Context-aware LLM extraction and controlled-vocabulary normalization of GEO transcriptomic sample metadata

The vast archives of modern biology contain millions of snapshots of life at the molecular level, capturing how genes turn on and off in cells from healthy people and those with disease. These records, stored in public repositories like the Gene Expression Omnibus, are meant to be a shared library where scientists can search for patterns across thousands of studies. However, the usefulness of this library has long been held back by a simple problem: the notes attached to each sample are written in a chaotic mix of human languages. One researcher might describe a sample as "breast tumor," while another writes "mammary carcinoma," and a third uses a cryptic abbreviation that only makes sense within their specific lab. Because these descriptions are unstructured and inconsistent, finding all the samples related to a specific disease or tissue type is like trying to find a single book in a library where every title is written in a different dialect and scattered across the shelves.

To solve this, Mateusz Szczepaniak and Jonathan Wren at the Oklahoma Medical Research Foundation have built a new system that acts as a universal translator for these biological records. They developed a locally running artificial intelligence pipeline designed to read the messy, free-text notes attached to over 800,000 human samples and convert them into clean, standardized labels. The system focuses on three critical pieces of information: the tissue the sample came from, the condition of the patient or organism, and any treatments the sample received. Instead of just reading the text, the system understands the context of the entire study, allowing it to fill in missing details that a single note might leave out. Once it extracts this information, it links every label to a controlled vocabulary, ensuring that "breast cancer" and "mammary tumor" are recognized as the same thing, while also grouping together terms that do not yet have a standard name.

The researchers tested their system on a massive dataset of human samples and found that it could successfully identify the tissue type in nearly every case where the information was available, achieving an accuracy of nearly 99.7 percent on a carefully checked benchmark. For identifying the disease or condition, the system was correct about 95 percent of the time. Perhaps most importantly, the system did not just guess; it used the broader context of the study to resolve ambiguities. If a single sample note said "control group" without specifying the disease, the system looked at the study's overall design and the labels of the other samples in that same experiment to correctly infer the missing context. This step allowed the system to recover missing information for more than three-quarters of the tissue labels that were initially marked as unspecified.

A key feature of this work is how it handles the terms that do not fit into existing medical dictionaries. Rather than forcing a strange or unique term into a category where it does not belong, the system groups these out-of-vocabulary concepts together based on their similarity. This preserves the specific details of the original research while still making the data searchable. The team also ensured that the system could run on local computers without needing to send sensitive data to external servers, making the process transparent and reproducible. By converting millions of unstructured notes into structured, searchable data, this pipeline removes a major barrier to scientific discovery. It allows researchers to easily combine data from different studies to find new insights into how diseases work, turning a fragmented collection of notes into a coherent, usable resource for the entire scientific community.