So hello, my name is Jonathan Lin, and I'm currently a second year vitreoretinal surgery fellow at Stanford University. I'm excited to share with you some research that I've been working on alongside my co-first author, Dr. Samson Maderaso, under the mentorship of Dr. Vinit Mahajan and Dr. Nima Gapor that we presented this past week at ARVO 2026. In the modern era, omics have been increasingly used to advance our understanding of human disease. These techniques are extremely powerful because they enable us to identify the tens to hundreds of genes, gene transcripts, proteins, metabolites, and other analytes that are perturbed in disease states. Nowadays, these techniques have become increasingly sensitive and advanced to the point that we can now detect levels of thousands of proteins from a human liquid biopsy as small as 50 to 100 microliters in volume. One challenge though that has emerged is a difficulty in identifying which of these many proteins are the most clinically significant or targetable from a therapeutic standpoint.

As such, we're often forced to rely on relatively arbitrary full change or false detection rate cutoffs, which is a rather unsophisticated approach. In our study, we sought to determine whether incorporating large scale electronic health record or EHR data into proteomics analysis may enhance our ability to rationally prioritize protein biomarkers that may be the most clinically relevant. We chose to focus on diabetic retinopathy and there's a need for novel therapies for diabetic retinopathy because existing therapies treat only certain features of diabetic retinopathy and do not work optimally for all patients. To perform this study, we used a large database of EHR data from nearly 320,000 patients being treated at Stanford. From 101 of these patients, we obtained an aqueous humor liquid biopsy at the time of plain ocular surgery and performed high resolution proteomic profiling using a platform called SOMA Scan.

By jointly analyzing EHR data with proteomics data using a multimodal AI approach, we were able to rationally prioritize proteins that were most linked to the EHR elements as well as identify proteins that are linked with specific disease features such as diabetic macular edema. And we were able to validate these findings using an independent cohort. Taken together, our study highlights the power of using multimodal AI to perform integrative analysis of EHR and proteomics data for enhancing proteomics analysis. Thank you for your attention.