(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



Analysis of the evolution of resistance to multiple antibiotics enables prediction of the Escherichia coli phenotype-based fitness landscape [1]

['Junichiro Iwasawa', 'Department Of Physics', 'Graduate School Of Science', 'University Of Tokyo', 'Tokyo', 'Tomoya Maeda', 'Graduate School Of Agriculture Research', 'Faculty Of Agriculture', 'Hokkaido University', 'Sapporo']

Date: 2022-12 

The fitness landscape represents the complex relationship between genotype or phenotype and fitness under a given environment, the structure of which allows the explanation and prediction of evolutionary trajectories. Although previous studies have constructed fitness landscapes by comprehensively studying the mutations in specific genes, the high dimensionality of genotypic changes prevents us from developing a fitness landscape capable of predicting evolution for the whole cell. Herein, we address this problem by inferring the phenotype-based fitness landscape for antibiotic resistance evolution by quantifying the multidimensional phenotypic changes, i.e., time-series data of resistance for eight different drugs. We show that different peaks of the landscape correspond to different drug resistance mechanisms, thus supporting the validity of the inferred phenotype-fitness landscape. We further discuss how inferred phenotype-fitness landscapes could contribute to the prediction and control of evolution. This approach bridges the gap between phenotypic/genotypic changes and fitness while contributing to a better understanding of drug resistance evolution.

Funding: This study was supported in part by JSPS KAKENHI (17H06389 and 19H05626 to C.F.; https://www.jsps.go.jp/english/index.html ), JST ERATO (JPMJER1902 to C.F.; https://www.jst.go.jp/erato/en/index.html ). J.I. was supported by a Grant-in-Aid for the Japan Society for Promotion of Science Fellows (JP18J21942). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

( A ) The experimental conditions used in the current study. We prepared seven starting strains, i.e., the parent E. coli strain, evolved under TET (TETE4, TETE6), evolved under KM (KME1, KME5), and evolved under NFLX (NFLXE4, NFLXE6). Starting from these strains, we performed laboratory evolution under three drugs (TET, KM, and NFLX). ( B ) Schematic figure describing the protocol of the laboratory evolution experiment with simultaneous IC 50 measurements along the evolutionary trajectories in the present study. For every passage, IC 50 was quantified for eight drugs, and cells were transferred from the well with the highest selection-drug concentration where the cell concentration exceeded a threshold. ( C ) The time series of IC 50 values during the evolution of the KME5 strain under TET, for example. Data of four independent culture lines are overlaid. The resistance values for TET, KM, NFLX, SS, PLM, NQO, SDC, and MMC are shown. All time-series data in the present study are presented in S1 Fig . The data underlying this figure can be found in S1 Data . KM, kanamycin; MMC, mitomycin C; NFLX, norfloxacin; NQO, 4-nitroquinoline 1-oxide; PLM, phleomycin; SDC, sodium dichromate dihydrate; SS, sodium salicylate; TET, tetracycline.

In this study, we aimed to construct an empirical fitness landscape for the evolution of antibiotic resistance in Escherichia coli using phenotypes as its basis. Our basic strategy was to densely observe multiple phenotypes and their corresponding fitness (i.e., drug resistance) along different trajectories of evolution ( Fig 1A and 1B ). The observation of trajectories under different antibiotics, starting from different locations in phenotypic space, could allow us to sample sufficient phenotypes to help build the phenotype-fitness landscape. The critical feature of our work is that we used eight different antibiotic resistance values (IC 50 ) as probes for fitness, and their principal components as phenotypes ( Fig 1B ). As previous studies have suggested, the antibiotic resistance space corresponds to a subspace of the gene expression space, making it a good candidate for probing evolutionary dynamics [ 14 , 18 ]. In addition, antibiotic resistance values can be measured in a high-throughput manner [ 34 ], enabling the efficient sampling of phenotypes during the course of laboratory evolution.

Fitness (or adaptive) landscapes are frequently invoked in experimental and theoretical studies as they offer a basis for the predictability and convergence of evolution [ 26 , 27 ]. Thus, constructing empirical fitness landscapes based on experimental data may be a suitable approach for predicting evolution. Most previous studies on empirical fitness landscapes were based on fitness assays of large mutant libraries for a gene of interest, which revealed broad interactions between mutations and the nonlinearity of the underlying landscapes [ 4 , 6 , 7 , 28 , 29 ]. Despite our accumulating knowledge on how specific mutations affect fitness, we are still far from constructing a comprehensive fitness landscape capable of predicting the evolutionary process of an entire cell. This difficulty is because of the high dimensionality of genotypic space, which makes it difficult to acquire sufficient data for prediction [ 2 , 30 ]. Parallel laboratory evolution experiments have shown that different single nucleotide changes can underlie similar phenotypic changes, implying the existence of multiple paths in the genotypic space to reach a fitted phenotype [ 14 , 31 – 33 ]. However, phenotypes often exhibit repeatable features in laboratory evolution, implying the low dimensionality of phenotypic space for evolution [ 2 , 8 , 15 – 18 ]. Previous studies thus suggest that the fitness landscape may be more traceable when using phenotypes rather than high-dimensional genotypes as its basis.

Laboratory evolution experiments, high-throughput sequencing, and phenotyping combined with data analysis have heralded a new age for evolutionary biology [ 1 – 3 ]. Massively parallel laboratory evolution experiments have revealed an unprecedented amount of information on evolution, including epistatic interactions in genotypic space [ 4 – 7 ], genotype–phenotype mapping [ 8 – 14 ], and the existence of repeatable features in evolutionary processes [ 8 , 15 – 17 ]. Furthermore, quantitative and theoretical modelling have enabled us to build fitness models based on genotypic and phenotypic data [ 2 , 18 – 20 ], enabling us to explore the origins of the constraints underlying evolution [ 21 – 25 ]. Despite numerous attempts to describe evolutionary dynamics, predicting and controlling evolution remains difficult.

Results

Observing trajectories in the PCA space We performed principal component analysis (PCA) of the resistance profiles for the 44 trajectories to investigate the evolutionary trajectories in the eight-dimensional resistance space. Because the dynamic ranges of IC 50 varied among the eight antibiotics, we normalized them and set the mean and standard deviation of the IC 50 variation among the 44 trajectories over the 27 days of evolution to (mean, standard deviation) = (0,1) before applying PCA. In Fig 3A and 3B, examples of evolutionary trajectories are highlighted in two-dimensional PCA space (explained variance ratio PC1:41%, PC2:23%). While most of the strains took similar trajectories, indicating convergence in the phenotypic space, several conditions (such as strains under TET evolution starting from KME1 and KME5) exhibited different evolutionary paths (Fig 3A). This divergence for TET evolution and the underlying genetic backgrounds will be discussed later. To interpret the phenotypes in PCA space, we plotted the IC 50 values for each antibiotic (Fig 3C–3J). These plots allowed us to visualize phenotypes with high fitness for the corresponding drugs. For example, the phenotypes with high KM resistance were located on the lower-right side of the PCA space, indicating that PC1 is positively correlated with KM resistance (Fig 3D). This tendency is also consistent with the evolutionary trajectories under KM selection (Fig 3B). The results indicate that the phenotypes that were highly resistant to each drug occupied different regions of PCA space. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 3. Evolutionary trajectories in two-dimensional PCA space. A total of 44 trajectories (11 conditions with four replicates each) in an eight-dimensional resistance space were projected into two-dimensional space using PCA. The variance explained by the first two components was 41% and 23%, respectively. (A) Evolutionary trajectories under TET selection. Trajectories starting from the parent and KM-evolved strains are highlighted by colors, while the gray lines show the trajectories in other conditions. The black and yellow circles represent the state of starting strains and the strains on the last day, respectively. (B) Evolutionary trajectories under KM selection. Trajectories starting from the parent and TET-evolved strains under KM selection are highlighted in color. The black triangles in (A) and (B) show the peaks of the inferred phenotype-fitness landscape for TET and KM, respectively, and will be explained later in Fig 4. (C–J) IC 50 values for each drug mapped on the two-dimensional PCA space. The colors represent the log-transformed IC 50 values (log 2 μg/mL). The data underlying this figure can be found in S1 Data. KM, kanamycin; PCA, principal component analysis; TET, tetracycline. https://doi.org/10.1371/journal.pbio.3001920.g003

Inference of the phenotype-fitness landscapes Based on the analysis shown in Fig 3, we inferred the phenotype-fitness landscape of drug resistance on the two-dimensional PCA plane by smoothing the IC 50 data via a Gaussian kernel (see Materials and methods for details). Fig 4 shows the inferred phenotype-fitness landscapes for the eight antibiotics used in this study. The inferred landscapes have a small number of peaks for each landscape, suggesting low ruggedness [4,7]. Note that the high noise frequency in the original IC 50 time series, compared with the mutation acquisition frequency, implies that the noise stems from measurement errors rather than from genotypic ruggedness. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 4. Estimated phenotype-fitness landscape. (A–H) The phenotype-fitness landscape obtained using Gaussian smoothing for eight drugs is presented. The resistance levels were normalized to a maximum value of one (see Materials and methods). (I) Differences in the evolutionary directions between the simulated gradient from the phenotype-fitness landscape (θ landscape ) and experimental observation (θ v ). For comparison, the difference with a random direction (θ random ) sampled from a uniform distribution [0,2π) is also plotted. The data underlying this figure can be found in S1 Data. KM, kanamycin; MMC, mitomycin C; NFLX, norfloxacin; NQO, 4-nitroquinoline 1-oxide; PLM, phleomycin; SDC, sodium dichromate dihydrate; SS, sodium salicylate; TET, tetracycline. https://doi.org/10.1371/journal.pbio.3001920.g004 We first evaluated whether the landscape could represent the observed evolutionary trajectories by analyzing the relationship between the gradients in the landscape and the direction of phenotypic changes. As the raw evolutionary trajectories were subject to experimental noise, we applied a moving average of six time points to the trajectories using a triangular window. The direction provided by the landscape θ landscape (x) at location x is calculated as follows: (1) where F(x) represents the altitude of the phenotype-fitness landscape and ∂/∂x PC1,2 is the gradient along the PC1,2 axis (Fig 4I). θ v (i), the evolutionary direction in the experimental observations, was calculated from the coordinates of the ith time point in the smoothed trajectories x PC1,2 (i) using (2) In Fig 4I, we plotted the distribution of |θ v −θ landscape | where θ landscape was calculated for each corresponding point along the observed trajectories. For comparison, we also plotted |θ v −θ random |, where θ random was sampled from a uniform distribution [0,2π). As shown in Fig 4I, the directions of evolution provided by the inferred landscape θ landscape had a significant correspondence with the observed trajectories θ v (p = 9.23×10−26, one-sided Mann–Whitney U test), suggesting that the inferred landscapes could reproduce the observed evolutionary trajectories. Note that given the fact that each step along the trajectories was defined by time (day) and not by a step in the genotypic space, there were moments where the trajectories stayed at a certain point and where θ v deviated from the inferred landscape θ landscape . In addition, the inferred phenotype-fitness landscape was calculated based only on the first two principal components, and the remaining 36% of the variance was not considered. Although we compared the inferred landscape with the trajectories used for the inference process, the two limitations raised above suggest that the correspondence between θ v and θ landscape was not trivial. The predictability of the inferred landscapes was evaluated by applying train/test splitting to the whole dataset (S2 Fig). Here, the training dataset was obtained by randomly excluding one out of the four replicates for each experimental condition, resulting in a total of 33 lines, and the excluded 11 lines were used for the test dataset. The phenotype-fitness landscape was inferred using the training data (S2A–S2C Fig) and was compared with the evolutionary directions (θ landscape,train ), with experimental observations based on the test data (θ v,test ) (S2D Fig). The landscape gradients and experimental observations showed significant agreement. Note that this verification for predictability is only possible for the landscape of TET, KM, and NFLX since the laboratory evolution experiments were performed only under these three drugs. However, the landscape properties of the other five drugs might also have similar predictive power due to the tight interconnectedness of drug resistance acquisition relations, which is known as cross-resistance and collateral sensitivity [11,12,14,18]. Overall, the results suggest a correspondence between the landscape-based properties and the experimental observations, indicating the validity and predictive power of the inferred landscapes.

Genotypes underlying two peaks in TET fitness landscape One interesting finding in the previous section is that there were two distinct peaks in the TET fitness landscape (Fig 4A). This structure originated from differences in evolutionary trajectories, starting from different initial strains. As shown in Fig 3A, most of the trajectories that started from KM evolved strains evolved toward the upper minor peak to increase TET resistance. In contrast, the phenotypes of other TET-resistant strains (e.g., TETE4 and TETE6 starting strains; see Fig 3B) were close to the major peak. This result demonstrates the historical dependence of the evolution of TET resistance. To clarify the genetic mechanism of this historical dependence, we resequenced eight TET-evolved strains starting from KM-evolved strains (four from KME1 and four from KME5) to identify mutation fixation during evolution under TET selection. As shown in Table 1, 7/8 of the resequenced strains had null mutations in the coding region of lon, which encodes Lon protease. These seven strains coincided with the strains that evolved toward the minor peak in the landscape, suggesting that the minor peak observed in Fig 4A corresponds to resistance acquisition caused by a lon mutation. It has been previously shown that the lon− mutant shows a low level of multidrug resistance, considering TET, ampicillin, chloramphenicol, and erythromycin (see Table 1 in [35] and [36,37]). We also confirmed the effect of the lon mutation on the parent MDS42 strain, showing that the lon− mutant exhibited a 2-fold increase in IC 50 for TET [14]. This study also showed that the lon− mutation caused sensitivity to MMC, which is consistent with the fact that the minor peak was located in the MMC-sensitive region (Fig 3J). PPT PowerPoint slide

PNG larger image

TIFF original image Download: Table 1. Mutations identified in KM evolved strains under TET selection. https://doi.org/10.1371/journal.pbio.3001920.t001 Interestingly, while lon mutations were identified in most of the TET-evolved strains starting from the KM-resistant phenotype, these mutations were not observed under TET selection when it started from the MDS42 parent strain [14]. In the latter case, mutations in the acrR regulator, which can activate the expression of the acrAB efflux pump genes, leading to TET resistance, were commonly identified. Here, the increase in IC 50 owing to the acrR− mutation was significantly larger than that due to the null mutation of lon [14]. This raises the following question: What makes the lon mutation special for TET resistance when starting from a KM-evolved strain? In E. coli, the uptake of aminoglycosides as KM is positively correlated with the proton motive force (PMF) [38,39]. Indeed, the KM-resistant phenotype in KME1 and KME5 strains was achieved by decreasing PMF to suppress the uptake of KM by null mutations of cyo genes whose products are involved in the electron transfer system (the identified mutations in KME1 and KME5 are presented in Supplementary Data 3 in [14]). Simultaneously, these cyo mutations result in hypersensitivity to several antibiotics, including TET [14]. This hypersensitivity was caused by the decreased activity of the AcrAB efflux pump, as this pump is a proton antiporter, and its activity is positively correlated with PMF [40,41]. The trade-off between KM and TET resistance is based on these PMF-dependent changes in the efflux and uptake activities of drugs [11,42]. In the present study, the KM evolved strains with cyo mutations did not acquire acrR mutations under TET selection, probably because of the decreased fitness gain of acrR mutations after the fixation of the cyo mutations. Specifically, increasing the activity of the AcrAB efflux pump by the acrR mutation is difficult under the decreased activity of the electron transfer system by cyo mutations. As an alternative strategy, in KME1- and KME5-evolved strains, lon mutations increased TET resistance. Although the detailed mechanism by which lon mutations contribute to TET resistance is unclear, one possible mechanism can be related to the stabilization of the transcriptional activator MarA, which is the substrate of the Lon protease [36,43], and subsequent activation of the marRAB operon. It was recently shown that the activation of the marRAB operon can cause antibiotic resistance through the acidification of the cytoplasm [44]. This previous study also showed that acidification-based drug resistance was independent of changes in PMF. These results suggest that this acidification-related resistance phenotype is a possible strategy for KME1- and KME5-evolved strains to achieve TET resistance under decreased PMF through cyo mutations. It should be noted that one out of eight TET-evolved strains starting from KM-evolved strains had no mutation in lon. The evolutionary trajectory toward this strain exhibited a different direction of phenotypic change (denoted by the arrowhead in Fig 3A) compared with the other seven trajectories. We confirmed that this strain had acquired mutations in acrA, rpoA, and atpB. The identified V45I mutation in acrA corresponded to the MP domain of AcrA, which transmits conformational changes in AcrB to TolC [45]. Given these results, we speculate that this strain acquired resistance to TET by improving the activity of the AcrAB-TolC efflux pump through an acrA mutation and not through the lon mutation.

[END]
---
[1] Url: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3001920

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/