Comprehensive machine learning-driven platform infers key tumor characteristics from blood-derived cfRNA
Authors
- BostonGene Corp., Waltham, USA
Abstract
Recent progress in high quality sequencing of circulating nucleic acids makes liquid biopsy an efficient approach to monitor tumor evolution and therapy response. Cell-free RNA (cfRNA) from blood and other biofluids contains a tumor-derived fraction [1], offering a minimally invasive tool to characterize tumor-related transcriptomic states. Here, we present a comprehensive machine learning (ML)-driven platform for analysis of blood-derived cfRNA to infer clinically important features and biomarkers of malignancies.
Methods:
cfRNA was extracted from 4 mL of double-spun plasma (n = 232 healthy and n = 92 breast, 36 lung, 23 pancreatic, and 17 colorectal cancer cases). NGS libraries were prepared according to the Agilent XT HS2 protocol using the V8+UTR exome-wide panel. Pisces 5.2 and samtools mpileup tools were used to call tumor-specific mutations from cfRNA. Abundance of transcripts from cancer-specific signatures was analyzed using gene set enrichment analysis (GSEA) and single-sample GSEA. ML decision tree-based models were trained on artificial data generated from open source bulk RNA-seq data from cancer cells, tissues, and sorted cells collected across the GEO database. Model testing was performed on real cfRNA sequences (n = 232 healthy, n = 168 cancer cases).
Results:
We developed robust protocols for plasma-derived cfRNA extraction and NGS library preparation for reproducible interpatient and intrapatient cfRNA transcriptome profiling (Figure 1). cfRNA profiles from cancer patients contained mRNA transcripts carrying tumor-specific hotspot mutations demonstrating a tendency to moderate positive correlation between tumor and cfRNA variant allele frequencies (VAFs; R = 0.41, p = 0.064; Figure 2), and profiles were also enriched with epithelial, epithelial-mesenchymal transition, senescence, and angiogenesis signatures (Figure 3). We employed an ML-driven approach to infer tumor-specific characteristics from the tumor-derived cfRNA fraction for breast, colorectal, lung, and pancreatic cancers. ML models trained with artificial cfRNA transcriptomes accurately detected the status of breast cancer (AUC = 0.73 ± 0.05, n = 153), tumor microenvironment fibrosis (AUC = 0.80 ± 0.07, n = 44), predicted PD-1 (AUC = 0.71 ± 0.03, n = 78), and liver metastasis (AUC = 0.70 ± 0.02, n = 143) when tested in clinical patient samples (Figure 4).
Conclusions:
The presented cfRNA-based platform offers unprecedented insight into the tumor biology compared to liquid biopsy assays used in current clinical practice. The proposed platform is universal and can potentially characterize any tumor-associated process accompanied by transcriptomic changes reflected in the cfRNA fraction.
Latest publications
