There is a growing awareness that tumor-adjacent normal tissues used as control samples in cancer studies do not represent fully healthy tissues. Instead, they are intermediates between healthy tissues and tumors. The factors that contribute to the deviation of such control samples from healthy state include exposure to the tumor-promoting factors, tumor-related immune response, and other aspects of tumor microenvironment. Characterizing the relation between gene expression of tumor-adjacent control samples and tumors is fundamental for understanding roles of microenvironment in tumor initiation and progression, as well as for identification of diagnostic and prognostic biomarkers for cancers.
To address the demand, we developed and validated TranNet, a computational approach that utilizes gene expression in matched control and tumor samples to study the relation between their gene expression profiles. TranNet infers a sparse weighted bipartite graph from gene expression profiles of matched control samples to tumors. The results allow us to identify predictors (potential regulators) of this transition. To our knowledge, TranNet is the first computational method to infer such dependencies.
We applied TranNet to the data of several cancer types and their matched control samples from The Cancer Genome Atlas (TCGA). Many predictors identified by TranNet are genes associated with regulation by the tumor microenvironment as they are enriched in G-protein coupled receptor signaling, cell-to-cell communication, immune processes, and cell adhesion. Correspondingly, targets of inferred predictors are enriched in pathways related to tissue remodelling (including the epithelial-mesenchymal Transition (EMT)), immune response, and cell proliferation. This implies that the predictors are markers and potential stromal facilitators of tumor progression. Our results provide new insights into the relationships between tumor adjacent control sample, tumor and the tumor environment. Moreover, the set of predictors identified by TranNet will provide a valuable resource for future investigations.
In oncological studies, control samples are usually biopsied from tumor-adjacent normal tissue. However, there is an increasing understanding that such samples represent a state that is intermediate between tumor and normal, and is influenced by environmental factors common to tumor and normal tissues, and by tumor microenvironment. Therefore, uncovering the relation between gene expressions across control and tumors samples can inform us about the roles of microenvironment in tumor initiation and progression. Here we present a predictive model, TranNet, to study the functional relationship between matched control and tumor samples. TranNet infers a transition function from gene expression in a control sample to that in the matched tumor sample. Simultaneously, the method identifies a set of genes that are predictors of this transition. To our knowledge, TranNet is the first computational method to infer such dependencies.
Our results demonstrated that TranNet efficiently captured the relation between tumors and their microenvironment, generating important implications for the detection, diagnosis, and prognosis of cancers.
Citation: Amgalan B, Day C-P, Przytycka TM (2023) Exploring tumor-normal cross-talk with TranNet: Role of the environment in tumor progression. PLoS Comput Biol 19(9): e1011472. https://doi.org/10.1371/journal.pcbi.1011472
Editor: Giulio Caravagna, University of Trieste Faculty of Mathematics Physics and Natural Sciences: Universita degli Studi di Trieste, ITALY
Received: July 11, 2023; Accepted: August 23, 2023; Published: September 18, 2023
This is an open access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication.
Data Availability: The authors confirm that all data underlying the findings are fully available without restriction. There are no primary data in the paper; all software is available at https://github.com/ncbi/TranNet.
Funding: This research was supported by the Intramural Research Program of the National Library of Medicine (T.M.P and B.A), and Intramural Research Program of the Center for Cancer Research, National Cancer Institute (C.-P.D). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests.
In multi-stage carcinogenesis theory, mutations accumulate to perturb the cell regulatory program, eventually causing cell transformation. These perturbations interact with other factors such as micro and macro environment or preexisting health conditions. Although mutations are central for the emergence of cancer, much of the understanding of the disease comes from studies of the cancer-related changes in gene expression. Cancer is characterized by dysregulated functions of many cellular processes including proliferation, cell-cell interactions, chromatin organization, DNA repair, and others. Yet many of these alterations are not mechanically linked to specific mutations, but are driven by changes in gene expression . One of the emerging concepts is that cancer progression is facilitated by increased cell plasticity, which allows cancer cells to switch dynamically between a differentiated states in response to stress . Cancer cell plasticity has been linked to the epithelial-to-mesenchymal transition (EMT) which has been shown to respond to microenvironmental signals or cancer therapy [3–8]. In addition, it has been estimated that at least 25% of cancers are associated with chronic inflammation [9, 10]. Last but not least, a recent study suggested that the tumor microenvironment and the microenvironment of control sample are strongly dependent . In fact, tumor-related alteration of adjacent tissue are believed to contribute to postoperative cancer recurrences that occur in up to a third of patients . This is not a surprise, as tumors and their adjacent normal tissues also share some exogenous exposures including environmental factors, such as smoking, diet, and genetic variations. However could gene expression from a control sample inform on tumor state? Which molecular pathways and functions in tumor are associated with gene expression changes in normal tissue? What can they teach us about tumor progression?
To address these questions, we developed the Transition Network model (TranNet), a computational approach to study the relation between the gene expression patterns in matched normal and tumor tissues. Focusing on tumor genes (defined as genes differently expressed between tumor and control samples), TranNet infers a transition function from gene expression in a control sample to an estimate of gene expression in the matched tumor sample. Simultaneously, the method infers a set of genes that are predictors of this transition (Fig 1). These predictor genes might but do not have to regulate this transition, and might instead reflect expression changes due to factors common to control and tumor samples. Functional analysis of the predictors and their target genes helps to shed light on possible sources of this relation.
TranNet constructs a transition function from gene expression in control samples to gene expression in the matched tumor samples. Simultaneously, the method infers a set of predictors in the control tissue that used for computing the transition and target genes in the tumor tissue that are influenced by the predictors. After validating the method, functional analysis of the predictors and their target genes is used to shed light on possible sources of the relation between gene expression in control and tumor samples.
Computationally, TranNet utilizes a network construction based on a sparse estimation of partial correlation  that uses an l1 norm constraint to ensure the selection of the most informative predictors (Fig 2). Recognising that the expression of the genes which did not pass the p-value threshold to be included in the set of tumor genes might also contribute to gene expression in tumor, we extended the network by including additional nodes, principal components, representing the informative trends on the expression data of these genes. We refer to the genes and principal components whose activity in control samples influence gene expression in tumor tissue samples as predictors.
Input matrices X and Y represent the expressions of genes in control (X, panel (A1)) and tumor (Y, panel (A2)) samples across p patients (rows). Genes (gi, i = 1, …n) which are differentially expressed between control and tumor samples (referred to as tumor genes) are represented by individual nodes (columns from 1 to n) and the rest of the genes are represented by principal components of their expression (columns from pc1 to pck) representing the major trends of the expression. The principal components which differentiate between normal and tumor tissue samples are considered as additional variables in the multi-variate analysis. The distribution trend of the expression of a differentially expressed gene gi is visualized in the left of the yellow panel. As illustrated in a 2-dimensional space in the right of the yellow panel, although neither nor is differentially expressed between control and tumor tissue samples, their joint distribution is differentiated along the axis of principal component pc1 showing a similar trend to the DE genes while pc2 is not differentiated (B). The transition map is a linear operator defined by matrix M computed to minimise Y’s representation error subject to a sparsity constraint explained in the main text (C). The result is summarised as a bipartite network representing explanatory influences from control samples to tumors (D). Predictive potential of a gene represents its total contribution to the transition network (E).
TranNet opens a new way to explore the relationship between gene expression in adjacent control samples and tumors. Our results indicate that the former can provide information about the latter and link, at least in part, inferred relations to tumor environment. As elaborated in Section ‘Discussion and conclusions’, not all inferred associations are assumed to be tumor drivers. Yet, as predictors of the expression of tumor genes, they are markers of other tumor-related processes including tumor-environment interaction and are important in this respect. Indeed, using TranNet we identified a set of genes that can serve as the predictors of tumor-environment relation as well as genes and pathways involved in this interaction. Taken together, this work offers a computational method to infer the relation between the gene expression patterns in matched normal and tumor tissues, and provided a new understanding of the relation between tumor, normal, and the environment.
The transition network model (TranNet)
An overview of the TranNet method is presented in Fig 2. Two matrices representing gene expressions in normal (A1) and tumor (A2) tissues of cancer patients are provided as the input. Tumor genes, defined as genes differentially expressed between control and tumor samples, are represented as network nodes while the expression of the rest of the genes is represented by meta-nodes corresponding to the principal components of the expression of these remaining genes. Each gene or principal component is represented by two nodes—one for each condition. The conceptual idea of including the principal components in the analysis is illustrated in (B). The transition matrix from normal to tumor state is obtained by solving the problem in (C) and the bipartite network formed by the transition matrix represents information flow from normal to tumor tissues (D). Finally, the strength of the total influence of a node in normal tissue on gene expression in tumor is quantified by predictive potentials (PP) (E).
Inference of the transition network
Let X and Y be the matrices describing gene expression in tumor and control samples respectively for p patients. The first n columns of each matrix represent the expression of the n tumor genes followed by k meta nodes corresponding to the principal components summarising gene expression trends of the remaining set of genes.
Specifically, xij and yij denotes the expression of gene gj (or pcj−n if j > n) in the ith patient in tumor and normal respectively where the “expression” of a principal component in patient i is explained below (see also Fig 2B right, for the description of differentially expressed (DE) principal component).
Assuming that X represents regulatory influences (or markers of such influences) and Y their targets, the weight matrix M describing transformations between them can be written as (1) More precisely, for patient i, the expression value yif of gene gf or principal component pcf in the patient’s tumor tissue can be approximated by a linear combination of the expressions of the n genes and k meta nodes in the patient’s normal tissue where denotes the transition weights from the normal tissue expression of the n genes and k principal components to the tumor tissue expression of gene gf or principal component pcf. Thus, our goal is to minimize the least square error subject to a unit l1 norm constraint on m(⋅)f as follows: (2) The l1 norm constraint on the edge weights allows for selecting the strongest transition effects on a given target node. Hence, this regularization acts to avoid over-fitting issues and only non-zero mjf, selected for gf, denotes the transition effect from gj. In this setting, for every target node, the optimization problem in Eq (2) (a constrained version of Eq (1)) searches over all possible combinations of transition effects from the predictor nodes and selects the best combination with their optimal weights to explain the activity of the target node .
To explain the role of principal components we shall recall that a fundamental assumption for multivariate analysis is that there are no unobserved factors affecting both explanatory and response variables globally. The expressions of the genes that were not selected as tumor genes (DE genes) might have such influence. To adjust for such effects on the transition mapping, we include meta nodes (adjustment variables in ) that represent the principal components of the expression data for these non-tumor genes (see Section ‘Materials and methods’). By the definition, principal components are vector representations of the general trends in a multi-dimensional data . Here data points represent all the samples (control and tumor) and each point is a vector representing gene expression (Fig 2B right). Embedding this data using principal components as reference axis, we say that a given principal component is differentially expressed between the control and tumor samples if the coordinates of the tumor samples on the axis defined by this principal component are significantly larger or smaller than the coordinates of the control samples (Fig 2B right).
Finally, we comment on the optimization problem. Although the least square minimization in Eq (2) is convex, the l1 norm constraint is non-smooth and derivative-based optimality conditions such as Lagrangian multipliers and Karush–Kuhn–Tucker (KKT) conditions are not, in general, directly applicable. While a coordinate descent algorithm with convex penalties  is commonly used to approximate this non-smooth problem, there is no optimal strategy for tuning parameters controlling the strength of the penalty term . Similarly complementing the standard quadratic programming formulation, the soft shrinkage method decomposes the l1 norm into 2n+k inequality constraints . However, handling 2n+k constraints is not practical for the large-scale problem. To overcome this issue, we implemented the projected gradient method  that converges to the optimal solution for the non-smooth constrained problem.
TranNet aims to identify predictors (genes and principal comments) whose expression in one condition (here normal) predicts gene expression in a different condition (here tumor). We model the information flow from normal to tumor tissue by the transition matrix in Eq (1), obtained by solving the non-smooth convex optimization problem in Eq (2). The total contribution of an individual gene or a principal component to the transition is captured by the non-zero transition weights from the corresponding node to its selected targets in the network, and the total incoming effect on each target is normalized by the l1 norm constraint in Eq (2). We define the predictive potential (PP) of such predictor as the summation of the absolute weights of the outgoing edges from the node representing the corresponding gene/principal component in normal tissue. (3) Thus, the predictors can be prioritized with respect to their predictive potential defined in Eq (3).
Validation of the TranNet model
We applied the TranNet model to the TCGA gene expression data for five different cancer types, focusing on solid tumors selected based on availability of matched control-tumor samples: Breast Cancer (BRCA), Lung Adenocarcinoma (LUAD), Lung squamous cell carcinoma (LUSC), Prostate Adenocarcinoma (PRAD), and Liver Hepatocellular Carcinoma (LIHC) (see Section ‘Materials and methods’ for data processing). To validate TranNet model, we first show that the impact of gene expression in normal tissue on gene expression in tumor is higher than expected by chance (predictability). Next, we show that the network edges inferred by TranNet are enriched in functional interactions. Subsequently, we analyse gene-wise prediction accuracy. These validations in BRCA are collected in Fig 3, and extended figures including the validation results for all the five cancers are provided in Figs A and B and C and D in S1 Text. Finally, we confirm that the results are stable, thus the conclusions of this study do not depend on the threshold used for selecting the gene set for analyzing.
Sample-wise predictivity of TranNet was evaluated based on a leave-one-out test (blue-shaded, Section ‘Expression of genes in control samples informs gene expression in tumor’); Interactions inferred by TranNet were tested with enrichment of the functional interactions (green-shaded, Section ‘The transition edges inferred by TranNet are enriched in functional interaction edges’); Gene-wise prediction accuracy of TranNet was evaluated by the prediction error (mean absolute error) on the targets. (grey-shaded, Section ‘Expression of genes in control samples informs gene expression in tumor’).
Expression of genes in control samples informs gene expression in tumor
The motivation for the TranNet model is the hypothesis that gene expression in control samples carry relevant information on gene expression in matched tumors. In order to confirm that this is indeed the case, we used a leave-one-out strategy (see Section ‘Materials and methods)’ to test if the model can predict the expression of tumor genes from the expression of predictors better than expected by chance. The paired-sample T-test is used to compare the results (prediction errors) of two predictions: the first based on the real expression data and the second based on permuted data. In addition, we tested how prediction accuracy depends on the number of used predictors assuming that the markers are selected in the decreasing order of their PP values. The prediction accuracy varies with the selection of different percentages of the top PP scoring predictors (see Patient-wise predictability in Fig 3 and Fig A in S1 Text, and Table 1). We note that, similar to the other biological networks, the distribution of both degrees and predictive potentials in the TranNet networks is characterised by a large number of predictors with very small predictive potentials and small proportion of predictors with high potentials (see Fig E in S1 Text and S1 File). In particular, the sum of predictive potentials for the top predictors (Table 1) were 42.2% (BRCA), 46.6% (LUAD), 50.7% (LUSC), 56.5% (PRAD) and 29.5% (LIHC) of the total predictive potentials over all predictors in the respective cancers.
The paired-sample T-test was used to compare the sample-wise prediction errors on real and permuted response data. This test was performed for the different numbers of top PP scoring predictors. The selected predictors corresponding to the optimal cut-off are provided with their PP scores in S1 File. The bottom rows report result using top 30% of predictors. The top rows report the results for the cancer-specific optimised set of predictors (See Fig 3 and Fig A in S1 Text).
As expected, not all tumor genes are well predicted. However, for the top 20% of predicted genes, the coefficients of determination are above 0.6 and the Pearson correlation coefficients are close to 0.9 in all cancers expect BRCA (Fig B in S1 Text). The relatively worse performance of the method for BRCA is likely to be related to a smaller impact of environmental factors on breast cancer relative to the environmental impact on lung, liver and prostate cancers. Importantly, the error distributions are highly non-symmetric and negatively skewed towards small errors as visualized in Fig 3 and Fig C in S1 Text, suggesting that the expression of only some tumor genes can be predicted from the expression in control samples. The GO terms enriched for the gene lists sorted by the prediction errors on the genes (tumor tissue) are summarized in S2 File and discussed in Section ‘Pathway enrichment of genes identified as targets of predictors revealed tumor-stroma interactions’.
The transition edges inferred by TranNet are enriched in functional interaction edges
TranNet infers a sparse bipartite graph modelling the transition from the expression of predictors to the expression of tumor genes in matched tumor samples. Since these edges encode potential regulatory influences (or markers associated with such influences), we expect that they should be enriched for functional interactions. To test this, we computed the overlap of the TranNet edges with the edges in the human functional interaction network  (see Section ‘Materials and methods’) and tested whether this overlap is larger than expected by chance. To this end, we constructed 1,000 random networks by permuting target nodes of the inferred transition network without changing the topology of the network. The enrichment of the functional interactions  for BRCA is provided in Fig 3, and the performances for all the five cancers are summarized as Fig D in S1 Text. These results confirm that TranNet infers relations consistent with the functional interactions .
Stability of TranNet
Since the networks inferred by TranNet are sparse (see the table in Fig E in S1 Text) and inferred from expression data using some cut-offs (see Section ‘Materials and methods’), it is important to test if an alternative selection of tumor genes would not lead to very different results. To test this, we computed the transition matrix for an alternative definition of tumor (DE) and Non-DE genes where using a q-value cut-off of 0.001 in the T-test for differential expression. Then, we compared the optimal set of predictors for the main set with the same number of top predictors computed for the alternative set. The Jaccard similarity and hyper-geometric test results between the two selected sets of predictors showed very high agreements as summarized in Table 2.
Overlap in the selected predictors based on the networks constructed form two different selection of tumor genes. Main set is described in Section ‘Materials and methods’ and the Alternative set was selected with a different threshold as described in Section ‘Stability of TranNet’. We selected the same number of top predictors from each list based on the number of optimal predictors for the Main set. The overlap reports the number of common predictors, Jaccard Index: Jaccard Index similarity between the two sets of the highest scoring predictors and HG p-value is hyper-geometric p value for the enrichment of the predictors obtained with the Alternative set in the Main set.
Insights into the relation between gene expression in tumor and in matched control samples
After validating TranNet model, we investigated the properties of the inferred interactions between control and tumor tissue samples. As noted above, the gene expressions are not predicted equally well for all genes. However, the error distributions are highly non-symmetric and negatively skewed towards small errors as visualized in Fig 3 for BRCA (the error distributions for all the five cancers are provided as Fig C in S1 Text). We started by investigating which biological pathways are enriched within the genes with better predictions. Towards this end we performed enrichment analysis for the gene list ranked by prediction accuracy. Strikingly, the genes whose expression was predicted with higher accuracy were enriched with specific groups of GO categories related to: environment, cell communication, immune response, signalling, and cell cycle (see the complete list of enriched GO terms in S2 File). Particularly, the enriched pathways in BRCA included cell-to-cell signalling, ion transport, and regulation of hormone level. Pathways in LUAD and LUSC had many terms related to cilia as expected due to the known impact of smoking on the properties of ciliated cells . Pathways in LUAD were generally related to cell cycle. Finally pathways enriched in LIHC included, in addition to the pathways related to cell cycle, pathways related to detoxification and taxis and thus related to liver-specific relation with environment .
Interestingly, the third principal components PC3 in LUAD and PRAD, are significantly differentiated between control and tumor samples with q-values 2.10E-05 for LUAD and 4.72E-05 for PRAD (see Table A in S1 Text) and have relatively high predictive potentials ranked 104-th in the LUAD predictors and 37-th in the PRAD predictors (see S1 File). This demonstrates that expression of genes that did not pass the significance level of differential expression have effects on the differentially expressed genes. As for the targets genes of these components, both are enriched with terms related to metabolism and DNA repair, and Spliceosome (see Table A in S1 Text) while also containing cancer type specific two terms.
Predictor genes are enriched in pathways that imply the features of tumor microenvironment
To find the properties of the inferred predictor genes, we used GOrilla  to perform a GO enrichment analysis. For each cancer, we identified predictor genes, ranked them by their predictive potential (PP) scores, and performed a ranked list enrichment analysis . We considered all three levels: enrichment of biological process (P), molecular function (F) and cellular component (C) (See Fig 4).
Rows correspond to enriched terms, while the color represents enrichment p-value. GO terms enriched in at least three cancers are shown. The complete lists of GO enrichment for the lists of predictor genes sorted by their PP scores are provided as in S3 File.
The predictor genes for all the five cancer types are enriched in pathways that imply the feature of the microenvironment around the tumors. For the Biological Process, they include humoral immune response, cell-cell signaling and cell communication, and G-protein coupled receptor. Molecular functions were enriched with cell communication (receptor regulator activity and receptor ligand activity), hormone activity, and inflammation (cytokine activity, serine hydrolase activity, serine-type peptidase activity). With respect to the terms related to cellular component, all cancers were enriched in pathways related to extracellular space and membrane. When compared among individual cancer types, the predictor genes have shown distinct features of tumor microenvironment. This is illustrated by the top 10 predictor genes in each cancer type (Table 3, the top 10 cut-off was selected based on 50% dropout in predictive potential). Unprecedented 85% of these genes have been identified in literature as cancer related and generally associated with tumor progression. For BRCA, the predictors include inflammation-induced genes (CCDC102B , SPOCD1 , EPYC , GLYATL1, SLC6A9 ), immune cell markers (CLEC4G  and BATF4 ), and neurogenesis genes (ADRA2B  and NNAT ). For LUAD, in addition to inflammation-induced genes (LYPD1 ), immune cell markers (CD244 ), and neural genes (WDR62 ), the predictors also include invasiveness-related genes (ELMO1 , LY6K , TMEM52 , TSACC, PCDH7). Similarly, the predictors of LUSC include to inflammation-induced genes (ENTPD6 , SLC51B , GNLY, SLCO1C1 , and B3GAT1 ), invasiveness-related genes (CITED1, OR2W3 , RASGEF1A ), and neural genes (GAL and SH2D5). Importantly, most of these genes are involved in chronic obstructive pulmonary disease (COPD). Taken together, these results suggested that BRCA, LUAD, and LUSC exhibited inflammatory neuroepithelial reactive stroma that are regulated by different mechanisms in immune responses and tissue regeneration. Also shown in Fig 4, the predictors of PRAD and LIHC were enriched in distinct pathways from other cancer types. Their features are demonstrated by the function of the top-ranked predictor genes. For PRAD, the majority of the top-ten predictors are neurogenesis genes (NTRK2 , LRRC7, LMX1B, ROPN1B); the others include genes involved in inflammatory response (TMEM178A) and androgen receptor regulation (PRMT5 ). This is relevant to the intense neuroepithelial interactions in prostate cancer. For LIHC, the predictors include genes responsive to DNA damage (AUNIP, CHEK2, CENPI) and hypoxia (FLVCR1), and two genes involved in liver disease such as fatty liver and hepatitis (GSTA4, ATP2A1 , RNF125 ). They reflect the metabolic modification in the tumor microenvironments.
The lists of the top 10 genes with the highest predictive potentials (Eq (3)) in each cancer together with references to sample papers discussing their roles in cancer. Stars denote references to papers that point to relations with the same cancer type. The predictive potentials of the top predictors for the five cancers are provided in S1 File.
Pathway enrichment of genes identified as targets of predictors revealed tumor-stroma interactions
To investigate how the microenvironments might regulate the tumors, we identified the pathways enriched (see Section ‘Materials and methods’ for the pathway enrichment) by the tumor genes (“targets”; Matrix Y in Fig 2B) associated with the top-ranked predictor genes in Table 3.
As shown in Fig 5, the targets in BRCA and LUAD are more enriched in cell proliferation-related pathways (the first group in Fig 5). In LUSC, the majority of targets are enriched in tissue remodeling and immune response pathways (the second and third groups in Fig 5, respectively). In PRAD, in addition to immune response and tissue remodeling, the targets are mostly enriched in neural and hormone pathways (fourth and fifth groups in Fig 5, respectively). In LIHC, the targets are enriched mostly in cell proliferation and tissue remodeling pathways. These results reflect the distinct signaling for tumor initiation and expansion in each cancer type. For example, studies have shown that the occurrence of LUSC and LIHC are frequently preceded with Chronic Obstructive Pulmonary Disease (COPD) and hepatitis, respectively, and the tissue injury and remodeling in such chronic inflammation promote tumorigenesis. Interestingly, the same types of pathways in tumors could be associated with different types of predictors, depending on cancer type. For example, cell proliferation-related pathways in tumors are associated with inflammation-related predictors for BRCA (CCDC102B and GLYATL1) and LUSC (SLCO1C1, SLC51B, and BATF2); with invasiveness predictors for LUAC (LY6K and TMEM52); with neurogenesis predictors (LRRC7, LMX1B, ROPN1B) in PRAD; and with hypoxic (FLVCR1) and DNA-damage response (AUNIP) predictors for LIHC. These observations suggest that tumor growth is stimulated by distinct microenvironmental factors in each cancer type.
Kegg pathways enriched for the targets of the top 10 regulators (Table 3) in each cancer type. Rows correspond to the predictor genes while columns represent the pathways enriched for more than 10 predictor genes over all the five cancers. The complete lists are combined in S4 File.
Auto-predictors and their properties
Interestingly, we noted that the top predictors do not include genes that have edges leading from a gene in control tissue to the same gene in tumor. We refer to such genes as auto-predictors and asked, if such auto-predictors have properties distinct form the properties of the top predictors. To answer this question, in each cancer, we have identified all auto-predictors (see S5 File for the list of auto-predictors). Note that since all network edges are between differentially expressed genes only, this excludes genes whose expression is simply the same in normal and tumor tissues. Not surprisingly, gene expression in control and tumor samples for auto-predictors tends to be significantly correlated (S5 File). Strikingly, in contrast to the predictors with high predictive potential, GO enrichment analysis of auto-predicting genes revealed relations to oxidation (including hemoglobin complex) and detoxification (Fig 6, S6 File). In particular, the terms enriched for the auto-predictors did not include terms directly related to tissue remodelling. While both, the top predictors and auto-predictors, contained terms related to immune response, the terms for auto-predictive genes were more strongly focused on Major Histocompatibility Complex (MHC)-related immune response (Fig 6, and S6 File). Methods and data used for the functional enrichment are described in Section ‘Materials and methods’.
Discussion and conclusions
The regulation of tumor progression is not fully understood. Recent studies demonstrated that control samples representing adjacent normal tissue in solid tumor studies differ form healthy tissue . We hypothesised that genes whose expression in control samples is predictive of the expression in tumor samples might provide insights into this relation.
We developed and validated a new computational method, TranNet, which is able to identify genes in adjacent normal tissue that are predictive (and potentially facilitating) regulation of a tumor by the tumor environment (Fig 4). In this study we focused on tumor-related genes for both: the predictor and the target gene sets. It is possible that the expression of tumor-unrelated genes in control tissue might be informative about tumor progression. However narrowing down our analysis to these genes allowed us to focus on the most direct tumor driving relationships. At the same time the variables representing principal components allowed us to capture the remaining relationships. In particular the information about sex has not contributed to the results (Fig F in S1 Text) while nearly all the top marker genes (Table 3) identified by this approach are associated with cancer progression.
We found that the functional properties of the predictor genes are consistent with the mechanisms of regulation by the environment as summarised in Fig 7. In addition, many of the identified predictors are previously recognized markers of the microenvironment-mediated tumor progression. One limitation of this study is that while TranNet identifies predictors of such external influences on a tumor it does not establish causality.
TranNet predictors (light blue) in normal tissue predict environment-related changes in tumor (light grey) (A). Cancer-specific functional enrichment of predictor genes (light blue) and inferred environmentally modulated changes in tumor (light gray) (B).
It is important to recognize that some external factors that might impact both normal tissues and tumors and do not necessarily cause tumor progression directly. Even though, many of these factors interact with and/or regulate tumor and are interesting in this respect COPID and hepatitis are likely examples. There is also an association between the incidence of a wide variety of malignancies and diabetes. Identification of such interactions is of fundamental importance for preventing cancer and development of novel treatments.
The concept of tumor environment, has many layers. Previous studies distinguished tumor’s micro and macro environments to capture the effects of the closest neighborhood (such as normal cells, molecules, and blood vessels that surround and feed a tumor cell) and more distant influences respectively . Here we also allow yet another level of influences: organismal environment and germline variations. Our approach provides a proof of principle that control samples can be used to gain insight into the impact of external factors, jointly referred here as the environment, on tumor and tumor progression. The enrichment of the predictor genes in processes associated with immune response suggest that some of these predictors might be in fact sensors of common exposure to chronic infection in normal and tumor tissues. Indeed, it is recognized that cigarette smoking induces lung inflammation, leading to the changes in cellular composition and functions, such as a reduction in the number and properties of ciliated cells. In future studies, it would be interesting to extend the model to allow for a separation of different types of external contributions. In general, it might be difficult to fully deconvolute such processes, as environmental contributors are mostly unknown. However, some environmental factors are mutagenic and in such cases the exposure of the organism to an environmental factor can be estimated by the strength of a corresponding mutational signature [87, 88]. These potential extensions aside, the results obtained by TranNet already provided new insights into the relation between tumor, adjacent control samples, and tumor environment. These relationships might have important implications for the detection and diagnosis of cancers at early stage.
Materials and methods
Gene expression data
Batch effect removed TCGA gene expression data  were downloaded from GitHub repository (https://github.com/mskcc/RNAseqDB) on April 5th, 2022. Requiring matched normal and cancer samples, five types of EMT cancers: Breast Cancer (BRCA), Lung Adenocarcinoma (LUAD), Lung squamous cell carcinoma (LUSC), Prostate Adenocarcinoma (PRAD), and Liver Hepatocellular Carcinoma (LIHC) were selected for the analysis and only patients that had both normal and tumor gene expressions were kept for each cancer type. We excluded genes which are expressed in neither normal (control) nor tumor samples, and the expression level of a gene is assumed as significant if 80% or more of its expression values over samples are greater than a threshold (see Fig G in S1 Text). We then selected the genes that are differentially expressed between control and tumor samples, and expression data for the remaining genes (non-differentially expressed) are compressed in the principal components that preserve 99% of covariate information of the data . From these principal components (PCs), only differentiated (differentially expressed) components are assumed as additional meta nodes in the network, representing the presence of potential covariates in the transition network, coming from the non-differentially expressed genes . The conceptual idea for defining the differentially expressed principal components and for including them as additional variables in the presenting analysis is described in Section ‘Inference of the transition network’ and Fig 2B. Finally, activity (expression of genes and magnitudes of PCs) of each node is standardized with Z-score to bring dissimilar features on a similar scale. The data cohorts for all the five cancers are provided as Table 4.
The leave-one-out test was performed as follows. For each patient i, remove the corresponding expression values xi(⋅) (a row) and yi(⋅) (a row) from the data X and Y. A transition matrix M(−i) is then computed from the remaining data (such as X(−i) × M(−i) ≈ Y(−i)). Then, check if the prediction error on the true response yi(⋅) is less than the error on the random response , such that , where is obtained as permuting yi(⋅) ( is denoted as the absolute summation over the prediction errors on the genes and principal components). A T-test for paired samples was performed to check if this prediction error is less than the random error. The prediction accuracy varies depending on the number of selected predictors as shown in Fig 3 for BRCA, and Fig A in S1 Text for all the five cancers. The number of predictors for each cancer was decided based on different cut-off percents of top genes and principal components with the highest predictive potentials. The test results for the five cancers are summarized in Table 1.
We used a functional network  of human genes for disease studies downloaded from HumanNet v3 (https://www.inetbio.org/humannet). This data set contains 18,459 protein coding genes and 977,495 interactions inferred from various datasets and functional networks. The interactions corresponding to our genes sets are used as the edge validation sets in the analysis of the cancers. The characteristics of the gene expression data and edge validation set are summarized in Table 4.
DE genes (tumor genes): differentially expressed genes selected based on T-test q-value < 0.01; Non-DE Genes: genes whose expressions are not differentiated between control and tumor samples; DE PCs: principal components which are differentiated between control and tumor samples representing potential impacts of Non-DE Genes (T-test q-value) on the transition mapping; TranNet: Pairs of Nodes in the Transition Network (DE Genes + DE PCs); Patients: Patients control and tumor samples are available (Clinical cohorts of the patients are summarized in Table B in S1 File); HN Genes: genes in both DE Genes and HumanNet-FN; HN Edges: Edges within HN Genes in the functional network.
Functional enrichment in ranked lists of genes
Enrichment of GO term biological processes, molecular functions and cellular components was performed using GOrilla  that identifies enriched GO terms in ranked lists of genes. This tool employs a flexible threshold statistical approach to discover GO terms that are significantly enriched at the top of a ranked gene list. The functional enrichment in ranked lists of genes was performed for the target genes sorted by prediction errors (Section ‘Insights into the relation between gene expression in tumor and in matched control samples’) and the predictor genes sorted by predictive potentials (Section ‘Predictor genes are enriched in pathways that imply the features of tumor microenvironment’).
Functional enrichment for subsets of genes
Enrichment of Kegg Pathways (Section ‘Pathway enrichment of genes identified as targets of predictors revealed tumor-stroma interactions’) and GO terms (Section ‘Auto-predictors and their properties’) was performed with a hypergeometric test followed by a Benjamin–Hochberg test for multiple comparison corrections using Kegg pathways  and GO terms . All the genes analysed in the study were used as the background, and Kegg pathways and GO terms enriched with q − value < 0.01 are reported in this study.
S1 Text. Supplemental text.
Tables A and B, Figs A and B and C and D and E and F and G.
S1 File. Supplemental tables.
Lists of the selected high scoring predictor genes and principal components with their predictive potentials for the five cancers.
S2 File. Supplemental tables.
Lists of GO terms enriched for the lists of genes sorted based on prediction errors for each of the five cancers (Genes whose tumor expressions are predicted with smaller errors are listed in higher ranks in the list): Enrichment of biological process, molecular function and cellular component are provided in separate sheets in the same file.
S3 File. Supplemental tables.
Lists of GO terms enriched for the lists of genes sorted based on their predictive potential (PP) scores for each of the five types of cancers: Enrichment of biological process, molecular function and cellular component are provided in separate sheets in the same file.
S4 File. Supplemental tables.
Lists of enriched Kegg pathways for the target genes of the top 10 PP scoring predictors for the five cancers: Enriched pathways and the corresponding p-values.
S5 File. Supplemental tables.
Lists of the auto-predictor genes with their auto-interaction weights, correlation between their normal and tumor tissue expressions and the corresponding p-values.
- 1. Sager R. Expression genetics in cancer: shifting the focus from DNA to RNA. Proc Natl Acad Sci USA. 1997 Feb;94(3):952–955. pmid:9023363
- 2. da Silva-Diz V, Lorenzo-Sanz L, Bernat-Peguera A, Lopez-Cerda M, oz P. Cancer cell plasticity: Impact on tumor progression and therapy response. Semin Cancer Biol. 2018 Dec;53:48–58. pmid:30130663
- 3. Shen S, Clairambault J. Cancer cell plasticity: Impact on tumor progression and therapy response. F1000Res. 2020; 9.
- 4. Yuan S, Norgard RJ, Stanger BZ. Cellular Plasticity in Cancer. Cancer Discov. 2019 Jul;9(7):837–851. pmid:30992279
- 5. Huang S. Genetic and non-genetic instability in tumor progression: link between the fitness landscape and the epigenetic landscape of cancer cells. Cancer Metastasis Rev. 2013 Dec;32(3-4):423–448. pmid:23640024
- 6. Huang S. Tumor progression: chance and necessity in Darwinian and Lamarckian somatic (mutationless) evolution. Prog Biophys Mol Biol. 2012 Sep;110(1):69–86. pmid:22579660
- 7. Wang M, Zhao J, Zhang L, Wei F, Lian Y, Wu Y, et al. Role of tumor microenvironment in tumorigenesis. J Cancer. 2017 Feb;8(5):761–773. pmid:28382138
- 8. Quail DF, Joyce JA. Microenvironmental regulation of tumor progression and metastasis. Nat Med. 2013 Nov;19(11):1423–1437. pmid:24202395
- 9. Balkwill F, Mantovani A. Inflammation and cancer: back to Virchow?. Lancet. 2001 Feb;357(9255):539–545. pmid:11229684
- 10. Coussens LM, Werb Z. Inflammation and cancer. Nature. 2002 Dec;420(6917):860–867. pmid:12490959
- 11. Aran D, Camarda R, Odegaard J, Paik H, Oskotsky B, Krings G, et al. Comprehensive analysis of normal adjacent to tumor transcriptomes. Nat Commun. 2017 Oct;8(1):1017. pmid:29057876
- 12. Mehlen P, Puisieux A. Metastasis: a question of life or death. Nat Rev Cancer. 2006 Jun;6(6):449–458. pmid:16723991
- 13. Amgalan B, Lee H. DEOD: uncovering dominant effects of cancer-driver genes based on a partial covariance selection method. Bioinformatics. 2015 Aug;31(15):2452–2460. pmid:25819079
- 14. Jablonski KP, Pirkl M, Ćevid D, Bühlmann P, Beerenwinkel N. Identifying cancer pathway dysregulations using differential causal effects. Bioinformatics. 2022 Mar;38(6):1550–1559. pmid:34927666
- 15. Ringnér M. What is principal component analysis?. Nat Biotechnol. 2008 Mar;26(3):303–304. pmid:18327243
- 16. Kim J, Kim Y, Kim Y. A gradient-based optimization algorithm for lasso. Journal of Computational and Graphical Statistics. 2008 17(4):994–1009.
- 17. Friedman J, Hastie T, Tibshirani R. Regularization Paths for Generalized Linear Models via Coordinate Descent. J Stat Softw. 2010 30(1):1–22. pmid:20808728
- 18. Donoho DL, Johnstone JM. Ideal spatial adaptation by wavelet shrinkage. biometrika. 1994 81(3):425–455.
- 19. Gafni EM, Bertsekas DP. Two-metric projection methods for constrained optimization. SIAM Journal on Control and Optimization. 1984 22(6):936–964.
- 20. Kim CY, Baek S, Cha J, Yang S, Kim E, Marcotte EM, et al. HumanNet v3: an improved database of human gene networks for disease research. Nucleic Acids Res. 2022 Jan;50(D1):D632–D639. pmid:34747468
- 21. Leopold PL, O’Mahony MJ, Lian XJ, Tilley AE, Harvey BG, Crystal RG. Smoking is associated with shortened airway cilia. PLoS One. 2009 Dec;4(12):e8157. pmid:20016779
- 22. Grant D. M. Detoxification pathways in the liver. J Inherit Metab Dis. 1991 14(4):421–430. pmid:1749210
- 23. Eden E, Navon R, Steinfeld I, Lipson D, Yakhini Z. GOrilla: a tool for discovery and visualization of enriched GO terms in ranked gene lists. BMC Bioinformatics. 2009 Feb;10(48). pmid:19192299
- 24. Si J, Guo R, Xiu B, Chi W, Zhang Q, Hou J, et al. Stabilization of CCDC102B by Loss of RACK1 Through the CMA Pathway Promotes Breast Cancer Metastasis via Activation of the NF-kB Pathway. Front Oncol. 2022 Jul;12:927358. pmid:35957886
- 25. Sato K, Masuda T, Hu Q, Tobo T, Kidogami S, Ogawa Y, et al. Phosphoserine Phosphatase Is a Novel Prognostic Biomarker on Chromosome 7 in Colorectal Cancer. Anticancer Res. 2017 May;37(5):2365–2371. pmid:28476802
- 26. Zhang J, Zhang S, Zhou Y, Qu Y, Hou T, Ge W, et al. KLF9 and EPYC acting as feature genes for osteoarthritis and their association with immune infiltration. J Orthop Surg Res. 2022 Jul;17(1):365. pmid:35902862
- 27. Carmans S, Hendriks JJ, Thewissen K, Van den Eynden J, Stinissen P, Rigo JM, et al. The inhibitory neurotransmitter glycine modulates macrophage activity by activation of neutral amino acid transporters. J Neurosci Res. 2010 Aug;88(11):2420–2430. pmid:20623529
- 28. Dominguez-Soto A, Aragoneses-Fenoll L, Martin-Gayo E, Martinez-Prats L, Colmenares M, Naranjo-Gomez M, et al. The DC-SIGN-related lectin LSECtin mediates antigen capture and pathogen binding by human myeloid cells. Blood. 2007 Jun;109(12):5337–5345. pmid:17339424
- 29. Roy S, Guler R, Parihar SP, Schmeier S, Kaczkowski B, Nishimura H, et al. Batf2/Irf1 induces inflammatory responses in classically activated macrophages, lipopolysaccharides, and mycobacterial infection. J Immunol. 2015 Jun;194(12):6035–6044. pmid:25957166
- 30. Rivero EM, Martinez LM, Bruque CD, Gargiulo L, Bruzzone A, Lüthy IA. Prognostic significance of α and β2-adrenoceptor gene expression in breast cancer patients. Br J Clin Pharmacol. 2019 Sep;85(9):2143–2154. pmid:31218733
- 31. Nass N, Walter S, Jechorek D, Weissenborn C, Ignatov A, Haybaeck J, et al. High neuronatin (NNAT) expression is associated with poor outcome in breast cancer. Virchows Arch. 2017 Jul;471(1):23–30. pmid:28540450
- 32. Fu XW, Song PF, Spindel ER. Role of Lynx1 and related Ly6 proteins as modulators of cholinergic signaling in normal and neoplastic bronchial epithelium. Int Immunopharmacol. 2015 Nov;29(1):93–98. pmid:26025503
- 33. Sun L, Gang X, Li Z, Zhao X, Zhou T, Zhang S, et al. Advances in Understanding the Roles of CD244 (SLAMF4) in Immune Regulation and Associated Diseases. Front Immunol. 2021 Mar;12:648182. pmid:33841431
- 34. Chen JF, Zhang Y, Wilde J, Hansen KC, Lai F, Niswander L. Microcephaly disease gene Wdr62 regulates mitotic progression of embryonic neural stem cells and brain size. Nat Commun. 2014 May;5:3885. pmid:24875059
- 35. Li H, Yang L, Fu H, Yan J, Wang Y, Guo H, et al Association between Gαi2 and ELMO1/Dock180 connects chemokine signalling with Rac activation and metastasis. Nat Commun. 2013 4:1706. pmid:23591873
- 36. AlHossiny M, Luo L, Frazier WR, Steiner N, Gusev Y, Kallakury B, et al. Ly6E/K Signaling to TGFβ Promotes Breast Cancer Progression, Immune Escape, and Drug Resistance. Cancer Res. 2016 Jul;76(11):3376–3386. pmid:27197181
- 37. Ehrlich KC, Lacey M, Ehrlich M. Epigenetics of Skeletal Muscle-Associated Genes in the ASB, LRRC, TMEM, and OSBPL Gene Families. Epigenomes. 2020 Jan;4(1). pmid:34968235
- 38. Zhuang Z Gao C. Development of a Clinical Prognostic Model for Metabolism-Related Genes in Squamous Lung Cancer and Correlation Analysis of Immune Microenvironment. Biomed Res Int. 2022; 6962056. pmid:36110123
- 39. Berg T, Myrbäck TH, Olsson M, Seidegard J, Werkstrom V, Zhou XH, et al. Gene expression analysis of membrane transporters and drug-metabolizing enzymes in the lung of healthy and COPD subjects. Pharmacol Res Perspect. 2014 Aug;2(4):e00054. pmid:25505599
- 40. Ridder DA, Lang MF, Salinin S, Roderer JP, Struss M, Maser-Gluth C, et al. TAK1 in brain endothelial cells mediates fever and lethargy. J Exp Med. 2011 Dec;208(13):2615–2623. pmid:22143887
- 41. Trimarco JD, Nelson SL, Chaparian RR, Wells AI, Murray NB, Azadi P, et al. Cellular glycan modification by B3GAT1 broadly restricts influenza virus infection. Nat Commun. 2022 Oct;13(1):6456. pmid:36309510
- 42. Gao M, Kong W, Huang Z, Xie Z. Int J Mol Sci. Nat Commun. 2020 Apr;21(8).
- 43. Wuttig D, Baier B, Fuessel S, Meinhardt M, Herr A, Hoefling C, et al. Gene signatures of pulmonary metastases of renal cell carcinoma reflect the disease-free interval and the number of metastases per patient. Int J Cancer. 2009 Jul;125(2):474–482. pmid:19391132
- 44. Chow R, Wessels JM, Foster WG. Brain-derived neurotrophic factor (BDNF) expression and function in the mammalian reproductive Tract. Hum Reprod Update. 2020 Jun;26(4):545–564. pmid:32378708
- 45. Beketova E, Owens JL, Asberry AM, Hu CD. PRMT5: a putative oncogene and therapeutic target in prostate cancer. Cancer Gene Ther. 2022 Mar;29(3-4):264–276. pmid:33854218
- 46. Kong J, Sun S, Min F, Hu X, Zhang Y, Cheng Y, et al. Integrating Network Pharmacology and Transcriptomic Strategies to Explore the Pharmacological Mechanism of Hydroxysafflor Yellow A in Delaying Liver Aging. Int J Mol Sci. 2022 Nov;23(22). pmid:36430769
- 47. Toyoda H, Kumada T, Kiriyama S, Tanikawa M, Hisanaga Y, Kanamori A, et al. Association between hepatic steatosis and hepatic expression of genes involved in innate immunity in patients with chronic hepatitis C. Cytokine. 2013 Aug;63(2):145–150. pmid:23673288
- 48. Pieper W, Ignatov A, Kalinski T, Haybaeck J, Czapiewski P, Nass N. The predictive potential of Neuronatin for neoadjuvant chemotherapy of breast cancer. Cancer Biomark. 2021 32(2):161–173. pmid:34092612
- 49. Ali MM, Di Marco M, Mahale S, Jachimowicz D, Kosalai ST, Reischl S, et al. LY6K-AS lncRNA is a lung adenocarcinoma prognostic biomarker and regulator of mitotic progression. Oncogene. 2021 Apr;40(13):2463–2478. pmid:33674747
- 50. Santoni MJ, Kashyap R, Camoin L, Borg JP. The Scribble family in cancer: twentieth anniversary. Oncogene. 2020 Nov;39(47):7019–7033. pmid:32999444
- 51. Zhang G, Shang H, Liu B, Wu G, Wu D, Wang L, et al. Increased ATP2A1 Predicts Poor Prognosis in Patients With Colorectal Carcinoma. Front Genet. 2020 Jun;13:661348.
- 52. Green AR, Krivinskas S, Young P, Rakha EA, Paish EC, Powe DG, et al. Loss of expression of chromosome 16q genes DPEP1 and CTCF in lobular carcinoma in situ of the breast. Breast Cancer Res Treat. 2009 Jan;113(1):59–66. pmid:18213475
- 53. Wu Q, Wang D, Zhang Z, Wang Y, Yu W, Sun K, et al. DEFB4A is a potential prognostic biomarker for colorectal cancer Oncol Lett. 2020 Oct;20(4):114. pmid:32863927
- 54. Suzuki C, Takahashi K, Hayama S, Ishikawa N, Kato T, Ito T, et al. Identification of Myc-associated protein with JmjC domain as a novel therapeutic target oncogene for lung cancer. Mol Cancer Ther. 2007 Feb;6(2):542–551. pmid:17308053
- 55. Zhang Y, Wei H, Fan L, Fang M, He X, Lu B, et al. CLEC4s as Potential Therapeutic Targets in Hepatocellular Carcinoma Microenvironment. Front Cell Dev Biol. 2021 9:681372. pmid:34409028
- 56. Burnett RM, Craven KE, Krishnamurthy P, Goswami CP, Badve S, Crooks P, et al. Organ-specific adaptive signaling pathway activation in metastatic breast cancer cells. Oncotarget. 2015 May;6(14):12682–12696. pmid:25926557
- 57. Xiao YF, Li BS, Liu JJ, Wang SM, Liu J, Yang H, et al. Role of lncSLCO1C1 in gastric cancer progression and resistance to oxaliplatin therapy. Clin Transl Med. 2022 Apr;12(4):e691. pmid:35474446
- 58. Faltermeier CM, Drake JM, Clark PM, Smith BA, Zong Y, Volpe C, et al. Functional screen identifies kinases driving prostate cancer visceral and bone metastasis. Proc Natl Acad Sci USA. 2016 Jan;113(2):E172–181. pmid:26621741
- 59. White DL, Li D, Nurgalieva Z, El-Serag HB. Genetic variants of glutathione S-transferase as possible risk factors for hepatocellular carcinoma: a HuGE systematic review and meta-analysis. Am J Epidemiol. 2008 Feb;167(4):377–389. pmid:18065725
- 60. Yang ML, Zhang JH, Li S, Zhu R, Wang L. SLC13A4 Might Serve as a Prognostic Biomarker and be Correlated with Immune Infiltration into Head and Neck Squamous Cell Carcinoma. Pathol Oncol Res. 2021 Nov;27:1609967. pmid:34840533
- 61. Masjedi S, Zwiebel LJ, Giorgio TD. Olfactory receptor gene abundance in invasive breast carcinoma. Sci Rep. 2019 Sep;9(1):13736. pmid:31551495
- 62. Pan J, Dai Q, Xiang Z, Liu B, Li C. Three Biomarkers Predict Gastric Cancer Patients’ Susceptibility To Fluorouracil-based Chemotherapy J Cancer. 2019 Jun;10(13):2953–2960. pmid:31281472
- 63. Zhang K, Zhao Z, Yu J, Chen W, Xu Q, Chen L. LncRNA FLVCR1-AS1 acts as miR-513c sponge to modulate cancer cell proliferation, migration, and invasion in hepatocellular carcinoma J Cell Biochem. 2018 Jul;119(7):6045–6056. pmid:29574975
- 64. Wang J, Shidfar A, Ivancic D, Ranjan M, Liu L, Choi MR, et al. Overexpression of lipid metabolism genes and PBX1 in the contralateral breasts of women with estrogen receptor-negative breast cancer Int J Cancer. 2017 Jun;140(11):2484–2497. pmid:28263391
- 65. Vaes RDW, Reynders K, Sprooten J, Nevola KT, Rouschop KMA, Vooijs M, et al. Identification of Potential Prognostic and Predictive Immunological Biomarkers in Patients with Stage I and Stage III Non-Small Cell Lung Cancer (NSCLC): A Prospective Exploratory Study. Cancers (Basel). 2021 Dec;13(24).
- 66. Zheng Y, Ming P, Zhu C, Si Y, Xu S, Chen A, et al. Hepatitis B virus X protein-induced SH2 domain-containing 5 (SH2D5) expression promotes hepatoma cell growth via an SH2D5-transketolase interaction. J Biol Chem. 2019 Mar;294(13):4815–4827. pmid:30659097
- 67. Shen J, Terry MB, Gammon MD, Gaudet MM, Teitelbaum SL, Eng SM, et al. IGHMBP2 Thr671Ala polymorphism might be a modifier for the effects of cigarette smoking and PAH-DNA adducts to breast cancer risk Breast Cancer Res Treat. 2006 Sep;99(1):1–7. pmid:16752224
- 68. Pang B, Sui S, Wang Q, Wu J, Yin Y, Xu S. Upregulation of DLEU1 expression by epigenetic modification promotes tumorigenesis in human cancer. J Cell Physiol. 2019 Aug;234(10):17420–17432. pmid:30793303
- 69. Ramello MC, Núñez NG, Tosello Boari J, Bossio SN, Canale FP, Abrate C. et al. Polyfunctional KLRG-1+CD57+ Senescent CD4+ T Cells Infiltrate Tumors and Are Expanded in Peripheral Blood From Breast Cancer Patients. Histopathology. 2020 Jul;12:713132.
- 70. Ding N, Li R, Shi W, He C. CENPI is overexpressed in colorectal cancer and regulates cell migration and invasion. Gene. 2018 Oct;674:80–86. pmid:29936263
- 71. Wang Z, Pei H, Liang H, Zhang Q, Wei L, Shi D, et al. Construction and Analysis of a circRNA-Mediated ceRNA Network in Lung Adenocarcinoma. Onco Targets Ther. 2021 Jun;14:3659–3669. pmid:34135596
- 72. Cheng J, Li Y, Wang X, Dong Z, Chen Y, Zhang R, et al. Response Stratification in the First-Line Combined Immunotherapy of Hepatocellular Carcinoma at Genomic, Transcriptional and Immune Repertoire Levels. J Hepatocell Carcinoma. 2021 Oct;8:1281–1295. pmid:34737983
- 73. Wu W, Warner M, Wang L, He WW, Zhao R, Guan X, et al. Drivers and suppressors of triple-negative breast cancer. Proc Natl Acad Sci USA. 2021 Aug;118(33). pmid:34389675
- 74. Lulli M, Del Coco L, Mello T, Sukowati C, Madiai S, Gragnani L, et al. DNA Damage Response Protein CHK2 Regulates Metabolism in Liver Cancer. Cancer Res. 2021 Jun;81(11):2861–2873. pmid:33762357
- 75. Liu D, Yang Y, Yan A, Yang Y, et al. SPOCD1 accelerates ovarian cancer progression and inhibits cell apoptosis via the PI3K/AKT pathway. Onco Targets Ther. 2020 Jan;13:351–359. pmid:32021280
- 76. Shinmura K, Kato H, Kawanishi Y, Igarashi H, Inoue Y, Yoshimura K, et al. WDR62 overexpression is associated with a poor prognosis in patients with lung adenocarcinoma. Mol Carcinog. 2017 Aug;56(8):1984–1991. pmid:28277612
- 77. Cantelli G, Orgaz JL, Rodriguez-Hernandez I, Karagiannis P, Maiques O, Matias-Guiu X, et al. TGF-β-Induced Transcription Sustains Amoeboid Melanoma Migration and Dissemination. Curr Biol. 2015 Nov;25(22):2899–2914. pmid:26526369
- 78. Kodama T, Kodama M, Jenkins NA, Copeland NG, Chen HJ, Wei Z. Ring Finger Protein 125 Is an Anti-Proliferative Tumor Suppressor in Hepatocellular Carcinoma. Cancers (Basel). 2022 May;14(11). pmid:35681566
- 79. Lin Y, Zhou X, Peng W, Wu J, Wu X, Chen Y, et al. Expression and clinical implications of basic leucine zipper ATF-like transcription factor 2 in breast cancer. BMC Cancer. 2021 Sep;21(1):1062. pmid:34565331
- 80. Zhang Z, Ji W, Huang J, Zhang Y, Zhou Y, Zhang J, et al. Characterization of the tumour microenvironment phenotypes in malignant tissues and pleural effusion from advanced osteoblastic osteosarcoma patients. Clin Transl Med. 2022 Nov;12(11):e1072. pmid:36305631
- 81. Meng M, Wu YC. LMX1B Activated Circular RNA GFRA1 Modulates the Tumorigenic Properties and Immune Escape of Prostate Cancer. J Immunol Res. 2022; 7375879. pmid:35832649
- 82. Ma C, Kang W, Yu L, Yang Z, Ding T, et al. AUNIP Expression Is Correlated With Immune Infiltration and Is a Candidate Diagnostic and Prognostic Biomarker for Hepatocellular Carcinoma and Lung Adenocarcinoma. Front Oncol. 2020 Dec;10:590006. pmid:33363020
- 83. Duss S, Brinkhaus H, Britschgi A, Cabuy E, Frey DM, Schaefer DJ, et al. Mesenchymal precursor cells maintain the differentiation and proliferation potentials of breast epithelial cells. Breast Cancer Res. 2014 Jun;16(3):R60. pmid:24916766
- 84. Chen Y, Shen L, Chen B, Han X, Yu Y, Yuan X, et al. The predictive prognostic values of CBFA2T3, STX3, DENR, EGLN1, FUT4, and PCDH7 in lung cancer. Ann Transl Med. 2021 May;9(10):843. pmid:34164477
- 85. Ngoc PCT, Tan SH, Tan TK, Chan MM, Li Z, Yeoh AEJ, et al. Identification of novel lncRNAs regulated by the TAL1 complex in T-cell acute lymphoblastic leukemia. Leukemia. 2018 Oct;32(10):2138–2151. pmid:29654272
- 86. Rutkowski MR, Svoronos N, Perales-Puchalt A, Conejo-Garcia JR, et al. The Tumor Macroenvironment: Cancer-Promoting Networks Beyond Tumor Beds. Adv Cancer Res. 2015 May;128:235–262. pmid:26216635
- 87. Alexandrov LB, Kim J, Haradhvala NJ, Huang MN, Tian Ng AW, Wu Y, et al. The repertoire of mutational signatures in human cancer. Nature. 2020 Feb;578(7793):94–101. pmid:32025018
- 88. Kim YA, Hodzic E, Amgalan B, Saslafsky A, Wojtowicz D, Przytycka TM. Mutational Signatures as Sensors of Environmental Exposures: Analysis of Smoking-Induced Lung Tissue Remodeling. Biomolecules. 2022 Sep;12(10). pmid:36291592
- 89. Wang Q, Armenia J, Zhang C, Penson AV, Reznik E, Zhang L, et al. Unifying cancer and normal RNA sequencing data from different sources. Sci Data. 2018 Apr;5:180061. pmid:29664468
- 90. Kanehisa M, Goto S. KEGG: kyoto encyclopedia of genes and genomes. Nucleic Acids Res. 2000 Jan;28(1):27–30. pmid:10592173
- 91. Carbon S, Ireland A, Mungall CJ, Shu S, Marshall B, Lewis S, et al. AmiGO: online access to ontology and annotation data. Bioinformatics. 2009 Jan;25(2):288–289. pmid:19033274