Artificial intelligence is starting to influence pharmaceutical development, with potential to streamline drug manufacturing and help researchers address formulation challenges that can derail otherwise promising candidates, industry leaders said at Cleveland Clinic’s AI Summit for Healthcare Professionals.
During the discussion, panelists described how AI may enable scientists to investigate more complex problems, accelerate laboratory workflows, and support work that was previously impractical. They also cautioned that generative AI can be unreliable and does not substitute for the scientific expertise, clinical judgment, and critical thinking required for safe drug development.
Jame Abraham, MD, FACP, chair of Cleveland Clinic’s Department of Hematology and Medical Oncology, moderated the panel. Speakers included Daniel Kinney, MBA, vice president of global commercial data and AI at Johnson & Johnson; Enoch Huang, PhD, former vice president of machine learning and computational sciences at Pfizer; and Thomas Fuchs, MD, chief AI officer at Eli Lilly.
Asked where AI has delivered the greatest measurable benefit at Lilly, Dr. Fuchs pointed to manufacturing. The company used hybrid deep-learning models to create digital twins that could accelerate the drying of active pharmaceutical ingredients. That improvement allowed “hundreds of millions of doses” to reach patients sooner and “paid manyfold for the computer build-out” required to support the work, Dr. Fuchs said.
He envisions even greater long-term potential in drug discovery, where researchers will use specialized AI tools to design proteins and analyze genetic sequences. Combined with automated laboratories, these tools could create a continuous process for designing, producing, testing, and evaluating potential medicines.
“That will change how we make medicines, and that will have enormous impact,” Dr. Fuchs said. However, he added, realizing that promise will take years and require more than training language models on previously published research.
AI can also help pharmaceutical companies analyze data on a much broader scale, Kinney said. Although health care and drug development have long relied on data, a shortage of data scientists who can write code and conduct analyses has often limited how quickly companies can draw insights.
Recent advances in large language models and systems in which multiple AI agents work together may help ease that constraint, he said. These systems could handle some coding and data analysis, allowing researchers to study more questions and examine the findings in greater depth.
“It allows us to ask so many harder and better questions, get to the third, the fourth question, and drive deeper insights that ultimately are going to improve everything that we do,” Kinney said.
Dr. Huang described how machine learning helped Pfizer researchers address the viscosity of an experimental trispecific antibody for an autoimmune condition. High viscosity can make an antibody difficult to inject through a narrow needle. However, measuring it requires large amounts of highly purified material and a time-consuming laboratory test.
Researchers adapted a convolutional neural network, a type of AI commonly used for image recognition, to analyze the distribution of electrical charges on the antibody’s surface, Dr. Huang said. The model detected patterns linked to viscosity that experts could not readily identify.
The system did more than speed up the design stage. It helped the team avoid repeated rounds of design and testing, each of which could take six to nine months. The researchers completed the work in a single cycle.
“We haven’t had a viscosity issue since deploying this algorithm,” Dr. Huang said. The trispecific antibody subsequently advanced to a phase 3 trial for atopic dermatitis, he added.
Panelists said pharmaceutical companies and health systems will likely use a mix of AI tools developed in-house and purchased from outside vendors. Kinney said organizations need internal experts who can evaluate these products and determine which ones are worth buying.
Dr. Huang called this an “inside-out strategy.” A core team with expertise in machine learning, data management, software development, and applied science can evaluate vendors’ claims and compare products objectively. The organization can also develop its own system if working with an outside company is too expensive.
Dr. Fuchs offered a similar principle: “You should always buy what accelerates you, but you should build what differentiates you.” Health systems have clinical data and expertise that technology companies lack, he said, and should use those resources to develop tools tailored to their needs.
The panelists also emphasized the limits of current generative AI. Dr. Huang recalled testing AI systems on mathematical and visual-recognition problems and receiving answers that were confident but wrong. The experience reinforced the need to understand which tasks these systems handle well and where they fall short.
“I’m not saying I’m assuming it’s wrong, but I’m trying to learn the boundaries for personal life and professional life,” he said.
Large language models may still help scientists challenge their assumptions, Dr. Huang said. In virtual laboratory discussions, researchers can assign AI agents roles such as skeptic, devil’s advocate, or supporter. The agents can weigh evidence for and against a hypothesis and help teams recognize potential biases.
Dr. Fuchs cautioned that language models predict likely sequences of words rather than understand the physical or biological world as humans do. They can help with education, communication, and organizing information, but remain unreliable when addressing questions at the limits of scientific knowledge.
“You have to learn where they work, where they don’t work,” he said. “It’s very important what the limits are of language models.”
