Fig. 3: Comparison of epistasis measures using simulated data from the NK fitness model.
From: Resolving discrepancies between chimeric and multiplicative measures of higher-order epistasis
A A fitness landscape f simulated following the NK fitness model with “ruggedness” parameter K = 1 contains only pairwise interactions. These interactions are represented with an interaction graph G. The 3-way multiplicative measure \({\epsilon }_{ijk}^{M}=0\) equals zero for all loci triples (i, j, k). However, if the triple (i, j, k) forms a triangle in the graph G (shown as a dashed red line), then the 3-way chimeric epistasis measure \({\epsilon }_{ijk}^{C}\) is non-zero and incorrectly indicates the presence of a higher-order interaction. B The fraction of non-zero (K + 2)-way interactions ("fraction of non-zero higher-order interactions'') identified by the multiplicative measure ϵM (orange) and the chimeric measure ϵC (blue) across 100 fitness landscapes f simulated according to the NK fitness model with ruggedness parameter K, with error bars indicating standard deviation across simulated instances. The fitness landscape f contains at most (K + 1)-way interactions, but the chimeric measure ϵC spuriously detects many non-zero (K + 2)-way interactions.
