CPSM: R-package of an Automated Machine Learning Pipeline for Predicting the Survival Probability of Single Cancer Patient

Article activity feed

1. 

   GigaScience

   Aug 7, 2026

   AbstractAccurate survival prediction is vital for optimizing treatment strategies in clinical practice. The advent of high-throughput multi-omics data and computational methods has enabled machine learning (ML) models for survival analysis. However, handling high-dimensional omics data remains challenging.This study introduces the Cancer Patient Survival Model (CPSM), an R package developed to provide individualized survival predictions through a fully integrated and reproducible computational pipeline. The CPSM package encompasses nine modules that streamline the survival modeling workflow, organized into four key stages: (1) Data Preprocessing and Normalization, (2) Feature Selection, (3) Survival Prediction Model Development, and (4) Visualization. The visual tools facilitate the interpretation of survival predictions, enhancing …

   AbstractAccurate survival prediction is vital for optimizing treatment strategies in clinical practice. The advent of high-throughput multi-omics data and computational methods has enabled machine learning (ML) models for survival analysis. However, handling high-dimensional omics data remains challenging.This study introduces the Cancer Patient Survival Model (CPSM), an R package developed to provide individualized survival predictions through a fully integrated and reproducible computational pipeline. The CPSM package encompasses nine modules that streamline the survival modeling workflow, organized into four key stages: (1) Data Preprocessing and Normalization, (2) Feature Selection, (3) Survival Prediction Model Development, and (4) Visualization. The visual tools facilitate the interpretation of survival predictions, enhancing clinical decision-making. By providing an end-to-end solution for multi-omics data integration and analysis, CPSM not only enhances the precision of survival predictions but also aids in discovering clinically relevant biomarkers.

   This work has been peer reviewed in GigaScience (see https://doi.org/10.1093/gigascience/giag067), which carries out single-anonymized peer review. These reviews are published under a CC-BY 4.0 license and were as follows:

   Reviewer 3：

   （Previous submission）

   The manuscript presents a well-designed R package that provides an end-to-end solution for individualized cancer survival prediction, integrating gene expression and clinical data into a single workflow. The article is clearly written, methodologically thorough, and demonstrates the software's strong utility for both research and potential clinical applications.

   Major Remarks:

   1. The current evaluation relies on a single random train/test split for each TCGA dataset. Given the very small test cohorts (13-17 patients), performance estimates (C-index, MAE, accuracy) are unstable and may not reflect true generalization. Please consider adopting more robust evaluation strategies such as k-fold cross-validation, repeated random sub-sampling, or nested CV. This would strengthen confidence in the reported results and reduce the risk of overfitting. Also I am concerned about repeatability of Lasso, given the number of features.
   2. Training metrics are often extremely high while test metrics drop substantially. This pattern suggests overfitting, likely due to high dimensionality (60,000+ features) and small sample sizes. The authors should acknowledge this explicitly and discuss strategies to mitigate overfitting (e.g., stricter regularization, dimensionality reduction, feature stability analysis).
   3. The manuscript applies CPSM to GBM, LAML, and PAAD. However, the rationale for selecting these specific cancer types is not discussed. It would strengthen the work if the authors explained whether the choice was driven by data availability, clinical relevance, or to showcase CPSM's applicability across tumor types with different characteristics. Without justification one could argue that those 3 were cherry picked to show the advantage of CPSM, especially since for other cancer types, e.g. BRCA the number of available cases is much larger and for this reason they make better candidates for a study that requires train/test splits.
   4. Although the title and abstract frame CPSM as a multi-omics integration tool, all demonstrations and evaluations are conducted exclusively with RNA-seq expression data (FPKM) in combination with clinical features. No examples of proteomics, metabolomics, or other omics integration are shown. This creates a mismatch between the stated scope (multi-omics) and the evidence provided. The authors should either include at least one additional omics dataset to demonstrate multi-omics integration, or adjust the framing to emphasize that CPSM is currently optimized for transcriptomics + clinical data, while being adaptable to other omics in principle.
   5. The manuscript compares CPSM's functionality with existing packages (glmnet, MTLR, randomForestSRC, rms, etc.), but no direct head-to-head performance comparison is presented. Including such benchmarking would make the performance gains of CPSM more convincing.

   Minor:

   1. The CPSM package requires users to run many separate functions in sequence. While this demonstrates flexibility, it may be overwhelming for non-expert users. Wrapping the most common steps into a single high-level function would make the package more user-friendly while still allowing advanced users to call individual functions.
   2. The vignette uses built-in example data, which is useful for illustration. However, it would be more impactful to include a real-life example starting from external data (loading data and creating SummarizedExperiment) and then walking through only the essential analysis steps. This would better reflect typical user workflows.
   3. The vignette could highlight key results more clearly, e.g., "This step produces the list of selected genes and their coefficients", so users understand what to look for after each step.
   4. The CPSM package vignette does not demonstrate best practices for robust model evaluation (e.g., cross-validation or repeated splits). Including a short demonstration or guidance would help users avoid overfitting and improve reliability of results. 5)Some minor language errors: "HK drafted and manuscript and PD and US refined the drafted manuscript.", "This function employ normalize.quantiles…"

   （Resubmitted manuscript）

   The revised manuscript shows clear effort in restructuring parts of the text and expanding several methodological descriptions. However, most of the major concerns raised in the previous review remain unresolved, and several central methodological limitations persist.

   1. The evaluation strategy continues to rely on a single random train/test split for each cancer type, with very small test sets (13-17 samples). As noted previously (Major Remark 1), such a design does not provide reliable performance estimates, especially for metrics as variable as C-index or MAE with small n. The supplementary material also indicates that no form of stability assessment has been introduced. It also reiterates the use of LASSO without addressing its known instability in high-dimensional settings and the authors do not discuss how the default 10-fold cross-validation used in glmnet (via nfolds) may further affect feature selection variability in such small cohorts
   2. The overfitting concerns (Major Remark 2) remain and are even more apparent in the extended results. Training performance is often extremely high, sometimes essentially perfect for risk-group prediction, while test performance declines sharply. Although some additional explanatory text was added, there is still no substantive discussion of why these models overfit (tens of thousands of features, small cohorts, noisy FPKM data) or how this might be mitigated. The supplement also demonstrates substantial instability in selected features (e.g., only three of ten GBM LASSO genes present in the CGGA dataset), further indicating that the models are sensitive to sampling noise. Combined with the extremely small test cohorts, this limits the interpretability of all reported metrics.
   3. The rationale for selecting only GBM, LAML, and PAAD is still insufficient (Major Remark 3). The supplement's comparison with the PAWPH method makes this even clearer: both CPSM and PAWPH perform reasonably well on LAML, but fail or perform inconsistently on GBM and PAAD. Without evaluation on larger TCGA cohorts (e.g., BRCA, LUAD), it remains difficult to understand whether these results generalize or are driven by these smaller datasets. The choice of these three tumour types still appears somewhat arbitrary.
   4. While the manuscript and supplement together now contain a functional comparison table and a PAWPH-based experiment, this still does not amount to a meaningful benchmark against existing survival modelling tools (Major Remark 5). Widely used baselines - Cox with elastic net, RSF, MTLR, or modern deep learning frameworks are not compared directly. Moreover, the PAWPH comparison highlights performance failures in GBM and PAAD, which does not support the claim that CPSM provides superior prediction accuracy.
   5. Some of the minor issues from the earlier review, concerning the package documentation were addressed (external data, information on the outputs), however the package vignette still does not demonstrate best practices for robust model evaluation (Minor remark 4). The package also still requires many separate function calls without an easier, high-level wrapper, which is acceptable but will limit the packages usability. Additionally, I find it confusing that the authors use their own package versioning in the text, which differs from the versioning used in Bioconductor. I have to assume that the article's v1.1.4 corresponds to Bioconductor's 1.3.0. Overall, although the manuscript is more detailed and somewhat clearer than before, the fundamental methodological issues remain unresolved. Substantial revisions would be needed before the results could be considered robust.

   Read the original source
2. 

   GigaScience

   Aug 7, 2026

   AbstractAccurate survival prediction is vital for optimizing treatment strategies in clinical practice. The advent of high-throughput multi-omics data and computational methods has enabled machine learning (ML) models for survival analysis. However, handling high-dimensional omics data remains challenging.This study introduces the Cancer Patient Survival Model (CPSM), an R package developed to provide individualized survival predictions through a fully integrated and reproducible computational pipeline. The CPSM package encompasses nine modules that streamline the survival modeling workflow, organized into four key stages: (1) Data Preprocessing and Normalization, (2) Feature Selection, (3) Survival Prediction Model Development, and (4) Visualization. The visual tools facilitate the interpretation of survival predictions, enhancing …

   AbstractAccurate survival prediction is vital for optimizing treatment strategies in clinical practice. The advent of high-throughput multi-omics data and computational methods has enabled machine learning (ML) models for survival analysis. However, handling high-dimensional omics data remains challenging.This study introduces the Cancer Patient Survival Model (CPSM), an R package developed to provide individualized survival predictions through a fully integrated and reproducible computational pipeline. The CPSM package encompasses nine modules that streamline the survival modeling workflow, organized into four key stages: (1) Data Preprocessing and Normalization, (2) Feature Selection, (3) Survival Prediction Model Development, and (4) Visualization. The visual tools facilitate the interpretation of survival predictions, enhancing clinical decision-making. By providing an end-to-end solution for multi-omics data integration and analysis, CPSM not only enhances the precision of survival predictions but also aids in discovering clinically relevant biomarkers.

   This work has been peer reviewed in GigaScience (see https://doi.org/10.1093/gigascience/giag067), which carries out single-anonymized peer review. These reviews are published under a CC-BY 4.0 license and were as follows:

   Reviewer 2：

   （Previous submission）

   This manuscript developed an R package to predict survival and risk levels of individual patient. Several core bioinformatics functions were included. TCGA datasets were used as examples to illustrate the functions and package. There are a few questions or comments listed below.

   1. The integration of multi-omics data is a strong motivation for this R package, according to the Introduction and Title. However, I could not find any relevant functions or examples in Method and Result. I am not sure if multi-omics data could be handled. If not, the authors should redefine their scope. If yes, more descriptions and examples should be provided.
   2. How does the normalization work for train and test data? Is it performed separately on each set or jointly on the combined data?
   3. Examples using only TCGA data are not sufficient to demonstrate the effectiveness of the R package. There are many other publicly available datasets beyond TCGA. Please include additional examples using non-TCGA data.
   4. The validity of the Cox proportional hazards (CoxPH) model depends on several assumptions. The R package should include a model diagnostics function to assess model fit. Additionally, alternative approaches should be provided for cases where these assumptions are not met.
   5. For the visualization, does the R package provide function/argument to customize the style of figure, such as font size, line color, etc.?

   （Resubmitted manuscript）

   Thanks for addressing the comments. There is one previous comment have not been fully addressed, which is regarding the model assumption exam functions. The CoxPH model assumption function is standard and should be added.

   Read the original source
3. 

   GigaScience

   Aug 7, 2026

   AbstractAccurate survival prediction is vital for optimizing treatment strategies in clinical practice. The advent of high-throughput multi-omics data and computational methods has enabled machine learning (ML) models for survival analysis. However, handling high-dimensional omics data remains challenging.This study introduces the Cancer Patient Survival Model (CPSM), an R package developed to provide individualized survival predictions through a fully integrated and reproducible computational pipeline. The CPSM package encompasses nine modules that streamline the survival modeling workflow, organized into four key stages: (1) Data Preprocessing and Normalization, (2) Feature Selection, (3) Survival Prediction Model Development, and (4) Visualization. The visual tools facilitate the interpretation of survival predictions, enhancing …

   AbstractAccurate survival prediction is vital for optimizing treatment strategies in clinical practice. The advent of high-throughput multi-omics data and computational methods has enabled machine learning (ML) models for survival analysis. However, handling high-dimensional omics data remains challenging.This study introduces the Cancer Patient Survival Model (CPSM), an R package developed to provide individualized survival predictions through a fully integrated and reproducible computational pipeline. The CPSM package encompasses nine modules that streamline the survival modeling workflow, organized into four key stages: (1) Data Preprocessing and Normalization, (2) Feature Selection, (3) Survival Prediction Model Development, and (4) Visualization. The visual tools facilitate the interpretation of survival predictions, enhancing clinical decision-making. By providing an end-to-end solution for multi-omics data integration and analysis, CPSM not only enhances the precision of survival predictions but also aids in discovering clinically relevant biomarkers.

   This work has been peer reviewed in GigaScience (see https://doi.org/10.1093/gigascience/giag067), which carries out single-anonymized peer review. These reviews are published under a CC-BY 4.0 license and were as follows:

   Reviewer 1：

   （Previous submission）

   This paper focused on the development of a new R/Bioconductor package that is able to provide feature selection and survival prediction for omics datasets with clinical survival outcomes. The strength of the package is that it provides an end-to-end solution for users, starting from data normalisation to results presentation (visualisation). It is also a comprehensive package in terms of survival model fitting, feature selection, performance evaluation and visualisation. However, I found the pipeline and methods included in the package may have issues that need to be fixed prior to applying the package for any analytical purpose. I think the "robust method" statement might be overclaimed.

   I have the following comments/questions to the authors.

   1. Could you include the brier score (maybe the one from the survAUC package) as the evaluation metric besides C-index and MAE?
   2. When calculating the PI score, are all censored and non-censored data used to create the binary "high-risk" and "low-risk" group in training? If censored data is included in the calculation, adjustment should be applied for the contribution of the censored samples, i.e. could be a similar idea with the inverse probability of censoring rate (IPCW). If not, I assume only very few samples are used in this calculation when the censoring rate in the data is high, say, up to 70%, which is also common in "omics' survival data.
   3. Since MTLR is used for prediction, it can also be used for feature selection. Why was the LASSO method being picked (and the univariate Cox) for feature selection?
   4. It is known that "omics" data contains covariates that are highly correlated with each other. Therefore, I am not sure of the application of multiple univariate Cox models for feature selection.
   5. It seems that from reading both the manuscript and the supplementary material that in the case studies, there is no repeated training and testing data split. There are also no cross-validation performances. I am not sure how reliable the reported results are based on only one random split of data.
   6. Presentation of the result tables can be improved. Outstanding performances mentioned in the results section in Table 1 and Table 2 could be highlighted in bold. Names of method 1 to 4 could be more informative, for instance, method 3 could be renamed as Method "PI_Cli".
   7. In Table 3, Glmnet can produce predicted survival probability for each individual in the test data. I am not sure what does the "No" mean under the "Single-patient survival probability prediction" column. MTLR should be able to run on p>n (high-dim) datasets. I am not sure why there is a "No" under the last column.
   8. In terms of the robustness, how does the current pipeline method compare to existing robust methods? For example, how does the feature selection performance and the prediction performance compare to "pawph" and the sure independence screening?

   References: Luo B, Gao X, Halabi S. Penalized weighted proportional hazards model for robust variable selection and outlier detection. Stat Med 2022;41:3398-420. https://doi.org/10.1002/sim.9424. Fan J, Feng Y, Wu Y. High-dimensional variable selection for Cox's proportional hazards model. Inst Math Stat (IMS) Collect 2010;6:70-86. https://doi.org/10.1214/10-IMSCOLL606.

   Read the original source
4. 

   Version published to 10.1101/2024.11.14.623597 on bioRxiv

   Nov 15, 2024