Back

Auto-MedCalc: Automated Biomarkers Discovery and Risk Score Generation with AI Agents

Ding, S.; Bhattacharya, S.; Butte, A.

2025-07-16 bioinformatics
10.1101/2025.07.10.664265 bioRxiv
Show abstract

Identifying biomarkers and generating risk scores are usually essential tasks in many biomedicine and clinical scenarios. However, this is a highly hypothesis-driven and experience-dependent process requiring extensive experiments as well. How to discover biomarkers from multi-modality, multi-source data and generate risk scores with higher precision motivates us to design the framework Auto-MedCalc, a data-driven pipeline to automatically identify biomarkers and generate risk scores for diagnosis. Auto-MedCalc is a multi-agent AI system with large language models and computational tools. Three representative studies of transplantation from the public ImmPort data portal are used for the experiment. Auto-MedCalc achieved 0.93, 0.88, and 0.88 ROC-AUC on the rejection prediction after kidney, liver, and heart transplants, surpassing the best expert-designed medical calculators by 50%, 4.7%, and 39%, respectively. Auto-MedCalc also validated the previously found human biomarkers and discovered new biomarkers, {beta}-Glucuronidase and Alpha-1-microglobulin, in heart transplants, which hadnt attracted much attention before. The superior performance from experiments shows that Auto-MedCalc can be used for both clinical applications and scientific discovery.

Matching journals

The top 10 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.