A ReAct Agentic AI System for Natural Language Querying and Statistical Analysis of The Cancer Genome Atlas Clinical Data

Hospitals and research centers generate vast amounts of information about patients, from their medical history to the specific genetic makeup of their tumors. One of the most significant collections of this data is the Cancer Genome Atlas, a massive digital library that holds clinical records for more than 11,000 patients across 33 different types of cancer. For decades, scientists have relied on this resource to find patterns that could lead to better treatments, but accessing the information has been a difficult task. The data is stored in complex file structures and scattered across many different formats, requiring researchers to write specialized computer code just to ask simple questions. This barrier means that valuable insights often remain locked away, accessible only to those with advanced programming skills, while doctors and clinicians who could use the data immediately are left without a direct way to query it.

To solve this problem, researchers have developed a new type of computer system designed to understand natural language and perform statistical analysis on this medical data. Instead of forcing a human to learn a complex programming language, this system allows a user to simply ask a question in plain English, such as inquiring about the survival rates of patients with a specific treatment. The system acts as an autonomous agent, a digital assistant that can think through a problem step-by-step. When a question arrives, the agent does not guess the answer; instead, it uses a large language model to reason about what tools it needs to find the truth. It selects from a set of eight different computational tools, which might include functions to extract specific data points, calculate averages, or compare groups of patients to see if a difference is statistically significant.

The process works by having the agent break a complex question into smaller steps. If a user asks about the survival time of patients with a certain tumor, the agent might first retrieve the relevant patient records, then calculate how long they lived, and finally run a statistical test to see if the result is reliable. Crucially, the system checks its own work. It compares its findings against a curated, trusted version of the data known as the Clinical Data Resource to ensure accuracy. If the initial attempt fails or if the data looks incomplete, the agent adapts its approach, trying a different tool or a different way of looking at the numbers until it arrives at a correct, clinically meaningful answer. This cycle of reasoning, acting, and verifying allows the system to handle difficult questions that would stump a standard computer program or a simple chatbot.

To test how well this system works, the researchers created a rigorous evaluation called TCGA-Agent-Bench. This test included 440 different questions ranging from simple lookups of a single patient's record to complex comparisons between large groups of people. The questions were graded on five levels of difficulty, and the answers were checked against the trusted Clinical Data Resource to see if the system got the numbers right and if the medical context was complete. The results showed that the agentic system was highly effective, answering 93.4% of the questions correctly. It performed perfectly on simple lookups and achieved a 99.1% success rate on questions about groups of patients. When compared to other methods, such as a fixed set of rules or a standard language model that does not use tools, this new system was significantly more accurate. A standard model that simply reads the data without reasoning through the steps managed only 81.8% accuracy, while a system that tries to retrieve information without a structured reasoning loop dropped to 66.9%.

The study also revealed where the system adds the most value. While the trusted Clinical Data Resource alone could answer most of the questions, there were specific areas where it fell short, such as details on drug treatments, specific tumor measurements, and biological markers. When the researchers tested the system on 26 questions that required these missing details, the full agentic system answered 100% of them correctly, whereas relying on the trusted resource alone would have yielded only 3.8% correct answers. This demonstrates that the system's ability to pull data from the raw, complex files is essential for a complete picture. Further analysis showed that the reasoning loop—the part where the agent thinks about its next step and checks its work—was the most important factor, boosting accuracy by 9.1% and improving the completeness of the answers by 22 percentage points.

The findings suggest that the power of this system comes not from the size of the computer model itself, but from how it is built to use tools and verify its own results. By combining the ability to understand human language with the precision of statistical software and the discipline of self-checking, the system makes a vast, difficult-to-use medical database accessible to anyone who can ask a question. It provides a way to get accurate, auditable answers with clear sources, turning a library of complex files into a conversational partner for medical discovery. This approach offers a path forward for analyzing clinical data, proving that the right architecture can unlock insights that were previously hidden behind a wall of technical complexity.