Experimental evolution of diverse Escherichia coli metabolic mutants identifies genetic loci for convergent adaptation of growth rate
Fig 7
Description and validation of the k-nearest neighbors (KNN) model.
(A) Illustration of our KNN model for estimating the (unknown, indicated by the question mark) growth rate associated with a given expression profile from the known growth rates for different expression profiles in its neighborhood. Circle centers and sizes indicate gene expression profiles (gi) and growth rate (fi), respectively. (B) Validation of the KNN model using the gene expression data from all 19 strains for which we measured both gene expression and growth rate in triplicate (underlined strains in Table 1). The KNN analysis was trained on a dataset comprising (i) the data curated in [59], (ii) our transcripomic and growth rate data, and (iii) randomly generated pseudo-profiles (see Materials and Methods for details) for a total of 11,275 datapoints. The dataset was partitioned into fifths that preserved the growth rate distribution and each combination of four fifths was used to train a KNN model that predicted the growth rate of the left-out fifth. For comparison with (C) only predictions regarding our transcriptomic growth data are retained. The figure shows good agreement between the resulting predicted growth rates and the experimentally measured growth rates (R2 = 0.84, p-value < 10−38), where the line of perfect agreement (dashed), 10% error (dark grey) and 25% error (light grey) are included as a reference. (C) Prediction of our measured growth rates by the KNN analysis when trained on publicly available data only. Results are presented for dataset D2 in S9 Table (R2 = 0.11, p-value < 0.015). Colors and dashed line have the same meaning as in (B). The accuracy in (C) is lower than in (B) because, with only a limited number of experiments to use for training, outlier transcriptional states remain under-sampled making it hard to resolve the growth rate of these states.