Distinguishing between cancer-related mutations and clonal hematopoiesis using cell-free RNA (cfRNA) expression levels in a machine learning model.

Abstract

3043 Background: Liquid biopsy is currently considered an important part of the clinical practice of oncology. However, proper interpretation of somatic mutations detected in liquid biopsy remains a challenge. Distinguishing cancer-related mutations from mutations resulting from clonal hematopoiesis of indeterminate potential (CHIP) can be difficult and a source of misinterpretation of liquid biopsy findings. We explored using cell-free RNA (cfRNA) profiling as a means for distinguishing between mutations detected as CHIP vs mutations detected as cancer. Methods: cfDNA and cfRNA from 102 patients with confirmed solid tumors, 93 patients with hematologic neoplasms and 40 patients with CHIP abnormalities were sequenced using a panel of 1501 gene for RNA and 284 genes for DNA. The solid tumors included lung, breast, ovarian, and colorectal. The hematologic neoplasms included lymphoid and myeloid neoplasms. Using a machine learning algorithm, we first selected the relative genes that distinguish between two classes using K-fold cross-validation (K=12). The selected genes were used to predict one class from the other using naïve Bayesian classifier, but we applied geometric mean naïve Bayesian (GMNB) to compensate for the expected underflow. Results: Using machine learning with cfRNA, we were able to distinguish between cases with confirmed cancer-related mutations and CHIP mutations with area under the curve (AUC) of 0.808 (95% CI: 0.720-0.895). Precision, recall and f1-score were 0.86, 0.422 and 0.566 for CHIP and 0.359, 0.825 and 0.500 for solid tumor, respectively. The machine learning required 70 genes for this classification. Leave-one-out (LOO) validation showed AUC of 0.766 (95% CI: 0.671-0.859). Distinguishing between hematologic neoplasms and CHIP was achievable using the expression of 30 genes. The AUC for this model was 0.787 (95% CI: 0.696-0.879). Precision, recall and f1-score were 0.906, 0.516, and 0.658 for CHIP and 0.438, 0.875, and 0.583 for hematologic neoplasms, respectively. Validation using LOO showed AUC of 0.736 (95% CI : 0.638-0.835). The top 10 genes that their expression was critical for distinguishing between CHIP and solid tumors were EIF4E, HNF1A, HOOK3, HIPK2, IL1RAP, BRAF, DNAJB1, GADD45B, H2AX, and HDAC6. The top 10 genes for distinguishing hematologic cancers from CHIP were PBX1, CAMTA1, PFDN5, PCBP1, SDC4, PRKACA, NT5C2, NBR1, RPS21, and AHCYL1. Conclusions: Analyzing cfRNA expression levels when used with machine learning may add another level of confidence to the ability of interpreting mutations detected in liquid biopsy testing, especially when these mutations are at low levels. This is particularly important in liquid biopsy testing for minimal residual disease (MRD) when a tissue baseline sample is not tested (tumor-agnostic).