Lishuang Qi

Critical limitations of prognostic signatures based on risk scores summarized from gene expression levels: a case study for resected stage I non-small-cell lung cancer

Most of current gene expression signatures for cancer prognosis are based on risk scores, usually calculated as some summaries of expression levels of the signature genes, whose applications require presetting risk score thresholds and data normalization. In this study, we demonstrate the critical limitations of such type of signatures that the risk scores of samples will change greatly when they are normalized together with different samples, which would induce spurious risk classification and difficulty in clinical settings, and the risk scores of independent samples are incomparable if data normalization is not adopted. To overcome these limitations, we propose a rank-based method to extract a prognostic gene pair signature for overall survival of stage I non-small-cell lung cancer. The prognostic gene pair signature is verified in three integrated data sets detected by different laboratories with different microarray platforms. We conclude that, different from the type of signatures based on risk scores summarized from gene expression levels, the rank-based signatures could be robustly applied at the individualized level to independent clinical samples assessed in different laboratories.

Individualized identification of disease-associated pathways with disrupted coordination of gene expression

Current pathway analysis approaches are primarily dedicated to capturing deregulated pathways at the population level and cannot provide patient-specific pathway deregulation information. In this article, the authors present a simple approach, called individPath, to detect pathways with significantly disrupted intra-pathway relative expression orderings for each disease sample compared with the stable, normal intra-pathway relative expression orderings pre-determined in previously accumulated normal samples. Through the analysis of multiple microarray data sets for lung and breast cancer, the authors demonstrate individPaths effectiveness for detecting cancer-associated pathways with disrupted relative expression orderings at the individual level and dissecting the heterogeneity of pathway deregulation among different patients. The portable use of this simple approach in clinical contexts is exemplified by the identification of prognostic intra-pathway gene pair signatures to predict overall survival of resected early-stage lung adenocarcinoma patients and signatures to predict relapse-free survival of estrogen receptor-positive breast cancer patients after tamoxifen treatment.

Autophagy-related prognostic signature for breast cancer.

Autophagy is a process that degrades intracellular constituents, such as long‐lived or damaged proteins and organelles, to buffer metabolic stress under starvation conditions. Deregulation of autophagy is involved in the progression of cancer. However, the predictive value of autophagy for breast cancer prognosis remains unclear. First, based on gene expression profiling, we found that autophagy genes were implicated in breast cancer. Then, using the Cox proportional hazard regression model, we detected autophagy prognostic signature for breast cancer in a training dataset. We identified a set of eight autophagy genes (BCL2, BIRC5, EIF4EBP1, ERO1L, FOS, GAPDH, ITPR1 and VEGFA) that were significantly associated with overall survival in breast cancer. The eight autophagy genes were assigned as a autophagy‐related prognostic signature for breast cancer. Based on the autophagy‐related signature, the training dataset GSE21653 could be classified into high‐risk and low‐risk subgroups with significantly different survival times (HR = 2.72, 95% CI = (1.91, 3.87); P = 1.37 × 10−5). Inactivation of autophagy was associated with shortened survival of breast cancer patients. The prognostic value of the autophagy‐related signature was confirmed in the testing dataset GSE3494 (HR = 2.12, 95% CI = (1.48, 3.03); P = 1.65 × 10−3) and GSE7390 (HR = 1.76, 95% CI = (1.22, 2.54); P = 9.95 × 10−4). Further analysis revealed that the prognostic value of the autophagy signature was independent of known clinical prognostic factors, including age, tumor size, grade, estrogen receptor status, progesterone receptor status, ERBB2 status, lymph node status and TP53 mutation status. Finally, we demonstrated that the autophagy signature could also predict distant metastasis‐free survival for breast cancer.

An individualised signature for predicting response with concordant survival benefit for lung adenocarcinoma patients receiving platinum-based chemotherapy.

Background:For lung adenocarcinoma (LUAD) patients receiving platinum-based adjuvant chemotherapy (ACT), predictive signatures extracted from survival data solely are not directly associated with platinum response. Another limitation of reported signatures, commonly based on risk scores summarised from gene expressions, is that they could not be applied directly to samples measured by different laboratories due to experimental batch effects.Methods:Using 60 samples of LUAD patients receiving platinum-based ACT in TCGA, we pre-selected gene pairs whose within-samples relative expression orderings (REOs) were significantly associated with both pathological response and 5-year survival, from which we selected an optimal signature whose within-samples REOs could identify responders with improved 5-year survival rate.Results:A predictive signature consisting of three gene pairs was developed. In an independent data set integrated from five small data sets, the predicted responders had a significantly higher 5-year survival rate than the predicted non-responders if and only if they received platinum-based ACT (log-rank P=0.0006). The predicted responders showed a 22% absolute benefit of platinum-based ACT in 5-year survival rate compared with untreated patients (log-rank P=0.0019).Conclusions:The REO-based signature can individually predict response to platinum-based ACT with concordant survival benefit directly for LUAD samples measured by different laboratories.

Differential expression analysis at the individual level reveals a lncRNA prognostic signature for lung adenocarcinoma

BackgroundDeregulations of long non-coding RNAs (lncRNAs) have been implicated in cancer initiation and progression. Current methods can only capture differential expression of lncRNAs at the population level and ignore the heterogeneous expression of lncRNAs in individual patients.MethodsWe propose a method (LncRIndiv) to identify differentially expressed (DE) lncRNAs in individual cancer patients by exploiting the disrupted ordering of expression levels of lncRNAs in each disease sample in comparison with stable normal ordering. LncRIndiv was applied to lncRNA expression profiles of lung adenocarcinoma (LUAD). Based on the expression profile of LUAD individual-level DE lncRNAs, we used a forward selection procedure to identify prognostic signature for stage I-II LUAD patients without adjuvant therapy.ResultsIn both simulated data and real pair-wise cancer and normal sample data, LncRIndiv method showed good performance. Based on the individual-level DE lncRNAs, we developed a robust prognostic signature consisting of two lncRNA (C1orf132 and TMPO-AS1) for stage I-II LUAD patients without adjuvant therapy (P = 3.06 × 10−6, log-rank test), which was confirmed in two independent datasets of GSE50081 (P = 1.82 × 10−2, log-rank test) and GSE31210 (P = 7.43 × 10−4, log-rank test) after adjusting other clinical factors such as smoking status and stages. Pathway analysis showed that TMPO-AS1 and C1orf132 could affect the prognosis of LUAD patients through regulating cell cycle and cell adhesion.ConclusionsLncRIndiv can successfully detect DE lncRNAs in individuals and be applied to identify prognostic signature for LUAD patients.

A rank-based transcriptional signature for predicting relapse risk of stage II colorectal cancer identified with proper data sources.

The irreproducibility problem seriously hinders the studies on transcriptional signatures for predicting relapse risk of early stage colorectal cancer (CRC) patients. Through reviewing recently published 34 literatures for the development of CRC prognostic signatures based on gene expression profiles, we revealed a surprising phenomenon that 33 of these studies analyzed CRC samples with and without adjuvant chemotherapy together in the training and/or validation datasets. This data misuse problem could be partially attributed to the unclear and incomplete data annotation in public data sources. Furthermore, all the signatures proposed by these studies were based on risk scores summarized from gene expression levels, which are sensitive to experimental batch effects and risk compositions of the samples analyzed together. To avoid the above-mentioned problems, we carefully selected three qualified large datasets to develop and validate a signature consisting of three pairs of genes. The within-sample relative expression orderings of these gene pairs could robustly predict relapse risk of stage II CRC samples assessed in different laboratories. The transcriptional and functional analyses provided clear evidence that the high risk patients predicted by the proposed signature represent patients with micro-metastases.

A landscape of synthetic viable interactions in cancer

&NA; Synthetic viability, which is defined as the combination of gene alterations that can rescue the lethal effects of a single gene alteration, may represent a mechanism by which cancer cells resist targeted drugs. Approaches to detect synthetic viable (SV) interactions in cancer genome to investigate drug resistance are still scarce. Here, we present a computational method to detect synthetic viability‐induced drug resistance (SVDR) by integrating the multidimensional data sets, including copy number alteration, whole‐exome mutation, expression profile and clinical data. SVDR comprehensively characterized the landscape of SV interactions across 8580 tumors in 32 cancer types by integrating The Cancer Genome Atlas data, small hairpin RNA‐based functional experimental data and yeast genetic interaction data. We revealed that the SV interactions are favorable to cells and can predict clinical prognosis for cancer patients, which were robustly observed in an independent data set. By integrating the cancer pharmacogenomics data sets from Cancer Cell Line Encyclopedia (CCLE) and Broad Cancer Therapeutics Response Portal, we have demonstrated that SVDR enables drug resistance prediction and exhibits high reliability between two databases. To our knowledge, SVDR is the first genome‐scale data‐driven approach for the identification of SV interactions related to drug resistance in cancer cells. This data‐driven approach lays the foundation for identifying the genomic markers to predict drug resistance and successfully infers the potential drug combination for anti‐cancer therapy.

Identification of driver copy number alterations in diverse cancer types and application in drug repositioning

Results from numerous studies suggest an important role for somatic copy number alterations (SCNAs) in cancer progression. Our work aimed to identify the drivers (oncogenes or tumor suppressor genes) that reside in recurrently aberrant genomic regions, including a large number of genes or non‐coding genes, which remain a challenge for decoding the SCNAs involved in carcinogenesis. Here, we propose a new approach to comprehensively identify drivers, using 8740 cancer samples involving 18 cancer types from The Cancer Genome Atlas (TCGA). On average, 84 drivers were revealed for each cancer type, including protein‐coding genes, long non‐coding RNAs (lncRNA) and microRNAs (miRNAs). We demonstrated that the drivers showed significant attributes of cancer genes, and significantly overlapped with known cancer genes, including MYC, CCND1 and ERBB2 in breast cancer, and the lncRNA PVT1 in multiple cancer types. Pan‐cancer analyses of drivers revealed specificity and commonality across cancer types, and the non‐coding drivers showed a higher cancer‐type specificity than that of coding drivers. Some cancer types from different tissue origins were found to converge to a high similarity because of the significant overlap of drivers, such as head and neck squamous cell carcinoma (HNSC) and lung squamous cell carcinoma (LUSC). The lncRNA SOX2‐OT, a common driver of HNSC and LUSC, showed significant expression correlation with the oncogene SOX2. In addition, because some drivers are common in multiple cancer types and have been targeted by known drugs, we found that some drugs could be successfully repositioned, as validated by the datasets of drug response assays in cell lines. Our work reported a new method to comprehensively identify drivers in SCNAs across diverse cancer types, providing a feasible strategy for cancer drug repositioning as well as novel findings regarding cancer‐associated non‐coding RNA discovery.

The influence of cancer tissue sampling on the identification of cancer characteristics.

Cancer tissue sampling affects the identification of cancer characteristics. We aimed to clarify the source of differentially expressed genes (DEGs) in macro-dissected cancer tissue and develop a robust prognostic signature against the effects of tissue sampling. For estrogen receptor (ER)+ breast cancer patients, we identified DEGs in macro-dissected cancer tissues, malignant epithelial cells and stromal cells, defined as Macro-Dissected-DEGs, Epithelial-DEGs and Stromal-DEGs, respectively. Comparing Epithelial-DEGs to Stromal-DEGs (false discovery rate (FDR) < 10%), 86% of the overlapping genes exhibited consistent dysregulation (defined as Consistent-DEGs), and the other 14% of genes were dysregulated inconsistently (defined as Inconsistent-DEGs). The consistency score of dysregulation directions between Macro-Dissected-DEGs and Consistent-DEGs was 91% (P-value < 2.2 × 10−16, binomial test), whereas the score was only 52% between Macro-Dissected-DEGs and Inconsistent-DEGs (P-value = 0.9, binomial test). Among the gene ontology (GO) terms significantly enriched in Macro-Dissected-DEGs (FDR < 10%), 18 immune-related terms were enriched in Inconsistent-DEGs. DEGs associated with proliferation could reflect common changes of malignant epithelial and stromal cells; DEGs associated with immune functions are sensitive to the percentage of malignant epithelial cells in macro-dissected tissues. A prognostic signature which was insensitive to the cellular composition of macro-dissected tissues was developed and validated for ER+ breast patients.

Deconvolution of the gene expression profiles of valuable banked blood specimens for studying the prognostic values of altered peripheral immune cell proportions in cancer patients.

Background The altered composition of immune cells in peripheral blood has been reported to be associated with cancer patient survival. However, analysis of the composition of peripheral immune cells are often limited in retrospective survival studies employing banked blood specimens with long-term follow-up because the application of flow cytometry to such specimens is problematic. The aim of this study was to demonstrate the feasibility of deconvolving blood-based gene expression profiles (GEPs) to estimate the proportions of immune cells and determine their prognostic values for cancer patients. Methods and Results Here, using GEPs from peripheral blood mononuclear cells (PBMC) of 108 non-small cell lung cancer (NSCLC) patients, we deconvolved the immune cell proportions and analyzed their association with patient survival. Univariate Kaplan-Meier analysis showed that a low proportion of T cells was significantly associated with poor patient survival, as was the proportion of T helper cells; however, only the proportion of T cells was independently prognostic for patients by a multivariate Cox regression analysis (hazard ratio = 2.23; 95% CI, 1.01–4.92; p = .048). Considering that altered peripheral blood compositions can reflect altered immune responses within the tumor microenvironment, based on a tissue-based GEPs of NSCLC patients, we demonstrated a significant association between poor patient survival and the low level of antigen presentation, which play a critical role in T cell proliferation. Conclusions These results demonstrate that it is feasible to deconvolve GEPs from banked blood specimens for retrospective survival analysis of alterations of immune cell composition, and suggest the proportion of T cells in PBMC which might reflect the antigen presentation level within the tumor microenvironment can be a prognostic marker for NSCLC patients.