MyeGPT: an AI agent for Multiple Myeloma

Today, advancements in our understanding of cancer biology are increasingly attributed to large-scale clinical-molecular datasets. The case in point for multiple myeloma, the second-most prevalent haematological malignancy, is the CoMMpass study, a dataset with the paired clinical and sequencing data of 1,143 patients. Given its complexity, the multi-omics data of CoMMpass demands programming skills which imposes a hurdle for experimental myeloma researchers who want to validate their hypotheses on population data. The rise of agentic AI over the past few years presents unparalleled opportunities to bridge this technical gap. We propose MyeGPT (Myeloma Generative Pretrained Transformer), an AI bioinformatician for multiple myeloma that relies on the CoMMpass dataset as its ground truth. MyeGPT converts natural language queries such as 'What are the characteristics of patients who relapse after induction therapy' or 'Compare the overall survival of high vs normal NSD2 expression' into de novo analyses backed on real data, then pro-actively generates plots to visualize the results. We develop a set of evaluation questions based on CoMMpass, complete with scoring criteria, and ran benchmarks to identify the best choice for LLMs and text-embedding models. We package MyeGPT as a ready-to-use browser application, enabling CoMMpass-grounded hypothesis validation from a smartphone.