A robust mRNA signature obtained via recursive ensemble feature selection predicts the responsiveness of omalizumab in moderate‐to‐severe asthma

Sarah Kidwai, Pietro Barbiero, Irma Meijerman, Alberto Tonda, Paula Perez‐Pardo, Píetro Lió, Anke H. van der Maitland‐Zee, Daniel L. Oberski, Aletta D. Kraneveld, Alejandro Lopez‐RinconClinical 2023 · DOI

TL;DR

This paper introduces a novel machine learning approach combining Recursive Ensemble Feature Selection (REFS) and Logic Explainable Networks (LEN) to identify a robust mRNA signature that predicts how well patients with moderate-to-severe asthma will respond to omalizumab treatment. The identified 4-gene signature achieved a high cross-validated accuracy of 0.975, offering a promising tool for personalized treatment decisions in asthma care.

What Problem Does This Paper Address?

Moderate-to-severe asthma affects millions worldwide and is often poorly controlled by standard inhaled corticosteroids and long-acting beta-agonists. Omalizumab, an anti-IgE biological therapy, is a valuable treatment option for these patients, but a significant portion do not benefit from it. Currently, reliable biomarkers to predict omalizumab responsiveness are limited, leading to potential trial-and-error treatment approaches, increased healthcare costs, and delayed symptom relief. The paper aims to address this gap by identifying a precise and predictive set of biomarkers from mRNA expression data to guide personalized omalizumab therapy.

Methods

The study employed two advanced machine learning (ML) techniques:

  • Recursive Ensemble Feature Selection (REFS): This algorithm was used to reduce a large set of gene expression data (28,402 genes) down to a compact, highly predictive signature. REFS combines the results of eight different classifier algorithms, ranking genes based on their importance across these diverse models. This ensemble approach aims to provide a more robust and less biased feature selection compared to single-algorithm methods.
  • Logic Explainable Networks (LEN): Complementing REFS, LEN was used to generate interpretable rules from the gene expression data. Unlike REFS, which focuses on feature selection, LEN performs both classification and feature selection simultaneously, providing logical explanations for its predictions. This interpretability is crucial for clinical adoption, allowing healthcare professionals to understand why a prediction is made.

The study utilized a two-stage process: first, REFS identified a 5-gene signature, and then LEN was applied to the same dataset to find overlapping predictive genes and generate explanatory rules.

Datasets

The study utilized a publicly available dataset from the Gene Expression Omnibus (GEO) with the accession code GSE134544. This dataset comprises:

  • 239 whole blood mRNA samples from 40 moderate-to-severe adult asthmatic patients.
  • Of these, 30 were classified as responders to omalizumab and 10 as non-responders.
  • 17 non-asthmatic healthy controls were also included for comparison, particularly for the LEN analysis.

The mRNA expression data was collected one week before the start of omalizumab treatment. This dataset was previously used in a study by Upchurch et al. (2020), providing a basis for comparison.

Key Insights and Ideas

  • The Power of Ensemble and Explainable ML: The combination of REFS and LEN proved highly effective. REFS identified a highly accurate, compact gene signature, while LEN provided crucial interpretability, making the findings more clinically relevant.
  • A Compact Signature is Sufficient: The study demonstrated that a small set of genes (4-5) can be highly predictive of omalizumab response, challenging the notion that complex, large gene sets are always necessary.
  • Biological Context is Key: Linking the identified genes to known biological pathways (mucosal immunity, cellular energy, airway remodeling) provided valuable insights into the potential mechanisms underlying omalizumab response and non-response.
  • Potential for Misdiagnosis: The identification of LOC100131780, linked to primary ciliary dyskinesia (PCD), raises the intriguing possibility that some patients diagnosed with severe asthma might actually have PCD, a condition that could mimic asthma symptoms and affect treatment response.

Key Results

  • High Predictive Accuracy: The REFS algorithm reduced the initial 28,402 genes to a 5-gene signature with a cross-validated accuracy of 0.975.
  • Identified Genes: The REFS algorithm identified five genes: CCDC113, SLC26A8, PPP1R3D, CLEC4C, and LOC100131780.
  • Overlapping Signature: The LEN algorithm identified four of these genes (CCDC113, SLC26A8, PPP1R3D, and LOC100131780) as also being predictive. These four genes were consistently upregulated in non-responders and downregulated in responders.
  • Biological Associations: The identified genes are associated with mucosal immunity, cellular energy, airway hypersensitivity, and remodeling. For instance, CCDC113 is linked to ciliary function and mucociliary clearance, while CLEC4C is involved in anti-viral immunity.
  • LEN Rules: The LEN algorithm generated three rules that predict omalizumab responsiveness, based on the expression levels of these genes relative to healthy controls.

Limitations and Future Work

  • Validation Cohort: The authors acknowledge the lack of an independent cohort dataset to validate their findings, which is a crucial step for clinical translation.
  • Uncharacterized Genes: The biological functions of some identified genes (e.g., LOC100131780) are not fully characterized, limiting the depth of interpretation.
  • Mixed Phenotypes: The study did not investigate the potential impact of mixed asthma phenotypes (e.g., Type 2 and non-Type 2 inflammation) on omalizumab response.
  • Future Directions:
    • Prospective Validation: Future studies should validate the computational approach in larger, prospective cohorts.
  • Clinical Utility: Testing the clinical utility of a simple blood test for these mRNA biomarkers is recommended.
  • Comparative Studies: Investigating which biological medication (omalizumab vs. other biologics) is best suited for individual patients based on these biomarkers.
  • Further Research: Exploring the interplay between asthma severity, gender, and reproductive health (given the SLC26A8 finding) and investigating conditions mimicking asthma symptoms like PCD.
  • In Vitro Experiments: Conducting in vitro experiments with whole blood samples could provide broader insights into gene function and response mechanisms.

Summary updated

Abstract

Both computational methods show 4 identical genes as predictors of omalizumab response in moderate-to-severe asthma patients. The obtained high accuracy indicates that our approach has potential in clinical settings. Future studies in relevant cohort data should validate our computational approach.
BibTeX
@inproceedings{kidwai2023robust,
  author = {Sarah Kidwai and Pietro Barbiero and Irma Meijerman and Alberto Tonda and Paula Perez‐Pardo and Píetro Lió and Anke H. van der Maitland‐Zee and Daniel L. Oberski and Aletta D. Kraneveld and Alejandro Lopez‐Rincon},
  title = {{A robust mRNA signature obtained via recursive ensemble feature selection predicts the responsiveness of omalizumab in moderate‐to‐severe asthma}},
  booktitle = {Clinical},
  year = {2023},
  doi = {10.1002/clt2.12306},
}
Figure