Abstract 2425: AI-powered approaches accurately predict t(11;14) positive multiple myeloma from H&E-stained histologic sections by identifying regions demonstrating lymphoplasmacytic cytology

Abstract Venetoclax, a highly selective oral BCL-2 inhibitor, has demonstrated efficacy in patients with t(11;14) multiple myeloma (MM). Fluorescence in situ hybridization (FISH) is the mainstay methodology to determine t(11;14) status, but requires significant quantities of bone marrow tissue or aspirate and is low-throughput. Herein, we: (i) describe an artificial intelligence (AI)-based model that accurately predicts t(11;14) status from routine H&E-stained MM whole slide images (WSIs) and (ii) demonstrate that the biological signal underlying the model predictions is driven by regions containing the lymphoplasmacytic phenotype of t(11;14) positive MM cells. H&E-stained MM WSIs (n=231) with known t(11;14) status were split into training/validation/test sets (60/20/20), and CNN-based tissue and cell segmentation models were trained from expert pathologist annotations to identify high density regions of MM cells. Additive multiple instance learning (aMIL) models with 5-fold cross validation were then trained using embeddings from pathology universal transformer (PLUTO) foundation model backbone [1], to predict t(11;14) status. To interpret the biological signal underlying aMIL model predictions, nuclear features [2] for MM cells and sparse autoencoder (SAE) dimensions [3] (based on PLUTO embeddings) were extracted from the model-predicted t(11;14) positive and negative regions. Mann-Whitney U test was used to compare distribution of nuclear features between t(11;14) positive and negative regions. Activation frequency (AF) delta (AF in positive images - AF in negative images) was used to find the SAE dimensions most active in t(11;14) positive regions. Validation and held-out test sets were leveraged to evaluate the aMIL model’s performance for predicting t(11;14) status, resulting in AUROC of 0.813 and 0.849, respectively. Upon investigating the regions predicted as t(11;14) positive, analyses revealed that these contained MM cell nuclei with higher circularity, solidity and less variability in shape and size (p-values<0.001 for all comparisons) compared to those predicted as t(11;14) negative. Additionally, SAE dimensions capturing circular cells (SAE-2202) and lymphocytes (SAE-1567, SAE-1355) were significantly more active in model-predicted t(11;14) positive images compared to the predicted negative images (AF delta- SAE-2202: 0.73, SAE-1567: 0.61, SAE-1355: 0.57). t(11;14) status can be predicted with high accuracy using routine H&E WSI. Model predictions rely on biologically meaningful signals based on regions containing MM cells with lymphoplasmacytic morphology. This technology has the potential to facilitate screening for t(11;14)-directed MM clinical trials. References [1] arXiv:2405.07905 (2024). [2] npj Precis. Onc. 8, 134 (2024). [3] arXiv:2407.10785 (2024). Citation Format: Neel Patel, Raymond Biju, Syed Ashar Javed, Sandrine Degryse, Lara Murray, Xiao Zhou, Jeremy A. Ross, William Wijaya, Pedro Munoz, Aditee Shrotre, Joann Palma, Patrick Calpazi, Yan Li, Pok Fai Wong, Francine Chen, Kenneth Emancipator, Kevin Kolahi. AI-powered approaches accurately predict t(11;14) positive multiple myeloma from H&E-stained histologic sections by identifying regions demonstrating lymphoplasmacytic cytology [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 2425.