Pairwise Versus Multiple Global Network Alignment

This abstract is based on the following paper: Vijayan, Vipin, Shawn Gu, Eric T. Krebs, Lei Meng, and Tijana Milenkovic. Pairwise Versus Multiple Global Network Alignment. IEEE Access 8 (2020): 41961--41974. Proteins, the major macromolecules of life, interact with each other to carry out cellular functioning. Thus, analyses of protein-protein interaction (PPI) networks can yield important insights into biological function, disease, and evolution. While biotechnological advancements have made PPI network data available for many species, functions of many proteins in many of these species remain unknown. One way to uncover these functions is to transfer biological knowledge from a well-studied species to a poorly-studied one. Genomic sequence alignment, which has revolutionized our biomedical understanding, can be used for this purpose. However, sequence alignment has a major drawback: it does not consider interactions between proteins (which are ultimately what carry out function). So, biological network alignment (NA) can be used in a complementary fashion to predict protein functional knowledge that sequence alignment alone cannot predict. Specifically, NA compares PPI networks of different species to find regions of their similarity (or conservation), thus allowing for the transfer of functional knowledge across conserved network (rather than just sequence) regions. Like genomic sequence alignment, NA can be local or global. Just as the recent trend in the NA field, we also focus on global NA, which can be pairwise (PNA) and multiple (MNA). While PNA aligns two networks, MNA can align more than two networks at once. Since MNA can capture conserved network regions between more networks than PNA, it is hypothesized that MNA leads to deeper biological insights compared to PNA. However, due to different outputs of PNA and MNA, a PNA method is only compared to other PNA methods, and an MNA method is only compared to other MNA methods. Comparison of PNA against MNA must be done to evaluate whether MNA indeed yields more biologically meaningful alignments than PNA, as only this would justify MNA's higher computational complexity. We introduce a framework that allows for this. We evaluate eight prominent PNA and MNA methods, on synthetic and real-world biological networks, using topological and functional alignment quality measures. We compare PNA against MNA in both a pairwise (native to PNA) and multiple (native to MNA) manner. PNA is expected to lead to higher-quality alignments than MNA under the pairwise evaluation framework. Indeed, this is what we find. MNA is expected to lead to higher-quality alignments than PNA under the multiple evaluation framework. Shockingly, we find this not always to hold; PNA is often better than MNA in this framework, depending on the choice of evaluation test. Thus, we believe that any new MNA methods should be compared not just to existing MNA methods, but also to existing PNA methods using our evaluation framework, to properly judge the quality of alignments that they produce. Also, we confirm empirically that PNA is faster than MNA in both evaluation frameworks. These results indicate that currently, MNA offers little advantage over PNA; in order for MNA to gain an advantage, a drastic redesign of MNA's current algorithmic principles might be needed.