Identifying cancer prognosis genes through causal learning Authors: Siwei Wu, Chaoyi Yin, Yuezhu Wang, Huiyan Sun Published: 2025 ABSTRACT AbstractAccurate identification of causal genes for cancer prognosis is critical for estimating disease progression and guiding treatment interventions. In this study, we propose CPCG (Cancer Prognosis’s Causal Gene), a two-stage framework identifying gene sets causally associated with patient prognosis across diverse cancer types using transcriptomic data. Initially, an ensemble approach models gene expression’s impact on survival with parametric and semiparametric hazard models. Subsequently, an iterative conditional independence test combined with graph pruning is utilized to infer the causal skeleton, thereby pinpointing prognosis-related genes. Experiments on transcriptomic data from 18 cancer types sourced from The Cancer Genome Atlas Project demonstrate CPCG’s effectiveness in predicting prognosis under four evaluation metrics. Validations on 24 additional datasets covering 12 cancer types from the Gene Expression Omnibus and the Chinese Glioma Genome Atlas Project further demonstrate CPCG’s robustness and generalizability. CPCG identifies a concise but reliable set of genes, obviating the need for gene combination enumeration for survival time estimation. These genes are also proved closely linked to crucial biological processes in cancer. Moreover, CPCG constructs a stable causal skeleton and exhibits insensitivity to the order of data shuffling. Overall, CPCG is a powerful tool for extracting cancer prognostic biomarkers, offering interpretability, generalizability, and robustness. CPCG holds promise for facilitating targeted interventions in clinical treatment strategies. Introduction Identifying cancer prognosis biomarkers is crucial for guiding clinical treatment intervention. With the advent of high-throughput sequencing technologies, transcriptomic data have emerged as a rich resource for identifying genes and gene patterns closely associated with important biological processes of tumor and patient survival [1–3]. In conventional prognosis analysis, researchers typically employ accelerated failure models or proportional hazard models, with overall survival time or hazard as target variables, combined with the least absolute shrinkage and selection operator (LASSO) model to unveil key prognostic genes for cancer [4–6]. The most commonly used COX-LASSO is a semiparametric approach to identify cancer prognosis–related genes [7, 8], but there is no causality between these genes and cancer prognosis [9]. Besides, machine learning (ML)–based and deep learning (DL)–based computational methods are also utilized to identify biomarkers and gene patterns, sometimes serving as an end-to-end strategy to predict survival [10–14]. However, the limited interpretability, robustness, and generalizability for end-to-end learning hinder their adoption in clinical application. In contrast to generalized correlation-based methods (ML, DL, etc.) that primarily rely on statistical metrics to screen variables highly associated with overall survival, causality-based approaches are designed to establish causal graphs or quantify causal effects to reveal causal relationships between variables, thereby identifying genes that are directly linked to survival outcomes. Unlike correlation, a causal relationship is stable and robust. Some variables may exhibit associations because they are both influenced by a third factor, or the association exists due to data distribution bias. In such scenarios, correlation may disappear or alter when the data distribution changes, often resulting in poor generalizability of the identified prognostic genes on independent test sets, limiting their clinical utility. Conversely, variables with a causal relationship typically remain stable amidst such disturbances [15]. Current methods for identifying causal genes can be roughly categorized into two types: causal structure learning–based method [16–18] and causal effect estimation–based ones [19, 20]. The former typically identifies causal genes linked to disease variables through causal feature selection algorithms, aiming to identify features causally related to the target variable [21]. These methods often rely on Bayesian networks and the theory of Markov boundaries [22–24]. Bayesian networks are probabilistic graphical models that describe causal relationships between variables using nodes (variables) and directed edges (relationships) [25]. The Markov boundary defines the local causal structure as the minimal set of variables that can best predict the target variable, encompassing the target variable’s parent nodes (direct causes), child nodes (direct effects), and spouse nodes (other parents of its children nodes) [25]. The stable impact of causal features on target variable across diverse data contexts ensures that models constructed with these features possess enhanced explanatory power [15] and robustness [26, 27], thereby substantially augmenting the predictive capacity of the models. Therefore, the primary objective in identifying the causal genes for cancer prognosis lies in determining the local causal structure of cancer prognosis. However, it challenges by the identification of the direction of edges within the Markov equivalence class [28]. On the other hand, causal effect estimation–based feature screening estimates the pure effect of one or a set of gene perturbations on core biological processes in cancers by balancing the distribution of all the confounders. Subsequently, genes exhibiting strong causal effects on the target variable are identified as causal genes [19, 20]. However, these causal feature selection methods encounter challenges when applied to prognostic gene recognition tasks, particularly due to right-censored survival data. Additionally, the presence of high-dimensional data poses significant computational hurdles in learning the causal skeleton [29, 30]. Consequently, there is a need for innovative methods that can simultaneously address these challenges. In this study, we propose the CPCG (Cancer Prognosis’s Causal Gene), a two-stage framework for identifying prognostic causal genes. In the first stage, CPCG aims to obtain a simplified gene dataset that is significantly correlated with prognosis, to facilitate subsequent causal skeleton construction. We employ an ensemble strategy, using both parametric and semiparametric hazard models, to model the association between gene expression and patient prognosis. In the second stage, we utilize these genes with overall survival time as the target variable. Through iterative conditional independence tests for each variable pair, we roughly construct a causal skeleton, selecting genes directly linked to overall survival time as prognosis causal genes. To assess predictive performance, we use a variety of metrics and validate generalizability across diverse datasets. We also explore the biological functions of these genes. In summary, we aim to develop a simple yet effective, interpretable, and highly generalizable method for identifying prognostic genes, which is crucial for clinical applications. Materials and methods Method overview In this study, we propose a framework called CPCG that integrates survival analysis techniques with censored data and causal feature selection methods to identify causal genes related to cancer prognosis from high-dimensional transcriptome data. CPCG is a two-stage identification scheme. In Stage I, to obtain a comprehensive yet relatively concise set of credible candidate genes, we employ both parametric and semiparametric strategies to obtain candidate genes while taking hazard as the target variable. Hazard encompasses survival time and survival status, effectively handling censoring in survival data. The distinction between these two strategies lies in the data types and the way that hazard is calculated so that the genes obtained by the two strategies can effectively complement each other. The union of these two gene sets forms the candidate gene pool for subsequent causal gene identification. In Stage II, we deploy a causal feature selection method that constructs a causal skeleton linking cancer prognosis–related genes and overall survival time through death cases data through iterative conditional independence tests, marginal independence tests, and graph pruning. These steps significantly eliminate confounding factors that interfere with causal inference. Genes directly linked to overall survival time are ultimately identified as the causal genes for cancer prognosis. Figure 1 illustrates the workflow of CPCG. Framework of CPCG for identifying cancer prognosis genes. In Stage I, we aim to screen candidate genes by using parametric and semiparametric strategies, focusing on hazard as the target variable. In each strategy, the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${\lambda}_t^p$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${\lambda}_t^{sp}$\end{document} represent the hazard calculated by different populations and models, and the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\log \left({\lambda}_t^p\right)$\end{document} in Stage I is transformed by the exponential proportional hazard model. The P-value of logrank test (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${P}_{LR}$\end{document}) is used to filter out irrelevant genes. In Stage II, CPCG identifies causal genes related to cancer prognosis from the candidate genes through iterative conditional independent test and graph pruning. Stage I: Obtain a collection of potential prognosis genes To obtain concise but credible candidate genes for downstream analysis, the selection and processing of hazard, especially for right-censored survival data, is a crucial yet challenging step. We consider the complex relationships among hazard, overall survival time, and gene expression from different perspectives to help identify candidate prognostic genes. In the parametric model, we focus solely on the death cases data, eliminating the need to address the censor issue. The hazard \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${\lambda}_t^p$\end{document} is calculated using LifeTable as Formula (1) through death cases’ survival data where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${d}_t^p$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${n}_t^p$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${\lambda}_t^p$\end{document} represent the number of deaths, the number of people still alive, and the hazard at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $t$\end{document} in this dataset, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $p$\end{document} indicates that these variables are calculated in the parametric model. Initially, we exclude a minimal number of samples with outlier survival time to ensure that the patients’ hazard follows an exponential distribution. We employed the widely used exponential proportional hazard model (the model’s details are in the Supplementary Methods section) to establish the relationship between hazard and gene expression level [6]. The relationship formula between hazard and influencing gene are as follows: where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${\lambda}_0^p$\end{document} represents the baseline hazard, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $X$\end{document} represents each to-be-detected gene, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\beta$\end{document} represents the impact of detected gene on patients’ hazard. In this model, baseline hazard \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${\lambda}_0^p$\end{document} is a constant, allowing us to apply a logarithmic transformation directly to this formula as follows: According to Formula (3), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\log \left({\lambda}_0^p\right)$\end{document} is a constant, so there is a clear linear relationship between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\log \left({\lambda}_t^p\right)$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $X$\end{document}. Using the univariate least squares regression model, we quantify the impact of gene expression level on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\log \left({\lambda}_t^p\right)$\end{document}, ranking genes based on their R-square value in decreasing order, and selecting the top \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $k$\end{document} genes as gene set I. Before assessing the impact of genes on hazard, we filter out genes irrelevant to cancer prognosis, removing outlier samples from the data. When evaluating each gene, for maximizing the utilization of patient data, we divide the samples into the high-expression group and low-expression group, two groups based on the median gene expression levels. Subsequently, we employ the logrank test [31] to determine significant differences in survival time between two groups. This step effectively excludes irrelevant genes, ensuring that the remaining genes are closely associated with cancer prognosis. In the semiparametric model, as it does not depend on specific hazard distributions, we use all patient’s data to calculate the hazard. This approach allows CPCG to gain a comprehensive understanding of the population’s survival. The hazard \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${\lambda}_t^{sp}$\end{document} is determined using the LifeTable as Formula (4) through all patients’ survival data, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${d}_t^{sp}$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${n}_t^{sp}$\end{document} represent the number of deaths and the number of people still alive at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $t$\end{document}, respectively.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $sp$\end{document} indicates that these variables are calculated in the semiparametric model. Similarly, we first employed the logrank tests to exclude irrelevant genes by using all patients’ data. As there is still no effective semiparametric model to quantify the impact of each influencing gene on hazard, to simplify, we directly used the R-square values obtained from linear univariate least squares regression to assess the impact of gene expression levels on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${\lambda}_t^{sp}$\end{document}. This method is used to rank genes in all patients’ data, and the top \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $k$\end{document} genes are selected as gene set II. In the CPCG approach, the parametric model is particularly adept at precisely quantifying the impact of gene expression levels on cancer prognosis. This model operates under the assumption that a patient’s hazard conforms to an exponential distribution, which allows for a detailed and accurate assessment of gene influence within a specific framework. On the other hand, the semiparametric model excels in extrapolating candidate genes across a broader range of patient data without relying on such specific distributional assumptions. This flexibility enables the identification of genes that may not conform to the strict assumptions of the parametric model, thus broadening the scope of potential prognostic markers. By combining the candidate gene sets derived from both models, our approach yields a more comprehensive gene list. This list includes genes that are explicable by the exponential proportional hazards model, as well as those that are not bound by data distribution assumptions. Consequently, the genes identified through CPCG exhibit heightened statistical robustness and are adaptable across diverse datasets. Furthermore, this collaborative approach facilitates the identification of genes that may have been overlooked when utilizing a single method. Stage II: Identify causal gene of cancer prognosis We take the union of gene set I and gene set II to form a pool of candidate genes for further identification of causal genes affecting cancer prognosis. Utilizing data from deceased patients, we employ conditional independence tests and graph pruning techniques to construct a causal skeleton aimed at identifying the causal genes from the candidate pool. Initially, all pairs of variables, including candidate genes and survival time, are connected to form an undirected graph. Subsequently, we conduct joint marginal independence tests and conditional independence tests (the descriptions of conditional independence test are in the Supplementary Methods) to partly ascertain direct relationships between each pair. Edges between variables showing only conditional dependence are then removed. We determine the conditional independence among variables based on their partial correlation coefficient, which refers to the correlation between two variables after eliminating the influence of other variables. The formula for the partial correlation coefficient is as follows: where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $i$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $j$\end{document} represent two arbitrary variables, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $k$\end{document} represents conditional variables, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $c$\end{document} represents any subset of other remaining variables, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${\rho}_{i,j\mid k}$\end{document} represents the partial correlation coefficient between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $i$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $j$\end{document} under condition \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $k$\end{document}. We consider \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${\rho}_{i,j\mid k}=0$\end{document} as variables \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $i$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $j$\end{document} being conditionally independent. After obtaining the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${\rho}_{i,j\mid k}$\end{document}, we use the Fisher-Z test [32] to determine if it is 0. We initially employed Fisher-Z transformation to convert the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\rho$\end{document} into a normal distribution: Then, we calculate the standard error of the correlation coefficient: where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $n$\end{document} represents the total number of all variables, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\left|k\right|$\end{document} is the quantity of conditional variables. Ultimately, we ascertain whether the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${\rho}_{i,j\mid k}$\end{document} is zero by applying Formula (8): where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\Phi \left(\ \right)$\end{document} is the cumulative distribution function of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathcal{N}\left(0,1\right)$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\alpha$\end{document} represents significance level. Once the formula is satisfied, we treat variables \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $i$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $j$\end{document} as conditionally independent concerning \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $k$\end{document}, and edges between variables \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $i$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $j$\end{document} are then removed. By employing conditional independence testing and edge pruning iteratively until no further edges can be removed from the graph, we construct a skeleton that represents the potential causal relationships among variables. Genes directly linked to overall survival time within this skeleton are identified as the causal genes influencing cancer prognosis. If a gene and survival time are linked through intermediary genes, their variations usually do not directly impact each other. We exclude the edge linking them as including these genes often leads to unstable and biased predictions. We do not distinguish whether the first-order neighbors of survival time are causes or results as we believe these genes can effectively reflect and predict prognosis in either scenario. Data collection and preprocessing We collect data from The Cancer Genome Atlas Project (TCGA), which involves 7371 patients across 18 distinct cancer types throughout the body (Fig. 2A), including a total of 2654 (36%) death cases and 4717 (64%) survival cases. Specifically, we mainly focus on cancer types with a death count exceeding 100 within the TCGA dataset. This threshold is set to ensure that the statistical analyses conducted would be robust and meaningful. In addition to these datasets, we include cancer types such as ESCA, UCEC, LAML, MESO, and SARC. These are chosen for their moderate mortality population, ranging between 50 and 100 deaths. The inclusion of these datasets allows us to assess the generalizability and effectiveness of our CPCG approach across a variety of mortality scales. To ensure the reliability and clinical relevance of the identified causal genes, we only focus on protein-coding genes and filter out genes with an average raw count below 10. The number of genes used for each cancer is detailed in Fig. 2A. In data preprocessing, we standardize the transcriptome data using transcripts per million (TPM) normalization. Overview of cancer types and datasets. (A) depicts 18 distinct cancer types from TCGA, organized in a table format by cancer onset location. It includes cancer type, abbreviation, number of genes, number of patients, number of death cases, number of survival cases, and survival distribution. (B) shows 24 independent datasets across 12 cancer types from GEO and CGGA. It lists cancer type, data sources, number of patients, the number of death cases, the number of survival cases, and survival distribution. CGGA_325 and CGGA_693 are from CGGA, while others are from GEO. To assess the generalization and robustness of our method, we collect an additional 24 independent datasets from 12 cancer types in the Gene Expression Omnibus (GEO) and 2 datasets from 1 cancer type in the Chinese Glioma Genome Atlas Project (CGGA) (Fig. 2B). These datasets encompass 4625 patients, with 2103 (45%) death cases and 2522 (55%) survival cases. Here, from these two databases, we found 6 datasets with over 100 deaths, another 6 datasets containing a moderate population of death cases, and additionally, 12 sets with <50 deaths. We expect these common cancer datasets can validate the relationship between the genes identified by CPCG and cancer prognosis in populations with large, medium, and small death cases. Gene expression levels from the GEO are stored in microarray data, and we apply a log2 transformation to the data. The CGGA data, obtained through high-throughput transcriptome sequencing, are standardized using TPM normalization. Evaluation metrics In this study, we use four major metrics to evaluate the prognosis prediction performance of CPCG, the first two common metrics are: 1) P-value of logrank test (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${P}_{LR}$\end{document}): The logrank test is a standard statistical method used to compare the survival time distributions of two groups. The resulting P-value can indicate the statistical significance of observed survival differences between patients from groups of different gene expression levels and evaluate the impact of genes on patient prognosis, with a smaller P-value suggesting a more significant distinction. 1) P-value of logrank test (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${P}_{LR}$\end{document}): The logrank test is a standard statistical method used to compare the survival time distributions of two groups. The resulting P-value can indicate the statistical significance of observed survival differences between patients from groups of different gene expression levels and evaluate the impact of genes on patient prognosis, with a smaller P-value suggesting a more significant distinction. 2) Concordance Index of Cox regression (C-index): The C-index measures the agreement between the predicted prognosis of patients, as estimated by the Cox model, and the actual event occurrence. A higher C-index indicates that the Cox model has better predictive capability in terms of patients’ hazard. 2) Concordance Index of Cox regression (C-index): The C-index measures the agreement between the predicted prognosis of patients, as estimated by the Cox model, and the actual event occurrence. A higher C-index indicates that the Cox model has better predictive capability in terms of patients’ hazard. Additionally, we introduced two derivative metrics: 3) Proportion of genes with significant P-values in the logrank test (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_L$\end{document}): Given the substantial variance in the number of output genes between CPCG and COX-LASSO, direct comparison of their performance based on all P-values in logrank tests is challenging. Therefore, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_L$\end{document} is devised to comprehensively represent the proportion of variables with significant P-values from both methods, with a higher \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_L$\end{document} indicating that there are more genes from the same set that have a significant impact on patient prognosis. This facilitates a direct comparison of the significance of identified variables. Their formula is as follows: 3) Proportion of genes with significant P-values in the logrank test (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_L$\end{document}): Given the substantial variance in the number of output genes between CPCG and COX-LASSO, direct comparison of their performance based on all P-values in logrank tests is challenging. Therefore, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_L$\end{document} is devised to comprehensively represent the proportion of variables with significant P-values from both methods, with a higher \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_L$\end{document} indicating that there are more genes from the same set that have a significant impact on patient prognosis. This facilitates a direct comparison of the significance of identified variables. Their formula is as follows: 4) Proportion of significant genes in Cox regression (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document}): A robust regression model should encompass variables that are genuinely linked to the target variable and exhibit significant P-values in regression while avoiding multicollinearity issues. Within the regression variables, a higher proportion of significant variables indicates the reliability of the selected predictors. Thus, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} represents the proportion of independent variables (identified prognosis-associated genes) in the Cox regression model that significantly influence survival time among the predicted prognosis-associated genes. A higher \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} implies that the Cox model encompasses more genes that truly impact patient survival time, and these genes play a substantial role in prognosis prediction. Their formula is as follows: 4) Proportion of significant genes in Cox regression (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document}): A robust regression model should encompass variables that are genuinely linked to the target variable and exhibit significant P-values in regression while avoiding multicollinearity issues. Within the regression variables, a higher proportion of significant variables indicates the reliability of the selected predictors. Thus, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} represents the proportion of independent variables (identified prognosis-associated genes) in the Cox regression model that significantly influence survival time among the predicted prognosis-associated genes. A higher \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} implies that the Cox model encompasses more genes that truly impact patient survival time, and these genes play a substantial role in prognosis prediction. Their formula is as follows: For detailed formulas and further explanations, please refer to the Supplementary Methods section. Experimental settings In Stage I, we establish a P-value threshold of .01 in the logrank test and filter out genes with P-values exceeding .01. For the parametric and semiparametric models, we select the top 100 genes as candidate prognosis-related genes. Moving to Stage II, we utilize the Fisher Z test to evaluate the significance of conditional independence tests, with the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\alpha$\end{document} set to 0.05. Results Causal genes identified by CPCG demonstrate strong discriminability in survival time By applying CPCG across 18 cancer types in TCGA, we pinpoint the core set of prognosis-related genes for each cancer type. Specifically, we identify 1 gene in ESCA; 2 genes in LIHC; and 3 genes in BLCA, COAD, LAML, and STAD. Moreover, we identify 4 genes in GBM, HNSC, KIRC, LUAD, LUSC, MESO, SARC, and UCEC; 5 genes in BRCA, LGG, and SKCM; and 6 genes in OV. Notably, our approach consistently reveals fewer than 10 causal genes for each cancer type, with 15 cancers demonstrating a stable gene set size of 3–5 genes. To further confirm these concise gene sets owing the capability of blocking the correlation between irrelevant genes and patient prognosis in methodology, after controlling these causal genes within the candidate set, PoI (The proportion of genes that become independence) and PoD (The proportion of genes that exhibt diminished correlation) are summarized in Table 1 (others in Supplementary Table 1): where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $gene{s}_C$\end{document} represents candidate genes. Table 1 displayed the results of representative genes from each cancer, and it can be observed that most gene can isolate the relationship between candidate genes and cancer prognosis; notably, 8 genes’\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PoI$\end{document} and 30 genes’ \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PoD$\end{document} are all exceeding 50%. These findings underscore the efficacy of the causal genes in isolating the association between irrelevant genes and patient prognosis while also diminishing the correlation of certain unrelated genes with survival time. Despite the limited number of causal genes identified, their association with patients’ survival time remains robust and is not disrupted by other genes. Table 1The impact of causal genes on the relationship between patients’ survival and other genes.Cancer typeGene \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\boldsymbol{PoI}$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\boldsymbol{PoD}$\end{document} Cancer typeGene \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\boldsymbol{PoI}$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\boldsymbol{PoD}$\end{document} BLCAGCNA66.67%63.95%LIHCRARS176.65%0.56%BRCALDLRAD338.89%53.37%LUADDENND1A41.30%24.31%COADABHD865.71%23.57%LUSCDUSP2241.18%34.09%ESCARBM2733.33%13.33%MESOLMNB246.28%53.16%GBMMED1048.62%36.00%OVGTF2F243.53%18.97%HNSCCD591.00%96.95%SARCB3GALT432.99%26.42%KIRCGDF513.08%27.23%SKCMVPS13C58.70%84.66%LAMLCBR176.79%12.78%STADDAB2IP48.28%71.51%LGGCRTAC133.58%38.95%UCECSLC16A132.94%43.09% The impact of causal genes on the relationship between patients’ survival and other genes. To determine the discriminability of these causal genes for all patients’ survival time in each cancer, we divided all patients into high and low expression groups based on genes’ optimized cut-off (the calculation of optimal cut-off is detailed in Supplementary Methods), and plotted KM (Kaplan–Meier, [33]) curves for representative genes from each cancer in Fig. 3 (others in Supplementary Fig. 1). Among all causal genes, 15 genes’ \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${P}_{LR}$\end{document} values are between 1E-03 and 1E-04, 11 genes’ \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${P}_{LR}$\end{document} values are between 1E-04 and 1E-05, and 12 genes’ \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${P}_{LR}$\end{document} values are below 1E-05, which indicate that the association between these genes and patients’ survival time has a very high statistical significance. Notably, the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${P}_{LR}$\end{document} of some genes are insignificant with outlier samples, but it can be observed that the KM curves from high and low expression groups of them exhibited significant differences within the period before the outlier time, which also demonstrates that these genes have a strong association with cancer prognosis. Specifically, the significant variations in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${P}_{LR}$\end{document} of XPO6 (from 1.78E-02 to 1.06E-02), TMEM164 (from 2.77E-02 to 2.23E-02), DDIT4L (from 2.07E-02 to 8.99E-03), DNASE1L3 (from 9.18E-02 to 3.79E-02) between datasets with and without outlier samples demonstrate the different effect of extreme survival times on P-values. In the parametric model of Stage I, CPCG obtains the candidate gene set I by using patients’ clinical data without outlier survival time, which enables its capacity to discover genes whose P-value of the logrank test is greatly affected by the outlier value, thus effectively reducing the bias of extreme survival times in identifying causal genes. Discriminability of causal genes on survival time. It illustrates the KM curves for high and low gene expression groups of representative causal genes from each cancer type. The red line represents the high-expression group, and the blue line represents the low-expression group. It also shows P-values from logrank tests conducted on patients’ overall survival times. CPCG can offer competitive cancer prognosis prediction with fewer genes We use Cox regression to assess how well the identified causal genes predict cancer prognosis in TCGA data, comparing them with genes from COX-LASSO in terms of output gene numbers and significance in regression (Fig. 4). For COX-LASSO, we use 10-fold cross-validation to find the best λ that minimizes the Cox model error. In Fig. 4A, we compare the gene numbers identified by both methods across various cancers and noticed that CPCG shows less fluctuation on this. Notably, COX-LASSO fails to identify genes in BRCA, ESCA, OV, and STAD, whereas our method remains effective, indicating CPCG’s stronger stability and applicability. Comparative analysis of CPCG and COX-LASSO. (A) We compare the number of genes identified by both methods across 18 cancer types. (B) It shows the C-index of Cox models based on genes identified by both methods. (C) It presents the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} of the two Cox models. Blue bars represent CPCG results, and pink bars represent COX-LASSO results. Each bar’s dark-colored regions indicate genes with significant regression parameters, while light-colored regions indicate nonsignificant ones. The total area of each bar is equal. (D) It displays the C-index of the two Cox models at different time points. Only results for 14 cancers are shown as COX-LASSO failed to identify cancer-related genes in BRCA, ESCA, OV, and STAD. We further evaluate the identified prognosis-associated genes using the C-index and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} matrices in TCGA data, as depicted in Fig. 4B and C. By fitting Cox regression models with these genes, CPCG achieves high C-index values in LAML (0.713), LGG (0.826), MESO (0.783), and SARC (0.722), with most genes significantly contributing to survival time. The C-index values of both methods were similar in LGG, LIHC, LUAD, LUSC, and UCEC within the margin of error (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\pm 0.03$\end{document}). Under the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} matrix, CPCG outperforms COX-LASSO in several cancers, suggesting that our method’s identified genes have a more substantial effect in Cox regression. These results highlight that a more concise gene set from our method can competitively predict cancer prognosis in Cox regression. To validate the predictive ability of both models at different time points, we randomly split patients into training (50%) and validation (50%) sets. We fit the models using bootstrap cross-validation with 500 iterations, except for MESO where we reduce iterations to 150 due to its small sample size. We calculate the C-index of the optimal models at different time points in the validation set and plot the C-index curves in Fig. 4D. Notably, genes identified by CPCG show higher C-index values compared to COX-LASSO in several cancers, demonstrating better predictive ability for cancer prognosis at each time point despite the smaller gene set size. We assess the performance of Cox regression models using causal genes by plotting Receiver Operating Characteristic (ROC) curves for 1-year, 3-year, and 5-year patient survival (see Supplementary Fig. 2). Notably, the AUC values in GBM, LAML, LGG, and MESO exceed 0.8, while in BRCA, KIRC, LUAD, SARC, and UCEC, they surpass 0.7, indicating strong predictive capabilities of these genes. Furthermore, calibration curves are plotted for 1-year, 3-year, and 5-year frames to evaluate predictive accuracy and model calibration (see Supplementary Fig. 3). These curves compare predicted event probabilities with actual event occurrences. Notably, within an acceptable error, (i) the calibration curves for 1-year survival in BRCA, ESCA, GBM, HNSC, LAML, LUSC, MESO, OV, SKCM, STAD, and UCEC; (ii) for 3-year survival in BLCA, BRCA, ESCA, HNSC, LAML, LUAD, LUSC, MESO, SARC, SKCM, STAD, and UCEC; and (iii) for 5-year survival in BLCA, BRCA, COAD, HNSC, LGG, LUSC, MESO, OV, SARC, SKCM, STAD, and UCEC closely align with the diagonal, indicating these models’ high predictive ability. Notably, calibration curves in BRCA, HNSC, LUSC, MESO, SKCM, STAD, and UCEC are particularly close to the diagonal, providing evidence that predictions from Cox models constructed with causal genes closely match actual outcomes. These experiments demonstrate that CPCG reliably identifies causal genes across diverse cancer types, yielding a concise gene set that enables accurate cancer prognosis predictions using Cox models. To investigate the association between genes identified by COX-LASSO and cancer prognosis, we use conditional independence tests to verify whether there is a relationship and, if so, whether it is direct or indirect. Initially, we merged the gene sets derived from CPCG and COX-LASSO, considering that causal genes might block the impact of COX-LASSO’s genes on patient prognosis. Subsequently, we computed the partial correlation coefficient for each COX-LASSO’s gene with patient survival time, after accounting any subset of other genes. We make following statistical analysis: the number of genes identified by COX-LASSO (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $No{G}_{CL}$\end{document}), the number of genes originally related to survival (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $No{O}_{CL}$\end{document}), the number of genes that become independent with patient prognosis (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $No{I}_{CL}$\end{document}) and the conversion ratio (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\frac{No{I}_{CL}}{No{O}_{CL}}$\end{document}) across 18 cancer types as detailed in Supplementary Table 2. The results indicate that a significant portion of genes identified by COX-LASSO exhibit conditional independence regarding cancer prognosis. Even for COAD, GBM, KIRC, LGG, LIHC, LUAD, LUSC, MESO, and UCEC, the conversion ratio reaches 100%. This observation suggests that the impact of COX-LASSO’s genes on cancer are mainly indirect as these effects can be mediated or isolated by other genes. CPCG performs well on other independent testing data To assess the generalizability of CPCG on independent testing data, we validate the relationship between the expression of our predicted prognosis causal genes and patients’ survival across 24 datasets across 12 cancer types from GEO and CGGA. For each dataset, we categorize cancer samples into high and low expression groups based on gene expression levels and plotted KM curves for comparison (see Fig. 5, with additional curves in Supplementary Fig. 4). Significant differences observed in these curves highlight the strong discriminative power of these causal genes in predicting survival time on independent testing data. Discriminability of causal genes on patients’ survival time in independent testing data. It displays the KM curves for high and low gene expression groups of representative causal genes from each cancer. The data sources are: BLCA: GSE32894, BRCA: GSE31448, COAD: GSE38832, GBM: CGGA_325, HNSC: GSE65858, LIHC: GSE10143, LUAD: GSE157009, LUSC: GSE157010, OV: GSE63885, SARC: GSE30929, SKCM: GSE54467, STAD: GSE26253. We conduct logrank tests on these independent datasets to compare CPCG with COX-LASSO under \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_L$\end{document} metric (Supplementary Table 3). The results indicate that our method’s \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_L$\end{document} value is comparable to COX-LASSO’s, particularly excelling in BRCA, OV, and STAD, while COX-LASSO yields no results. Moreover, in GSE32894 (BLCA), GSE48075 (BLCA), GSE162228 (BRCA), GSE31448 (BRCA), GSE42568 (BRCA), GSE29621 (COAD), GSE38832 (COAD), CGGA_693 (GBM), GSE74187 (GBM), GSE65858 (HNSC), GSE10143 (LIHC), GSE31210 (LUAD), GSE157010 (LUSC), GSE30161 (OV), GSE63885 (OV), GSE30929 (SARC), and GSE26253 (STAD), our method outperforms COX-LASSO in terms of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_L$\end{document}. Notably, the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_L$\end{document} from GSE32894 (BLCA), CGGA_693 (GBM), GSE10143 (LIHC), and GSE30929 (SARC) is equal to 1, which means all causal genes from these four cancers can strongly distinguish the statistical significance of observed survival differences in validation data. These results indicate that our causal genes exhibit good predictive ability across various datasets from different platforms. Additionally, we assess the C-index of the causal genes identified by CPCG for their predictive accuracy on survival times across independent datasets (Supplementary Table 4). The predictive performance aligns well with that observed on the TCGA datasets, with an acceptable margin of error in 18 out of 24 independent datasets. These datasets include GSE48075 (BLCA), GSE162228 (BRCA), GSE42568 (BRCA), GSE29621 (COAD), GSE38832 (COAD), CGGA_325 (GBM), GSE74187 (GBM), GSE65858 (HNSC), GSE10143 (LIHC), GSE15654 (LIHC), GSE31210 (LUAD), GSE41271 (LUAD), GSE26939 (LUSC), GSE30161 (OV), GSE63885 (OV), GSE53118 (SKCM), GSE54467 (SKCM), and GSE26253 (STAD). Notably, the causal genes achieve high C-index values in GSE32894 (0.744\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\pm$\end{document}0.041) and GSE31448 (0.760\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\pm$\end{document}0.010). This further underscores the robustness and generalizability of these causal genes in predicting cancer prognosis across diverse cancer types. Biological relevance of identified causal genes The set of 68 prognosis causal genes identified by CPCG in TCGA data across 18 cancer types holds significant biological relevance. Upon reviewing the literature, we discovered that 31 out of these 68 genes are intricately linked to the respective cancers they are associated with, as detailed in Supplementary Table 5. Through gene function analysis, we observed that among these 31 causal genes: five genes (PINX1, CYP2E1, MYLIP, DUSP22, VPS13C) may participate in regulating the migration and invasion capabilities of cancer cells, affecting the metastasis of cancer to distant sites [34–38]; four genes (TSPAN13, MYLIP, DUSP22, DNASE1L3) play roles in regulating the proliferation and apoptosis of cancer cells; six genes (TSPAN13, MT-C01, ABAT, MYLIP, CYP2C18, KCNQ1) are crucial in cancer occurrence and development, potentially involved in regulating abnormal cell proliferation and transformation [36, 39–43]; six genes (CD5, RARS1, DNASE1L3, NOX4, CLIC2, CD1C) may mediate the interaction between cancer and immune responses, impacting the recognition and immune attack on cancer cells [44–49]; and the expression level of 15 genes (PINX1, MED10, RPL39L, CD5, MS4A1, RPL3, DNASE1L3, BCL3, CBR, CD33, VPS13C, CLIC2, SLC16A1, MS4A1, PILL1) are associated with cancer prognosis, and some serve as prognostic markers [34, 38, 45, 47, 49–58]. CPCG has a low sensitivity to the input order of the data In computational methodology, the construction of a causal skeleton by CPCG is dependent on a series of conditional independence tests between variables. However, the accuracy of these tests hinges on correctly capturing the variables’ conditional independence relationships in preceding tests as altering the order of variables in the tests leads to variations in the captured relationships and yields different outcomes. Consequently, the conditional independence tests in Stage II are supposed to be sensitive to the input order of data, impacting the inference of causal structures. To assess the sensitivity of CPCG to data input order, we conduct shuffling experiments in Stage II of our method. Genes obtained from Stage I are ranked in descending order based on their R-square values from least squares regression for overall survival time. These genes are then input into Stage II in both sorted and randomly shuffled orders (repeated over five iterations), respectively. Through the experiment results, we observe that despite variations in data input order across experiments, the performance of Cox regression, particularly under C-index and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} evaluations, remained stable in predicting cancer prognosis (Fig. 6). Notably, CPCG consistently identified the same gene set in over half of the experiments across 12 cancer types, including BLCA, BRCA, COAD, ESCA, GBM, HNSC, LAML, LIHC, OV, SKCM, STAD, and UCEC. In some cases, such as COAD and STAD, shuffling the input order even had no discernible impact on the results. For other cancer types, the C-index and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} values from the six experiments were similar. We suspected that this may be attributed to the accurate candidate gene sets obtained in Stage 1. Performance of Cox models constructed from shuffling experiments. For each cancer, the upper bar graph represents the models’ C-index, and the lower bar graph represents the models’ \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document}. “H2L” denotes models based on sorted data, while “rand_1” to “rand_5” represent models based on shuffled data. In practical application, we recommend inputting the genes obtained from Stage I into Stage II in descending order according to the correlation strength with the overall survival time. This is primarily because genes with higher correlation imply their higher impact on patients’ overall survival time, potentially serving as causal genes. Prioritizing these genes can expedite the construction of the causal skeleton. CPCG doesn’t need to consider Markov equivalence class and V-structure In the previous section, we take genes directly linking to survival time in the causal skeleton as causal genes. However, a question arises regarding whether including spouse genes or solely utilizing parental genes of survival time would enhance the predictive performance of Cox models. In this context, we conducted two additional experiments to validate this assumption. Unlike condition-based causal feature selection methods, which treat variables within the Markov boundary through causal structure learning as causal features (refer to Markov equivalence class and V-structure in Supplementary Methods for detailed explanations). CPCG does not consider spouse genes as we believe that they do not directly influence cancer prognosis. To investigate whether incorporating spouse genes improves the predictive ability of Cox models, we first identify variables sharing a common child node with overall survival time but not directly linked to each other as spouse variables. Using the causal skeletons from Result 3.5 as a basis, we then compared the C-index and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} of Cox regressions before and after including spouse genes (see Supplementary Fig. 5). Our analysis reveals that incorporating spouse genes does not significantly enhance the C-index of Cox regression models. In fact, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} shows a decreasing trend and is lower compared to models without spouse genes. Across experiments in COAD, ESCA, HNSC, LGG, and LIHC, the Cox regression model excluding spouse genes consistently exhibits similar C-index values and higher \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document}. These findings suggest that including spouse genes does not notably improve and may even reduce the predictive ability of Cox models. In our second experiment, we delved into the issue of Markov equivalence classes, specifically exploring the necessity of inferring edge directions for variable pairs and only outputting the parental genes of overall survival time. To assess whether constructing a Cox model solely based on parental genes yields superior predictive performance, we utilized various causal direction inference methods, including ANM [59], BivariateFit, CDS [60], IGCI [61], GES [62–65], GIES [62–65], LiNGAM [66], and PC [67], to determine the directionality of relationships among variables. By integrating the outcomes of these methods and identifying parental genes based on the direction of edges pointing toward survival time in at least four techniques, we conduct statistical analyses on the C-index and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} of Cox regressions before and after incorporating child genes (Supplementary Fig. 6). Our findings reveal that among the cancers where parental genes are successfully identified, integrating child genes into the Cox model significantly improved the C-index with minimal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} loss. This enhancement is observed in BLCA, BRCA, COAD, ESCA, LAML, LUSC, OV, SARC, and SKCM. Notably, in all COAD experiments and partial ESCA, GBM, KIRC, MESO, OV, SARC, and UCEC experiments, the Cox model incorporating child genes exhibited a higher or equal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} compared to the model with only parental genes while also demonstrating a significantly higher C-index. These results highlight that utilizing genes directly linked to overall survival time as causal genes can notably enhance the predictive ability of Cox models for cancer prognosis with minimal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $PS{P}_C$\end{document} loss. On the other hand, despite the integration of multiple methods, the precise causality between genes and cancer prognosis remains challenging to determine conclusively. Therefore, we propose to take a balanced approach without excessive emphasis on predicting cancer prognosis solely based on parental genes of survival times in the causal skeleton. Discussion and conclusion We proposed a two-stage framework CPCG to identify genes associated with cancer prognosis from the perspective of causal as causation is more stable than correlation. The experiments demonstrate that the genes obtained through CPCG exhibit stability, reliability, good reusability, and enhanced interpretability in cancer prognosis prediction, aiding in more effective interventions for diseases. Before causal inference, integrating parametric and semiparametric components to process hazard and construct a one-to-one relationship between each gene and cancer prognosis for preliminary screened causal genes helps precisely narrow down the set size of potential causal genes and avoids the influence of high-dimensional features on correlation measurement, undoubtedly facilitating the efficiency and effectiveness of subsequent causal feature selection. This is because obtaining causal structures is highly challenging in high-dimensional but relatively low-sample scenarios as it can be computationally complex and, in some cases, even intractable. Therefore, precise initial screening of candidate genes is crucial. The conditional independence test used in Stage II of CPCG can effectively solve the multicollinearity issue. We also demonstrated that within the structural learning module, CPCG is insensitive to the input order of the data, which is also attributed to the precise selection of causal genes obtained from Stage I. Furthermore, we show that our method does not need to consider Markov equivalence classes, which is a challenging issue in causal structure learning. These all make using CPCG involve fewer considerations. Our method undoubtedly eliminates spurious relationships between genes and survival, some of which are effects manifested after multiple mediating variable transmissions and some are due to common causes. Hence, it helps tumor systems biologists and clinical practitioners address long-standing issues. Traditional biomarker identification methods based on feature selection also face certain gaps when it comes to practical clinical application, one of which is how to select a final, small set for optimal predictive performance. Typically, the best set is obtained through enumeration. Our method can acquire a small set of causal genes, without the need for further time-consuming enumeration. Through a literature review, it also demonstrates the biological association between identified causal genes and cancer progress and malignancy. Moreover, we investigated and found that 13 out of the 68 genes are related to secreted proteins, with proteins encoded by 11 genes easily detectable in blood (CYP2E1, H2AW, CD5, OAF, CBR1, CRTAC1, RARS1, KIRREL3, CYP2C18, SNAP25, CD33) [68–70]; these further underscored the potential of these causal genes as clinical prognosis biomarkers that the proteins they encode can be detected in the body fluids of cancer patients. There is also room for further improvement in our approach. Currently, CPCG focuses only on transcriptomic data; if other multi-omics data, as well as imaging data, become a larger sample size, we will continue exploring causal associations with other features related to prognosis. Besides, In the semiparametric part of Stage I, we simply evaluate the impact of each gene on hazard using the least squares method. Future research should consider integrating more nonlinear regression methods to better quantify the gene–hazard relationship. In conclusion, effective methods for identifying prognosis-related genes from a causality perspective are currently lacking. CPCG is anticipated to be a valuable tool for efficiently identifying a concise combination of causal genes that exhibit stability, reliability, and good reusability and provide significant assistance in predicting cancer prognosis. The methodology and design logic of CPCG can also serve as a crucial reference for robust biological feature selection, such as features used for early diagnosis and drug efficacy identification. Key Points To identify genes that are causally associated with cancer prognosis, we propose CPCG (Cancer Prognosis’s Causal Gene), a new framework that integrates rigorous candidate gene set filtering and causal structure learning to detect the genes directly influencing or being influenced by survival time. To identify genes that are causally associated with cancer prognosis, we propose CPCG (Cancer Prognosis’s Causal Gene), a new framework that integrates rigorous candidate gene set filtering and causal structure learning to detect the genes directly influencing or being influenced by survival time. The genes identified by CPCG exhibit robustness, generalizability, and interpretability in predicting cancer prognosis. They perform well on independent test sets and remain resilient to noise interference. The genes identified by CPCG exhibit robustness, generalizability, and interpretability in predicting cancer prognosis. They perform well on independent test sets and remain resilient to noise interference. CPCG obtains a condensed gene set, predominantly encoding proteins that can enter the bloodstream. Additionally, existing literatures support that these genes are closely tied to essential cancer processes. All these make CPCG promising to practical applications and effective intervention target design. CPCG obtains a condensed gene set, predominantly encoding proteins that can enter the bloodstream. Additionally, existing literatures support that these genes are closely tied to essential cancer processes. All these make CPCG promising to practical applications and effective intervention target design. CPCG is also expected to serve as a reference for identifying other biomarkers, from the perspective of causality. CPCG is also expected to serve as a reference for identifying other biomarkers, from the perspective of causality. Supplementary Material