Integrating Biological Covariates into Gene Expression-Based Predictors of Radiation Sensitivity

Abstract

The use of gene expression-based classifiers has resulted in a number of promising potential signatures of patient diagnosis, prognosis, and response to therapy. However, these approaches have also created difficulties in trying to use gene expression alone to predict a complex trait. A practical approach to this problem is to integrate existing biological knowledge with gene expression to build a composite predictor. We studied the problem of predicting radiation sensitivity within human cancer cell lines from gene expression. First, we present evidence for the need to integrate known biological conditions (tissue of origin, RAS, and p53 mutational status) into a gene expression prediction problem involving radiation sensitivity. Next, we demonstrate using linear regression, a technique for incorporating this knowledge. The resulting correlations between gene expression and radiation sensitivity improved through the use of this technique (best-fit adjusted R² increased from 0.3 to 0.84). Overfitting of data was examined through the use of simulation. The results reinforce the concept that radiation sensitivity is not driven solely by gene expression, but rather by a combination of distinct parameters. We show that accounting for biological heterogeneity significantly improves the ability of the model to identify genes that are associated with radiosensitivity.

Figure 1

Investigation of building predictors for radiation sensitivity in 48 cell lines. (a) Classification accuracy of radiation sensitivity predictor built from 48 cell lines, using different numbers of features in the regression model. (b) Classification accuracy of radiation sensitivity predictor built from 48 cell lines, using different types of normalization. MAS5.0 and MAS4.0 algorithms generated the most accurate predictors. (c) Classification accuracy of radiation sensitivity predictor built from 48 cell lines, using different types of classification algorithms, including linear regression, least median, and SMO.

Figure 2

Biological characteristics differ when considering 35 cell lines versus an expanded set of 48 cell lines. (a) The proportion of RAS wild-type cell lines increased (60% to 69%). (b) The proportion of p53 wild-type cell lines increased (26% to 35%). (c) Tissue of origin of cell lines did not change significantly. (d) Venn diagram showing the lack of concordance in correlation when using a test for correlation (p < 0.05) using only RAS mutant or RAS wt cell lines in the 35-cell line set. Only 16 probesets were found correlated in both sets.

Figure 3

Adj-R² values for linear equations fitting SF2 on 48 cell lines. Adj-R² values increase systematically as more covariates are included in the linear model.

Figure 4

Change in Adj-R² values obtained by incorporating interaction terms in linear model for either biological indicators or random variables. Linear models were created for each gene, incorporating the three biological indicators. Differences in Adj-R² values were computed for each experiment between using an additive model and an interaction-based model. The mean difference in R² was recorded when adding interactions terms to the individual linear models.