The study team built a Human Protein Network (HPN) containing 17,302 proteins and 492,195 relationships, then distilled that graph into 20 numerical, graph-derived features to encode topology. A convolutional-neural-network ensemble — described by the authors as the functional-distance model (FDM) — was trained on known drug-target-indication relationships to learn patterns of functional proximity in the network. Investigators reported resubstitution metrics from the training set: an apparent balanced accuracy of 0.811 and a Matthews correlation coefficient (MCC) of 0.627. The authors explicitly stated those values describe resubstitution performance and are not independent estimates of generalization.
Using the trained FDM, the group conducted a COPD-centred network search. The search evaluated 2,645 proteins in the proximity of COPD-associated nodes and kept 224 candidates that exceeded an FDM score threshold greater than 0.50. The pipeline therefore shifted from a broad interactome to a focused set of network-prioritised proteins that the authors considered plausible mechanistic links to specific comorbid states.
After network prioritisation, the authors used independently assembled public patient-derived transcriptomic data to test candidate expression. They applied simple one-feature classifiers for each candidate protein and used cross-validation on saved predictions to assess discriminative signal between COPD samples and samples annotated with specific COPD-associated comorbidities. The manuscript describes this as a two-stage strategy: first, a network-based functional prioritisation; second, empirical discrimination in real-world molecular data. The source did not report detailed information on classifier hyperparameters, dataset sizes, or exact cross-validation folds in the abstract; those methodological specifics are not provided there.
The exploratory panel that emerged from the two-stage pipeline contained five primary nominations. Investigators reported the following gene-to-comorbidity associations from their analyses: GTF2I for bacterial infection, UBXN7 for type 2 diabetes, LIFR for hypertension, SRGN for lung neoplasm, and FGF4 for pneumonitis. These proteins were retained after both the network FDM score filter and subsequent transcriptomic discrimination. The authors framed these molecules as hypotheses for comorbidity-specific markers rather than established diagnostic or prognostic biomarkers.
For the saved cross-validation predictions on the transcriptomic classifiers, the authors reported balanced accuracy values ranging from 0.838 to 0.970 and MCC values ranging from 0.726 to 0.941 across the nominated candidates. The manuscript notes that the FDM training-set metrics (balanced accuracy 0.811; MCC 0.627) describe resubstitution performance and therefore cannot substitute for an independent generalization assessment. Investigators emphasised that prospective longitudinal cohorts are required before any claim about progression prediction or clinical utility can be made. The abstract did not provide independent test-set results, external validation cohorts, or calibration metrics; the source did not report those details in the abstract.
The authors reported that SRGN (serglycin) has direct experimental support in non-small-cell lung cancer, citing experimental literature as context for that nomination. For the remaining candidates — GTF2I, UBXN7, LIFR, and FGF4 — the study team noted varying degrees of indirect mechanistic support but characterised them as hypothesis-generating. Data indicated that these proteins emerged from an orthogonal approach in which network-based prioritisation and patient-derived molecular discrimination constrain one another, but the abstract makes clear that mechanistic validation beyond existing indirect literature is still required.
The authors declared the work's main contribution to be methodological: a reusable framework for generating comorbidity-specific biomarker hypotheses from an index disease state. They emphasised limitations that restrict immediate clinical translation. Specifically, the manuscript states that prospective longitudinal cohorts are necessary before any claims of progression prediction or clinical utility can be made. The abstract did not report detailed cohort demographics, timing of sample collection, or prospective validation plans; those elements are therefore not available from the source abstract. Readers should treat the nominated proteins as exploratory leads that require experimental follow-up and independent validation prior to clinical application.