Abstract 7210: Predicting drug response heterogeneity in AML: novel AI-enabled proteomics model for a prospective clinical trial

Abstract Acute myeloid leukemia (AML) is an aggressive blood cancer with a significant chance of relapse and only a 17% 5-year survival rate for patients refractory to first-line therapy. Current therapies, even when combined with newer agents, have not significantly enhanced long-term survival for many AML patients. One major barrier in the development of effective therapies for AML and their successful deployment for patients, has been the lack of accurate model systems to identify and predict therapeutic efficacy in patient subsets. Current AML cell lines do not predict drug sensitivity because the cell lines harbor strong driver alterations that occur in only a small fraction of patients and their biology has drifted from that of an AML patient. With state-of-the-art culture conditions, we have created over 50 patient-derived hematopoietic stem/progenitor cell (HSPC) models and have measured cellular response across 40 AML therapeutics. Leveraging our pipeline for LC-MS/MS and innovative data integration, we have created a patient landscape with embedded cell models for prospective therapeutic positioning. HSPCs were evaluated using a panel of cell surface markers, and preliminary results show marginal drift from the original patient samples. Unbiased proteomics were generated with peripheral blood samples from 100 AML patients and all ex-vivo models. Pathway overrepresentation and correlation analysis are used to confirm that patient-derived HSPCs are closer in biological context to patient subpopulations than commercially available cell lines. Additionally, global proteomics data from over 2500 samples from 11 public and 5 private datasets across 8 different cancer indications were collected to perform a molecular functional analysis to guide a transfer learning model for matching patients to ex vivo HSPCs. A custom deep learning architecture was built to (1) define global cancer mechanism enrichments of the AML samples and (2) integrate paired ex vivo treatment response. To drive a prospective clinical trial at FHCC, our model was used to select a minimal set of proteins that can accurately separate patient subsets for which there are differential sensitivity profiles across 4 AML therapeutics. For this trial, 40 patients with relapsed/refractory AML or other high-grade myeloid neoplasms will be enrolled and randomized to one of two arms. Arm A: receive standard of care hypomethylating agent therapy (20 patients). Arm B: undergo an ex vivo MS-based targeted assay on the peripheral blood to predict response to various treatment regimens; if a drug is identified through the assay, patients will receive that treatment (20 patients). The primary objective of the trial is to assess the feasibility of a single MS-based assay for the selection of the ideal AML therapeutic for each patient to be used in combination with azacitidine to achieve response, extend survival, and improve quality of life. Citation Format: Pilgrim Jackson, James T. Sorrentino, Mahan Abbasian, Antonious Koller, Salvador Martinez-Bartolome, Derek L. Stirewalt, Mary-Elizabeth Percival. Predicting drug response heterogeneity in AML: novel AI-enabled proteomics model for a prospective clinical trial [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 7210.