Lagrangian Relaxation Applied to Sparse Global Network Alignment

Data on molecular interactions is increasing at a tremendous pace, while the development of solid methods for analyzing this network data is lagging behind. This holds in particular for the field of comparative network analysis, where one wants to identify commonalities between biological networks. Since biological functionality primarily operates at the network level, there is a clear need for topology-aware comparison methods. In this paper we present a method for global network alignment that is fast and robust, and can flexibly deal with various scoring schemes taking both node-to-node correspondences as well as network topologies into account. It is based on an integer linear programming formulation, generalizing the well-studied quadratic assignment problem. We obtain strong upper and lower bounds for the problem by improving a Lagrangian relaxation approach and introduce the software tool natalie 2.0, a publicly available implementation of our method. In an extensive computational study on protein interaction networks for six different species, we find that our new method outperforms alternative state-of-the-art methods with respect to quality and running time.

💡 Research Summary

The paper addresses the problem of global pairwise network alignment (GNA), a task that seeks a one‑to‑one mapping between the vertices of two biological interaction graphs while simultaneously rewarding the preservation of edges. The authors formulate GNA as an integer linear program (ILP) that generalizes the well‑studied quadratic assignment problem (QAP). In this formulation, binary variables xᵢₖ indicate whether vertex i of the first graph is aligned to vertex k of the second graph, and auxiliary binary variables yᵢₖⱼₗ represent the product xᵢₖ·xⱼₗ, i.e., whether the pair (i, j) is aligned to the pair (k, l). The objective function consists of a sum of node similarity scores c(i,k) and edge similarity scores w(i,k,j,l). Constraints enforce that each node participates in at most one match, turning the classic assignment constraints of QAP into inequalities appropriate for a matching problem.

To obtain tractable bounds, the authors apply Lagrangian relaxation: they dualize the linking constraints that enforce yᵢₖⱼₗ = yⱼₗᵢₖ using multipliers λᵢₖⱼₗ. The resulting Lagrangian dual (LD) decomposes into |V₁|·|V₂| independent maximum‑weight bipartite matching subproblems, each solvable by the Hungarian algorithm. For dense graphs the overall complexity is O(n⁵), but by exploiting the sparsity typical of protein‑protein interaction (PPI) networks the authors achieve O(n⁴ log n) using a successive‑shortest‑path variant.

The LD provides an upper bound Z_LD(λ) on the optimal alignment score, while the primal solution x obtained from the same λ yields a feasible alignment whose score Z_lb(λ) serves as a lower bound. The gap Z_LD – Z_lb can be reduced by iteratively updating λ. Two complementary strategies are proposed:

1. Subgradient Optimization – a classic method where λ is updated in the direction of the subgradient g(λ) = yᵢₖⱼₗ – yⱼₗᵢₖ, scaled by the current duality gap. The step size α is adaptively halved or doubled based on recent improvements, and the process stops when subgradients vanish or a time/iteration limit is reached.

2. Dual Descent – an extension of the approach by Adams and Johnson for QAP. Here the authors examine the dual of the matching subproblems (variables α, β for the bipartite matching and µ, ν for the edge‑pair subproblems) and compute their slacks π and γ. λ is then updated using a convex combination of γ and the slacks, ensuring that the new λ does not violate feasibility of the dual variables. This method guarantees that the upper bound never increases and often yields a much tighter bound in far fewer iterations.

The algorithm alternates or combines these two update schemes, achieving rapid convergence in practice. The authors implement the full pipeline in a publicly available tool called natalie 2.0. The software accepts a node similarity matrix C and an edge‑weight tensor W, and allows users to prune unlikely node pairs based on external evidence (e.g., BLAST e‑values, Gene Ontology similarity), thereby making the alignment graph sparse and further speeding up the matching step.

Experimental evaluation uses six real PPI networks (human, yeast, fly, etc.) ranging from a few thousand to tens of thousands of proteins. The authors compare natalie 2.0 against two state‑of‑the‑art global alignment methods: IsoRank, which solves a spectral relaxation of the same objective, and Graal, which aligns graphlets. Performance is measured by (i) the number of conserved edges and (ii) functional coherence of aligned protein clusters assessed via GO term enrichment. natalie 2.0 consistently discovers 10–20 % more conserved edges than the competitors and achieves higher GO coherence scores. In terms of runtime, natalie 2.0 is 2–3 times faster than IsoRank and about 1.5 times faster than Graal on the same hardware. Notably, when only subgradient optimization is used, thousands of iterations are required; the addition of dual descent reduces this to a few hundred iterations, dramatically improving practical efficiency.

The paper’s contributions are threefold: (1) a rigorous ILP formulation of global network alignment that captures both node and edge similarity; (2) a novel combination of Lagrangian relaxation, subgradient optimization, and dual descent that yields strong, provable bounds while remaining computationally feasible for large, sparse biological networks; (3) an open‑source implementation that outperforms existing methods on real‑world data. Limitations include the current restriction to pairwise one‑to‑one alignments (extension to many‑to‑many or multi‑network alignment is left for future work) and the assumption of non‑negative edge weights (negative scores would require additional handling).

In summary, the authors provide a mathematically solid, algorithmically efficient, and empirically validated framework for global network alignment, advancing the state of the art in comparative network biology and offering a versatile tool for future studies involving large‑scale interaction graphs.