From Data to Discovery: AI’s Expanding Role in Cancer Research

As technology advances, so does scientific progress. Artificial intelligence (AI) methods have rapidly progressed beyond data classification and prediction. These new capabilities make the use of AI tools a promising way to reshape cancer research and precision oncology.

In a plenary session at the American Association for Cancer Research (AACR) 2026 annual meeting, researchers shared insights into foundation models, which can analyze and train on large datasets and power chatbots, AI agents, and explored the benefits and challenges of bringing AI tools into the clinic.

AI Tools for Transforming the Discovery Process

Two decades ago, early classical machine learning methods enabled researchers to analyze structured, tabular data; however, more complex data necessitated different models—deep learning could extract such patterns. Foundation models, AI systems that are trained on massive datasets, represent the next step in this evolution as they can be repurposed for diverse applications, such as predicting protein structure, mutations, and gene regulation.

Foundational Models for Understanding Cells, DNA, and Protein Function

This motivated Bo Wang, a computer scientist at the University of Toronto, to build a foundation model for single-cell research using a generative pre-trained transformer (GPT) architecture across a repository of over 33 million cells. In 2024, he and his colleagues revealed scGPT, which could be used to pull biological insights concerning cells and genes, including cell-type annotation, genetic perturbation prediction, gene network inference, and more.¹

Building upon this use of descriptive datasets, Wang turned his focus to building other models to learn from massive causal datasets and predict non-scalable models. In March 2026, he and his team released a bioRxiv paper on X-Cell, a foundation model that predicts how gene expression changes under perturbations.² The dataset used to train X-Cell consisted of almost 26 million perturbed single-cell transcriptomes across seven screens. “This is probably the largest high-quality genome-wide perturbation datasets out there,” explained Wang.

In addition to X-Cell, Wang highlighted the next level of understanding what specific biology the foundation models teach—this requires reasoning. They developed BioReason, which paired a foundation model for DNA sequences and a reasoning language model, to understand biological sequences and what they mean.³ They tested BioReason on disease pathway predictions, which got better and better at understanding biology.

But what about protein function? While other models like AlphaFold solve the problem of protein structure, Wang wanted to extend the capability of these models to also help biologists better annotate protein function. Through similar iterations of assigning tokens for protein sequences, gene ontology terms, and language, BioReason-Pro could predict protein function.⁴ They then put it to the test against 27 invited protein experts and biologists to evaluate how good the model’s prediction was against human annotation. Not only was BioReason-Pro faster in generating predictions, but Wang also emphasized that “80 percent of the time, we actually find that the prediction from BioReason-Pro is actually higher quality than human annotators.”

The Development of AI Agents

While foundational models can power large language models like ChatGPT, Gemini, and Claude, their usefulness in biomedicine remains limited. They rely heavily on user prompts and lack true autonomy, making them less suited to the iterative and complex nature of scientific research.

Jure Leskovec, a computer scientist at Stanford University, has been exploring how to bridge this gap by developing copilots and co-scientists that “allow us to do science better, faster, and more reliably.” He pointed out that the current workflow of researchers is fragmented, where scientists read papers, collect data, run experiments, order lab materials, analyze the data, and draft a protocol. This pace is slow and creates a bottleneck. Thus, AI agents built on foundation models offer a way to streamline this complexity.

In the summer of 2025, Leskovec and his colleagues introduced Biomni, an open-source, general-purpose biomedical agent.⁵ “Now we have…an integrated biology environment, which is basically this user interface in which a scientist can operate, upload the data, talk to the [AI] agent, brainstorm to the agent, and the agent can then take actions, visualize the data, and be a partner in the scientific workflow,” he explained. Through this method, the human scientists remain “at the steering wheel,” with the AI serving as a collaborative partner.

Biomni can scan scientific literature to identify relevant databases, tools, and software created by other researchers, expanding what the scientist can readily access. At the same time, acknowledging the limitations of large language models—such as their tendency to produce unreliable or fabricated outputs—Leskovec emphasized features like asking clarifying questions, supporting brainstorming, and providing transparent self-evaluations. These iterations enable the agent to adapt to a researcher’s preferences over time.

Leskovec added that the system had already proven useful across multiple applications, from designing cloning protocols for plasmids to generating hypotheses from large datasets. In addition, he noted that the AI agent works well alongside robotic or robot-operated wet labs. “I want to convey here that the future of science is agenic, and that virtual AI scientists are going to help us speed up biomedical research and help scientists in everyday research,” said Leskovec.

AI Models for Clinical Diagnosis and its Scalability

Building Foundational Models for Improved Cancer Diagnosis and Treatment Prediction

Faisal Mahmood, a computer scientist at Harvard Medical School, works on developing foundation models that integrate histopathology, genomics, radiology, and clinical data to improve cancer diagnosis and treatment prediction.

He remarked that one needed about 10,000 whole slide images from pathology to train an AI model for useful clinical performance. This led him to find a more data-efficient way to do so. In one instance, Mahmood and his colleagues applied this method to cancers with an unknown primary site to see if they could predict the origins of the tumors from a histological image alone.

Mahmood and his team have developed different models: TITAN, for representing the entirety of a whole slide image into a singular representation; THREADS, for contrasting hematoxylin and eosin images with the corresponding genomic and molecular data.^6,7 By being able to parse patient data into a singular vector, Mahmood explained, “that can then be used for a lot of different downstream applications.” These applications can be predictive and be related to diagnosis, prognosis, and treatment response.

The Successes and Challenges of AI Implementation in the Clinic

The next challenge is to figure out how to take these technological advancements and deliver them to patients in the clinic. Suchi Saria, a computer scientist at Johns Hopkins University, described treatment as reactive and random, where a clinician must do their best to quickly gather information about the patient’s history and provide care. This can result in a lot of variability in the care that clinicians give. According to Saria, proactive, evidence-based treatment is a solution, and AI can make the delivery of this possible.

Ideally, real-time clinical AI tools can augment the care team’s effort, pulling from current and past encounters and synthesizing the information in a way that can attract a clinician’s attention to something they may have missed. However, Saria remarked that the current AI tools are not built for this. One example is the alerting system for sepsis in hospitals; the old-school rules-based or analytic alerting systems often miss the signs or flag sepsis after treatment has already started, so there is very little clinical adoption.

How can researchers build better AI tools that clinicians will adopt? Saria pointed out that there are several different hurdles to AI adoption: instances where patient symptoms exhibit heterogeneity and the data needs to encompass all of the variations, as well as ensuring that the information is being processed in real-time and relayed back to the clinician.

Circling back to her example of sepsis, Saria described implementing an AI system in the Cleveland Clinic to enable early identification, treatment, documentation, and abstraction. This type of architecture enabled the system to route each patient sample to a specialized model that could best interpret the data signal.

When implemented, the team saw a vast improvement of almost 60 percent over the existing tool and enabled faster identification. “Patients are getting treatment earlier, and simultaneously what we’re observing is shorter length of stays and reduced mortality. So, this is extremely exciting, and the ability to be able to deploy these tools to…drive action is ultimately what impacts patient outcomes.” Saria added that she hopes this work can bring value to diverse populations and settings.

As scientists build AI tools to better capture cancer biology, their potential spans the full research pipeline from early hypothesis generation to precision oncology. With continued validation and clinical integration, these advances could help redefine the future of cancer research and treatment.