Early prognosis of respiratory virus shedding in humans

Abstract

This paper addresses the development of predictive models for distinguishing pre-symptomatic infections from uninfected individuals. Our machine learning experiments are conducted on publicly available challenge studies that collected whole-blood transcriptomics data from individuals infected with HRV, RSV, H1N1, and H3N2. We address the problem of identifying discriminatory biomarkers between controls and eventual shedders in the first 32 h post-infection. Our exploratory analysis shows that the most discriminatory biomarkers exhibit a strong dependence on time over the course of the human response to infection. We visualize the feature sets to provide evidence of the rapid evolution of the gene expression profiles. To quantify this observation, we partition the data in the first 32 h into four equal time windows of 8 h each and identify all discriminatory biomarkers using sparsity-promoting classifiers and Iterated Feature Removal. We then perform a comparative machine learning classification analysis using linear support vector machines, artificial neural networks and Centroid-Encoder. We present a range of experiments on different groupings of the diseases to demonstrate the robustness of the resulting models.

Pathway Expression Analysis

“…Our preliminary experiments addressing time-evolving human clinical data provide some insight into the pathway activity for humans infected with respiratory viruses. Our 4 to 2, 4 to 3 and 6 to 1 classification experiments in the results are designed to mimic the experiments that were done by Aminian et al 3 . For a direct comparison of best CPE results between this paper to the gene expression results using SVM from Aminian et al see Table 3.…”

“…In this paper we find that pathway expression methods generally outperform the gene expression methods on the GSE73072 data set with analogous feature selection techniques. For example, we use pathway expression methods to improve upon most of the classification results found in Aminian et al at the 25 to 32 hour time bin 3 . Pathways which are selected using pathway expression methods tend to be distinct from the pathways selected by known pathway ranking methodologies, specifically CePa and ORA.…”

“…In this paper we will run our experiments on a human microarray data set with multiple respiratory viruses and studies, the GSE73072 data set 2 . Previous work on these respiratory virus data used machine learning (ML) models, e.g., neural networks, support vector machines (SVM), centroid encoders (CE), and spectral gene network analysis, to identify discriminatory biomarkers within early shedders challenged with influenza and subsequently classify those subjects [3][4][5][6] .…”

“…O'Hara et al found that gene expression data from some pathways like B Cell Maturation and Activation + Cell Adhesion Molecules were used to produce 100% classification accuracy for certain experiments with this data set but analyses with pathway expression has never been done 6 . This work builds on the early detection analysis done by Aminian et al 3 by utilizing "pathways expression" in place of gene expression. While eigengenes, from weighted gene correlation network analysis (WGCNA 21 ), is a known method which produces what we would call pathway expression data, to the best of our knowledge, there is no application of eigengenes in downstream ML experiments and feature extraction.…”

“…By using pathway expression as an alternative to gene expression, we extract pathways as features using Sparse Support Vector Machines (SSVM) and subsequently rank pathways by their weight in the SSVM model. As a metric for success, we compare our classifications results using pathway expression data against the classification results of Aminian et al who use gene expression data 3 . We do this by extracting the top pathways for each pathway expression method on training data, described below, restricting the test pathway expression data to those top pathways, and finally classifying controls vs. shedders (from 4 evenly spaced time bins within 32 hours after infection) in the test pathway expression data via SVM.…”