Latent-space embedding of expression data identifies gene signatures from sputum samples of asthmatic patients

Shaoke Lou, Tianxiao Li, Daniel Spakowicz, Xiting Yan, Geoffrey Chupp, Mark Gerstein

Abstract

Background: The pathogenesis of asthma is a complex process involving multiple genes and pathways. Identifying biomarkers from asthma datasets, especially those that include heterogeneous subpopulations, is challenging. Potentially, autoencoders provide ideal frameworks for such tasks as they can embed high-dimensional gene expression data into a low-dimensional latent space in an unsupervised fashion, enabling us to extract distinguishing features from expression data. Results: Here, we developed a framework combining a denoising autoencoder and a supervised learning classifier to identify gene signatures related to asthma severity. Using the trained autoencoder with 50 hidden units, we found that hierarchical clustering on the low-dimensional embedding corresponds well with previously defined and clinically relevant clusters of patients. Moreover, each hidden unit has contributions from each of the genes, and pathway analysis of these contributions shows that the hidden units are significantly enriched in known asthma-related pathways. We then used genes that contribute most to the hidden units to develop a secondary random-forest classifier for directly predicting asthma severity. The feature importance metric from this classifier identified a signature based on 50 key genes, which can predict severity with an AUROC of 0.81 and thus have potential as diagnostic biomarkers. Furthermore, we could use these key genes for successfully estimating, via support-vector-machine regression, the FEV1/FVC ratios across patients, achieving pre- and post-treatment correlations of 0.56 and 0.65, respectively (between predicted and observed values). Conclusion: We found that the denoising autoencoder framework can extract meaningful patterns corresponding to functional gene groups and patient clusters from the gene expression of asthma patients. Specifically, from the top-weighted gene set in the hidden units, we identified 50 genes that are predictive of asthma severity and other relevant clinical traits. Furthermore, we found that these genes play central roles in biological networks and pathways. Thus, we believe that they could provide a potential source for biomarkers for asthma severity.

Introduction

Asthma is a complex, heterogeneous condition characterized by various biological mechanisms, or endotypes. Identifying reliable biomarkers for disease severity is complicated by the high dimensionality of transcriptomic data—often exceeding 20,000 genes—and the inherent noise in clinical sputum samples. Traditional methods, such as Differential Expression (DE) analysis, often focus on genes with the largest magnitude of change, which may not necessarily be the primary drivers of the disease or the most effective indicators of severity.

To address these challenges, this study introduces dAsthma, a machine-learning framework designed to navigate clinical genetic noise. By utilizing denoising autoencoders, the researchers aimed to extract stable, non-linear features that better categorize patients and predict disease progression.

Methods

The dAsthma framework employs a two-stage approach to analyze gene expression data from induced sputum samples (originally sourced from Yan et al., 2015), comprising profiles for 22,148 genes.

Unsupervised Feature Extraction

The core of the framework is a one-layer Denoising Autoencoder (DAE). Unlike standard autoencoders, a DAE is trained to reconstruct original data from inputs intentionally corrupted with random noise. This process forces the model to learn robust, essential patterns, compressing the high-dimensional input into a 50-unit "latent space."

Supervised Learning and Network Analysis

Following feature extraction, the weights from the DAE were used to identify influential genes. These genes informed a Random Forest classifier to predict asthma severity (mild vs. severe) and SVM/LASSO models to estimate clinical lung function, specifically the FEV1/FVC ratio. To ensure biological relevance, the identified genes were mapped onto protein-protein interaction networks (STRING database) to assess their "centrality" or connectivity within regulatory systems.

Results and Discussion

Latent Space and Clinical Subtypes

The 50 hidden units generated by the DAE were found to correspond closely with known clinical Transcriptomic Endotypes of Asthma (TEAs). This indicates that the unsupervised model successfully captured the underlying biological structure of the disease without prior labeling.

Predictive Performance

The dAsthma framework significantly outperformed traditional methods. The identified 50-gene signature achieved an average AUROC of 0.81 in predicting asthma severity, whereas the gene set identified by standard DE analysis achieved an AUROC of only 0.54, barely exceeding random chance. Furthermore, the model successfully estimated FEV1/FVC ratios, with correlations of 0.56 (pre-treatment) and 0.65 (post-treatment) between predicted and observed values.

Biological Hub Genes

Interestingly, the 50 genes identified by dAsthma showed minimal overlap with those found through standard DE. However, network analysis revealed that the dAsthma genes possessed much higher centrality. These "hub" genes act as critical nodes in biological networks, influencing multiple pathways even if their individual expression levels do not show extreme fluctuations. Pathway enrichment analysis confirmed that these hidden units were significantly associated with asthma-related immune responses and antigen processing.

Conclusion

The denoising autoencoder framework provides a powerful tool for extracting meaningful biological patterns from noisy clinical data. By identifying a 50-gene signature that serves as a central hub in biological networks, this research offers a promising source for biomarkers of asthma severity. While the results are compelling, future work should focus on validating these findings in larger cohorts and conducting wet-lab experiments to confirm the functional roles of these identified genes in disease progression.

Summary updated

Abstract

Abstract Background: The pathogenesis of asthma is a complex process involving multiple genes and pathways. Identifying biomarkers from asthma datasets, especially those that include heterogeneous subpopulations, is challenging. Potentially, autoencoders provide ideal frameworks for such tasks as they can embed high-dimensional gene expression data into a low-dimensional latent space in an unsupervised fashion, enabling us to extract distinguishing features from expression data. Results: Here, we developed a framework combining a denoising autoencoder and a supervised learning classifier to identify gene signatures related to asthma severity. Using the trained autoencoder with 50 hidden units, we found that hierarchical clustering on the low-dimensional embedding corresponds well with previously defined and clinically relevant clusters of patients. Moreover, each hidden unit has contributions from each of the genes, and pathway analysis of these contributions shows that the hidden units are significantly enriched in known asthma-related pathways. We then used genes that contribute most to the hidden units to develop a secondary random-forest classifier for directly predicting asthma severity. The feature importance metric from this classifier identified a signature based on 50 key genes, which can predict severity with an AUROC of 0.81 and thus have potential as diagnostic biomarkers. Furthermore, we could use these key genes for successfully estimating, via support-vector-machine regression, the FEV1/FVC ratios across patients, achieving pre- and post-treatment correlations of 0.56 and 0.65, respectively (between predicted and observed values). Conclusion: We found that the denoising autoencoder framework can extract meaningful patterns corresponding to functional gene groups and patient clusters from the gene expression of asthma patients. Specifically, from the top-weighted gene set in the hidden units, we identified 50 genes that are predictive of asthma severity and other relevant clinical traits. Furthermore, we found that these genes play central roles in biological networks and pathways. Thus, we believe that they could provide a potential source for biomarkers for asthma severity.
BibTeX
@article{lou2020latent,
  author = {Shaoke Lou and Tianxiao Li and Daniel Spakowicz and Xiting Yan and Geoffrey Chupp and Mark Gerstein},
  title = {{Latent-space embedding of expression data identifies gene signatures from sputum samples of asthmatic patients}},
  journal = {BMC Bioinformatics},
  year = {2020},
  doi = {10.1186/s12859-020-03785-y},
}
Figure