Here we discover forms of GWAS that go beyond univariate tests, considering the context of the SNP to help in its significance as biomarker.


Many approaches use networks to give each SNP a context. Specifically the protein-protein interaction network (PPIN) is widely used in biomarker discovery, where we place an edge if two proteins interact and we map high-throughput data of the gene to the corresponding protein (Robinson et al., 2017). The goal is to find interesting regions in the network, which result in a shorter list of biomarkers to further study.

  • Guilt by association approaches: non-hit genes with a good, albeit non-significant score which are associated to many hits are considered hits themselves.
  • Feature extraction from the PPIN: some approaches extract a similarity matrix based on adjacency, common neighbours, shortest path, diffusion kernel, etc. which then use to weight the scores.
  • Subgraph of hits: the subgraph is expected to contain hits with high scores, but also some hits with lower scores. Although the latter might add a suboptimal weight, they might allow for the inclusion of other high-score genes through their edges.

Guilt by association

Undirected graphical models

Robinson et al. propose using undirected graphical models a.k.a. Markov random fields (Robinson et al., 2017). They try to guess the true label $X_i$ by minimizing an energy function

where $i$ and $j$ are vertices, $(i,j)$ are connected pairs of vertices, and $l$ and $k$ are labels. This formula contains two kind of potentials:

  • Unary potentials $u_{i;l}$ that only depend on the score $z_i$ for that node and the probability density function for a label $l$ $\pi_l$
  • Pairwise potentials $w_{ij;lk}$ that penalizes neighbours not sharing label

    where $e_{ij}$ is the weight of the edge.

They tested their approach on an RNAi screen to identify genes implicated in DNA repair. For each RNAi they calculated a $\phi$-score from the fluorescence value of the reporter gene OGG1-GFP. They use STRING to compile a PPI network, in a weighted graph were the edge weight measures the evidence.


  • Robinson, S., Nevalainen, J., Pinna, G., Campalans, A., & Radicella, J. P. (2017). Incorporating interaction networks into the determination of functionally related hit genes in genomic experiments with Markov random fields, 170–179.