Figure 5: Multidimensional models are sensitive and phenotype-specific.

(a) Sensitivity (top) and specificity (bottom) of each SWLR and BF model. SWLR models are more sensitive and more specific than the best possible single-feature model. Genotypes are ordered from most identifiable to less identifiable, from each group (refer to Fig. 3a). (b) SWLR model for a178 distinguishes only a178 populations, and barely population a163. Bar plots represent the average a178 average phenotypic probability of each population. (c) Matrix representing the performance of each SWLR model and the detectability of each mutant under other models. Coloured dots represent cases where the specific tested population (rows) scored with an average phenotypic probability above 0.5 under the tested model (columns). The size of the dot, as well as its colour, represents the average probability resulting from each model. Grey dots represent an average probability below 0.5. Dots plotted on the diagonal line show that each model is able to identify its own population with a probability >0.5 (except for two very subtle mutants), while the probabilities for wild type are all below 0.5 (no highlighting). Rows highlighted in green are severe mutant populations. Columns highlighted in purple are models for severe mutants. (d) Models for a178 detect a178 populations. (e) Identification of the population that each model best identifies (besides itself) reveals phenotypic similarities and corroborates known relationships. Empty circles represent cases where some genetic or phenotypic relationship is known. Coloured nodes are instances where known genetic relationships are confirmed in the network analysis. (a176-R2 is a different imaging set from a176, several generations later, a178(x3) is a three times outcrossed a178). Sample sizes, statistical analysis for model creation and feature selection are listed in Supplementary Information.