Network-guided identification of candidate biomarkers for COPD-associated comorbidities using deep learning and real-world transcriptomic data
Chronic obstructive pulmonary disease (COPD) frequently coexists with infections, metabolic disease, cardiovascular disease, inflammatory lung injury and lung cancer. Molecular markers derived by comparing a comorbidity with healthy controls need not identify the transition from an established COPD state to that comorbidity. We therefore developed a two-stage computational strategy that asks a narrower question: which proteins can distinguish COPD samples from samples annotated with a specific COPD-associated comorbidity? First, a Human Protein Network (HPN; 17,302 proteins and 492,195 relationships) was encoded by 20 graph-derived features. A convolutional-neural-network ensemble, termed the functional-distance model (FDM), was trained on known drug-target-indication relationships. Stored training-set predictions yielded an apparent balanced accuracy of 0.811 and Matthews correlation coefficient (MCC) of 0.627; these values describe resubstitution performance and are not an independent estimate of generalization. The COPD-centred network search evaluated 2,645 nearby proteins and retained 224 candidates using a COPD FDM score above 0.50. Candidate expression was then tested in independently assembled public patient-derived transcriptomic data using simple one-feature classifiers and cross-validation. The resulting exploratory panel contained GTF2I for bacterial infection, UBXN7 for type 2 diabetes, LIFR for hypertension, SRGN for lung neoplasm and FGF4 for pneumonitis. Across the saved cross-validation predictions, balanced accuracy ranged from 0.838 to 0.970 and MCC from 0.726 to 0.941. SRGN has direct experimental support in non-small-cell lung cancer, whereas the remaining candidates have varying indirect mechanistic support and should be regarded as hypotheses rather than established disease biomarkers. This work demonstrates an orthogonal discovery framework in which network-based functional prioritization and patient-derived molecular discrimination constrain one another. Its main contribution is methodological: it provides a reusable strategy for generating comorbidity-specific biomarker hypotheses from an index-disease state. Prospective longitudinal cohorts are required before claims of progression prediction or clinical utility can be made.