Teaching AI to Read Liquid Biopsies

Researchers explore whether large language models can support cfRNA biomarker discovery

June 26, 2026
2 min

The Pathologist

Overview

A study published in Nature Communications evaluates the potential of large language models (LLMs) to identify diagnostic biomarkers from cell-free RNA (cfRNA) data. LLMs showed variability in performance compared to traditional methods, indicating the necessity for further validation before clinical application.

Background

Liquid biopsies, particularly those analyzing cfRNA, offer a minimally invasive approach to diagnostic biomarker discovery. The complexity of cfRNA datasets presents challenges in identifying clinically useful biomarkers. The integration of artificial intelligence, specifically LLMs, may enhance the biomarker discovery process, but their reliability compared to established methods is not fully established.

Data Highlights

No numerical data provided in the source material.

Key Findings

LLMs outperformed random gene selections in identifying diagnostic biomarkers from cfRNA data.
Performance was strongest in the tuberculosis dataset, with some LLM-generated panels comparable to traditional methods.
Results were more modest in Kawasaki disease and multisystem inflammatory syndrome in children (MIS-C), and weakest in myalgic encephalomyelitis/chronic fatigue syndrome (ME/CFS).
LLMs frequently selected genes associated with immune and inflammatory pathways.
Inconsistent performance in independent biomarker discovery workflows was noted, not surpassing established machine learning approaches.
Rigorous validation of LLM-generated biomarker signatures is necessary before clinical application.

Clinical Implications

The findings indicate that LLMs can assist in generating hypotheses for biomarker discovery, but should be used alongside traditional bioinformatics and statistical methods.