Abstract 896: Predicting metastatic transcriptomes of patient tumors with deep learning

Abstract

Abstract Metastases are the primary cause of cancer-related death, and improving the means of predicting and targeting their development is one of the major goals in cancer research. While surgical resection and neo-adjuvant therapy can cure well-confined primary tumors, our ability to effectively treat cancer is largely dependent on our capacity to interdict the process of metastasis. The recent accumulation of ‘omics data from metastatic tumors provides an unprecedented opportunity to develop machine learning models to predict the molecular changes during metastasis and explore the patterns of metastasis formation. With this in mind, we developed MetMapper, a deep learning model trained on primary and metastatic tumors from > 13 000 patients integrating data from 11 published data resources. MetMapper can predict the transcriptomic changes of a primary tumor when it metastasizes to different distant organs. The results were extensively validated using transcriptomics data from matched primary and metastatic tumor biopsies extracted from the same patients. Furthermore, MetMapper’s predictions revealed that the non-random patterns of cancer metastases can be partly explained by the degree of transcriptome reprogramming needed during metastasis: primary tumors tend to metastasize to organs that require minimal changes to their transcriptomes. Using MetMapper, we derived a metastatic potential score for patient tumors and demonstrate that this score can be used to stratify patients into high and low survival groups across different indications. The predicted metastatic potential of patient tumors significantly correlates with experimentally characterized metastatic potential of cancer cell lines. Additionally, by performing in-silico perturbations of genes and oncogenic pathways that can alter the metastatic potential of patient tumors, we identified genomic features that are highly associated with metastases to specific organs, some of which were reported by existing pan-cancer clinical sequencing studies. Our results demonstrate the utility of MetMapper as a novel AI-powered methodology for investigating mechanisms and patterns of metastatic dissemination, as well as forecasting metastatic outcomes of patient tumors. Citation Format: Gang Li, Evan Béal, Dean Sumner, Giorgio G. Galli, Viviana Cremasco, Joshua M. Korn, Frank Dondelinger, David Ruddy, Audrey Kauffmann, Slavica Dimitrieva. Predicting metastatic transcriptomes of patient tumors with deep learning [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 896.