Comprehensive transcriptomic analysis of cell lines as models of primary tumor samples across 22 tumor types

Katherine Yu, Bin Chen, Dvir Aran, Joseph K. Charalel, Atul J. Butte, Theodore C. Goldstein, Marina SirotabioRxiv 2018 · DOI

Abstract

Cancer cell lines are indispensable tools in oncology, yet their ability to accurately represent the biological complexity of primary tumors is often questioned. This study presents a pan-cancer transcriptomic analysis comparing approximately 9,000 profiles from The Cancer Genome Atlas (TCGA) with the Cancer Cell Line Encyclopedia (CCLE) across 22 tumor types. By adjusting for tumor purity—accounting for the presence of non-cancerous immune and stromal cells in primary samples—we evaluated the fidelity of cell lines as models. Our findings reveal consistent upregulation of cell-cycle pathways in cell lines and identify significant discrepancies in specific models, including colorectal lines that were misidentified as epithelial cells. We propose the TCGA-110, a new panel of representative cell lines, to improve the translatability of in vitro findings.

Introduction: The Translation Gap

In cancer research, cell lines are the workhorses of the laboratory. They are cost-effective, immortalized, and essential for high-throughput drug screening. However, a notorious "translation gap" exists: many therapeutic candidates that show promise in a petri dish fail to demonstrate efficacy in human clinical trials. A major contributor to this failure is the poor selection of cell line models. If a cell line does not accurately reflect the molecular landscape of the primary tumor it is meant to represent, the resulting data may be fundamentally misleading.

Previous efforts to validate these models were often limited by small sample sizes or a failure to account for the "messiness" of primary tumors. Unlike the pure populations found in cell cultures, primary tumors are a complex mixture of malignant cells, infiltrating immune cells, and connective tissue (stroma). This study addresses these limitations by performing a massive "reality check" on 666 cell lines against nearly 9,000 primary tumor samples.

Methodology: Adjusting for Purity

To ensure a rigorous comparison, the researchers integrated transcriptomic (RNA-seq) data from the TCGA and CCLE repositories. The analysis involved several critical computational steps:

  1. Batch Correction: Using the RUVseq package, the team removed technical noise inherent in data aggregated from different laboratories and sequencing runs.
  2. Purity Adjustment: This was the study's critical innovation. Using four distinct methods (including DNA methylation and pathology slide analysis), the researchers estimated the proportion of malignant cells in each TCGA sample. They then used linear regression to "subtract" the transcriptomic signals originating from immune and stromal cells, allowing for a direct comparison between the cancer cells in the tumor and the cancer cells in the flask.
  3. Correlation and Pathway Analysis: Spearman correlations were calculated based on the 5,000 most variable genes, and Gene Set Enrichment Analysis (GSEA) was used to identify biological pathways that differed between the models and the primary samples.

Key Insights

The "Culture Shock" Phenomenon

Across all 22 cancer types, a universal trend emerged: cell lines consistently showed an upregulation of cell-cycle and proliferative pathways, such as E2F and Myc targets. This suggests a form of "culture shock" or selective pressure; the in vitro environment favors cells that can divide rapidly on plastic, a trait that may not be as dominant in the complex microenvironment of a human organ.

Identity Crises in Cell Lines

A case study of colorectal cancer revealed startling discrepancies. Three specific cell lines (Hs 675.T, Hs 698.T, and Hs 255.T) were found to correlate poorly with primary colorectal tumors. Further investigation suggested these lines were not derived from epithelial cancer cells at all, but were actually fibroblasts (connective tissue cells) that had likely outgrown the cancer cells during the initial culturing process.

Tissue-Specific Fidelity

The study found that liquid cancers, such as leukemia and lymphoma, have the most representative cell line models, with median correlation coefficients reaching as high as 0.75. This is likely because the suspension environment of a laboratory flask more closely mimics the natural "floating" state of blood cells compared to the structural complexity of solid organs like the liver or brain (where correlations were lower, e.g., 0.50 for Cholangiocarcinoma).

The TCGA-110 Panel

One of the most significant outputs of this research is the proposal of the TCGA-110. For decades, the NCI-60 panel has been the industry standard for drug testing. However, this study demonstrates that the NCI-60 is increasingly outdated. By selecting the five most representative cell lines for each of the 22 tumor types based on transcriptomic correlation, the authors created the TCGA-110—a panel that is significantly more biologically relevant to modern patient data.

Conclusion and Future Directions

While transcriptomic similarity is a powerful metric, it is only one layer of biology. Future work must integrate proteomics, mutational data, and epigenetics to provide a truly holistic view of model fidelity. Furthermore, while the TCGA-110 offers a more representative toolkit, experimental validation is still required to prove that these high-fidelity lines better predict clinical drug responses.

To assist the broader research community, the authors have launched a web application that allows scientists to evaluate and select the most appropriate cell lines for their specific cancer of interest, potentially narrowing the gap between the lab bench and the patient bedside.

Summary updated

Abstract

Abstract Cancer cell lines are commonly used as models for cancer biology. While they are limited in their ability to capture complex interactions between tumors and their surrounding environment, they are a cornerstone of cancer research and many important findings have been discovered utilizing cell line models. Not all cell lines are appropriate models of primary tumors, however, which may contribute to the difficulty in translating in vitro findings to patients. Previous studies have leveraged public datasets to evaluate cell lines as models of primary tumors, but they have been limited in scope to specific tumor types and typically ignore the presence of tumor infiltrating cells in the primary tumor samples. We present here a comprehensive pan-cancer analysis utilizing approximately 9,000 transcriptomic profiles from The Cancer Genome Atlas and the Cancer Cell Line Encyclopedia to evaluate cell lines as models of primary tumors across 22 different tumor types. After adjusting for tumor purity in the primary tumor samples, we performed correlation analysis and differential gene expression analysis between the primary tumor samples and cell lines. We found that cell-cycle pathways are consistently upregulated in cell lines, while no pathways are consistently upregulated across the primary tumor samples. In a case study, we compared colorectal cancer cell lines with primary tumor samples across the colorectal subtypes and identified three colorectal cell lines that were derived from fibroblasts rather than tumor epithelial cells. Lastly, we propose a new set of cell lines panel, the TCGA-110, which contains the most representative cell lines from 22 different tumor types as a more comprehensive and informative alternative to the NCI-60 panel. Our analysis of the other tumor types are available in our web app ( http://comphealth.ucsf.edu/TCGA110 ) as a resource to the cancer research community, and we hope it will allow researchers to select more appropriate cell line models and increase the translatability of in vitro findings.
BibTeX
@misc{yu2018comprehensive,
  author = {Katherine Yu and Bin Chen and Dvir Aran and Joseph K. Charalel and Atul J. Butte and Theodore C. Goldstein and Marina Sirota},
  title = {{Comprehensive transcriptomic analysis of cell lines as models of primary tumor samples across 22 tumor types}},
  year = {2018},
  doi = {10.1101/422592},
  howpublished = {bioRxiv},
}
Figure