Abstract 2430: Artificial intelligence enables prediction of MET amplification & associated morphologic features from H&E-stained NSCLC specimens

Abstract Background: MET receptor tyrosine kinase, when upregulated by mutation, fusion, or gene amplification, acts as an oncogenic driver in many cancers, including non-small cell lung cancer (NSCLC). Tumors with MET gene amplification (MET-amp) exhibit distinct clinical behavior. Thus, distinguishing MET-amp from non-MET-amp cases, typically determined by fluorescent in situ hybridization (FISH), is crucial. Here, we introduce an artificial intelligence (AI) model to predict MET-amp status directly from whole slide images (WSIs) of routine hematoxylin and eosin (H&E)-stained histologic sections and reveal elevated densities of highly pleomorphic (HP) tumor cells in MET-amp NSCLC. Methods: Additive multiple instance learning (aMIL) models were trained with 5-fold cross-validation to predict MET-amp, as determined by FISH, from H&E-stained WSIs from multiple cancers (Ntotal=1, 384, Namp=545) using embeddings from the pathology universal transformer (PLUTO) foundation model backbone1, 2. Model performance was assessed on a held-out test set of NSCLC H&E-stained WSIs (N=170) using the area under the receiver operating characteristic curve (AUROC) for FISH status comparison. PLUTO embeddings were also used to train a logistic regression classifier to identify HP tumor cells in WSIs. These embeddings were extracted from pathologist-annotated HP tumor cells (n=1, 725) and non-HP tumor cells (n=4, 872) in NSCLC WSIs. The trained model was deployed on embeddings extracted from all tumor cells in each WSI, and probability scores were converted to HP/non-HP cell predictions using a cutoff of 0.5. The proportion of predicted HP cells to total tumor cells was computed for each NSCLC WSI (Ntotal=262, Namp=131). Results: The aMIL MET-amp prediction model achieved an AUROC of 0.74 when deployed on the NSCLC test set. The HP cell classifier consistently and accurately identified HP tumor cells on a test set for this task (n=340 HP tumor cells and n=778 non-HP tumor cells), achieving sensitivity=98% and specificity=84%. When deployed on NSCLC H&E-stained WSIs, the model-predicted proportion of HP cells was significantly elevated in MET-amp cases (p=0.0004, Mann-Whitney U test). The proportion of HP cells was also predictive of MET-amp in cases with adenocarcinoma (AUROC=0.65). Conclusions: We demonstrate that an aMIL model using embeddings from a pathology foundation model accurately predicts MET-amp directly from H&E. MET-amp was further shown to be associated with an increased number of HP cells and is consistent with the report that MET-amp tumors are more likely to be poorly differentiated3. This work may enable a screening method to identify patients likely to have MET-amp from routinely collected H&E slides. 1. Juyal, D, et al. (2024) arXiv:2405.07905 2. Javed, SA et al. (2022) arXiv:2206.01794 3. Overbeck, Y, et al. (2020) Transl Lung Cancer Res. 9:603-16 Citation Format: Nhat Le, Ylaine Gerardin, Syed Ashar Javed, Jennifer Hipp, Jacqueline Brosnan-Cashman, Miles Markey, Lara Murray, Bahar Rahsepar, Aditee Shrotre, Adam Luo, William Wijaya, Dawn Spelke, Pok Fai Wong, Joshua Hernandez, Sheila Bheddah, Ramya Seshadri, Steven Chirieleison, Patrick Caplazi, Yan Sun, Athan Vasilopoulos, Amita Mistry, Yan Li, Francine Chen, Peter Ansell, Kenneth Emancipator, Kevin Kolahi. Artificial intelligence enables prediction of MET amplification & associated morphologic features from H&E-stained NSCLC specimens [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 2430.