Abstract 2384: Improved non-invasive gene expression inference using advanced machine learning

Abstract Background: Gene expression profiling is crucial for understanding cancer dynamics and development, aiding in the identification of molecular subtypes, predicting patient outcomes, and discovering therapeutic biomarkers. Traditional profiling methods rely on invasive tissue specimens analyzed through RNA-seq or microarrays. These approaches can delay diagnosis due to scheduling challenges, pose risks of complications during and after surgery, and may be inaccessible to high-risk or frail patients. Recently, non-invasive gene expression inference from blood plasma has emerged as a promising alternative that analyzes fragmentation patterns of cell-free DNA molecules to infer gene expression (EPIC-Seq, Esfahani, Nat Biotechnol 2022). Further, we demonstrated that EPIC-Seq data can be effectively applied to lymphoma signatures derived from RNA-seq, yielding robust results (Mutter, ASH 2023). However, despite these advances, EPIC-Seq results cannot be directly compared to RNA-seq profiles on a single gene level. To address this challenge, we built a machine learning model that maps EPIC-Seq data into an RNA-seq-compatible format. Methods: We developed an advanced, generative AI framework to learn mappings between non-invasive EPIC-Seq from blood plasma and tumor RNA-seq from invasive tissue biopsies. The model leverages advanced AI techniques to identify patterns in the EPIC-Seq domain and translate them into equivalent representations in the tumor RNA-seq domain and is designed handle complex, high-dimensional data while preserving key biological and functional features. For this, we sequenced a total of 304 formalin-fixed paraffin embedded (FFPE) and blood plasma samples from 227 diffuse large B-cell lymphoma (DLBCL) (n=110) and mantle-cell lymphoma (MCL) (n=117) patients using both paired (DLBCL: n=33; MCL: n=44) and unpaired data (DLBCL: n=52 FFPE, 25 plasma; MCL: n=56 FFPE, 17 plasma). Results: Using a combination of paired and unpaired data we were able to robustly predict gene expression in RNA-seq space from non-invasive EPIC-Seq data. By splitting the cohorts in 80% training and 20% validation (n=88 patients, n=115 samples), we observed a high concordance of predicted gene expression and RNA-seq data in DLBCL samples (R=0.91). Similar accuracy was achieved in MCL (n=93 patients, n=129 samples) further confirming the model performance (R=0.94). In addition, applying the model trained on DLBCL data yielded high concordances in MCL data (R=0.68), emphasizing the robustness and applicability of the model to different lymphoma subtypes. Consistently, the model trained on MCL data achieved high concordance on DLBCL data (R=0.65). Conclusion: These findings suggest that EPIC-Seq can serve as a reliable surrogate for RNA-seq data, offering comparable gene expression values for enhanced diagnosis and personalized treatment strategies in lymphoma patients. Citation Format: Jurik A. Mutter, Amirsaman Ashtari, Nick Phillips, Benoit Tessoulin, Jan Boegeholz, Cedric Rossi, Steven Le Gouill, Maximilian Diehn, Mohammad Shahrokh Esfahani, Ash A. Alizadeh. Improved non-invasive gene expression inference using advanced machine learning [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 2384.