Transcriptional Profiles from Paired Normal Samples Offer Complementary Information on Cancer Patient Survival--Evidence from TCGA Pan-Cancer Data

Abstract

Although normal tissue samples adjacent to tumors are sometimes collected from patients in cancer studies, they are often used as normal controls to identify genes differentially expressed between tumor and normal samples. However, it is in general more difficult to obtain and clearly define paired normal samples, and whether these samples should be treated as "normal" due to their close proximity to tumors. In this article, by analyzing the accrued data in The Cancer Genome Atlas (TCGA), we show the surprising results that the paired normal samples are in general more informative on patient survival than tumors. Different lines of evidence suggest that this is likely due to tumor micro-environment instead of tumor cell contamination or field cancerization effect. Pathway analyses suggest that tumor micro-environment may play an important role in cancer patient survival either by boosting the adjacent metabolism or the in situ immunization. Our results suggest the potential benefit of collecting and profiling matched normal tissues to gain more insights on disease etiology and patient progression.

Figure 1. Adjacent normal samples’ transcriptional profiles offer different but clinically relevant information.

(a) Hierarchical clustering of tumor and normal samples of 60 breast cancer patients with 6,000 most varying genes. (b) Rand indexes comparing groupings based on clinical or genetic features with those based on gene expression data with the 6,000 most varying genes of each data type using k-means clustering with k = 5, where a higher Rand index value indicates a higher degree of similarity.

Figure 2. Normal samples consistently offer additional information to improve survival prediction performances across multiple cancer cohorts.

The figures show the boxplots comparing distributions of survival prediction Mean Cross Validation Error from 20 random runs for penalized cox regression model from different datasets for six cancer cohorts of the RNASeq data.

Figure 3. Normal samples provide an overall better prediction accuracy for patient survival and can be especially helpful for those patients with outlying survival prediction performances using tumor sample data alone.

(a) Heatmap of the consistency matrix from tumor samples in the 60-patient breast cancer data using Elastic Net Cox Regression. Red color represents low consistency, while yellow represents high consistency. Individual 49 is highlighted with blue arrow as being an outlier with generally poor consistency of prediction with real survival using tumor data. (b) Consistency matrix calculated from correspondent normal data. Individual 49 is highlighted also, and is generally with comparable consistency levels with others individuals and no longer outlying.

Figure 4. Gene set analyses show that metabolic, immune and cell growth related pathways are involved in boosting the survival related signals in adjacent normal samples.

(a) Heat map showing the enriched pathways with FDR less than 0.25 consistently found at least five times across 18 cancer cohorts by data types considerations. Cells are colored according to the enrichment level calculated by transforming FDR p values to z values. The red color shows pathways that are up regulated for longer-surviving patients, while the blue color shows pathways that are down regulated for longer-surviving patients. Each cell has three sub cells, representing tumor data, normal data, and fold change data respectively from up to down. X-axis labels are colored according to the pathway types, with red being metabolic related pathways, yellow being immune related pathways, and purple being cell growth related pathways. (b) Boxplots comparing the distribution of concordance rates for enriched pathways found for each pair of cancer cohorts using different data types and different FDR thresholds to select enriched pathways. (c) Density plots of Univariate Cox Regression p values (representing each gene set signature’s correlation with survival using different types of data) for the 60-patient breast cancer data. (d) Boxplots showing the distributions of the top 100 log ratios of the p values from Univariate Cox Regression between either normal vs. tumor (tumor associated) or tumor vs. normal (normal associated) for each gene set signature. This is also for the 60-patient breast cancer data set.