TL;DR
This paper presents a novel machine learning-based approach to accurately identify circulating tumor cells (CTCs) from blood samples, even when they exhibit diverse phenotypes. By integrating single-cell RNA sequencing data from various cancer types and training a classifier on both CTCs and peripheral blood mononuclear cells (PBMCs), the researchers achieve high accuracy in distinguishing these rare cells, paving the way for more effective cancer monitoring and diagnosis.
What Problem Does This Paper Address?
Metastasis, the spread of cancer from its primary site to distant organs, is responsible for the vast majority of cancer-related deaths. Circulating tumor cells (CTCs) are cancer cells that have detached from the primary tumor and are traveling through the bloodstream. Their detection and characterization offer a unique window into tumor evolution and therapeutic response. However, CTCs are extremely rare in the blood, making their isolation and analysis challenging. Traditional methods often rely on specific epithelial markers (like EpCAM), which are downregulated during the epithelial-mesenchymal transition (EMT) – a process crucial for cell migration and invasion. This means many CTCs, especially those undergoing EMT, can be missed. The paper aims to overcome this limitation by developing a method that can identify CTCs regardless of their specific phenotype, including those that have lost typical epithelial markers.
Methods
The study employs a multi-pronged approach:
- Integrative Analysis of CTC Transcriptomes: The researchers collected and integrated publicly available single-cell RNA sequencing (scRNA-seq) data from CTCs across various cancer types. This allowed them to study the collective gene expression patterns of these rare cells.
- Epithelial-Mesenchymal (E:M) Score: To quantify the degree of EMT in CTCs, they developed an "E:M score." This score is calculated by assessing the expression levels of known epithelial and mesenchymal markers. A high E:M score indicates a more epithelial phenotype, while a low score suggests a mesenchymal phenotype.
- Machine Learning Classifier: A key contribution is the development of a machine learning classifier. This classifier was trained using scRNA-seq data from both CTCs and peripheral blood mononuclear cells (PBMCs). The goal was to teach the algorithm to distinguish CTCs from the much more abundant immune cells in the blood, even when the CTCs exhibit diverse gene expression profiles.
- Label-Free Microfluidic Workflow: The study also validates a novel microfluidic workflow for CTC enrichment. This workflow combines size-based enrichment (ClearCell FX) with negative selection for immune cell markers (CD45 and CD31) using the Polaris system. This label-free approach aims to capture CTCs without relying on specific surface markers, thus being more inclusive of different CTC phenotypes.
Datasets
The study utilized several datasets:
- Publicly Available CTC Datasets: Approximately 500 single-cell transcriptomes from CTCs across five cancer types (breast, prostate, lung, pancreas, melanoma) were collated from 10 independent studies.
- Publicly Available PBMC Datasets: Expression profiles of peripheral blood mononuclear cells (PBMCs) were collected from six different studies.
- Newly Generated Data: Circulating breast tumor cells were captured from three HER2- breast cancer patients using the newly developed microfluidic system. This data was used to validate the performance of the trained classifier.
The paper does not introduce a new benchmark dataset but rather integrates existing ones and validates its findings on newly generated data from a novel workflow.
Key Insights and Ideas
- CTCs form a continuum of EMT: The analysis revealed that CTCs from different cancer types largely lie on a continuous spectrum of epithelial-to-mesenchymal transition (EMT). This highlights the dynamic nature of CTCs as they adapt for survival and dissemination.
- Inverse relationship between PD-L1 and MHC: Integrative analysis of CTC transcriptomes showed an inverse correlation between the expression of PD-L1 (a key immune checkpoint molecule) and MHC (major histocompatibility complex) molecules. This finding is significant for understanding how CTCs evade immune surveillance and has implications for immunotherapy strategies.
- Machine learning for unbiased CTC detection: The development of a machine learning classifier trained on both CTCs and PBMCs offers a powerful tool for unbiased CTC identification. This approach moves beyond marker-dependent methods, which can miss CTCs undergoing EMT.
- Label-free enrichment is feasible: The validated microfluidic workflow demonstrates the potential for label-free enrichment of CTCs, which is crucial for capturing the full spectrum of CTC phenotypes.
Key Results
- High accuracy classifier: The Gradient Boosting Machine (GBM) classifier achieved a mean accuracy of approximately 93% in distinguishing CTCs from PBMCs across various datasets.
- Validation of novel workflow: The newly developed ClearCell-Polaris workflow successfully enriched and identified CTCs from patient blood samples, with all captured cells being classified as CTCs by the trained model.
- Identification of potential novel markers: Beyond EpCAM, the study identified other surface protein-coding genes (e.g., ITGB5, TACSTD2, SLC39A6) that were differentially upregulated in CTCs over blood cells, suggesting their potential utility for marker-dependent capture.
Limitations and Future Work
- Dataset variability: The study acknowledges that the curated CTC datasets were of variable quality, which was addressed through preprocessing steps. However, inherent biases in data quality across different studies might still influence the results.
- Limited sample size for validation: While the novel workflow was validated on three patients, a larger cohort would be beneficial for more robust validation.
- Future directions: The authors suggest that with decreasing costs of single-cell gene expression screening, their proposed strategy for unbiased CTC detection and characterization could see high adoption rates. Further research could focus on expanding the marker panel for improved capture and exploring the clinical utility of this approach for patient stratification and treatment monitoring.
Summary updated