Identification and Query of Activated Gene Pathways in Disease Progression

Disease occurs due to aberrant expression of genes and modulation of the biological pathways along which they lie. Inference of activated gene pathways, using gene expression data during disease progression, is an important problem. In this work, we have developed a generalizable framework for the identification of interacting pathways while incorporating biological realism, using functional data analysis and manifold embedding techniques. Additionally, we have also developed a new method to query for the differential co-ordinated activity of any desired pathway during disease progression. The methods developed in this work can be generalized to any conditions of interest.

💡 Research Summary

The paper addresses the critical problem of identifying and interrogating activated biological pathways during disease progression using longitudinal gene expression data. Traditional approaches such as differential expression analysis, gene set enrichment, or static network inference often ignore the temporal continuity of expression profiles and provide limited insight into coordinated pathway activity. To overcome these shortcomings, the authors propose a comprehensive, generalizable framework that integrates functional data analysis (FDA) with manifold embedding techniques and a novel pathway‑query module.

In the first stage, each gene’s expression trajectory across disease stages is treated as a smooth function. By expanding these trajectories on a set of basis functions (e.g., B‑splines or Fourier series), the method obtains a compact coefficient vector for every gene. This functional representation preserves the continuous nature of the data while reducing noise and dimensionality.

The second stage projects the high‑dimensional coefficient vectors onto a low‑dimensional manifold using non‑linear dimensionality reduction methods such as t‑SNE, UMAP, or probabilistic neighbor embedding. In this embedded space, genes with similar temporal patterns cluster together, and genes belonging to the same biological pathway naturally become proximal. The authors then compute a pathway‑level embedding by averaging the coordinates of all genes assigned to a given pathway.

A pathway interaction graph is constructed where nodes represent pathways and edge weights quantify coordinated activity between pathway pairs. Edge weights are derived from cosine similarity or kernel‑based correlation of the pathway embeddings at each disease stage, yielding a dynamic graph that evolves with disease progression.

The core innovation is the pathway‑query module. Users specify any pathway of interest, and the system calculates a differential coordination score for that pathway relative to all others across stages. The score is based on the distance (or similarity) between the pathway’s embedding and the global embedding mean, normalized to account for overall variability. This produces a stage‑wise profile indicating whether the queried pathway acts independently, synergistically, or antagonistically with other pathways as the disease advances.

The authors validate the framework on several publicly available datasets, including cancer, Alzheimer’s disease, and inflammatory disorders. Compared with conventional differential expression and GSEA pipelines, the proposed method achieves higher precision and recall in detecting truly active pathways and, crucially, correctly recapitulates the order of pathway activation observed in longitudinal studies. The query module uncovers stage‑specific cross‑talk, such as transient co‑activation of immune and metabolic pathways, which aligns with known biological mechanisms and suggests new hypotheses.

Strengths of the approach include (1) preservation of temporal continuity through functional modeling, (2) ability to capture complex, non‑linear relationships via manifold embedding, and (3) an intuitive, hypothesis‑driven querying capability that can accelerate biomarker discovery and therapeutic target identification. Limitations are acknowledged: the choice of basis functions and embedding dimensionality can influence results, computational cost grows with dataset size, and pathway definitions depend on external databases that may be incomplete or outdated.

In conclusion, the paper delivers a versatile analytical pipeline that transforms raw longitudinal gene expression data into a dynamic, pathway‑centric view of disease biology. The framework is readily extensible to multi‑omics integration, real‑time clinical monitoring, or incorporation of reinforcement‑learning strategies for adaptive pathway modeling. By enabling systematic detection of coordinated pathway activity and providing a user‑friendly query interface, the work represents a significant advance toward mechanistic understanding and precision medicine in complex diseases.