A crowdsourced analysis to identify ab initio molecular signatures predictive of susceptibility to viral infection

Slim Fourati, Aarthi Talla, Mehrad Mahmoudian, Joshua Burkhart, Riku Klén, Ricardo Henao, Zafer Aydın, Ka Yee Yeung, Mehmet Eren Ahsen, Reem Almugbel, Samad Jahandideh, Xiao Liang, Torbjörn E. M. Nordling, M. Shiga, Ana Stanescu, Robert Vogel, Gaurav Pandey, Christopher Chiu, Micah T. McClain, Chris Woods, Geoffrey S. Ginsburg, Laura L. Elo, Ephraim L. Tsalik, Lara M. Mangravite, Solveig K. SiebertsbioRxiv 2018 · DOI

TL;DR

This paper leverages a crowdsourced analysis of gene expression data from human viral exposure studies to identify early molecular signatures that predict susceptibility to respiratory viruses. The findings demonstrate that even pre-exposure gene expression profiles can predict symptom development, highlighting the potential for early diagnostics and interventions.

What Problem Does This Paper Address?

Respiratory viruses are a major cause of illness, but individual responses to infection vary widely, from asymptomatic to severe disease. Understanding the factors that determine this variability is crucial for developing effective prevention and treatment strategies. While many studies focus on late-stage predictors of disease, this research aims to identify ab initio (from the beginning) molecular signatures – specifically, gene expression patterns in peripheral blood – that can predict an individual's susceptibility or resilience to viral infection before or very early after exposure. This could pave the way for proactive health management rather than reactive treatment.

Methods

The core of this study is the Respiratory Viral DREAM Challenge, a crowdsourced data analysis competition. Researchers from various teams were provided with gene expression data from seven human viral exposure experiments involving H1N1, H3N2, Rhinovirus, and RSV. The data included peripheral blood gene expression profiles collected from healthy volunteers at two key timepoints: prior to viral exposure (T₀) and up to 24 hours post-exposure (T₂₄). Teams were tasked with building predictive models for three outcomes: viral shedding, presence of symptoms, and symptom severity. The challenge utilized state-of-the-art analytical algorithms, and teams employed diverse approaches for data preprocessing, feature selection, and predictive modeling. The study also performed pathway enrichment analysis on the identified predictors to understand the underlying biological mechanisms.

Datasets

The study utilized data from seven human viral exposure experiments, comprising a combined total of 148 healthy volunteers. These experiments involved exposure to four different respiratory viruses: RSV, H1N1 (two studies), H3N2 (two studies), and Rhinovirus (two studies). The data included peripheral blood gene expression profiles (Affymetrix Human U133A 2.0 GeneChips), symptom scores (modified Jackson score), and viral shedding data. An additional, previously unavailable RSV human exposure experiment with 21 subjects was used as an independent test dataset. The collated seven datasets served as the training data, and a subset of 23 subjects' phenotypic data was withheld as a leaderboard test set.

Key Insights and Ideas

  • Pre-symptomatic prediction is possible: The most striking finding is that gene expression profiles before viral exposure can predict whether an individual will develop symptoms. This suggests that inherent biological characteristics play a significant role in determining susceptibility.
  • No single "magic bullet" predictor: While models could predict outcomes, there was little overlap in the specific gene features identified by different teams. This suggests that multiple combinations of genes can contribute to prediction, and the underlying biological processes are complex.
  • Pathway-level insights are crucial: Despite the lack of overlap in individual genes, the aggregated gene sets from successful models were enriched for common biological pathways. This highlights the importance of looking at biological processes rather than just individual genes for a comprehensive understanding.
  • Heme Metabolism as a key pathway: The Heme Metabolism pathway emerged as a significantly enriched pathway associated with a higher risk of developing symptoms. This pathway is involved in immune responses and iron homeostasis, suggesting a link between these processes and viral susceptibility.
  • Data preprocessing is critical: The study found that adequate data preprocessing was the analysis task most significantly associated with predictive ability, underscoring its importance in building accurate models.

Key Results

  • Models built on gene expression data from T₀ and T₂₄ timepoints were able to predict symptom presence and severity better than expected by random chance.
  • For symptom presence (SC2), the best models achieved AUROC scores of approximately 0.863 at both T₀ and T₂₄.
  • For symptom severity (SC3), the best models achieved correlation coefficients of approximately 0.490 (T₀) and 0.495 (T₂₄).
  • Pathway enrichment analysis identified HEME METABOLISM as the most significantly associated pathway with developing symptoms. Other pathways associated with susceptibility included inflammatory responses (KRAS SIGNALING, APOPTOSIS), while homeostatic pathways (OXIDATIVE PHOSPHORYLATION, MYC TARGETS) were associated with resilience.
  • The study found that approximately 25% of subjects were "inherently easy" to predict, while another 25% were "inherently difficult," suggesting a spectrum of individual predictability.

Limitations and Future Work

  • Clinical significance: While statistically significant, the predictive power of the best individual models fell short of immediate practical clinical significance. Further efforts are needed to improve accuracy.
  • Ensemble performance: Ensembles constructed from the challenge submissions did not outperform the best individual submissions, possibly due to the relatively small training set size, which may have led to overfitting.
  • Homogeneity of submissions: The relative homogeneity of submissions and the observation that the same subjects were misclassified by multiple teams suggest a potential plateau in predictive ability using gene expression alone.
  • Future Work: The authors suggest exploring integrative analyses that combine gene expression data with post-transcriptional data (metabolomics, proteomics) to improve accuracy. They also propose further investigation into the role of pathways like Heme Metabolism in viral susceptibility and the potential for developing diagnostic tests for baseline susceptibility. The Respiratory Viral DREAM Challenge itself is highlighted as a valuable framework for future research in this area.

Summary updated

Abstract

Abstract Respiratory viruses are highly infectious; however, the variation of individuals’ physiologic responses to viral exposure is poorly understood. Most studies examining molecular predictors of response focus on late stage predictors, typically near the time of peak symptoms. To determine whether pre- or early post-exposure factors could predict response, we conducted a community-based analysis to identify predictors of resilience or susceptibility to several respiratory viruses (H1N1, H3N2, Rhinovirus, and RSV) using peripheral blood gene expression profiles collected from healthy subjects prior to viral exposure, as well as up to 24 hours following exposure. This analysis revealed that it is possible to construct models predictive of symptoms using profiles even prior to viral exposure. Analysis of predictive gene features revealed little overlap among models; however, in aggregate, these genes were enriched for common pathways. Heme Metabolism, the most significantly enriched pathway, was associated with higher risk of developing symptoms following viral exposure.
BibTeX
@misc{fourati2018crowdsourced,
  author = {Slim Fourati and Aarthi Talla and Mehrad Mahmoudian and Joshua Burkhart and Riku Klén and Ricardo Henao and Zafer Aydın and Ka Yee Yeung and Mehmet Eren Ahsen and Reem Almugbel and Samad Jahandideh and Xiao Liang and Torbjörn E. M. Nordling and M. Shiga and Ana Stanescu and Robert Vogel and Gaurav Pandey and Christopher Chiu and Micah T. McClain and Chris Woods and Geoffrey S. Ginsburg and Laura L. Elo and Ephraim L. Tsalik and Lara M. Mangravite and Solveig K. Sieberts},
  title = {{A crowdsourced analysis to identify <i>ab initio</i> molecular signatures predictive of susceptibility to viral infection}},
  year = {2018},
  doi = {10.1101/311696},
  howpublished = {bioRxiv},
}
Figure