Pan-Cancer Analysis for Immune Cell Infiltration and Mutational Signatures Using Non-Negative Canonical Correlation Analysis

Simple Summary

Mutational signatures have been used to infer links between tumor development and mutational processes. However, the functional roles of mutational signatures in tumor cells and their influence on the tumor microenvironment remain unclear. This study investigates the impact of tumor microenvironments and immune activities on mutational signatures using integrative analysis of whole-exome sequencing and RNA sequencing data obtained from the same patients. This study applies non-negative canonical correlation analysis to the datasets and identifies the mutational signatures related to the immune cell infiltration in each tumor type. This approach helps to search for the survival and immunological aspects of the mutational signatures in tumors.

Abstract

Mutational signatures indicate the mutational processes and substitution patterns in cancer cell genomes. However, the functional consequences of mutational signatures remain unclear, and there have been no comprehensive systematic studies to examine the relationships between the mutational signatures and the immune cell infiltration. Here, the relationship between mutational signatures and immune cell infiltration using non-negative canonical correlation analysis based on 8927 patients across 25 tumor types was investigated. By inspecting mutational signatures with the maximal coefficients determined by the non-negative canonical correlation analysis, the study identified mutational signatures related to immune cell infiltration composed of tumor microenvironments. The analysis was validated by showing that the genes associated with the identified mutational signatures were linked to overall survival by a Kaplan–Meier curve and a log-rank test and were mainly related to immunity by gene set enrichment analysis. These results will help expand our knowledge of tumor biology and recognize the functional roles and associations of immune systems with mutational signatures.

1. Introduction

The causes of somatic mutations in patients with tumors are only partially known. For example, exogenous mutagens are introduced by tobacco smoking or ultraviolet exposure, and endogenous mutagens are introduced by spontaneous methylcytosine deamination in the 5${}^{\prime }$-CG dinucleotide context, resulting in C-to-T mutations [1,2]. Furthermore, DNA repair deficiency occurs through homologous recombination deficiency or DNA mismatch repair deficiency, and enzymatic DNA editing is caused by the functional activity of the APOBEC protein family induced by cytidine deaminase enzymes [3,4]. However, the complete identification of many mutational processes is very difficult.

To address these difficulties, mutation signatures representing specific patterns have been developed from a large catalog of mutations generated by several types of cancer genome sequencing [1,5,6]. Mutational signatures are distinctive combinations of somatic alteration resulting from a specific cause raised generally in the cancer genome. Single nucleotide variants (SNVs) are generally classified into six mutational spectra (C:G > A:T, C:G > G:C, C:G > T:A, T:A > A:T, T:A > C:G, T:A > G:C); in the trinucleotide context, 96 combinations of SNVs are possible. Tumor samples with 96 variables are then decomposed using a non-negative matrix factorization (NMF) method, and new matrices are generated using mutational signatures [7]. Mutational signatures have helped understand the mutational processes and causes of tumors; many cancer researchers have also used mutational signatures for their analyses [8,9,10].

Recent studies have revealed that the apolipoprotein B mRNA editing enzyme, catalytic polypeptide-like (APOBEC) mutational signature is potentially associated with immune responses in non-small cell lung cancer [11,12]. Two members of the APOBEC family, APOBEC3A and APOBEC3B, are known to contribute to mutations in cancers by generating C > T and C > G mutations in TCN trinucleotides [13,14]. The authors characterized the APOBEC mutational signatures related to the response to immune checkpoint blockade therapy. These studies were mostly conducted on non-small cell lung cancer and demonstrated the potential of APOBEC signatures. However, the functional characteristics of most mutational signatures remain unknown. Moreover, until now, to my knowledge, there have been no systematic analyses to examine the connections between the mutational signatures and the immune cell infiltration.

Here, this study investigated the immunological effects of mutational signatures. Firstly, the mutational signatures and the immune cell infiltration were inferred from somatic mutation and gene expression profiles in each tumor sample, respectively. Then, the relationships between the mutational signatures and the immune cell infiltration were examined based on canonical correlation analysis (CCA). CCA is a useful approach to determining the correlation between two multivariate datasets [15]. Both datasets are projected into lower-dimensional spaces, where their correlation is maximal. Generally, CCA is considered as a dimensionality reduction method [16,17], however, it has been applied for various purposes, including classification and clustering [18,19,20]. CCA has also been employed to solve many biological problems, mainly using the integrative analysis of multi-omics datasets [21,22,23]. As shown in some previous works using CCA, the estimated coefficients of the variables by the CCA can reveal the importance of the variables for another input dataset of the CCA. Thus, this study applied the CCA to infer the functional roles of each mutational signature. However, traditional CCA is sensitive to outliers and noise. In addition, the occurrence of negative values of the coefficients hinders the intuitive interpretation of the results [24]. This study used regularization strategies to solve the ill-posed problem and non-negative constraints that force all input elements and coefficients to be zero or positive values, as suggested by Sigg et al. [25]. Experiments with 8927 tumor patients across 25 tumor types in The Cancer Genome Atlas (TCGA) showed that these methods could be useful for identifying mutational signatures related to immune cell infiltration in each tumor type.

To date, potential implications of mutational signatures for their functional and immunological effects remain to be addressed in tumor patients. Hence, the main objective of this study is to decipher mutational signatures related to immune cell infiltration in each tumor type and to assess their functional and clinical effects.

2. Materials and Methods

2.1. Data Preparation

Somatic mutation calls and gene expression profiles for Genomic Data Commons (GDC) TCGA tumor samples were obtained from UCSC Xena [26]. UCSC Xena provide the somatic mutation data detected from whole-exome sequencing, and for this experiment, the somatic mutation called using MuTect2 was downloaded. The gene expression profiles were achieved from RNAseq datasets processed using the HTseq [27] pipeline. Although the downloaded expression levels were represented as fragments per kilobase of transcript per million fragments mapped (FPKM), these were transformed to transcripts per kilobase million (TPM) values for further analyses [28,29].

Among these TCGA tumor samples, the experiments were carried out with 8927 tumor patients across 25 major tumor types, wherein more than 100 patients had identical TCGA barcode IDs between somatic mutation data and gene expression profiles. The tumor types and numbers of samples used in the experiments are shown in

Table 1. Tumor types and the number of samples used in the experiments.

Tumor Type	Samples
Bladder urothelial carcinoma (BLCA)	409
Breast invasive carcinoma (BRCA)	974
Cervical squamous cell carcinoma and endocervical adenocarcinoma (CESC)	287
Colon adenocarcinoma (COAD)	397
Esophageal carcinoma (ESCA)	161
Glioblastoma multiforme (GBM)	150
Head and neck squamous cell carcinoma (HNSC)	494
Kidney renal clear cell carcinoma (KIRC)	333
Kidney renal papillary cell carcinoma (KIRP)	279
Brain lower grade glioma (LGG)	504
Liver hepatocellular carcinoma (LIHC)	360
Lung adenocarcinoma (LUAD)	510
Lung squamous cell carcinoma (LUSC)	490
Ovarian serous cystadenocarcinoma (OV)	273
Pancreatic adenocarcinoma (PAAD)	170
Pheochromocytoma and paraganglioma (PCPG)	179
Prostate adenocarcinoma (PRAD)	493
Rectum adenocarcinoma (READ)	134
Sarcoma (SARC)	236
Skin cutaneous melanoma (SKCM)	466
Stomach adenocarcinoma (STAD)	373
Testicular germ cell tumors (TGCT)	145
Thyroid carcinoma (THCA)	488
Thymoma (THYM)	119
Uterine corpus endometrial carcinoma (UCEC)	527

.

2.2. Estimation of Mutational Signatures and Immune Cell Infiltration

All subsequent analyses were performed using R 3.6.3. Among the somatic mutations, SNVs were only extracted to determine the mutational processes in each tumor sample, and the mutational process was inferred, using the R package deconstructSigs 1.8.0 [30], from the SNVs. deconstructSigs estimates the similarity to the known Catalogue of Somatic Mutations in Cancer (COSMIC) mutational process version 2 [31] in individual tumor samples. The approach uses a multiple linear regression model to iteratively explore coefficients of pre-defined mutational signatures that are restricted to positive values.

From the gene expression profile, the TPM values were used as an input of EPIC [32], a tool for estimating cell subtype abundance from bulk gene expression profiles. EPIC predicts the composition of each cell type, B cells, CD4 T cells, CD8 T cell, endothelial cells, macrophages, natural killer (NK) cells, cancer-associated fibroblasts (CAFs), and other cells. The experiments were performed using the R package EPIC 1.1.5.

2.3. Non-Negative CCA and Estimation of Coefficients

Figure 1 is an overview of the analysis of the relationships between the immune cell infiltration and mutational signatures. The basic methods and the estimation of the coefficients were proposed by Siggs et al., and the experiments were carried out using the R package nscancor 0.6.1–25 [25,33], followed by regularization using L2-norm. The non-negative CCA approach to capture the relationships between mutational signatures and immune cell infiltration is outlined as follows:

The traditional CCA finds relationships between two multidimensional variables to maximize their correlation with the linear combination as follows:

$$\underset{\omega ,\beta }{argmax}corr(\omega \mathrm{A},\beta \mathrm{M})$$

(1)

where A and M are the matrices of the cell subtype abundance and mutational signatures, respectively. $\omega$ and $\beta$ are coefficient vectors for the linear projections of the two variables. The main objective of the CCA is to detect the coefficient vectors $\omega$ and $\beta$ that maximize the correlation between $\omega$A and $\beta$M using linear projections. The optimal solution for CCA can be discovered by the direct transformation to a generalized eigenvalue problem [34].

However, in contrast to the conventional CCA, in this study, $\omega$ and $\beta$ must be equal to or larger than zero for the non-negative constraint. Thus, $\omega$ and $\beta$ were determined using an iterative regression solver based on a least-squares formulation that was extended using regularization techniques [25,35,36,37]. The optimization problem in Equation (1) can be reformulated equivalently using the following equation:

$$(\widehat{\omega },\widehat{\beta })=\underset{\omega ,\beta }{argmin}\left|\right|\omega \mathrm{A}-{\beta \mathrm{M}\left|\right|}^2.$$

(2)

Equation (1) was converted to a problem to search for a minimum distance in a projection space using the linear transformation of A and M, similar to finding the coefficients using the least-squares methods in a linear regression problem.

For a given value of $\widehat{\omega }$, Equation (2) can be optimized by finding the minimum mean-squared error for the regression coefficients $\beta$ as follows:

$$\widehat{\beta }=\underset{\beta }{argmin}\left|\right|\widehat{\omega }\mathrm{A}-\beta \mathrm{M}{\left|\right|}^2.$$

(3)

In Equation (3), $\widehat{\beta }$ is estimated using the following rescaling step:

$$\widehat{\beta }:=\frac{\widehat{\beta }}{\left|\right|\widehat{\beta }\mathrm{M}\left|\right|}.$$

(4)

For example, at the (t+1)th iteration step, ${\widehat{\omega }}^t$ is a constant, and a regression step is performed to estimate the coefficient ${\widehat{\beta }}^{t+1}$. Then, ${\widehat{\beta }}^{t+1}$ is considered a fixed constant, and the coefficient ${\widehat{\omega }}^{t+1}$ is determined to be similar to that in Equations (3) and (4), respectively. The procedure is iteratively continued until the coefficients $\omega$ and $\beta$ converge.

To apply a penalty term at the regression coefficients for regularization, Equation (3) is modified as follows:

$$\widehat{\beta }=\underset{\beta }{argmin}\left(\right||\widehat{\omega }\mathrm{A}-{\beta \mathrm{M}\left|\right|}^2+{\lambda }_1\sum _{g=1}^G\left|\right|{\beta }_g{\left|\right|}^2),$$

(5)

where ${\beta }_g$ is the coefficient of the gth component of CCA. Similarly, an optimal $\omega$ vector can be determined by adding the constraint as follows:

$$\widehat{\omega }=\underset{\omega }{argmin}\left(\right||\widehat{\beta }\mathrm{M}-{\omega \mathrm{A}\left|\right|}^2+{\lambda }_2\sum _{h=1}^H\left|\right|{\omega }_h{\left|\right|}^2),$$

(6)

where ${\omega }_h$ is the coefficient of the hth component of CCA represented as the $\omega$ vector. ${\lambda }_1$ and ${\lambda }_2$ are the shrinkage parameters for regularization in Equations (5) and (6).

A high coefficient for a mutational signature indicates that it is highly linked to the immune cell infiltration. In the experiments, many of the mutational signatures were observed as zero; therefore, mutational signatures with the number of non-zero values exceeding 20% of the total samples of the tumor type were used. The glmnet() function’s parameter alpha was specified as zero for Ridge regression, while the other parameters were left as defaults.

2.4. Finding Genes Correlated to Mutational Signatures and Survival Analysis

The Pearson correlation coefficient (PCC) was measured between the number of somatic mutations in a gene and the mutational signature value inferred by deconstructSigs. Genes with the highest PCC, whose value was more than 0.1 in each tumor type, were selected for further survival analysis. Overall survival (OS) information was obtained using UCSC Xena [26]. The samples were divided into two groups based on the gene expression levels of the highest correlated genes using the R package maxstat 0.7–25, which estimates a optimal cut point based on maximally selected rank statistics [38]. The statistical significance (p-value) of the survival analysis was calculated using a log-rank test.

2.5. Geneset Enrichment Analysis Using a Pre-Ranked Gene List

The GSEApreranked approach was used to identify genesets associated with the identified mutational signatures in each tumor type. Basically, gene set enrichment analysis (GSEA) is a method that is used to identify overrepresented genesets from gene expression data using the Kolmogorov–Smirnov statistic [39,40]. However, GSEApreranked executes GSEA against a ranked list of genes and calculates an enrichment score based on the pre-defined genesets. In this experiment, the ranking was measured by the PCC between the gene expression levels and the values of the mutational signatures estimated by deconstructSigs. The GSEApreranked experiment was carried out using the Java-based GSEA 4.2.3 tool (Broad Institute, Cambridge, MA, USA) with default parameters. For the genesets, c5.go.bp.v7.5.1 derived from Gene Ontology (GO) biological process (BP) were used, which were annotated in the Molecular Signatures Database (MSigDB) [41].

3. Results

3.1. Overall Representation of Mutational Signatures

This study first examined the prevalence of mutational signatures and cell subtype abundances in 25 TCGA tumor types using deconstructSigs and EPIC, respectively.

Table 2. Average occurrence of mutational signatures across 25 tumor types.

Mutational Signature	Average
Signature 1	0.1250
Signature 2	0.0605
Signature 3	0.0801
Signature 4	0.0392
Signature 5	0.0472
Signature 6	0.0174
Signature 7	0.0458
Signature 8	0.0403
Signature 9	0.0218
Signature 10	0.0106
Signature 11	0.0120
Signature 12	0.0203
Signature 13	0.0519
Signature 14	0.0083
Signature 15	0.0102
Signature 16	0.1335
Signature 17	0.0135
Signature 18	0.0169
Signature 19	0.0176
Signature 20	0.0167
Signature 21	0.0146
Signature 22	0.0069
Signature 23	0.0050
Signature 24	0.0095
Signature 25	0.0134
Signature 26	0.0175
Signature 27	0.0042
Signature 28	0.0136
Signature 29	0.0125
Signature 30	0.0282

shows the average occurrence of mutational signatures across 25 tumor types, and Figure 2 presents the tumor type-specific profiles. Mutation signatures 16 and 1 were predominantly represented, as shown in Table 2, with mean values of 0.1335 and 0.1250, respectively. Next, tumor type-specific over- and underrepresented mutational signatures were investigated, and the representatives were substantially different. Mutational signatures 16 and 1 were widely represented in many tumor types, with values greater than 0.1 for 17 (BLCA, BRCA, ESCA, GBM, HNSC, KIRC, KIRP, LGG, LIHC, LUSC, PAAD, PCPG, PRAD, SARC, TGCT, THCA, THYM) and 12 tumor types (BRCA, CESC, COAD, ESCA, GBM, HNSC, LGG, PAAD, PRAD, READ, STAD, UCEC), respectively. Furthermore, the number of tumor types with maximal values for mutational signatures 16 or 1 was eight (ESCA, HNSC, KIRC, KIRP, LIHC, PCPG, THCA, THYM for signature 16; COAD, GBM, LGG, PAAD, PRAD, READ, STAD, UCEC for signature 1). However, signature 3 was the best for BRCA, OV, SARC, and TGCT, and signature 4 was the best for LUAD and LUSC. Signature 3 is related to BRCA1 and BRCA2 mutations, which are generally observed in breast and ovarian tumors [42,43], whereas signature 4 is related to tobacco smoking [44]. Moreover, BLCA, CESC, and SKCM showed the best occurrence of signatures 13, 2, and 7, respectively, in contrast to other tumor types. Signatures 13 and 2 are known as APOBEC mutation signatures, whereas signature 7 is known as a UV signature.

Next, the study investigated which genes were commonly mutated and which sequence alteration was mainly observed. Supplementary Figure S1 shows PCCs between the mutational signatures and the highly mutated top 20 genes, and Supplementary Figure S2 represents the correlation between the mutational signatures and the sequence alteration. As expected, tumor-related genes such as TP53, PIK3CA, and KRAS were highly mutated in each tumor type, and the genes were positively correlated with certain mutational signatures. For example, in SKCM, signature 7 (UV signature) was the most representative (Figure 2), and the signature was positively correlated with most of the top 20 genes. However, this property was not observed in any of the tumor types. For instance, although GBM was overrepresented in signature 1 (aging signature), positively correlated genes among the top 20 genes were detected in signatures 10 and 14. When the genomic modifications were examined, the amount of PCC was not identical to the representative signatures (Supplementary Figure S2).

3.2. Mutational Signatures Associated with the Cell Subtype Composition

To the mutational process and cell subtype abundance datasets for each tumor type, this study applied non-negative CCA and identified the mutational signatures associated with immune cell infiltration based on a canonical correlation analysis. Across the 25 tumor types, signatures 2, 1, and 13, in descending order, were the most highly ranked mutational signatures. Among these, signatures 2 and 13 are APOBEC signatures, whose associations with immunity have been previously reported [11,12]. However, the absolute values of the three coefficients were small (0.0477, 0.0468, 0.0407), and it was difficult to clearly determine the activities of mutational signatures in immunity and tumor microenvironments by this experiment.

Therefore, the same experiments were performed for each tumor type. Supplementary Figure S3 represents the canonical correlation coefficients and their p-values in each tumor type. These were statistically significant in all tumor types (p-values range from $1.10\times {10}^{-14}$ to $3.14\times {10}^{-2}$). Figure 3 shows the coefficients of the first canonical component of the mutational signature dataset, and Supplementary Figure S3 is the coefficients of the second canonical component. Notably, the mutational signatures associated with immune cell infiltration identified by this non-negative CCA approach showed different patterns with the overall abundance of mutational signatures (shown in Figure 2) in each tumor type, as well as across the 25 tumor types (shown in Table 2).

In the experimental results, in LUSC, the mutational signature with the highest coefficients in the first canonical component of our CCA results was signature 13, well-known as an APOBEC signature, with a coefficient of 0.3348. The results were in line with previous studies that were mentioned above [11,12]. In addition, in the second canonical component, another APOBEC signature, signature 2, appeared to have the highest coefficient (0.3292) (Supplementary Figure S3). The APOBEC signatures were also observed as the highest coefficients in the first component of several tumor types, such as BRCA (signature 13, coeff. = 0.2253), HNSC (signature 2, coeff. = 0.3287), STAD (signature 2, coeff. = 0.3819), SKCM (signature 2, coeff. = 0.3233), SARC (signature 13, coeff. = 0.7919), as well as LUSC.

Next, we performed survival analysis based on the gene with the highest correlation to the identified mutational signatures. For several tumor types, all PCCs were very low (<0.1), and were thus excluded from the survival analysis. Figure 4 shows the Kaplan–Meier plots for each tumor type. The genes with the highest PCC in each tumor type were clearly different from the highly mutated genes represented in the row labels of Supplementary Figure S1. Although some genes, such as TP53, were observed among the highly mutated genes, most of the genes with the highest PCCs were not detected among the top 20 highly mutated genes. Survival analysis using Kaplan–Meier plots and log-rank tests showed statistically significant differences (p-value < 0.05) in many tumor types (FAT2 in BLCA, UTRN in BRCA, TP53 in COAD, DNAH5 in ESCA, PBRM1 in KIRC, EGFR in LGG, LYST in LUSC, and PKHDL1 in SKCM), and the p-value for CACNA1E in LIHC was also very close to 0.05. It was noted that TP53, PBRM1, and EGFR were involved in the Cancer Gene Census in the COSMIC database [45].

3.3. Genesets Associated with the Identified Mutational Signatures

To further validate the approach used, this study performed GSEA [40] against a pre-ranked list of genes (GSEApreranked) by the PCCs between the mutational signatures and the expression values of the genes. Figure 5 shows the results of GSEApreranked with pre-defined GO BP genesets for nine tumor types (BRCA, HNSC, LGG, LUAD, LUSC, PRAD, SKCM, THCA, and UCEC) with the largest number of samples (>450 samples), and Supplementary Tables S1–S50 show the complete results with FDR < 0.25 for all tumor types.

When firstly showing the enriched GO BP terms based on the positively correlated gene, in BRCA, immune-related terms such as adaptive immune response, antigen receptor-mediated signaling pathway, and immune response regulating cell surface receptor signaling pathway were mainly observed. These properties have also been observed in other tumor types. For example, in HNSC and LUSC, the GO BP terms with the lowest FDR values were adaptive immune response, antigen processing, and presentation of peptide antigen with FDR $<1.0\times {10}^{-6}$. For SKCM, terms such as natural killer cell-mediated immunity (FDR = 0.0094), positive regulation of natural killer cell-mediated cytotoxicity (FDR = 0.010), and regulation of natural killer cell-mediated immunity (FDR = 0.012) were among the top five GO BP terms. In addition, for THCA, GO BP terms, such as adaptive immune response, immune response regulating cell surface receptor signaling pathway, and leukocyte-mediated immunity, were uncovered with FDR $<1.0\times {10}^{-6}$. For LGG, brain-related terms, such as neurons or synapses, were primarily discovered, but some immune-related terms, including antigen processing and presentation of peptide antigen via MHC class I (FDR = 0.030) were also enriched.

In some tumor types, negatively correlated genes were mainly enriched in immune-related terms. In LUAD, the two terms with the lowest FDR values were the B cell receptor signaling pathway (FDR <$1.0\times {10}^{-6}$) and phagocytosis recognition (FDR = 0.0060). In addition, in UCEC, terms for GO BP such as adaptive immune response (FDR <$1.0\times {10}^{-6}$) and B cell receptor signaling pathway (FDR <$1.0\times {10}^{-6}$) were mainly observed.

In addition, GSEApreranked results using other tumor types not represented in Figure 5 also revealed that immune-related terms were highly enriched. For instance, the top-ranked GO BP terms for positively correlated genes in BLCA were the B cell receptor signaling pathway (FDR = 0.0036). In KIRC, terms such as negative regulation of immune response (FDR = $3.01\times {10}^{-5}$), positive regulation of interleukin 4 production (FDR = $3.12\times {10}^{-5}$), and negative regulation of immune system processes (FDR = $3.15\times {10}^{-5}$) were frequently observed. Furthermore, when searching for negatively correlated genes, for COAD, many of the enriched GO terms for negatively correlated genes were immune-related, including activation of the immune response, adaptive immune response, antigen receptor-mediated signaling pathway, and B cell-mediated immunity, whose FDRs were lower than $1.0\times {10}^{-6}$. As another example, STAD also revealed many immune-related terms with statistical significance, including B cell receptor signaling pathway (FDR $5.83\times {10}^{-5}$), B cell homeostasis (FDR = 0.0050), antigen processing and presentation of exogenous antigen (FDR = 0.0053), regulation of leukocyte adhesion to vascular endothelial cells (FDR = 0.0073), and antigen processing and presentation of exogenous peptide antigen (FDR = 0.0075).

4. Discussion

The analysis of the mutational signatures of tumor patients has expanded the existing understanding of the tumor genome and provided new insights into tumor initiation and progression. However, identifying the biological processes and functions underlying somatic mutagenesis remains challenging. This study searched for mutational signatures associated with immune cell infiltration in patients with tumors using non-negative CCA and verified that the approach would be useful to uncover their functional roles in survival and immunological contexts. To my knowledge, this study is the first work to systematically investigate the functional roles of the mutational signatures. Although some variables (mutational signatures) may represent negative effects on the immune cell infiltration, the non-negative constraints of the coefficients help to interpret the influences of the variables using the magnitude of the coefficient values, similar to NMF which has commonly been used in many biological data analyses [46,47,48].

The composition of mutational signatures in each tumor sample was estimated using deconstructSigs. Although the COSMIC mutational signature version 3 was constructed recently, deconstructSigs did not support version 3. Moreover, single base substitution (SBS) signatures were partitioned into 60 signatures, which is an excessive amount of variables compared to the number of samples, and it would lead to making the interpretation difficult. Mutational signature version 2 was therefore used in this experiment.

The estimation of cell subtype composition based on the bulk gene expression profiles can facilitate advanced analysis of heterogeneous tumor samples and has the potential to provide crucial insights into functional genomics research. Despite the innovative value of tumor genome analysis, the estimation results of these approaches are not reliable. Recently, Strum et al. conducted a systematic evaluation of computational methods used to assess the abundance of cell subtypes from bulk transcriptome data [49]. They showed that EPIC was one of the most reliable methods; thus, I chose EPIC to estimate the cellular composition.

As mentioned above, the relationships between the mutational signatures and immunity were studied in LUSC [11,12], which showed that the APOBEC signature was associated with the immune response. Furthermore, recently, the relationship between APOBEC mutagenesis and the immune response in breast cancer has been reported [50], and some studies have shown an association between APOBEC mutagenesis and immunity in HNSC [51,52]. In my experimental results, the APOBEC signatures were also observed as the highest coefficients in BRCA, HNSC, as well as LUSC. Moreover, the identified mutational signatures were distinct from the abundance of mutational signatures in each tumor type, Thus, these results verified that the CCA-based approach was valid for the identification of immunity-related mutational signatures.

A limitation of this study was that it did not use molecular features or clinical information such as tumor grade, histology, and lymph node involvement. For instance, breast tumors were subdivided into luminal A, luminal B, basal-like, Her2-enriched, etc. By subgrouping the samples based on oncological features or molecular properties, subsequent analysis might obtain informative functional characteristics of the mutational signatures.

Instead, this study carried out survival analysis using the correlated genes with the mutational signatures identified by non-negative CCA. Survival analysis is a useful data-driven method for clinical trials. In our experiment, in most of tumor types, the p-values were lower than or close to 0.05, as shown in Figure 4. Although three tumor types, HNSC, LUAD, and OV, showed relatively high p-values (>0.1), relative differences between the two groups were observed in the Kaplan–Meier plots. Especially, CASP8 in HNSC is a well-known cancer-related gene included in the Cancer Gene Census whose somatic mutation is predominantly observed in oral squamous cells. Although a very low p-value was not achieved in this survival analysis, the CASP8 gene plays an important role in programmed cell death, and many studies have demonstrated that this gene is associated with tumor prognosis and immune signatures [53,54]. Similarly, it has been reported that the BPTF gene, detected in LUSC, is related to prognosis and survival in lung adenocarcinoma, non-small cell lung cancer [55,56], and tumor microenvironments [57]. Moreover, TTN identified in OV is the longest gene in the human genome, and is thus usually excluded from cancer genome analyses because of its much higher chance of being mutated. The gene with the second highest PCC in OV was MACF1, contributing to tumor progression [58]; its clinical importance and relationship to survival of the gene was recently discovered in serious ovarian cancer [59]. Thus, it would be proven that our approach was reliable to search the mutational signatures related to immunity and survival outcome. Eventually, the study might potentially be applied in clinical applications to improve patient survival.

Tumor development and progression are associated with the accumulation of the mutations, but not all the sequence alterations cause tumors. Therefore, the identification of the driver mutations and the finding of their functional roles are some of the major problems in tumor research. However, the mutational signatures did not clearly explain the acquisition of a specific driver mutation according to the mutation-causing processes. Nevertheless, a recent study explored how the molecular influences of the passenger mutations and the mutational signatures are related to the molecular functional impact [60]. Likewise, my experimental results showed the functional aspects of the mutational signatures, particularly related to immune activities, even though I did not handle the tumor driver genes directly. Moreover, on the other hand, some studies have attempted to analyze the connection between the mutational processes and the gain of the driver mutations by positive selection during cancer evolution [61,62]. By utilizing the links of the mutational signatures and the driver or passenger mutations from these kinds of studies, it would be possible to enhance the knowledge of tumor initiation, evolution, and immune activity.

A limitation of this study was that it did not consider molecular features or clinical information such as tumor grade, histology, and lymph node involvement. For instance, breast tumors were subdivided into luminal A, luminal B, basal-like, Her2-enriched, etc. By subgrouping the samples based on oncological features or molecular properties, subsequent analysis might obtain informative functional characteristics of the mutational signatures. Thus, in future works, in-depth investigations on a single cancer type should be carried out to uncover the relationships between the mutational signatures and the tumor infiltration.

To fully grasp the potential of mutational signatures, additional studies are required to develop a mechanistic understanding of how mutational signatures form. Further integrative analyses using multi-omics data and clinical information may provide novel information or insights to help understand the roles of mutational signatures in tumor biology. For instance, it might be possible to understand why distinct mutational signatures are related to different prognoses and immune reactions and can help predict patient responses to drugs or immunotherapy by combining various clinical data. Moreover, it is also necessary to determine how mutational signatures are related to previously identified clinical and molecular biomarkers. Furthermore, in future works, in-depth investigations on a single cancer type should be carried out to uncover the relationships between the mutational signatures and the tumor infiltration.

5. Conclusions

In conclusion, this research investigated the mutational signatures associated with immune cell infiltration in non-negative CCA. It is the first comprehensive analysis to investigate the relationships between mutational signatures and their functional consequences. This approach will help examine the functional roles of mutational signatures by systematically applying other types of datasets. Thus, this study serves as a basis for utilizing mutational signatures as valuable information for both scientific research and clinical applications.