Q&A: What can AI realistically do for medical research right now?

HERSHEY, Pa. — A new wave of artificial intelligence (AI) tools is emerging to help speed up scientific discovery, especially in biomedical research.

By analyzing massive datasets, these tools could flag genes that shed light on how disease progresses or drugs that may lead to a potential new treatment, but any errors AI makes can carry serious consequences, according to Dajiang Liu, a University Distinguished Professor and director of artificial intelligence and biomedical informatics at Penn State College of Medicine.

In this Q&A, Liu — who also holds appointments in public health sciences and in molecular and precision medicine — discussed the promises and risks of this new generation of AI tools and what the next generation of researchers and healthcare professionals need to know.

Q: What makes biomedical research different from other sectors where AI has been adopted more quickly?

Liu: In many commercial settings, an AI error may lead to inconvenience, inefficiency or financial loss. In biomedical research, errors can affect scientific conclusions, drug development decisions or even patient care, making the stakes higher. It requires rigorous study design, careful definition of physical traits or phenotype, biological interpretation and independent validation.

Another key difference is that biology is extremely complex. Within the human body, disease is influenced by genetics, environment, behavior, immune function, aging, treatment history and many other factors that tangle together in ways we don’t fully understand. Biomedical data are often noisy, incomplete and gathered from different populations, technologies and health systems. Even experts may disagree on a patient’s diagnosis.

There are also important ethical and regulatory considerations. Biomedical AI often relies on sensitive patient data, so privacy, consent, fairness and responsible data use are central issues. For AI to be useful in this space, it not only has to be technically impressive but also scientifically rigorous, reproducible and trustworthy.

Q: How might tools like these fit into biomedical research workflows in academia or industry? What kinds of tasks are they most likely to help with right now?

Liu: In the near term, these tools are most useful as research accelerators. They can help researchers generate hypotheses, summarize existing literature, identify patterns in large datasets and prioritize genes, variants, pathways or compounds for follow-up study. They are designed to work across environments that scientists already use, including literature databases, coding notebooks, statistical tools and computing clusters.

For example, in genomics and biomedical informatics, AI can help integrate many layers of data, including genetic variation, gene expression, protein data, imaging and electronic health records. That can make it easier to identify biological mechanisms that may contribute to disease or treatment response. In drug discovery, AI may help with target identification, molecular design, toxicity prediction and repurposing existing drugs for new indications.

AI may change the way biomedical research is done, but it will not eliminate the need for scientific judgment. These predictions still need experimental and clinical validation. AI can help narrow the search space, but it does not remove the need for careful study design, biological expertise and validation in relevant models and patient populations.

Q: How important is access to large, high-quality datasets in this new AI landscape?

Liu: It is fundamental. AI models depend heavily on the data are that are used to develop them. The future of biomedical AI will depend on building datasets that are not only large but also carefully curated, representative and linked to meaningful biological and clinical information.

In biomedical research, large and high-quality datasets are important because many disease mechanisms are subtle and heterogeneous. Small or biased datasets may not capture the full complexity of human biology.

But even a large dataset can lead to unreliable conclusions. Data quality matters just as much. Researchers need accurate phenotypes, well-characterized clinical samples, standardized measurements and appropriate metadata.

Making sure the population is well-represented is also essential. If AI tools are trained primarily on data from one ancestry group, one health system or one type of patient population, the tools may not apply to the general population.

Q: As AI becomes integrated into biomedical research, what are other risks people should be aware of?

Liu: AI systems can produce results that appear sophisticated and convincing, even when they are wrong and biased. There are also risks related to reproducibility and transparency. Some AI models are difficult to interpret, and it may not always be clear why a model made a particular prediction. For biomedical research, we need tools that can be evaluated rigorously, benchmarked against existing approaches and tested across independent datasets.

Finally, there is a risk that AI could be used without enough expertise in the area. Biomedical questions require careful framing. A technically strong AI analysis can still be scientifically flawed if the question, data structure or causal interpretation is wrong.

AI is not magic. While getting answers from AI is easy, they may not be the right answers. These AI tools do not automatically produce cures, and they should not be viewed as a substitute for rigorous science. The value of AI will depend on how responsibly it is developed, tested and integrated into biomedical research and health care.

Q: What do these advancements in AI mean for training the next generation of scientists and clinicians? What skills will matter most in an AI-enabled research environment?

Liu: The researchers who succeed in this environment will be those who can combine computational tools with rigorous experimental design, domain expertise and a clear understanding of human health. Not everyone needs to become an AI engineer, but they do need enough AI literacy to understand what these tools can and cannot do.

Critical thinking will be more important than ever. Trainees need to know how to evaluate data quality, recognize bias, interpret model outputs and design appropriate validation studies. They also need a strong foundation in biology and medicine, because AI-generated predictions must be interpreted in the context of real biological mechanisms and real patients.

Communication and collaboration will also be essential. The most impactful work will often require teams that include clinicians, basic scientists, computational scientists, statisticians, ethicists and patients. Training programs should prepare students to work across disciplines and to use AI responsibly.