucinations. The system is designed to confine the large language model to a well-defined database of trusted information, tailored for the individual centers within the agency, Fakhouri said.

This allows staff to use the tool to access information needed for their unique roles. For example, staff members can quickly generate a summary of industry comments on a particular proposal, or the Office of New Drugs could turn to Elsa to quickly generate a history of regulatory submissions, which can span years.

While Elsa has been leveraged for these types of burdensome tasks, Fakhouri doesn’t believe the technology is used for final decision-making. 

Still, “staff can use the tools to augment the work that they're doing. We should all be happy about that,” she said.

Where the FDA’s AI program should go from here

Transparency about how the FDA is using AI in its processes is still lacking, Fakhouri said. 

“If the regulators are using AI in certain ways to augment reviewer work or to become an assistant to a reviewer, I think it's good practice for industry to know what these uses look like," Fakhouri said.

Detailing how the FDA is using AI could also foster collaboration with the industry. 

“If I am a sponsor or a CRO preparing a submission on behalf of a sponsor, I can prepare my submission with all the data labels and information that might be helpful for the reviewer and the [AI assistant] to be able to review my package,” she said.

Fakhouri said she expects the agency to move toward greater transparency over time. 

“But it would require someone in a leadership position at the agency to become aware of that and want to actually make it happen,” she added.

Fakhouri would also like to see the agency streamline AI-related rulemaking that impacts the industry, such as how AI tools are validated for clinical trials. 

Under Makary, policy changes were sometimes announced outside the traditional FDA guidance process and through journal articles or press conferences instead. But the agency appears to be returning to its former norms under Diamantas, who recently confirmed that informal statements made by the former commissioner don’t represent official policy. 

“They will go through the regular guidance and policy development processes,” Fakhouri said. 

While traditional rulemaking insulates pharma from uncertainty, it can also be a slower process that creates its own challenges as companies rapidly adopt different AI platforms

“The fastest guidance that you could put out would still probably take a year. A year in the age of AI is very slow,” she said. 

Finding a balance between structure and flexibility will be one of the FDA’s key policy challenges. 

“I don't necessarily know what the solution is, but more agile, flexible policy development would be needed, and perhaps more frequent communication between industry and the regulators,” she said.

AI adoption is still ticking up in pharma with 34% of companies leveraging the tech for specific functions, according to a recent GlobalData survey. Early drug development in particular has been the prime driver of AI adoption, with nearly 60% of respondents reporting they use the tech for discovery and target identification.

But pharma companies that want to take it further are often still unsure where they should deploy the technology. Only about 15% of global pharma and life sciences companies said they felt fully prepared to develop AI business models, according to a recent PwC survey, and many still don’t have a detailed AI plan sketched out. 

Not only that, AI hasn’t always fulfilled its grandiose promises. 

We’re still early in the evolution of AI and in many ways, its impact is just now being felt in pharma. But AI has been around long enough for companies to figure out where it might not live up to the hype. 

At this year’s annual BIO convention, I asked industry leaders about the various ways they’ve been disappointed in AI. Here are the use-cases they identified where AI still has plenty of room to grow. 

R&D timelines still lag

“In the early stages with antibody design or protein design, AI can shorten your development time from a year to maybe a few months. But this is not life-changing for the whole process. It's disappointing because the hope was taking development from one year to a week … that would be a big change. People adopted the technology, saw what it can do and thought it was incredible, right? So why is it still taking six months? The hope is that we will really shorten timelines to make it cheaper and faster.”
~Sharon Gour Arie, chief operating officer, AION Labs

It can’t create drugs from scratch

“We can start from a really good point with a sequence of a potential molecule for our antibody therapeutics, and then we can optimize it. If you start from a reasonably good sequence, you can get exceptional outcomes. What we have not seen with AI is going from scratch to a really good outcome. I think there's a lot of confusion about what generative means. If we start from a sequence and then we use AI to optimize it, then yes, it generates a new sequence. But that is not generative AI. So we can go from good to great. We still cannot go from nothing to great.”
~Marianne De Backer, CEO, Vir Biotechnology

Key word restrictions stymie research 

“If we have the word Ebola in our Excel sheet, you can't use Claude anymore. It just dies … because it thinks you're trying to generate an Ebola bioweapon. There's no path right now to remove those restrictions. The model is built. Those restrictions are important, and I would argue that's the right initial move. And the reality is, we could be using these tools to help us design things like clinical trials and that would be phenomenal. But if we want to use AI in biomedical research, we have to use these large language models, and those restrictions would have to be fixed in the new models.”
~Jared Bauer, CEO, Seek Labs

AI implementation needs work

“To me the disappointing thing is not about AI, it's about how companies are implementing AI. The CEO says, ‘AI is awesome, I want everyone to use AI, and I want to see ROI.’ And then the company rolls out Copilot or Claude and staff are supposed to figure it out. That doesn't work because most people will do the same thing, which is summarize documents or write stuff. It can't be like trickle-down AI on everyone, and then it's going to work.

It’s better if you have a citizen developer … or a plan where the company puts people from all the different divisions in one room, and asks them to come up with use-cases for tasks that are challenging, laborious, take a lot of time or can easily be automated. Then you build the solution.” 
~Tala Fakhouri, chief AI and regulatory strategy officer, Parexel

AI hasn’t unlocked the secrets of biology

“We know AI helps us work more quickly, more efficiently. There's nothing wrong with that because that's going to drive some benefit, maybe a productivity or an economic benefit. But I just feel that biology still has so many secrets. So how could we use the AI tools to help uncover some of those secrets? That's going to be the key to opening up novel ways to discover new drugs.”
~Lorna Ewart, chief scientific officer, Emulate

It’s getting in the way of big picture thinking

“I just finished reading Mustafa Suleyman's book called ‘The Coming Wave,’ and he was talking about the industrial revolution and how we thought it would improve our quality of life because the machinery would free up so much time that people would have a challenge figuring out what to do with their time. That didn't happen. It increased productivity but we're busier than ever. AI will take the load off stuff like writing reports, reading email, summarizing meetings … and we think we're going to have so much more time, but I think it's also creating the opposite situation.

There's a huge acceleration of information creation and proliferation across the board in our lives that AI helps enable. For example, there are tools that send all these automated emails. It creates a lot of noise, making it harder for me to find the real people to talk to. So I spend part of my day just filtering through the noise … and you can't step back and think big picture and get big ideas.”
~Svetlana Lucas, chief business officer, Scribe Therapeutics

Sponsored content

By EVERSANA

The pharmaceutical industry has moved beyond asking whether AI will reshape commercialization. That answer is already clear. The technology works, investment is widespread, and nearly every organization has begun to explore its potential.

What remains unresolved is far more practical—and far more urgent: who can actually make AI perform.

Not in a demo. Not in a pilot. But inside the conditions that define pharmaceutical commercialization.

That is why a new expectation is emerging—one that is already reshaping how leaders evaluate AI.

Don’t buy before you try.
And more importantly: your partners should prove it.

If your vendor cannot prove it in practice, they should not be in your strategy.

At EVERSANA, that expectation is already being met. Operating inside real commercialization environments, our platform moves from concept to compliant, activation-ready content in 30 minutes. Connecting strategy, omnichannel planning, and content creation in a single operating model rather than a series of disconnected steps.

That should reset the standard.

AI is not defined by what it claims to do. It is defined by what it produces in your environment. If a platform cannot demonstrate that it delivers usable, compliant outputs—and that those outputs translate directly into execution—it does not matter how advanced it appears. It is not ready to scale.

Across the industry, AI activity is accelerating. Teams are testing use cases, running pilots, and generating increasingly sophisticated outputs. But at the enterprise level, progress remains uneven. While adoption is widespread, McKinsey reports 88% of organizations report using AI in at least one function—but most have not yet embedded it deeply enough to realize enterprise-level value.

That gap between capability and performance now defines the market.

Organizations are already adjusting. The conversation is moving away from selecting AI based on potential and toward validating it based on results.

Because this is no longer a question of access to AI. The market already has it. The question is execution—who can turn AI into something that performs reliably, repeatedly, and at scale.

Capability alone is no longer enough. What matters is whether a platform can perform inside the realities of pharmaceutical commercialization.

In practical terms, that means the AI must operate end-to-end, not just at a single point in the workflow. It must produce outputs that are compliant from the start. It must integrate into how teams already work. And it must scale across brands and lifecycle stages without being rebuilt each time it is applied.

This is where agentic design becomes critical. AI agents must be built to operate within compliant workflows—not around them—with the right knowledge, tools, and guardrails to drive business-critical outcomes, not isolated tasks. Without that, AI remains fragmented, no matter how advanced the underlying model is.

This is what separates platforms that can be demonstrated from platforms that can be deployed.

At EVERSANA, we are already operating this model in practice—supporting over 10 companies across the commercialization spectrum, from launch through to fully in-market brands, executing across strategy, content, and activation in a single, integrated system.

This is not a pilot environment. It is not a contained use case.

It is a coordinated system of specialized AI agents working across strategy, content, and activation—trained for specific tasks, connected to shared knowledge and workflows, and designed to operate within the regulatory and commercial realities of pharma.

It is AI embedded directly into the commercialization workflow—producing work that moves from concept to execution in minutes while maintaining the compliance, control, and precision the industry requires.

Scaling AI in pharmaceutical commercialization is not about layering in more tools or experimenting with vendors that can show isolated capability. It requires a model that already works across strategy, content, compliance, and activation—and continues to perform as it expands.

The advantage will go to those who align with that model early.
Because the next phase of the market will be defined by execution—who can deliver faster, more consistently, and at scale.

AI leadership will be earned.

Earned through execution.
Earned through scale.
Earned through proof that stands up in real-world conditions.

Not claimed through positioning. Not assumed through investment. Not projected through roadmap.

Because in a market where AI is expected, execution is the only true differentiator.

And there will be no ambiguity about who is leading.

AI leadership will not be claimed. It will be earned.

AI has been touted as a tool to streamline drug R&D, which can cost $2.6 billion and take up to 15 years, according to the Journal of Medicinal Chemistry.

But whether AI will actually develop better drugs remains an open and complicated question. To date, few AI drugs have reached clinical trials. Those that have tend to see more success in phase 1 than traditional drugs. However, that advantage might not last in phase 2. 

The space has also been marked by high-profile setbacks, including Verge Genomics’s AI-identified ALS candidate, VRG50635, which failed to move beyond phase 1. 

While the overall picture is still emerging, a few closely-watched candidates offer some real-world insights into how AI-developed drugs are performing.

Insilico’s leading candidates

Insilico Medicine has become one of the highest-profile players in AI drug development and has struck a number of deals with Big Pharma companies looking to leverage its capabilities. Insilico also has its own candidates in the works as part of its expansive pipeline that includes over 40 programs in a wide swath of indications. 

Often considered pharma’s leading AI-designed drug candidate, the company’s rentosertib was discovered with an AI-powered biology tool that pointed to a novel mechanism called TNIK inhibition. The approach has the potential to slow or reverse the progression of idiopathic pulmonary fibrosis, a deadly disease that stiffens tissue in the lungs that has few effective treatment options. Even with existing drugs, IPF patients typically only live for two to four years after diagnosis.

Boehringer Ingelheim notched an FDA nod for an IPF drug called Jascayd last year, marking the first new approval for the condition in more than a decade. But Jascayd’s potential impact on the disease was dubbed by one analyst as “modest” because of its moderate effect on lung function and potential adverse interactions with other medications.

In a small phase 2 study, rentosertib improved lung function and is now moving into a 52-week phase 3 in China, which the company said marks a major milestone for the treatment and AI-driven drug discovery. 

“Rentosertib was not discovered by starting from a conventional target and simply screening more compounds. It came from a biology-first, aging-informed AI workflow that connected TNIK to fibrotic and inflammatory disease mechanisms, and then used generative chemistry to create a drug candidate with the properties required for clinical development,” the company said in a press release.

Insilico is also advancing another closely-watched drug called garutadustat, an oral PHD inhibitor for inflammatory bowel disease. The treatment, which is now in phase 2 in China and will be tested in patients with ulcerative colitis, has been touted as a potential best-in-class option by the company. The dual-mechanism drug, flagged by AI during a 12-month analysis, aims to both reduce inflammation and repair the intestinal damage that marks the condition.

Recursion overcomes setbacks and charges ahead 

Recursion Pharma has also become a leading company in AI drug development, but has grappled with a few stumbles in the clinic. 

Disappointing results for its lead drug asset for cerebral cavernous malformation and a setback with another drug targeting neurofibromatosis type II led the company to prune its pipeline last spring. But Recursion is still advancing a number of candidates, including REC-4881, a MEK inhibitor it in-licensed from Takeda Pharmaceuticals after AI identified the mechanism as a promising treatment for familial adenomatous polyposis, a condition marked by the development of hundreds or thousands of precancerous colorectal polyps.

Recursion announced results from the ongoing phase 1b/2 trial for the drug REC-4881, showing a 43% median reduction in polyp burden in treated patients over three months. Most patients in the trial continued to show reductions even after they were off the medication for 12 weeks — a promising sign for a condition with no existing treatment options. The trial is slated for completion in 2027.

Astellas Pharma has been on a quest to transform its approach to drug development in the last few years with the goal of ditching its long-held reliance on outsourcing and bringing more operations in-house. 

“We’ve created a team that is responsible for overseeing the development of every single asset end-to-end,” explained Tadaaki Taniguchi, who’s leading the reorganization of Astellas’ R&D program as the company’s chief research and development officer. 

Under the new model, Astellas began internalizing its global clinical operations and deploying its own research teams to run trials. And this major shift in operations wouldn’t have been possible without AI, Taniguchi said. 

While AI has a strong foothold in drug discovery, Astellas leaned into the technology’s capabilities in streamlining ho-hum tasks in clinical operations like protocol development, cross-border communication and patient outreach. But the “sweet spot” for the Japan-based pharma has been using AI to translate documents.

“We’re working in 15 to 20 different countries,” Taniguchi said. “If you are running trials globally, you need to translate the language properly — even just for creating language for patient materials. For a Japanese company, we now have more confidence that we can do this globally.” 

Internalizing clinical development has allowed Astellas to mostly cut out CROs and work more quickly and directly with site investigators to tackle pain points, which has “accelerated pipeline progress,” Taniguchi said. 

Although Astellas hasn’t publicly disclosed how much the new strategy has slashed costs, the effort contributed to a broader overhaul that’s saved the company $65 billion yen (about $406 million) over the last two years. Astellas also reported record-high revenue in 2025. 

The revamp is particularly apt as Astellas stares down a dreaded patent cliff for its blockbuster cancer med Xtandi, which will start losing protections in the U.S. next year.

Astellas won’t have any splashy late-stage readouts in the coming year to showcase its progress with its next up-and-comers. Instead, the company will be in execution mode while it advances a pipeline built around four main focus areas: immuno-oncology, gene therapy, blindness and regeneration and targeted protein degradation. 

But in 2027 and beyond, Astellas will begin to reveal if it’s made headway on a number of fronts, including a potential high-profile showdown with Revolution Medicines in the KRAS arena. 

Astellas’ rising stars

One of Astellas’ potential rising stars is its KRAS-targeting treatment setidegrasib. The company recently started a phase 3 trial for the protein degrader in KRAS G12D-mutated pancreatic ductal adenocarcinoma, which impacts 40% of patients with the disease. 

Revolution Medicines already has a notable lead in the push to bring a KRAS-targeting therapy to pancreatic cancer. The biotech recently published results it heralded as "unprecedented" for the condition that showed its candidate, daraxonrasib, nearly doubled the survival rate in a phase 3 study. 

But Astellas could still wind up with an edge in pancreatic cancer, Taniguchi said, because its candidate degrades, or destroys the protein, rather than just inhibiting it. That distinction could offer a better safety profile, which remains a notable challenge for Revolution, and is why it’s developing daraxonrasib to be used with milder forms of chemotherapy, according to Taniguchi. 

“Our candidate is easier to combine with different chemotherapies in this setting,” Taniguchi said. “So their strategy and our strategy are quite different.”

An early-stage trial for setidegrasib in lung cancer — another indication both Astellas and Revolution are targeting — demonstrated a positive safety profile. 

Topline results for setidegrasib’s phase 3 lung cancer trial are expected in 2028, according to the company. As the data for both indications roll in, the drug’s overall survival rate will be key to determining how well it performs compared to other candidates in the space, Taniguchi said. 

Another flagship candidate Taniguchi highlighted is Astellas’ claudin-targeting T-cell engager ASP2138, which is heading into phase 3 for gastric cancer.

The company is also pinning its future growth on a gene therapy candidate for Pompe disease and a cell therapy dubbed ASP7317 for geographic atrophy, an advanced form of dry age-related macular degeneration. The company recently presented positive early-stage data for ASP7317 and aims to move the candidate into phase 3 within the next year, Taniguchi said. 

“When I meet people in the field, they are so excited about this product,” he said.

Pharma companies already use AI for tasks like optimizing clinical trials, identifying drug targets and developing predictive models for risk detection. But AI’s next iteration promises to take the technology a step further with “agents” that act autonomously.

Among those at the forefront of this shift is Owkin, an AI company whose clients include Sanofi, Bristol Myers Squibb and Merck & Co. Its autonomous “AI scientist” can tackle complicated questions about topics like commercialization and drug discovery in a fraction of the time human researchers could. 

Pascal Weinberger, Owkin’s co-CEO, likens it to handing a thesis to a Ph.D. student and asking them to report back with their findings.  

“The mental model is: What if you gave every researcher in your company a data center full of genius Ph.D. students in their pocket?” Weinberger said.

AI agents operate, plan and adapt independently, or as McKinsey & Company recently described it, they change “AI’s role from tool to coworker.” 

Owkin’s tool, K Pro, uses agentic AI to take a complex, multipart pharma research question and deliver a comprehensive answer with annotated citations to back it up. Weinberger offers this example: Asking the tool to analyze the top indications being pursued in clinical trials for an antibody targeting CTLA-4, comparing them to gene expression across MOSAIC cancer types, and determining whether there’s an indication competitors are missing.

The AI scientist will answer the question step by step, conducting literature research, running gene expression analysis and data science models, and doing general web searches for other relevant information like press releases.

“Just answering a question like this would take a team traditionally weeks because you're doing a lot of research on different areas,” Weinberger said. But their tool can do it in hours or days and it provides the “scientifically validated plots that you would expect in an analysis like this.”

In addition to reviewing existing literature, the tool also draws on a patient dataset from more than 800 hospitals, including “gene data, pathology slices, insurance reports, anything,” Weinberger said. 

“All of that has some little piece of information on how the underlying biology actually works and it helps you model it,” he said.

‘Supercharging’ existing employees

Several pharma companies are already using agentic AI, including Novo Nordisk, which is working with NVIDIA to develop customized agents. Johnson & Johnson is using agents to determine certain chemical processes related to drug development. On the vendor side companies like IQVIA are developing customized AI agents that can handle a range of tasks. Owkin says some of its clients are already using its K Pro platform. 

But humans still play a large role in all of these situations.

“There’s this idea that over time, especially as the models get more powerful, you can remove the human from a lot of those steps because maybe an AI model can do the review better than a human can, or maybe an AI can come up with better research hypotheses,” Weinberger said.

In fact, these tools’ greatest value will likely come from performing tasks that are too complex or inefficient for humans, like analyzing unstructured datasets, according to McKinsey. While tools like K Pro can operate independently, Weinberger doesn’t believe it will dramatically reshape companies’ workforces and put data scientists and biomedical researchers out of a job.

Some hiring experts agree.

“I wouldn’t say that AI is necessarily replacing jobs one-for-one,” Jae Yoo, executive director of EPM Scientific, a recruitment firm specializing in pharma, biotech and R&D hiring, told PharmaVoice earlier this year. “It’s rehousing and reshaping the types of jobs that are now coming in.”

A recent poll of industry executives echoed that sentiment, with pharma’s C-suite leaders reporting they don’t think AI will lead to major job losses.

Weinberger said that’s true even within Owkin.

“We still have a big team of researchers that basically fuels the machine, building the skills, finding and verifying the data, finding and verifying the resource, asking the right questions,” he said.

Ultimately, humans will still be responsible for using AI-generated data to decide a company’s strategic direction. That’s why having the tool produce verifiable results is so crucial.

“It's important to our pharma customers that they can review this,” Weinberger said. “I want to know how this graph was produced because I'm going to make a million, if not hundred-million-dollar decision based on the analysis that I get here.”

Weinberger isn’t promising that some jobs won’t go away.

“If the world changes in a meaningful way, some people will inevitably lose their jobs,” he said. “I think the goal here is to accelerate progress.”

But Weinberger believes most displaced employees will move onto other positions because the task at hand is so massive. Millions of patients will still have unmet medical needs and the drug discovery process is still long and expensive.

He argues that if a tool like theirs can make one researcher or engineer 1,000 times more efficient, pharma companies likely won’t reduce staff and settle for the same number of discoveries and revenues they’ve had for years.

“That doesn't really happen in the world,” he said. “What happens is the pharma says, great, now I have 1,000 people that can do the work of 100,000 people or a million people, figuratively speaking, and I can grow my top line. I can solve more diseases. I can add more drugs to my portfolio.”

In focus with: Natalie Ma, chief business officer and co-founder at Deep Origin

Deep Origin’s vision: What if “virtual humans” and in silico trials could become stand-ins for real humans in clinical trials?

That’s the long-term vision of Deep Origin, a computational drug discovery company that combines molecular physics with AI to create predictive models for clinical trials.

Rather than failing faster, “the idea is to stop doing experiments that will fail,” said Ma.

“What we're after is the quality of the prediction itself. That’s the predictivity crisis, a systematic gap between what works in preclinical models and what happens in the clinic,” she said.

In addition to advancing its drug discovery platform with a variety of partners and working to better predict drug safety, Deep Origin is also simulating human biological systems at the cellular and organ levels to eventually build a complete “virtual human.”

But it’s a long, step-by-step journey to chase that larger goal. In the meantime, Deep Origin is tackling more bite-sized industry challenges.

“We don’t position our immediate work as ‘replacing’ clinical trials,” she said. “Right now, it’s about reducing late safety failures and animal use, and giving programs more confidence going into the clinic.”

Deep Origin received a $31.7 million contract from the Advanced Research Projects Agency for Health’s Catalyst program to replace animal testing with an FDA-qualifiable in silico prediction platform that will test how a medicine moves through and affects the human body.

Ma said it will bring its first organ models online later this year or early next year.

“The goal is to produce more accurate, human-relevant safety predictions before a molecule advances, bringing organ-level models online, tracing predicted drug-induced liver injury or tox signals back to specific off-target interactions and using those insights to redesign or de-prioritize risky compounds,” Ma said.

Why it matters: Drug developers are “consummate gamblers,” Ma said. They face a 90% clinical failure rate and still knowingly embark on a decade-plus, multimillion dollar odyssey to bring drugs to market.

Drug developers are burdened by other limitations, too, from relying on animal testing to navigating ultra-rare disease states with so few patients that traditional clinical trials are tough to pull off.

But Ma envisions a world where “we get predictive enough, you [can] run an in silico clinical trial, and clinical trials are a formality.”

“You can parameterize virtual humans for different genetic backgrounds and comorbidities, and over time that could inform in silico clinical trial concepts. In its final form, researchers will be able to add a small molecule structure and a dose to the virtual human and see the outcomes across organs,” she said. “But that’s a long regulatory horizon, gated by validation and dialogue with the FDA.”

Here, Ma expands on the company’s vision.

The interview has been lightly edited for brevity and style.

PHARMAVOICE: How are your partners currently using the platform?

NATALIE MA: We work with partners in three ways: discovery partnerships that put our systems and scientists against a partner's hardest programs; direct SaaS access to the platform and co-development of Deep Origin-originated programs. In practice, partners use our platform across the preclinical pipeline, from prioritizing targets by efficacy and toxicity risk, designing and filtering molecules in hit discovery, to prioritizing leads in optimization. We haven't done a rescue of a program after a toxicity signal or patient population segmentation, but we think our models could be adapted to these as well. 

Because most collaborations are confidential, we talk about the workflows and the impact rather than specific deal structures. A lot of the value comes from being able to increase success rate in the lab while still finding high-quality hits and then layering safety predictions on top, so what moves forward has a much better chance of working in humans, not just in animals.

We also see tremendous, bidirectional value in working directly with partner team scientists: they bring specific incredible knowledge, while we work to impart and educate on AI methods. It's incredibly important each team knows how to use, weigh and integrate the evidence or recommendations from our models, because despite anyone's claims, no model is correct 100% of the time, particularly because there is often context only the researchers know.  

Would each “virtual human” be an individual unique from each other? 

That’s the framework that underpins the “virtual human” work. We can run models on a part (a single type of cell or one organ), using parameters that are an average of a population, or as specific as a single individual. While we are focused currently on preclinical stage work, we have actually discussed with some clinicians and industry the creation of individualized models for patients to predict response to treatments. There's been a lot of interest! 

I want to note some distinction here with “digital twin," which usually connotes a “statistical average” of patient parameters to represent a patient. Those models are generally built on population statistics. In contrast, the virtual human is a system of interconnected models, modeling human biology across multiple organ systems, within which there are models of multiple cell types, within which there are models of gene, protein, macromolecule and small molecule interactions. There are representations of biology across scales to capture the nuance of individuals.

Is this idea of a “virtual human” being pursued by other companies/research institutions? How is yours different or similar?

To some extent yes, though less than I would expect, in part due to the large number of discrete disciplines the work spans.

From a perspective of reducing animal use and using new approach methodologies that are more human relevant, there are many in industry, academia and nonprofits pursuing this. This is wonderful for drug development because it means we can mature multiple human-relevant models at once to inform which molecules will succeed as drugs. It also works well for us as model builders, as they provide yet another source of data to create increasingly predictive models.

Computationally, the field is counterintuitively sparser, with much of it focused on building property-specific machine learning predictors (which tend to fail beyond their training data and don't get you mechanistic explainability) or foundational models of chemistry/biology (very black-box and limited in explainability) or in building models at one scale of biology (a virtual cell, which begs the question of what cell type and limits the questions you can ask of the system, or just focusing on docking, which is good but incomplete — a binder does not a drug make). 

We tie models together across scales — from quantum mechanical interactions and docking, through protein dynamics and cell‑state changes, up to organ‑level exposure and whole‑body physiology — so safety questions can be traced back to concrete mechanisms.

And because our models are grounded in physics, every prediction comes with an explanation a scientist can inspect and test, which is crucial for partners and regulators who have to act on the outputs of our work.

AI has already demonstrated its mettle as an effective tool for discovering novel drugs. Over two dozen candidates developed with AI are in clinical trials, according to analysis published late last year.

Now, scientists are looking to one of the next frontiers.

Rather than using AI to create the right drug, some new models are being trained to pinpoint the underlying triggers of disease, an approach that could reveal new targets for drugs to go after. In other words, rather than finding the keys to treat a disease — AI could help reveal which locks have gone unexplored.

“It’s about finding new targets altogether … new biology, new mechanisms for solving disease,” said Nick Naclerio, founding partner at Illumina Ventures. “That’s a much harder problem.”

Pulling off this feat will mean building models that unravel the complex intricacies of human biology.

“AI is trained on what’s in the literature, but there’s a finite understanding of how human biology works,” Naclerio said.

With that goal in mind, Illumina, which specializes in R&D technologies and diagnostics, announced a research consortium this year to build a comprehensive map of human disease biology.

Called the Billion Cell Atlas, the genome-wide perturbation dataset is being created with an alliance including industry bigwigs like Merck & Co., Eli Lilly and AstraZeneca. The goal is to build a “curated set of cell lines to drive drug target validation, train advanced AI models at scale, and advance research into fundamental disease mechanisms that had previously been out of reach,” according to the press release.

The Atlas model will develop over 200 cell lines relevant to a range of disease areas such as cancer and cardiometabolic and neurological conditions. Researchers then plan to use CRISPR technology to test what happens to billions of individual cells when thousands of different genes are switched on and off.

“Imagine you bought a new house and wanted to know what the light switches do,” Naclerio said. “In this model, each switch could impact thousands of systems … and you’ve got around 20,000 genes you’re going to turn on and off multiple times in hundreds of tissues. Now you’ve got an enormous dataset.”

The insights the model generates into human biology could then be used to identify drug targets with the help of AI, Naclerio said.

Illumina’s efforts add to a growing number of AI approaches to computational biology.

Google DeepMind’s AlphaFold platform has become known for its ability to predict complex protein structures — a breakthrough that snagged a Nobel Prize in 2024.

Other biotechs are pushing these bounds further. Australia-based Omnigeniq, for example, developed a platform that can visualize proteins in their ligand-ready state, which will help “engineer therapies that come from an understanding of the changes that lead to disease propagation,” the company’s CEO, Jordana Blackman, told PharmaVoice earlier this year. 

New disease-focused techniques from the academic arena could also set the stage for upending long-held approaches to drug R&D. 

Instead of the traditional method of testing one protein or drug at a time, an AI-driven tool developed by researchers at Harvard Medical School can now identify multiple disease drivers and then which genes could address them. This new model could steer drug development toward more effective targets, researchers reported in September.

“The next few years will be a race to build foundational models that can make leaps in generating new insights into understanding biology,” Naclerio said.

The data cycle

AI systems are only as good as the data that feeds them — and models for disease target prediction are no different. Pulling together those datasets will be a chief focus for the industry. 

“This is where the competition is right now,” Naclerio said. “Pharma is looking for big datasets to train models to find new targets.”

The current need for more data follows a technology development cycle that’s emerged recently.

“Over the last 20 years, the amount of data we could generate has gone up by many orders of magnitude. It got to the point where we had more data than we knew what to do with. Then came AI to ingest all that data,” Naclerio said. “And now we’re back around to having new tools like AI, but we need to ramp up experimentation to extract new data for them.”

Although the industry is not at the point where AI can “instantly find cures,” Naclerio said, the timing is almost right for the available data and toolsets to converge for the next leap forward.

“Over time, we are going to bend the curve of drug discovery productivity and see huge improvements,” he said.