Deep Learning Scans Gene Data to Reveal New Asthma Biomarkers

Asthma affects hundreds of millions of people worldwide, yet its underlying genetics remain stubbornly elusive. A new study published in Neural Computing and Applications by Md. Asim Iqbal, M. Sadanandam, and A. Kumaraswamy of Kakatiya University in Telangana, India, tackles this problem with an ambitious computational pipeline that combines deep learning, a nature-inspired optimization algorithm, and graph neural networks to mine public gene expression data for clues. The work, published on 3 October 2026, demonstrates how modern machine learning architectures can be chained together to sift through thousands of genes and surface a small handful of candidates that may drive the disease — and, crucially, may serve as targets for new diagnostics and therapies.

The starting point for the study is a familiar frustration in respiratory medicine. Asthma is a heterogeneous condition characterized by airway inflammation and obstruction driven by edema, bronchospasm, and excessive mucus production, but the genetic factors responsible for its pathogenesis remain largely unknown despite decades of progress. Reliable biomarkers are essential for diagnosis, for understanding disease mechanisms, and for predicting how individual patients will respond to treatment. Existing biomarkers capture only part of this picture, and the authors argue that new ones are needed to refine treatment strategies and guide therapeutic decisions over the long term.

To find them, the team turned to the Gene Expression Omnibus (GEO), the public repository maintained by the US National Center for Biotechnology Information. Three datasets — GSE64913, GSE67472, and GSE137268 — supplied the raw material. These datasets contain gene expression profiles sampled from the airway epithelium and from induced sputum of people diagnosed with asthma, providing complementary views of the tissue environments where the disease manifests. By analyzing samples drawn from these two sources, the researchers could look for genes whose activity correlates strongly with asthma status across different sampling contexts, a strategy designed to filter out artifacts tied to any single tissue type or collection protocol.

The analytical pipeline itself unfolds in three stages, each addressing a distinct computational challenge. The first stage uses a pretrained convolutional neural network (CNN) for feature extraction. CNNs, best known for their success in image recognition, apply layers of learned filters that abstract increasingly abstract representations from raw input. Here, the network converts high-dimensional gene expression profiles into new deep features — compact numerical descriptions that capture patterns of co-regulation and differential expression that simpler statistical tests might miss. Transfer learning from a pretrained model means the network arrives already equipped with general-purpose pattern detectors, which the authors then adapt to the specific structure of transcriptomic data.

Deep feature extraction, however, creates its own problem: dimensionality. Gene expression datasets typically measure tens of thousands of genes across comparatively few samples, and adding CNN-derived features can make the search space even more unwieldy. The second stage of the pipeline confronts this with a modified black winged kite (MBWK) algorithm, an enhanced version of a metaheuristic optimization method inspired by the hunting behavior of the black-winged kite, a bird of prey that hovers and strikes with precision. Metaheuristics of this kind explore large solution spaces by iteratively adjusting a population of candidate solutions, balancing exploration of new regions against exploitation of promising ones. In this application, the algorithm selects the optimal subset of features, discarding redundant or noisy dimensions while preserving the signals most informative for distinguishing asthmatic from healthy expression profiles.

The third and most novel stage applies an adaptive dynamic graph neural network (AD-GNN) to the optimized features. Graph neural networks represent data as networks of nodes and edges, and they are naturally suited to genomics because genes do not act in isolation — they operate within regulatory networks, signaling pathways, and protein interaction maps. An AD-GNN can adapt its structure dynamically as it learns, propagating information across the gene interaction graph so that the activity of any single gene is interpreted in light of its neighbors. This relational reasoning is what allows the model to identify hub genes: highly connected genes whose expression patterns sit at the center of asthma-associated network activity, rather than merely differing on average between patient groups.

The output of this three-stage process is a short list of candidate genes: CCDC167, SERPINB2, SEC14L1, and POSTN. Each has a plausible biological story. POSTN, which encodes the matricellular protein periostin, has long been associated with type 2 airway inflammation and epithelial remodeling in asthma, lending the pipeline immediate face validity. SERPINB2 belongs to the serpin family of protease inhibitors and is known to respond to inflammatory signaling in epithelial tissue. CCDC167, a coiled-coil domain-containing protein of less certain function, has recently attracted attention as a potential biomarker for airway inflammation, and its independent emergence here reinforces that hypothesis. SEC14L1, involved in lipid handling, rounds out the set. The authors report that the predicted genes demonstrate functional relevance across a variety of asthma-related tissues, supporting their candidacy as genuine disease markers rather than statistical coincidences.

What makes the study notable is less any single component than the way the components interlock. CNN feature extraction addresses the problem that raw expression values are noisy and high-dimensional; the MBWK optimizer addresses the problem that even informative features come bundled with redundancy; and the AD-GNN addresses the problem that genes must be understood in context. The authors report that the proposed model successfully identified both asthma biomarkers and treatment options as demonstrated by their experiments, positioning the framework as a template for biomarker discovery that could in principle be retrained on other diseases by swapping in different datasets. Because all three GEO datasets are publicly available, the entire approach is reproducible by any laboratory with the computational resources to run it.

The broader context is a rapidly growing effort to bring artificial intelligence into precision medicine for respiratory disease. Recent years have seen machine learning applied to cough sound analysis for automated asthma detection, to routine blood biomarkers for asthma prediction, to the prediction of allergen immunotherapy response, and to the identification of ferroptosis-related and cellular senescence-related genes in severe asthma. Omics approaches — genomics, transcriptomics, proteomics, metabolomics — generate data at a scale that outstrips classical statistical methods, and deep learning offers a way to extract signal from that abundance. The Kakatiya University study sits squarely within this movement, and its use of graph-based learning reflects a growing recognition that biological data are inherently relational.

Cautions remain, as they always do at this stage of translational research. Candidate biomarkers identified computationally must be validated experimentally — at the protein level, in independent patient cohorts, and ultimately in prospective clinical studies — before they can inform diagnosis or drug development. The authors themselves used publicly available, de-identified data with no direct human participation, which means the pipeline’s outputs are hypotheses to be tested rather than clinical conclusions. Still, the study offers a concrete demonstration that a carefully engineered chain of deep learning tools can compress a genome-wide search into a handful of testable targets. For a disease as heterogeneous and burdensome as asthma, tools that accelerate that search from years to months could meaningfully change how quickly new diagnostics and therapies reach patients, and the four genes highlighted here now join the growing list of leads awaiting experimental confirmation.

Subject of Research: Deep learning-based gene expression analysis for identifying diagnostic biomarkers and therapeutic targets in asthma

Article Title: Deep learning-based analysis of gene expression for identifying diagnostic biomarkers and therapeutic targets in asthma

Article References: Iqbal, M. A., Sadanandam, M., & Kumaraswamy, A. (2026). Deep learning-based analysis of gene expression for identifying diagnostic biomarkers and therapeutic targets in asthma. Neural Computing and Applications, 38(19), Article 768. https://doi.org/10.1007/s00521-026-12519-9

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12519-9

Keywords: asthma, deep learning, gene expression, biomarkers, graph neural networks, convolutional neural networks, GEO database, therapeutic targets, precision medicine, bioinformatics, airway epithelium, machine learning