Feature-weight based measurement of cancerous transcriptome using cohort-wide and sample-specific information

Abstract

Identifying cancerous samples or cells using transcriptomic data is critical for cancer related basic research, early diagnosis, and targeted therapy. However, the high transcriptional heterogeneity of cancers still hinders people from accurately recognizing cancerous transcriptome using bulk, single-cell, or spatial RNA-seq data. Here, we present a novel method named FWP (Feature Weight Pro) that helps measure cancerous transcriptome using transcriptomic data. The workflow of FWP is, first, to calculate feature weights using the training dataset, and then, for each sample in the testing dataset, calculate the feature-weight based final score by combining the cohort-wide and sample-specific information. Those two types of information are utilized through conducting weighted principal component analysis and calculating correlation perturbations. The effectiveness and superiority of FWP over other methods are shown by using bulk, single-cell, and spatial RNA-seq data of multiple cancer types. In addition, the high robustness and efficiency of FWP are also demonstrated by using different numbers of features and cells, respectively. FWP is available at https://github.com/jumphone/fwp.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13402-023-00879-6.

Introduction

Heterogeneity is a key characteristic of cancers [1, 2]. By comprehensively analyzing multi-omics bulk data of over 10,000 cancerous samples in TCGA (The Cancer Genome Atlas Program), researchers found that different types of cancers are shown to have high molecular heterogeneity, which tends to be dominated by the “cell-of-origin” [1], indicating the possibility of accurately identifying cancerous samples or cells using the prior knowledge of the matched cancer type (i.e., matched prior knowledge). However, it’s not always possible to get the matched prior knowledge, especially when conducting pan-cancer early diagnosis [3, 4] or investigating single-cell or spatial data with multiple distinguished cancer subtypes [5, 6]. Thus, a practical computational method for measuring cancerous transcriptome is supposed to be effective when using either the matched or un-matched prior knowledge.

To decipher the composition of bulk or spatial RNA-seq data, a well-acknowledged strategy is to conduct deconvolution analysis (e.g., EPIC, CIBERSORTx, and MuSiC) [7–9], which uses the cell-type specific patterns of training data to estimate the cell-type proportion (e.g., the proportion of cancerous cells) of testing data. Unlike other deconvolution methods, MuSiC doesn’t require an additional list of signature genes, which eliminates the procedure of fine-tuning parameters for getting appropriate signature genes. In addition, multiple benchmark studies have demonstrated the effectiveness and superiority of MuSiC over other methods [10, 11]. To identify cancerous cells using single-cell RNA-seq (scRNA-seq) data, a recent published method, named ikarus, uses machine learning to measure cancerous transcriptome [12], and shows much higher accuracy than existing cell-type annotation methods (e.g., SingleCellNet, scMRMA, and ACTINN) [13–15]. In sum, previous studies have already shown the effectiveness of their methods in deciphering the cell-type composition and identifying cancerous cells. However, those studies always use matched prior knowledge to demonstrate the strength of their methods, lacking the examination of utilizing un-matched prior knowledge for measuring cancerous transcriptome.

Here, we present FWP (Feature Weight Pro), a novel method for measuring cancerous transcriptome using bulk, single-cell, or spatial RNA-seq data. After learning feature (gene) weights from the training dataset, FWP calculates the feature-weight based final score by combining the global (cohort-wide) and local (sample-specific) information of the testing dataset. The global information is based on the weighted principal component analysis (PCA), while the local information is based on the perturbations of correlation between the feature weights and expression values of each testing sample. The mathematical model for calculating correlation perturbations is proposed by a study of building sample-specific networks [16], and is improved and accelerated by our recent study of predicting cell-type specific chromatin loops [17]. Unlike previous studies, we use both of the matched and un-matched prior knowledge as training data to show the effectiveness and superiority of FWP over other methods. The benchmark datasets include bulk, single-cell, and spatial RNA-seq data of multiple cancer types [5, 18, 19]. Besides, FWP shows high robustness and efficiency when using different numbers of features and cells, respectively.

Results

In this study, the benchmark datasets include three parts: (1) the bulk RNA-seq data of TCGA; (2) the scRNA-seq data of a breast cancer atlas [5] and a study of oligodendroglioma [18]; (3) the spatial RNA-seq data of the breast cancer atlas [5] and a study of pancreatic ductal adenocarcinoma (PDAC) [19]. The bulk part covers 15 cancer types with tumor and normal samples (Fig. 1A). For each cancer type, samples are randomly assigned as training or testing data. The training data is used for calculating feature weights of FWP and training models of other methods, while the testing data is used for validating the effectiveness of FWP and other methods. Samples in the first 4 out of 15 cancer types are labeled as the first dataset, which is used for building the computational model of FWP, while other cancer types are labeled as the second dataset, which is used for further testing the performance of FWP and other methods. It’s worth noting that the first dataset is not used as “training dataset”. They are randomly selected and used for building the computational framework (choosing mathematical formulas and hyperparameters) of FWP, but not calculating feature weights (calculating feature weights is the “training process” of FWP). The purpose of dividing cancer types into the first and second datasets is to test the systematic bias (over-fitting) caused by choosing mathematical formulas or hyperparameters (the number of PCs). The single-cell part contains breast cancer cells derived from 26 primary tumors with three subtypes of breast cancer (ER+, HER2+, and TNBC) [5]. The original study of Wu et al. has described the high heterogeneity of single-cell transcription profiles among different subtypes [5]. The single cells of oligodendroglioma are derived from 6 patients [18]. The spatial part contains spatial data of 6 (breast cancer) [5] and 3 (PDAC) [19] tumors. Supplementary Fig. 1 is provided for illustrating the detailed usage of benchmark datasets (details and results about datasets of oligodendroglioma and PDAC are in Supplementary Fig. 2).

Fig. 1.

The workflow of FWP. To get the feature weight of a given gene, we calculate the correlation coefficient between gene’s expression values and phenotype tags (Tumor: 1; Normal: 0) of training data and assign the correlation coefficient as the feature weight of the given gene. “cor” and “lm” are two functions in R. “cor” and “lm” are used for calculating pearson correlation coefficient and conducting linear regression, respectively. The detail about “correlation perturbation” is described in the studies of Liu et al. [16] and Zhang et al. [17]. Here, we assign all negative z-values as 0.

The workflow of FWP is illustrated in Fig. 1. To get the feature weight of a given gene, we calculate the correlation coefficient between gene’s expression values and phenotype tags (Tumor: 1; Normal: 0) of training data and assign the correlation coefficient as the feature weight of the given gene. When dealing with the testing data, FWP first assigns the correlation coefficient between gene’s expression values and feature weights as the original score, and then uses weighted PCA and correlation perturbations to calculate global and local scores, respectively. FWP finally uses linear regression to combine those two types of scores and outputs final scores. The first dataset described in Fig. 2A is used for testing the performance of origin, global, local, and final scores. As shown in Fig. 2B, FWP (final score) shows the most significant improvement comparing with the original score (AUC values and p-values are provided in Supplementary Figs. 3 and 4, respectively). The default number of used principal components (PCs) is 2, which is determined by screening the performance of FWP with different numbers of PCs (Fig. 2C). In addition, FWP shows high performance (AUC = 0.9) with only 10% genes (random selection), demonstrating the high robustness of FWP (Supplementary Fig. 5). In addition, the best threshold of final score for each cancer type is not a fixed value (Supplementary Fig. 6). Therefore, before making final decision, we recommend users to use FWP, unsupervised clustering, and marker identification methods at the same time.

Fig. 2.

The development and validation of FWP. (A) Details of bulk RNA-seq datasets used in this study. The first dataset is used for building the computational model of FWP, while the second dataset is used for testing FWP. “nSamples” indicates the number of samples. The full names of 15 cancer types can be found at the official website of The Cancer Genome Atlas (TCGA): https://portal.gdc.cancer.gov/. All expression profiles (FPKM) and phenotype labels (“Normal” or “Tumor”) are downloaded from UCSC Xena database [20] (https://xena.ucsc.edu/). (B) The performance of FWP in the first dataset. The p-value is calculated by using the one-tailed paired Wilcoxon test. The area under the ROC curve (AUC) is calculated using the package named “pROC” [21]. The heatmap is generated by using the package named “ComplexHeatmap” [22]. (C) Testing different number of principal components (PCs). “nPCs” indicates the number of principal components. (D) The performance of FWP in the second dataset. When using FWP, the nPCs is 2 (default parameter). (E) The performance of FWP in scRNA-seq data generated by Wu et al. [5]. “END”, “PVL”, “PLB”, “CAF”, and “EPI” stand for endothelial, perivascular-like, plasmablast, cancer-associated fibroblast, and epithelial cells, respectively. Because MuSiC is quite slow when processing individual cells, we conduct down-sample and use 5,000 cells (different random seeds show similar results, data not shown). “CESC” stands for cervical squamous cell carcinoma and endocervical adenocarcinoma. The scRNA-seq data is processed using Seurat [23] and BEER [24]. The “single-cell (BRCA)” in the bottom left panel is the same with the “Wu et al. single-cell (BRCA)” in the top left panel. (F) The performance of FWP in spatial RNA-seq data generated by Wu et al. [5]. “CID4465” is a sample ID. “TNBC” stands for triple negative breast cancer. The spatial RNA-seq data is processed using Seurat [23]

To test the performance of FWP in an unbiased way, we apply FWP and competing methods to the second dataset described in Fig. 2A. As shown in Fig. 2D, FWP achieves the highest averaged AUC value (0.881) comparing with MuSiC (0.850) and ikarus (0.700), suggesting the high effectiveness of FWP. Although the original score shows a similar averaged AUC value (0.866) with FWP, the improvement of FWP comparing with the original score is significant (p-value = 1e-04, Supplementary Fig. 7). In addition, we have also checked the performance of FWP using matched or un-matched prior knowledge. As shown in Supplementary Fig. 8, FWP consistently shows the highest averaged AUC values (matched: 0.970; un-matched: 0.872) in those two situations. Finally, FWP and competing methods are applied to the single-cell and spatial RNA-seq data of breast cancer. As shown in Fig. 2E and F, FWP gets the highest averaged AUC value (single-cell: 0.826; spatial: 0.724) comparing with the original score (0.798; 0.715), MuSiC (0.674; 0.604), and ikarus (0.730; 0.608), indicating the superiority of FWP over other methods in handling single-cell and spatial RNA-seq data. As shown inSupplementary Fig. 1, there are 15 and 90 tests for single-cell and spatial datasets of breast cancer, respectively. The training data of each test is based on only one cancer type in the TCGA dataset. As two examples of scRNA-seq data, the UMAP (Uniform Manifold Approximation and Projection) with cell-type labels and FWP’s final scores (train: CESC and BRCA) are shown in Fig. 2E and Supplementary Fig. 9. As two examples of spatial RNA-seq data, the images with cell-type labels and FWP’s final scores (train: CESC and BRCA) are shown in Fig. 2F and Supplementary Fig. 10. What’s more, we have tested the efficiency of FWP by using different numbers of cells (Supplementary Fig. 11). Users can get final scores of 5,000 cells in around 24 s and there is an almost linear relationship between the running time and the used number of cells, demonstrating the high efficiency of FWP.

Discussion

In conclusion, FWP offers an accurate measurement of cancerous transcriptome using bulk, single-cell, or spatial transcriptomic data. The computational framework of simultaneously using cohort-wide and sample-specific information is the key novelty of FWP. Unlike previous methods, FWP is shown to be effective when using either the matched or un-matched prior knowledge, suggesting the potential practical value of FWP in early diagnosis and exploring unknown cancer types (or distinguished cancer subtypes). Finally, the high robustness and efficiency of FWP makes it a stable and easy-to-use tool for dealing with large-scale input data. Nevertheless, there are two limitations of FWP: (1) FWP is now only based on transcriptome data (expression profiles). Identifying cancerous samples, cells, or spatial spots can be improved by integrating multi-omics data (e.g., genetic background of populations and copy number variations), which still needs further investigation. (2) FWP currently can only be used for measuring cancerous transcriptome but not identifying cancer cell types. In our future study, we plan to build a workflow for integrating FWP with other cell-type annotation method.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

F.Z. conceived and supervised the study. Q.W., F.Z., and J.S. carried out the data collection, data analysis, and result organization. J.S. and F.Z. drafted the manuscript. J.S., Q.W., and F.Z. revised and edited the manuscript. All authors read and approved the final manuscript.

Funding

National Natural Science Foundation of China (32100516), Shanghai Sailing Program (21YF1422600), Startup Fund for Young Faculty at SJTU (SFYF at SJTU) (21X010501077).

Data Availability

The source code and related data are available at https://github.com/jumphone/fwp.

Declarations

Competing interests

The authors declare no competing interests.