Abstract 2452: Path2Omics: An AI model integrating frozen slides to enhance transcriptomic predictions from tumor pathology images
Danh-Tai Hoang, Eldad David Shulman, Saugato Rahman Dhruba, Ranjan Kumar Barman, H. Lalchungnunga, Omkar Singh, MacLean P. Nasrallah, Eric A. Stone, Kenneth Aldape, Eytan Ruppin · Cancer Research · 2025
Abstract Background: Artificial intelligence (AI) has shown promise in predicting transcriptomics from tumor pathology slides. However, previous studies have relied on Formalin-Fixed Paraffin-Embedded (FFPE) slides, the clinical standard. No study has incorporated Fresh Frozen (FF) slides in model training for predicting gene expression. Methods: Here, we introduce Path2Omics, an AI framework that integrates both FFPE and FF slides to predict gene expression across 23 TCGA cancer types. For each cancer type, we trained two separate models, an FFPE model based on FFPE slides and an FF model based on FF slides. Our integrated model averages predictions from each component. Results: In five-fold cross-validation within the TCGA cohort, FFPE models achieved an average of 3, 503 well-predicted genes (defined as those with a correlation between predicted and actual values across samples above 0.4). The FF models, however, achieved double the predicted gene coverage, with an average of 7, 299 well-predicted genes. To evaluate cross-slide generalizability, we applied the FF model to FFPE slides, yielding an average of 3, 973 well-predicted genes. Conversely, the FFPE model applied to FF slides achieved 2, 786 well-predicted genes, 30% lower than the FF model’s performance on FFPE slides. In seven external test datasets, the integrated model averaged 4, 391 well-predicted genes, a 30% improvement over the traditional FFPE model alone. Using the predicted gene expression to predict treatment response in two TransNeo breast cancer cohorts - one with HER2-negative patients treated with chemotherapy (93 patients) and another with HER2-positive patients treated with chemotherapy plus trastuzumab (61 patients) - our models achieved an AUC of 0.82, accuracy of 0.76 and odds ratio of 7.6 in predicting patient response to chemotherapy, and an AUC of 0.68, accuracy of 0.74, and odds ratio of 5.4 in predicting response to trastuzumab. These results were comparable to those obtained using measured gene expression and outperformed the “direct” modes that used slides without predicted gene expression as an intermediate. Conclusions: Path2Omics is the first deep learning model integrating both FFPE and FF slides for training to predict transcriptomics. The integrated model remarkably improves on the performance of the model trained on FFPE slides, even when the test set has only FFPE slides. Our results show that the inferred gene expressions are as effective as the measured gene expressions in predicting treatment response, highlighting Path2Omics’ potential for clinical applications. Citation Format: Danh-Tai Hoang, Eldad D. Shulman, Saugato Rahman Dhruba, Ranjan Barman, H. Lalchungnunga, Omkar Singh, MacLean P. Nasrallah, Eric A. Stone, Kenneth Aldape, Eytan Ruppin. Path2Omics: An AI model integrating frozen slides to enhance transcriptomic predictions from tumor pathology images [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 2452.