Journal:PCM-SABRE: A platform for benchmarking and comparing outcome prediction methods in precision cancer medicine

Full article title	PCM-SABRE: A platform for benchmarking and comparing outcome prediction methods in precision cancer medicine
Journal	BMC Bioinformatics
Author(s)	Eyal-Altman, Noah; Last, Mark; Rubin, Eitan
Author affiliation(s)	Ben-Gurion University of the Negev
Primary contact	Email: eyalnoa at post dot bgu dor ac dot il
Year published	2017
Volume and issue	18
Page(s)	40
DOI	10.1186/s12859-016-1435-5
ISSN	1471-2105
Distribution license	Creative Commons Attribution 4.0 International
Website	http://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-016-1435-5
Download	http://bmcbioinformatics.biomedcentral.com/track/pdf/10.1186/s12859-016-1435-5 (PDF)

Abstract

Background: Numerous publications attempt to predict cancer survival outcome from gene expression data using machine-learning methods. A direct comparison of these works is challenging for the following reasons: (1) inconsistent measures used to evaluate the performance of different models, and (2) incomplete specification of critical stages in the process of knowledge discovery. There is a need for a platform that would allow researchers to replicate previous works and to test the impact of changes in the knowledge discovery process on the accuracy of the induced models.

Results: We developed the PCM-SABRE platform, which supports the entire knowledge discovery process for cancer outcome analysis. PCM-SABRE was developed using KNIME. By using PCM-SABRE to reproduce the results of previously published works on breast cancer survival, we define a baseline for evaluating future attempts to predict cancer outcome with machine learning. We used PCM-SABRE to replicate previous work that describes predictive models of breast cancer recurrence, and tested the performance of all possible combinations of feature selection methods and data mining algorithms that was used in either of the works. We reconstructed the work of Chou et al. observing similar trends – superior performance of Probabilistic Neural Network (PNN) and logistic regression (LR) algorithms and inconclusive impact of feature pre-selection with the decision tree algorithm on subsequent analysis.

Conclusions: PCM-SABRE is a software tool that provides an intuitive environment for rapid development of predictive models in cancer precision medicine.

Keywords: Breast cancer, data mining, reproducible research

Background

Predicting the outcome of cancer from gene expression data is a clinically important, computationally challenging task. For example, early-stage, estrogen-receptor-positive, HER2-negative breast cancer patients that are considered to be at low risk for recurrence can avoid chemotherapy, while patients at high or intermediate risk are treated with aggressive (and harmful) chemotherapy.

Efforts to stratify patients by risk of recurrence in other tumor types, and the ability to stratify patients by overall chances of survival are not as advanced. Moreover, the relative success in risk stratification for breast cancer patients has been challenged, proposing that it in fact stratifies patients into tumor subtypes, which can be achieved with much simpler tests.

As a result, a large number of papers have been published and are still being published where gene expression data is analyzed in order to construct models that predict cancer survival or cancer recurrence. Much of these efforts are concentrated on breast cancer, the second most commonly diagnosed cancer among American women (besides skin cancer). About 1 in 8 U.S. women (about 12 percent) will develop invasive breast cancer over the course of her lifetime, and similar rates are reported worldwide. Breast cancer is an attractive domain for risk stratification as it is estimated that resection is a sufficient treatment for 70 to 80 percent of the patients, while the remaining patients will develop advanced metastatic lesions, which are largely impossible to cure. Aggressive chemotherapy will reduce the chance of advanced metastasis for those patients in that situation, though it would be harmful and unnecessary therapy for those who aren't. Thus, great efforts have been invested in stratifying patients’ risk of recurrence.

Due to the importance of risk stratification in breast cancer, combined with its relatively high abundance, breast cancer is the type of tumor for which expression profiles of newly diagnosed patients are most abundant. Several works have been published that apply machine-learning techniques to this data for predicting cancer survivability. Unfortunately, we found it quite challenging to directly compare these works for the following reasons:

1. Incomplete specification of critical stages in the process of knowledge discovery, such as feature selection.
2. Differences in the measures used to evaluate models performance. Some only provide the overall accuracy of the proposed classifier, some offer only the area under a curve (AUC), while others provide no statistical measures and only present the Kaplan-Meier charts that visualize the survival curves based on predicted classes.
3. Different studies apply different inclusion/exclusion criteria with little or no overlaps between the patients considered.

Incomplete documentation of the analytic process is a common cause for irreproducibility of published results. We conclude that there is a need for a platform that would allow researchers to describe their analytic work in the field of risk stratification for cancer patients in a reproducible way that can be used for further investigation. Such a platform should allow the replication of previous works and methodologically evaluate the impact of alterations in one or more stages of the knowledge discovery process on its performance in the task of cancer survival prediction. Such a tool can help to understand and compare the current state of predictions for breast cancer, and if applied to new cancer types, to prevent the "Tower of Babel" situation that has emerged for breast cancer.

Implementation

We developed a platform that allows replicating, comparing and improving knowledge discovery pipelines for cancer survival predictions, and demonstrate its applicability for Breast Cancer (Fig. 1). PCM-SABRE (Precision Cancer Medicine - Survival Analysis Benchmarking, Reporting and Evaluation), was developed using KNIME (Konstanz Information Miner). KNIME is a modern, flexible and intuitive open-source data analytics platform that allows performing sophisticated statistics and data mining analysis to develop, among other things, predictive models. We chose KNIME since it is a popular, user-friendly software that does not require programing knowledge. Its node-based workflow structure allows easily assessing the impact of changing one knowledge discovery step (for example, data mining algorithm) on the predictive performance without changing any other steps of the workflow. Another major benefit of KNIME is the ability to create new nodes; this feature is particularly useful when a researcher is interested in integrating a new method he developed into an existing KNIME workflow.

Figure 1. Screenshot of PCM-SABRE

We designed PCM-SABRE workflow according to the common steps of knowledge discovery in data. First, the user can use a supplied dataset or load a new dataset. The dataset has to be a .csv file, in the form of a table in which the rows represent the patients and the columns represent clinical data, gene expression data, or any other types of variables. The dependent variable can be binary or continuous (it will be transformed into a binary variable) and needs to represent survival time (for example, relapse-free-survival time or death time).

The second Meta-node is the preprocessing step, where a binary dependent variable is created and patients with missing data or censored survival information are being filtered. We chose to use a default threshold of five years in order to split the continues survival variable into HIGH (t < 5 years) or LOW (t ≤ 5 years) risk, but this threshold is an input parameter that can be changed in a way that will be explained later. Missing values imputation is performed using random forest classification that builds a model using the non-missing rows and predicts the variable value for the missing rows. The default version of PCM-SABRE allows selecting patients according to their ER status and lymph node status, but the "Select Patients" Meta-node is optional and can be easily modified to meet other inclusion/exclusion criteria. The third Meta-node is the feature selection step, where the users can choose between two methods of feature selection (information gain or ANOVA) or add another feature selection method (from the available nodes in KNIME, using scripting or external tools). The fourth Meta-node is the modeling step, where we offer a choice of five well-known and relevant classifiers. The methods included in the out-of-the-box basic version of the workflow are described in Table 1.

Table 1. Machine learning methods available in PCM-SABRE Meta-node Method KNIME node Default parameters 1.1 Select Patients Estrogen Receptor (ER) status R script 1.2 Select Patients Lymph Node (LN) status R script 2.1 Feature Selection Information Gain (InfoGain) InformationGainCalculator (Community node – Palladian) Top 100 ranked 2.2 Feature Selection ANOVA One-way ANOVA Include genes with p-value < 1.0E-6 3.1 Modeling Logistic Regression (LR) Logistic (3.7) (Weka node) Ridge = 1.0E-8, 3.2 Modeling Random Forest (RF) Random Forest Learner Split criteria = Information Gain Ratio, Number of models = 350 3.3 Modeling Artificial Neural Network (ANN) PNN Learner (DDA) Theta Minus = 0.2, Theta Plus = 0.4 3.4 Modeling K-Nearest Neighbors (KNN) IBK (3.7) (Weka node) KNN = 15 3.5 Modeling Support Vector Machine (SVM) SVM Learner Kernel = RBF, sigma = 0.2

It should be noted that thanks to the design of KNIME, adding additional Modeling and Feature Selection methods involves just dropping additional nodes in the appropriate Meta-nodes and connecting them by drag-and-drop using the existing methods as templates. Our experience with experimental biologists suggests that any oncology researcher without programming capabilities can achieve this with little or no special training. Fig. 2 illustrates how the user can easily and quickly add additional classifier to the workflow: (1) double-click modeling → new model → cross-validation (2) delete the decision tree learner and predictor (3) choose from the Node Repository another learner and predictor nodes, then drag-and-drop them instead of the deleted nodes (4) connect the X-Partitioner node Training data output into the Learner node input, connect the Learner node PMML output into the PMML input of the Predictor node, connect the Predictor node to the X-Aggregator node, and connect the X-partitioner Test data output to the Predictor node.

Figure 2. Demonstration of drag-and-drop model replacement (Naïve Bayes instead of decision tree)

The fifth Meta-node is the evaluation step, which calculates the performance measures of different models (among them the accuracy and the area under the ROC). An important feature of PCM-SABRE is a .csv file (flow_variables.csv) that allows the user to control some default input parameters without the need to change these parameters inside the specific KNIME nodes. The controlled input parameters are: (1) Feature selection method (default = infoGain), ER status (default = all patients), Lymph node status (default = all patients) and the threshold for the binary survival variable (default = five years). Changing and adding another input parameter is simple and only requires filling cells in Excel. Additional details on how to use PCM-SABRE can be found in the user manual.

PCM-SABRE output includes, for each combination of a feature selection method with a classification algorithm, (1) performance measures, (2) ROC analysis and (3) a list of ranked features.

Results

We developed PCM-SABRE (available as Additional file 1) as a software system that allows for the comparison and improvement of expression-based predictive models of cancer patients. We used PCM-SABRE to replicate previous work that describes predictive models of breast cancer recurrence, and we evaluated the performance of all possible combinations of feature selection methods and data mining algorithms that was used in either of the works.

Using PCM-SABRE for replicating a previous work that utilizes machine learning to induce outcome prediction models

We first demonstrate the value of PCM-SABRE to investigators implementing new machine learning pipelines for breast cancer recurrence prediction by replicating the work of Chou et al. Our analysis reconstructs the paper to the best of our ability, with the following exceptions:

• We use KNIME rather than the original software (Clementine 10.1), and we use as input data a more current compendium of expression data (to be called the Györffy dataset for the rest of this paper). The dataset is available for download here: http://kmplot.com/analysis/index.php?p=download.
• The Györffy dataset originally contained 1809 examples (breast cancer patients) and 22,216 features (clinical features and probes expression level).
• A binary class attribute was created indicating whether the cancer recurred within five years or not.

To best reproduce the original work, we made the following modifications to the default out-of-the-box KNIME pipeline:

1. A preprocessing step was added that reproduces the preprocessing performed in the original paper. This step was conducted with a specialized R script written for this purpose. In this step, features were transformed from probe to gene level. After the transformation, the dataset contained 13,725 features.
2. In the preprocessing Meta-node, we removed lymph node positive patients and patients with follow-up time of less than five years (1219 patients remained).
3. Two new feature selection methods were added to the feature selection Meta-node (Fig. 3):

a. The Mann–Whitney U test was used for decreasing the number of genes from 13,725 to 100 exactly as described by Chou et al. The Mann–Whitney U non-parametric test, which is also known as the Wilcoxon rank sum test, tests for differences between two groups on a single, ordinal variable with no specific distribution. The U statistic of each group is calculated as a difference between the actual sum of ranks of the group observations and the sum expected value under the null hypothesis that the distribution of the ordinal variable in both groups is equal. (See Chou et al. for more details.)

b. A compound selection method was added, in which the results of the DT algorithm were used to determine which features will be retained for PNN and LR analysis.

   DA (Decision tree + Probabilistic neural network) DT + PNN → DA

   DL (Decision tree + Logistic regression) DT + LR → DL

Figure 1. Screenshot of PCM-SABRE

The classification performance results from PCM-SABRE and from the original paper are compared in Table 2. In contrast to the original work, PCM-SABRE reports that LR has the best performance. Moreover, both show a different trend when adding the DT feature selection methods. It is worth noting that the estimated accuracy reported by PCM-SABRE is higher than in the original work. This may be because a different dataset was used for the analysis.

Table 2. Predictive power (in terms of percent accuracy) of several feature selection methods combined with different classification models. AUC results are shown in parentheses. Prediction model PCM-SABRE pipeline Chou et al. MW U test Feature selection InfoGain ANOVA MW U test RF 76.52 (NA) 77.70 (NA) 76.10 (NA) NA LR 76.27 (73.0) 66.55 (62.49) 75.68 (70.95) 64.12 (58.96) PNN 76.52 (74.09) 76.27 (75.21) 74.58 (72.32) 69.54 (63.88) KNN 75.76 (67.78) 75.34 (68.48) 76.10 (70.30) NA SVM 72.64 (NA) 72.64 (NA) 72.64 (NA) NA DT 70.19 (60.59) 68.07 (61.53) 64.44 (57.34) 63.45 (56.90) DL NA NA 75.34 (71.71) 68.90 (61.66) DA NA NA 75.51 (72.23) 65.91 (61.65)

Using PCM-SABRE for optimizing and improving breast cancer outcome prediction

For the task of breast cancer outcome prediction, we again used the dataset published by Györffy et al. and conducted the preprocessing steps maintained above. Table 2 summarizes the performance of all combinations of feature selection methods and classification algorithms. LR, PNN, KNN, and DT performed better combined with the InfoGain feature selection method, in terms of Accuracy but not in terms of AUC. RF performed better combined with the ANOVA feature selection method and achieved the highest Accuracy (77.70%).

Notes

This presentation is faithful to the original, with only a few minor changes to presentation. In some cases important information was missing from the references, and that information was added. Some grammar were corrected when necessary. Some tables and figures were moved slightly to match up better with their text reference. What was citation #11 originally (link to the Györffy dataset) was removed as a citation and turned into an inline URL.