Host-response-based gene signatures for tuberculosis diagnosis: A systematic comparison of 16 signatures

Abstract

Background The World Health Organization (WHO) and Foundation for Innovative New Diagnostics (FIND) have published target product profiles (TPPs) calling for non-sputum-based diagnostic tests for the diagnosis of active tuberculosis (ATB) disease and for predicting the progression from latent tuberculosis infection (LTBI) to ATB. A large number of host-derived blood-based gene-expression biomarkers for diagnosis of patients with ATB have been proposed to date, but none have been implemented in clinical settings. The focus of this study is to directly compare published gene signatures for diagnosis of patients with ATB across a large, diverse list of publicly available gene expression datasets, and evaluate their performance against the WHO/FIND TPPs. Methods and findings We searched PubMed, Gene Expression Omnibus (GEO), and ArrayExpress in June 2018. We included all studies irrespective of study design and enrollment criteria. We found 16 gene signatures for the diagnosis of ATB compared to other clinical conditions in PubMed. For each signature, we implemented a classification model as described in the corresponding original publication of the signature. We identified 24 datasets containing 3,083 transcriptome profiles from whole blood or peripheral blood mononuclear cell samples of healthy controls or patients with ATB, LTBI, or other diseases from 14 countries in GEO. Using these datasets, we calculated weighted mean area under the receiver operating characteristic curve (AUROC), specificity at 90% sensitivity, and negative predictive value (NPV) for each gene signature across all datasets. We also compared the diagnostic odds ratio (DOR), heterogeneity in DOR, and false positive rate (FPR) for each signature using bivariate meta-analysis. Across 9 datasets of patients with culture-confirmed diagnosis of ATB, 11 signatures had weighted mean AUROC > 0.8, and 2 signatures had weighted mean AUROC ≤ 0.6. All but 2 signatures had high NPV (>98% at 2% prevalence). Two gene signatures achieved the minimal WHO TPP for a non-sputum-based triage test. When including datasets with clinical diagnosis of ATB, there was minimal reduction in the weighted mean AUROC and specificity of all but 3 signatures compared to when using only culture-confirmed ATB data. Only 4 signatures had homogeneous DOR and lower FPR when datasets with clinical diagnosis of ATB were included; other signatures either had heterogeneous DOR or higher FPR or both. Finally, 7 of 16 gene signatures predicted progression from LTBI to ATB 6 months prior to sputum conversion with positive predictive value > 6% at 2% prevalence. Our analyses may have under- or overestimated the performance of certain ATB diagnostic signatures because our implementation may be different from the published models for those signatures. We re-implemented published models because the exact models were not publicly available. Conclusions We found that host-response-based diagnostics could accurately iden...

Concise whole blood transcriptional signatures for incipient tuberculosis: A systematic review and patient-level pooled meta-analysis

“…This allowed us to perform a comprehensive, head-to-head analysis of candidate signatures for incipient TB for the first time, ensuring that each head-to-head comparison was performed on paired data. This contrasts with a recent head-to-head systematic evaluation that included only two of the eight best-performing signatures in our analysis, and compared performance for incipient TB in only one dataset over a 0-6 month time period 40 . Finally, our meta-analytic methods ensured a standardised approach to RNAseq data.…”

“…(CD1C + BLK)); and (2) as an SVM using the four constituent gene pairs, as previously described 40 . Since the former approach achieved marginally better 18…”

A generalizable 29-mRNA neural-network classifier for acute bacterial and viral infections

“…Historically, applying machine learning to diagnose acute infections using transcriptomic data has been confounded by technical and clinical heterogeneity in attempts to translate to realworld patient populations. For example, regression and decision tree classifiers trained using data collected on one type of microarray and tested in another type perform poorly, arguably at least in part due to inadequate cross-platform normalization 13,17 . Even models tested in data from the same technical platform can be prone to overfitting due to the lack of adequate representation of clinical heterogeneity in the training data 18 .…”

Reanalysis And Integration Of Public Microarray Datasets Reveals Novel Host Genes Modulated In Leprosy

“…The integration provided a comprehensive view of the differential expression profile of several genes involved in processes ranging from the innate and adaptive immune system to structural components and metabolic pathways. In addition, future datasets, whether from microarrays or RNA sequencing can also add information regarding the revealed genes and pathways and could even be aggregated to create gene expression signatures representative of leprosy [31].…”