A machine learning pipeline for discriminant pathways identification

Motivation: Identifying the molecular pathways more prone to disruption during a pathological process is a key task in network medicine and, more in general, in systems biology. Results: In this work we propose a pipeline that couples a machine learning solution for molecular profiling with a recent network comparison method. The pipeline can identify changes occurring between specific sub-modules of networks built in a case-control biomarker study, discriminating key groups of genes whose interactions are modified by an underlying condition. The proposal is independent from the classification algorithm used. Three applications on genomewide data are presented regarding children susceptibility to air pollution and two neurodegenerative diseases: Parkinson’s and Alzheimer’s. Availability: Details about the software used for the experiments discussed in this paper are provided in the Appendix.

💡 Research Summary

The paper introduces a modular machine‑learning pipeline designed to uncover molecular pathways that are differentially perturbed between case and control groups in high‑dimensional transcriptomic studies. The workflow consists of four sequential stages. First, a predictive model—either a classifier such as Spectral Regression Discriminant Analysis (SRDA) or a regression method with L1‑L2 regularization—is trained on the expression matrix using a rigorous Data Analysis Protocol (DAP) that includes nested cross‑validation. Feature ranking yields a gene signature; the size of the signature (k) is chosen to balance classification accuracy against signature stability. Second, the signature is fed into a pathway enrichment step (e.g., GSEA or GSA) to retrieve full biological pathways that contain the selected genes, thereby preserving functional context beyond the limited set of discriminative probes. Third, for each enriched pathway a separate co‑expression network is inferred for the case and control cohorts using algorithms such as Weighted Gene Co‑Expression Network Analysis (WGCNA) or ARACNE. To mitigate the under‑determination inherent in network reconstruction from few samples, only pathways with 4–1000 nodes are considered, and correlation thresholds are tuned to achieve a compromise between average Ipsen‑Mikhailov distance and network density. Finally, the two networks per pathway are compared using the Ipsen‑Mikhailov (IM) spectral distance, which measures global topological divergence by comparing Laplacian eigenvalue spectra. The IM distance is normalized, and pathways are ranked by this metric; an auxiliary score Δd (difference in weighted degree for each gene) is also reported to highlight individual gene‑level interaction changes. The authors demonstrate the pipeline on three publicly available GEO datasets: (1) a childhood air‑pollution cohort (GSE7543) where a 50‑probe signature achieved 76 % classification accuracy and highlighted apoptosis, skeletal development, and nervous system development pathways as most disrupted; (2) early and late stage Parkinson’s disease (PD) cohorts (GSE6613, GSE20295) where signatures of 70 and 90 genes yielded 62 % and 80 % accuracy respectively, and pathways related to oxidative stress, cytoskeleton organization, and synaptic transmission showed high IM distances; (3) early and late stage Alzheimer’s disease (AD) cohorts (GSE9770, GSE5281) where similar analyses identified alterations in synaptic plasticity, energy metabolism, and immune response pathways. In each case, the identified pathways align with known disease mechanisms, supporting the biological validity of the approach. The paper emphasizes the independence of the pipeline from any specific algorithmic component, allowing researchers to substitute classifiers, feature selectors, enrichment tools, network inference methods, or distance measures as needed. Limitations discussed include sensitivity to correlation‑based threshold selection during network construction and the interpretability challenges of spectral distances, which may require complementary visualization or statistical testing. Overall, the study provides a flexible, reproducible framework for integrating machine learning‑driven feature selection with global network comparison, offering a powerful means to pinpoint disease‑relevant pathway disruptions across diverse omics datasets.