Can a large language model take molecular findings across DNA, RNA, and protein and help connect them into a coherent biological interpretation? Here, Jason Powers presents findings from a proof-of-concept study by IQVIA Laboratories using multiomics data from patients with multiple myeloma and healthy donors, as shared at the AMP 2026 Meeting & Expo.
The following transcript has been edited for clarity.
Hi, my name is Jason Powers and I'm director of Bioinformatics at IQVA Laboratories. I'm presenting on AI augmented integration of multiomic information to reveal actionable oncologic signatures.
This study brings together two related trends – the growth of big data and the emergence of AI – and asks how AI can help us interpret these increasingly complex biological data sets.
Today, genomics, proteomics, metabolomics, flow cytometry, and other technologies can each provide a different view of the underlying biology. Bioinformatics can identify variants or disruptive molecular pathways, but integrating all of those findings into a cohesive biological story remains challenging. So, IQVIA Laboratories asked a simple question: could large language models (LLMs) lower the barrier to an initial integrated interpretation of multiomic data?
For this proof of concept, we analyze commercially acquired multiple myoma and healthy donor samples using whole exome sequencing, whole transcriptome sequencing, and OLink proteomics.
Importantly, we use whole blood and plasma rather than bone marrow. So, we were looking for peripheral signals of multiple myoma and feeding our observations into the LLM rather than directly profiling the tumor. We identified a pathogenic bracket 2 variant and a low frequency DMT3A variant consistent, we think, with clonal hematopoesis or CHIP variant. We also observed disruptions in W and B cell receptor signaling pathways at the RNA level and cytokine and TNF alpha and unfolded protein response pathways at the protein level.
We then gave those findings to Co-Pilot with very little additional context. We didn't tell it that the samples were for multiple myoma. We simply asked it to interpret the findings from the perspective of a pathologist or an oncologist. And from those findings, Co-Pilot, in its own words, identified a plasma cell neoplasm most likely. It also identified relevant treatment considerations, appropriately cautioned that regimen selection requires full clinical context, and suggested additional follow-up testing.
These conclusions were consistent with our molecular pathologist interpretation, providing an encouraging proof of concept. So, ultimately, this was an experiment asking a fairly simple question. Can an LLM take findings across DNA, RNA, and protein and help connect them into a coherent biological interpretation? In this case, the answer was yes. And we at IQVIA Laboratories believe it provides an early glimpse of the role AI can play in helping us to interpret these increasingly large and complex biological data sets.
Thank you.
