Fig 1.
Black arrows indicate outputs that are actually produced. Red arrows indicate theoretical output.
Fig 2.
Trade-off between precision and recall for selecting the subset of interactions significantly deviating from zero versus all interactions.
Top and bottom panels depict gain of precision and loss of recall, respectively. (a) glinternet; (b) xyz.
Fig 3.
Identification of epistasis for increasing numbers of observations of the pairwise double knockdown.
The number of additional main effects not overlapping with the set of interacting genes is fixed to 20. Results using (a) glinternet and (b) xyz.
Fig 4.
The distribution of the fraction of gene pairs stratified by ranges of observed double knockdowns.
Gene pairs with zero observations are not shown. (a) p = 100, n = 1000; (b) p = 1000, n = 10000.
Fig 5.
Identification of epistasis for varying effect size.
Using (a) glinternet and (b) xyz.
Fig 6.
Concordance between the sign of true and estimated epistasis.
The fraction of incorrectly identified signs between true and estimated epistasis for (a) glinternet and (b) xyz.
Fig 7.
Concordance between the sign of true and estimated epistasis.
The fraction of incorrectly identified signs between true and estimated epistasis for (a) glinternet and (b) xyz.
Fig 8.
Simulations repeated using xyz and larger data sets.
(a) number of observations of double knockdown. (b) Precision/recall/f1 by actual effect strength.
Fig 9.
Precision, recall, and F1 performance for varying numbers of synthetic lethal pairs, with additional background interactions, using xyz.
Neither side of the lethal interactions are used as main effects, and as far as lethal interactions are concerned, there is no hierarchy present.
Fig 10.
Performance on increasingly large data sets.
(a) glinternet and (b) xyz.
Fig 11.
Run time in seconds to find interactions on increasingly large data set.
(a) glinternet. (b) xyz. We compiled glinternet with OpenMP and ran with numCores = 10.
Fig 12.
Distribution of xyz failures.
Table 1.
Ten most significant predicted effects of siRNA perturbation screens, targeting all human kinases.