Patient-oriented gene set analysis for cancer mutation data

Abstract

Recent research has revealed complex heterogeneous genomic landscapes in human cancers. However, mutations tend to occur within a core group of pathways and biological processes that can be grouped into gene sets. To better understand the significance of these pathways, we have developed an approach that initially scores each gene set at the patient rather than the gene level. In mutation analysis, these patient-oriented methods are more transparent, interpretable, and statistically powerful than traditional gene-oriented methods.

Figure 1

Observed (blue) and expected number of altered samples (T_s) across the gene sets in the dataset from [4], as a function of the size of the gene set. The expected numbers are computed using the permutation null and denoted by E(T_s|H_s0). The values within two standard deviations of the E(T_s|H_s0) are also shown.

Figure 2

CAT plot comparing the patient-oriented methods to the gene-oriented method for the glioblastoma data from [4]. Each graph represents a pairwise comparison of two methods: The gene sets are ranked according to the P-value, a list of top gene sets is created at each rank, then the fraction of gene sets in the list common to both methods is graphed.

Figure 3

Power analysis. Plot of the average number of truly positive (spiked-in) gene sets included in the top X gene sets, as X varies. The red line indicates the ideal scenario over 100 simulation runs. Simulations use the permutation null (top panel) or the passenger null (bottom panel) data-generating mechanisms. The four patient-oriented methods return more true positives than the gene-oriented method, except when one focuses on short lists, which include sets that are relatively easy to detect. (Note that the overlap of the two methods which use the passenger null looks like a star (*).)

Figure 4

Calibration analysis. Plot of the average true fraction of false discoveries versus the average q-value over 100 simulation runs. The identity line indicates the ideal scenario of a perfect FDR-control rate; being above the identity line indicates anti-conservative behavior, and being below the identity line indicates conservative behavior. Simulations use the permutation null (top panel) or the passenger null (bottom panel). The gene-oriented method shows very anti-conservative behavior while the patient-oriented methods are generally calibrated or conservative.