I-Boost: an integrative boosting approach for predicting survival time with multiple genomics platforms

Abstract

We propose a statistical boosting method, termed I-Boost, to integrate multiple types of high-dimensional genomics data with clinical data for predicting survival time. I-Boost provides substantially higher prediction accuracy than existing methods. By applying I-Boost to The Cancer Genome Atlas, we show that the integration of multiple genomics platforms with clinical variables improves the prediction of survival time over the use of clinical variables alone; gene expression values are typically more prognostic of survival time than other genomics data types; and gene modules/signatures are at least as prognostic as the collection of individual gene expression data.

Keywords: Cancer genomics; Data integration; Gene modules; Variable selection.

Fig. 1

Simulation settings and results. a Prediction accuracy of LASSO, elastic net, I-Boost-CV, and I-Boost-Permutation measured by risk correlation under three different settings. b The average number of variables selected by the four methods under three different settings. Different types of the selected variables are represented by different colors. c MSE of the four methods under three different settings. The error is decomposed into errors of parameters for different data types, as represented by different colors. d Number of signal variables and distribution of signals across different data types for the three simulation settings. The number of signal variables is zero if the proportion of signals of the data type is 0%. Abbreviations are as follows: GeneExp represents individual gene expression, Module represents gene module, Clinical represents clinical variable, CNV represents copy number variant, Mutation represents somatic mutation, miRNA represents microRNA expression, and Protein represents protein expression

Fig. 2

Analysis results for the TCGA LUAD, KIRC, and pan-cancer data sets using LASSO and elastic net. Each row represents a particular combination of data types used as predictors, as indicated by the box on the left. Each dot is an average C-index value obtained by performing LASSO or elastic net on 30 training and testing data set pairs. See the caption of Fig. 1 for the abbreviations of the data types

Fig. 3

Analysis results for the TCGA LUAD, KIRC, and pan-cancer data sets using elastic net, I-Boost-CV, and I-Boost-Permutation. Each row represents a particular combination of data types used as predictors, as indicated by the box on the left. Each dot is an average C-index value obtained by performing elastic net, I-Boost-CV, or I-Boost-Permutation on 30 training and testing data set pairs. See the caption of Fig. 1 for the abbreviations of the data types

Fig. 4

NRI values for the TCGA LUAD, KIRC, and pan-cancer data sets using I-Boost-CV or I-Boost-Permutation. Each dot represents the average NRI between a model with both clinical and genomic variables (estimated by I-Boost-CV or I-Boost-Permutation) and the model with clinical variables only (estimated by maximum partial-likelihood estimation) over 30 training and testing data set pairs

Fig. 5

NRI values between models containing individual gene expression data and models containing gene modules under the TCGA LUAD, KIRC, and pan-cancer data sets. Each dot represents the average NRI obtained by fitting I-Boost-CV or I-Boost-Permutation on two sets of predictors over 30 training and testing data set pairs. The first set of predictors contains a combination of data types and gene modules; the second set of predictors contains the same combination of data types and individual gene expression data. A positive NRI represents better prediction using the model with gene modules

Fig. 6

Analysis results for the TCGA LUAD, KIRC, and pan-cancer data sets, using elastic net, I-Boost-CV, and I-Boost-Permutation on nested models. In the left panel, the leftmost dots are fixed at zero, and each remaining dot represents the average NRI obtained by fitting elastic net, I-Boost-CV, or I-Boost-Permutation over 30 training and testing data set pairs. Each dot except the leftmost dots represents the maximum NRI between a model that contains one more data type than the model corresponding to the dot on the left and the model corresponding to the dot on the left. Above each dot, the name of the additional data type is included. In the right panel, the average C-index values and the average numbers of selected variables for the models shown in the left panel are plotted. The arrows indicate the orders of models with respect to the number of data types they contain. See the caption of Fig. 1 for the abbreviations of the data types