Abstract 5401: Computational cancer cell gene expression deconvolution from tumor bulk RNA-seq via the machine learning algorithm Helenus

Abstract

Background: Biomarker gene expression is becoming more commonly utilized for clinical decision-making in oncology clinical practice. However, complex tumor tissue comprises a population of cancer cells (CC) and the tumor microenvironment (TME), causing expression signals belonging to the CC and TME calculated from bulk RNA-seq of the tumor tissue to be indistinguishable. To circumvent this, Helenus, a gene expression deconvolution tool, was developed to estimate TME-specific gene expression, consequently, providing precise CC-specific gene expression. Methods: Helenus performs the “subtraction” of TME gene expression from the total expression calculated from bulk RNA-seq of the tumor tissue. To accurately reconstruct the CC expression profile, LightGBM gene models were trained on artificial transcriptomes created from > 1,000 different solid tumor cancer cell lines and > 3,000 samples of various TME cellular proportions. The LighGBM gene models included genes expressed predominantly in the TME (e.g., CD3E), both the TME and the CC (e.g., BCL6), or in the CC (e.g., HER2). The input features included: 1) RNA percentages of TME cell types predicted by the cell deconvolution tool Kassandra (Zaitsev et al., 2022); 2) evaluation of TME target gene expression via the estimation of its weighted average expression profile in TME cell populations; and 3) a set of TME- and CC-specific genes. The resulting predictions were adjusted based on the CC cell fraction. To evaluate Helenus’ performance, CC and TME RNA were mixed at different ratios using various cancer cell lines and peripheral blood-derived TME cell populations and suspensions of tumor cells prepared from cancer tissue across multiple tumor purity dilutions. Results: Helenus deconvolution resulted in an increased concordance correlation value from 0.73 to 0.98 between the real gene expression profile of pure CC and the reconstructed CC expression from bulk RNA-seq. Helenus showed high concordance between the gene expression profile of sorted cancer cell lines and the deconvolved gene expression across a wide range of CC RNA concentrations (20-90%) mixed with imitated TME RNA at varying concentrations. Helenus demonstrated high performance calculating gene expression of multiple clinically relevant biomarkers in the TME:cancer cell line mixes: CD274 (PD-L1) (mean absolute error [MAE] ~3.5-fold reduction); HLA-A (~2.8-fold MAE reduction); MKI67 (Ki-67, ~2.2-fold MAE reduction), ERBB2 (HER2, ~1.7-fold MAE reduction). Helenus deconvolved CC expression and found significant correlations with CC gene amplifications and deletions (e.g., BCL-2, VNN3) independent of tumor purity (p < 0.003). Conclusion: Helenus, the CC gene expression deconvolution tool, was developed with high accuracy to contribute to tumor diagnosis, disease monitoring, treatment decisions, and clinically relevant biomarker identification. Citation Format: Valentina Beliaeva, Ekaterina Ivleva, Boris Shpak, Daniil Litvinov, Anastasia Zotova, Krystle Nomie, Daniiar Dyikanov, Alexander Kuznetsov, Maria Savchenko, Aleksandr Zaitsev, Nathan Fowler, Alexander Bagaev. Computational cancer cell gene expression deconvolution from tumor bulk RNA-seq via the machine learning algorithm Helenus. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5401.