Questions Answered: Generative AI in Bioinformatics

In this blog, Dr. Daoud Meerzaman answers questions about the use of generative AI in bioinformatics. Dr. Meerzaman oversees CBIIT’s Computational Genomics & Bioinformatics Branch (CGBB), where he and his team provide bioinformatics analysis support for life sciences as well as clinical and translational research conducted by intramural scientists.

Question: How are large language models (LLMs) being used in biomedical research? 

Answer:

Before answering that, it helps to define where LLMs fit in the broader AI landscape. Artificial intelligence is the broad umbrella. Within AI, there is machine learning, and within machine learning, there is deep learning. Generative AI is one class of deep learning approaches, and LLMs are a major category of generative AI designed to work with human language.

In biomedical research, LLMs are especially useful for knowledge synthesis, information extraction, and data structuring. We are dealing with an enormous volume of information: millions of publications, disease-gene associations, clinical trial records, pathology reports, and data embedded in electronic health records (EHRs). No individual researcher—or even a team of researchers—can manually absorb and organize all this information fast enough to keep pace with the field. LLMs can help by rapidly identifying patterns, summarizing findings, extracting relevant concepts, and organizing unstructured text into formats researchers can analyze more efficiently.

For example, in the context of clinical trials, an appropriately developed and validated LLM can help identify the trial type, summarize the therapeutic intervention, extract patient response information, and connect these elements to relevant literature. It can also help researchers compare trial criteria, identify biomarkers mentioned across studies, and generate structured summaries from complex documents.

That said, there is a major caveat: hallucination. Sometimes an LLM produces information that sounds plausible but is incorrect, incomplete, or entirely fabricated. One common example is a “Frankenstein citation,” where several citations are pulled together and labeled as a single citation that does not actually exist. This is a serious problem. That is why human supervision is absolutely essential. LLMs can accelerate work, but they should not be treated as autonomous sources of truth.

Question: What is retrieval-augmented generation (RAG), and how does it apply to biomedical literature analysis?

Answer:

RAG is a method for improving the quality and reliability of LLM responses. In simple terms, RAG works by connecting a language model to a trusted information source—such as a curated document repository, a biomedical literature collection, or an internal knowledge base—so that the model can retrieve relevant information before generating its answer.

Instead of relying only on what the model learned during training, RAG allows it to pull from real-time or institutionally curated data sources. These sources may be represented in a searchable vector database, which helps the system find semantically relevant content even when the wording differs from the question. In practice, that means the model is not searching random corners of the internet; it is grounding its response in data that your organization knows, trusts, and maintains.

This is especially important in medicine and biomedical research. If I ask a model a question about biomarker evidence, a disease mechanism, or treatment response, I want the answer anchored in authoritative, citable sources. RAG helps reduce hallucinations by forcing the model to retrieve evidence first and then generate a response based on that evidence. It does not eliminate errors completely, but it is one of the most practical and effective ways to make LLMs more reliable for scientific use. The CGBB Team and I used RAG briefly in a small project to map Common Data Elements (CDEs).

For biomedical literature analysis, RAG can support tasks such as identifying relevant publications on a gene or pathway, summarizing evidence for a candidate biomarker, comparing findings across studies, and providing traceable answers with source documents attached. That ability to trace an answer back to the original source is critically important in research settings. 

Question: What are the risks of using generative AI tools with sensitive research data? 

Answer:

There are several risks, and one of the most obvious is still hallucination. Even when a model sounds confident, it can be wrong. In research, that can mislead interpretation, introduce false leads, and waste valuable time and resources.

Another major risk is bias. A model can produce biased results if individuals trained it on data that overrepresents one racial/ethnic population and underrepresents others. In biomedical research, that is especially concerning because bias can affect how we characterize diseases, how we predict risk, and how we draw treatment-related conclusions. We must be very intentional about evaluating these systems across different populations.

A third concern is data exposure. Sensitive research data may include personally identifiable information, protected health information, unpublished findings, or proprietary data sets. If we enter  such data into an external platform without proper safeguards, they could be inadvertently exposed, retained, or used in ways the researcher did not intend. There is also the risk of malicious activity. Some adversaries may attempt to reverse-engineer a shared model or exploit vulnerabilities in a system to infer information that was previously used in training or prompting.

So, the risk profile is not just about accuracy. It is also about privacy, security, fairness, governance, and trust. That is why organizations working with biomedical data need strong guardrails around where researchers use these tools, what data are permissible, how people review the outputs, and whether the models operate in secure and compliant environments. Here at NCI CBIIT, my CGBB team mitigates risk first by being careful about the data we use. The data sets we work with do not include personally identifiable information or protected health information, which removes a major category of exposure risk. We use secure, approved computational environments; we limit the types of data that our AI tools can use; and we ensure that bioinformatics experts and domain scientists review our outputs before they influence research decisions. It is CGBB’s responsibility to help make these guardrails practical in day-to-day research. We support reproducible and auditable workflows, validate AI-assisted outputs, and evaluate models for potential bias—because even de-identified biomedical data can reflect gaps or inequities in studied populations.

Question: How is AI used in biomedical research?

Answer:

AI is operating across many areas of biomedical research, but one of the most exciting examples is the development of foundation models for digital pathology. Imagine a patient undergoes a biopsy, and that tissue is placed on a pathology slide. Traditionally, that slide is interpreted by highly trained pathologists. Today, with AI, we can digitize the slide and analyze it computationally against very large collections of previously annotated slides.

That opens up remarkable possibilities. By learning from vast numbers of examples, AI systems can detect visual patterns associated with different tumor types, subtypes, and outcomes. In research settings, these models can help classify cancer, distinguish among histologic features, and generate predictions related to prognosis or likely therapeutic response. Some of these capabilities genuinely feel like science fiction, but they are becoming increasingly practical.

What excites me most is the broader impact this could have. There are parts of the world where pathology slides may be available, but advanced sequencing technologies are not. AI-based slide analysis could help extend analytical capability to settings that do not yet have full access to sophisticated molecular infrastructure. That creates an opportunity to support global cancer research and, eventually, clinical decision support in more equitable ways.

My CGBB team and I are applying these AI capabilities to study cancers such as ovarian cancer, glioblastoma, and lung cancer. Our work draws on a range of publicly available foundation models for computational pathology—models such as UNI, CHIEF, TITAN, and Virchow—that have emerged from the research community to enable large-scale slide analysis. Using these models, we work to predict cancer type, estimate outcomes such as survival, and investigate associations with response to therapies. These systems are meant to support research and expert interpretation (not replace pathologists or clinicians), but they can significantly accelerate discovery.

Question: How is AI enabling multi-omics data integration? 

Answer:

This is one of the most important frontiers in cancer research. Cancer is not a single-modality disease. It is shaped by multiple interacting layers of biology—genomics, transcriptomics, proteomics, epigenomics, imaging, pathology, and clinical features. If we look at only one layer, we may miss critical signals. AI gives us a way to bring these different data types together into a unified analytical framework.

In our CGBB work, we may begin with a whole slide image from pathology and then layer it with genomic mutation data, gene expression data, and protein expression profiles. When these modalities are integrated effectively, the model can detect relationships that would be difficult to identify by analyzing each data set in isolation. This can improve the ability to predict tumor subtype, disease progression, survival, or drug response.

In some of our efforts, combining these modalities has improved predictive performance to around 80% accuracy in specific tasks. The exact performance depends on the disease, the quality of the data, the cohort size, and the question. But the overall principle is clear: a more complete biological picture usually leads to better-informed models.

This is why multi-omics integration matters so much. Cancer is biologically complex, and AI can help us model that complexity more realistically. Instead of asking what one data type says about a tumor, we can ask what the combined biological system reveals.

Have another question?

If we missed your question about generative AI in bioinformatics, email NCI CBIIT. We’ll connect you with a contact. If it is a commonly asked question, we’ll update the blog with an answer.

Author

Daoud Meerzaman, Ph.D.
Dr. Meerzaman is chief of NCI CBIIT’s Computational Genomics & Bioinformatics Branch. He and his team provide bioinformatics analysis support for life sciences and clinical and translational research for intramural scientists. This includes NCI’s Division of Cancer Epidemiology and Genetics and the Center for Cancer Research.