Cancer cell lines are widely used to study cancer biology and to test hypotheses in translational research [1, 2]. Numerous cancer cell lines are derived from tumors, and are used for various purposes such as understanding disease progression, developing diagnostics, and screening anti-tumor drugs. However, the relevance of a cell line is dependent on how closely it resembles the tumors being studied, while the in vitro cell culture environment differs from that of in vivo tumor tissue. Moreover, not all cancer cell lines have equal value or relevance as tumor models [3]. Therefore, relating tumors and cell lines, and recognizing their similarities and differences is of critical importance in translational research and understanding cancer biology.

It has been shown that genetic and epigenetic changes including gene mutations, deletions, amplifications, translocations and methylation status found in lung tumors are retained in lung cancer cell lines [4]. Prolonged cell culture is more likely to cause the secondary genomic changes such as copy number variation and gene expression [3]. Several previous studies have compared the genomic and transcriptomic differences between tumors and cell lines, but their sample size has been very limited [5–8]. Integrating disease tissue gene expression and drug gene expression profiled in cancer cell lines for therapeutic discovery has been extensively applied by our lab and others [9–13]. Rapid advances in the field of genomics have led to the generation of a high volume of molecular data across various tumors and cell lines. Of particular note, The Cancer Genome Atlas (TCGA) Research Network has characterized the genomic and gene expression profiles of over 10,000 human tissue samples cross 32 tumor types [14]. The Cancer Cell Line Encyclopedia (CCLE) provides the genomic and gene expression profiles of over 1000 cell lines [15]. Domcke et al. recently evaluated ovarian cell lines as tumor models by comparing their genomic profiles published in CCLE and TCGA, and they found that some rarely used cell lines might be more appropriate to study ovarian cancers [16]. To our knowledge, here we present the first comprehensive comparison of tumor and cell line samples using gene expression data specifically for liver cancer.

Liver cancer is the sixth most frequent cancer globally with higher occurrences in Asia and sub-Saharan regions and is the second leading cause of global cancer deaths. The incidence and mortality of liver cancer have increased in the United States and Europe in the past decade. It was projected to become the third leading cause of cancer-related death by 2030 in the USA [17]. While HCC is the most common type of liver cancer, the current effective treatments of HCC remain limited. HCC tumors are insensitive to conventional chemotherapy, although limited benefits have been shown from the targeted drug sorafenib on only a small subset of tumors [18]. Meanwhile, a large number of liver cell lines have been developed and used to screen compounds, and many drugs are showing promising in vitro models. However, that these best drug candidates fail to be effective clinically [19] indicates the necessity to examine the correlation between HCC tumors and cancer cell lines and elucidate their similarities and differences for more effective translation of bench-side discovery into clinical utility.

In this work, we compared 200 HCC tumor samples from TCGA to over 1000 cancer cell lines from CCLE using their gene expression profiles, and we found that most HCC cell lines are significantly correlated to the primary tumors in comparison to other cancer cell lines. However, from the set of 25 HCC cell lines, nearly half are not significantly correlated to the tumors. We also found that a small subset of the tumors is not significantly correlated to HCC cell lines. We further identified the differentially expressed genes between the tumor samples and cell lines, and we found that a substantial number of differentially expressed genes are located in the extra cellular space. The genes that are over-expressed in the tumor samples are related to immune response, drug metabolism and ABC transporter. This work helps guide us in optimal selection of cell lines and highlights the differentially expressed genes that play an essential role in translating from in vitro to in vivo and clinical research.

An increasing number of HCC cell lines have been developed for the study of cancer biology and translational research [29], however, only a few cell lines are often chosen for in vitro or in vivo studies due to limited experimental capability. Choosing appropriate cell lines for experimental testing is thus of critical importance. Unfortunately, crucial information about cell lines (e.g., metastasis status, hepatitis virus status, drug treatment, mutation) is often missing or ambiguous in cell line databases and relevant publications, which further confounds the selection process. As the genomic and transcriptomic profiles of cancer cell lines and tumors become more publically available, using gene expression to relate cancer cell lines to tumor samples becomes a more useful tool to inform cell line selection and has been applied to other cancers (e.g., ovarian cancer [16]). Although gene expression is not the only factor, it can be used to exclude certain cell lines that do not share similar profiles with the tumors being studied.

Our gene expression analysis of tumors and cell lines demonstrates that HCC cell lines are very tissue specific. They resemble the majority of HCC tumors using top 5000 varying genes. This conclusion still holds if top 1000 varying genes are used (data not shown). The four widely used cell lines (i.e., HepG2, HuH 7, Hep3B and PLC/PRF/5) are highly correlated to the tumors. However, nearly half of the cell lines are strikingly not significantly correlated to the tumors. Poor differentiation or genomic damage by chemotherapy drugs may account for the poor correlations of some cell lines.

Our analysis also shows that there is substantial variation of individual tumor samples in regards to their correlations with cell lines. Intrinsic heterogeneity among samples may account for the variation. Specifically, we identify 8 out of 200 tumor samples as outliers that have poor correlations with the known HCC cell lines. It may suggest that the drug tested in these cell lines will most likely have a different clinical effect in the outlier patients. It is interesting to further associate these tumors to clinical features so that this subset of tumors can be easily identified clinically and personalized treatments can be designed. However, the current 8 samples are insufficient to draw a robust conclusion. It may be also interesting to associate them to other genetic features.

Even though many HCC cell lines have overall high correlations with HCC tumors, a large number of genes are actually absent or lowly expressed in cell lines. A significant number of these genes are involved in disease development. Some genes are related to immune system and pharmacokinetics including drug metabolism and drug transportation. It's known that immune system and pharmacokinetics are related to drug response [30]. Inducing missing or low expressed genes in cell lines by external agents may be a way to improve cell line models [26, 31].

In summary, by leveraging publicly available transcriptomic data, we relate individual HCC tumors and cancer cell lines. We demonstrate that the most commonly used HCC cell lines closely resemble HCC tumors based on their gene expression, and we also identify nearly half of HCC cell lines do not resemble HCC tumors. Even in the highly correlated cell lines, a significant number of genes that are critical for disease development or drug response are absent or expressed at low levels, therefore, additional attention should be paid to these genes in translational research. Our future work includes covering more tumor samples from pubic sources (e.g., Gene Expression Omnibus) and more cell lines (the Sanger Cancer Cell Line project [32]), analyzing genetic features (e.g., mutation, copy number variation, single nucleotide polymorphism), and linking the subtype of patients classified by clinical features or molecular features to cancer cell lines. Finally, this work serves as a model analysis that can be extended to other tumor types of interest.