Inferring Disease and Gene Set Associations with Rank Coherence in Networks

A computational challenge to validate the candidate disease genes identified in a high-throughput genomic study is to elucidate the associations between the set of candidate genes and disease phenotypes. The conventional gene set enrichment analysis often fails to reveal associations between disease phenotypes and the gene sets with a short list of poorly annotated genes, because the existing annotations of disease causative genes are incomplete. We propose a network-based computational approach called rcNet to discover the associations between gene sets and disease phenotypes. Assuming coherent associations between the genes ranked by their relevance to the query gene set, and the disease phenotypes ranked by their relevance to the hidden target disease phenotypes of the query gene set, we formulate a learning framework maximizing the rank coherence with respect to the known disease phenotype-gene associations. An efficient algorithm coupling ridge regression with label propagation, and two variants are introduced to find the optimal solution of the framework. We evaluated the rcNet algorithms and existing baseline methods with both leave-one-out cross-validation and a task of predicting recently discovered disease-gene associations in OMIM. The experiments demonstrated that the rcNet algorithms achieved the best overall rankings compared to the baselines. To further validate the reproducibility of the performance, we applied the algorithms to identify the target diseases of novel candidate disease genes obtained from recent studies of GWAS, DNA copy number variation analysis, and gene expression profiling. The algorithms ranked the target disease of the candidate genes at the top of the rank list in many cases across all the three case studies. The rcNet algorithms are available as a webtool for disease and gene set association analysis at http://compbio.cs.umn.edu/dgsa_rcNet.

💡 Research Summary

**
The paper introduces rcNet (Rank Coherence in Networks), a novel network‑based framework for inferring associations between a set of candidate disease genes and disease phenotypes. Traditional gene‑set enrichment methods often fail when the gene list is short or poorly annotated because disease‑gene annotations are incomplete. rcNet overcomes this limitation by jointly exploiting three heterogeneous networks: (1) a gene‑gene interaction network (G), (2) a disease‑phenotype similarity network (P), and (3) a bipartite disease‑gene association network (A).

Given a query gene set g, rcNet first propagates the binary seed vector through G using label propagation (equivalently a random walk with restart) to obtain a smooth relevance vector ˜g. The propagation is governed by a parameter α that balances smoothness against fidelity to the seed. Analogously, for any candidate disease phenotype set p, a relevance vector ˜p is derived by propagating p through P with parameter β. Both propagations have closed‑form solutions: ˜g = (1‑α)(I‑αĜ)⁻¹g and ˜p = (1‑β)(I‑βĤ)⁻¹p, where Ĝ and Ĥ are row‑normalized adjacency matrices.

The core idea of rcNet is “rank coherence”: if p truly corresponds to the disease(s) underlying g, then genes with high scores in ˜g should be strongly connected to phenotypes with high scores in ˜p via the known associations A. This intuition is formalized in two ways. The primary formulation treats the problem as a ridge‑regression: minimize Ω(p)=‖A ˜p ‑ ˜g‖² + κ‖p‖². Substituting the closed‑form expressions for ˜g and ˜p yields a standard ridge‑regression objective whose solution is p* = (ĀᵀĀ + κI)⁻¹Āᵀ(I‑αĜ)⁻¹g, where Ā = (1‑β)A(I‑βĤ)⁻¹. The resulting p* is a real‑valued vector that approximates the binary phenotype indicator; the top‑scoring phenotypes are reported as the predicted disease(s).

A second, exhaustive strategy enumerates each phenotype individually. For each candidate phenotype j, a unit vector p_j is set, ˜p_j is computed, and a coherence score is evaluated either as Pearson correlation (rcNet corr) between A ˜p_j and ˜g, or as the negative squared difference Σ_i,j A_ij(˜p_i‑˜g_j)² (rcNet lap). The phenotype with the highest score is selected. This enumeration guarantees the exact optimum for a single‑phenotype query but incurs O(m³) computational cost (m = number of phenotypes) and does not naturally extend to multi‑phenotype predictions.

The authors evaluated rcNet on two versions of OMIM disease‑gene associations (May 2007 and May 2010). In leave‑one‑out cross‑validation, rcNet consistently outperformed state‑of‑the‑art methods such as CIPHER, PRINCE, and random‑walk‑with‑restart in terms of precision, recall, and area under the ROC curve. In a realistic “future‑prediction” test, where associations added after the 2007 snapshot were hidden, rcNet placed the newly discovered disease‑gene links within the top 5 % of its ranking far more often than competing methods.

Beyond OMIM, rcNet was applied to three independent case studies: (1) GWAS‑derived candidate genes for complex traits, (2) copy‑number‑variation regions implicated in disease, and (3) differentially expressed gene signatures from microarray experiments. In each scenario, the true disease phenotype (e.g., coronary artery disease for a GWAS set, schizophrenia for a CNV region, breast cancer for an expression signature) was ranked among the top 1‑3 predictions, demonstrating the method’s robustness to sparse and noisy gene lists.

Key strengths of rcNet include: (i) simultaneous exploitation of both gene‑gene and phenotype‑phenotype network topology, providing a global view of functional similarity; (ii) the use of label propagation scores that capture long‑range network effects, which is especially valuable when the query gene set is small; (iii) a mathematically principled ridge‑regression formulation that yields a closed‑form solution and can be extended with regularization or prior information. Limitations are primarily computational: matrix inversions scale cubically with the number of genes and phenotypes, which may become prohibitive for whole‑genome networks. Moreover, the current implementation focuses on binary phenotype selection; extending to multi‑label or hierarchical disease models would require additional modeling.

Future directions suggested by the authors involve (a) employing sparse matrix techniques or iterative solvers to reduce the O(n³ + m³) burden, (b) integrating graph neural networks to learn non‑linear transformations of the propagation scores, and (c) expanding the phenotype network with clinical metadata (treatment response, drug targets) to support precision‑medicine applications.

rcNet is made publicly available as a web tool (dgsa_rcNet) that accepts a gene list and returns a ranked list of disease phenotypes, facilitating rapid hypothesis generation for researchers working with high‑throughput genomic data.