1 Title: 1 Rational Design of Antibiotic Treatment Plans 2 3 Portia M. Mira 1 4 Kristina Crona 2 5 Devin Greene 2 6 Juan C. Meza 1 7 Bernd Sturmfels 3 8 Miriam Barlow 1 9 10 11 Institutional Affiliations: 12 1 School of Natural Science, University of California, Merced 13 2 Department of Mathematics, American University 14 3 Departments of Mathematics, Statistics, and EECS, University of California, 15 Berkeley 16 17 18 Key Words: 19 Adaptive Landscapes, Antibiotic Cycling, β-lactams, Antibiotic Resistance 20 21 Abstract: 22 23 24
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
Title: 1 Rational Design of Antibiotic Treatment Plans 2 3 Portia M. Mira1 4 Kristina Crona2 5 Devin Greene2 6 Juan C. Meza1 7 Bernd Sturmfels3 8 Miriam Barlow1 9 10 11 Institutional Affiliations: 12 1School of Natural Science, University of California, Merced 13 2 Department of Mathematics, American University 14 3 Departments of Mathematics, Statistics, and EECS, University of California, 15 Berkeley 16 17 18 Key Words: 19 Adaptive Landscapes, Antibiotic Cycling, β-lactams, Antibiotic Resistance 20 21 Abstract: 22 23 24
2
Introduction 25 Antibiotic resistance is an inevitable outcome whenever antibiotics are used. 26 There are many reasons for this: 1) As humans (also as eukaryotes), we are 27 vastly outnumbered by bacteria in nearly all measures, including total population 28 size, biomass, genetic diversity, emigration, and immigration [1]; 2) bacteria can 29 use horizontal gene transfer to share resistance genes across distantly related 30 species of bacteria, including non-pathogens [2]; 3) compared to humans, 31 bacteria have relatively few vulnerable target sites [3]; 4) microbes are the 32 sources of nearly all antibiotics that are used by humans [4]. Given the 33 overwhelming numbers of bacteria, the limited number of target sites, the 34 numerous ways that they can infect humans, and that they have been exposed to 35 naturally occurring antibiotics for billions of years, resistance to antibiotics used 36 by human populations is unavoidable. 37 38 Once resistance is present in a bacterial population, it is exceedingly difficult to 39 remove for several reasons. If any amount of antibiotic is present in the 40 environment, antibiotic resistance genes will confer a large fitness advantage [5], 41 and even when antibiotics are not present in an environment, the fitness costs for 42 carrying and expressing resistance genes are small to non-existent [6]. In 43 addition to it being difficult to remove antibiotics from the environment [7], if 44 humans were to completely abandon the use of antibiotics, resistance would 45 persist for years [8]. 46 47 Efforts to remove resistance genes from clinical environments by either 48 discontinuing or reducing the use of specific antibiotics for some period of time, 49 either through general reduction of antibiotic consumption or periodic rotations of 50 antibiotics (cycling) have not worked in any reliable or reproducible manner [9]; 51 indeed it would have been surprising if they had worked [10],[11]. 52 53 Since antibiotic resistance is unavoidable, it only makes sense to accept its 54 inevitability and do the best we can within that framework. A reasonable 55 approach is to rotate the usage of antibiotics. This has been implemented in 56 many ways and there are recent studies to model the optimal duration, mixing vs 57 cycling, and how relaxed antibiotic cycles may be and still function as planned 58 [12,13]. However, those models have not focused on developing a method for 59 creating the ideal succession of antibiotics. In a previous publication [14], we 60 proposed that susceptibility to antibiotics could be restored by rotating 61 consumption of multiple antibiotics that are a) structurally similar, b) inhibit/kill 62 bacteria through the same target site, and c) result in pleiotropic fitness costs that 63 reduce the overall resistance of bacteria to each other. We showed an anecdotal, 64 proof-of-principle example [14] of how this might work with a series of β-lactam 65 antibiotics in which some would select for new amino acid substitutions in the 66 TEM β-lactamase and others that would select reversions in TEM ultimately 67 leading back to the wild-type (un-mutated) state. 68
3
69 Our current work is to identify β-lactam treatment plans that are the most likely to 70 return a population expressing a small number of variant TEM genotypes to the 71 wild-type state. The wild type TEM-1 and a handful of its descendants confer 72 resistance to penicillins alone, while most of its descendants confer resistance to 73 either cephalosporins or penicillins combined with β-lactamase inhibitors 74 (inhibitor resistance), and a few confer resistance to both. Of the 194 clinically 75 identified TEM genotypes that encode unique amino acid sequences [15], 174 76 (89.7%) differ from the wild type TEM-1 by at most four amino acid substitutions 77 (see Table 1). Our choice of a system that includes four amino acid substitutions 78 is based upon an apparent threshold for amino acid substitutions among 79 functional TEM genotypes. The rarity of the co-existence of cephalosporin 80 resistance and inhibitor resistance and the fact that no single substitution confers 81 both phenotypes suggested that sign epistasis (i.e. reversals of substitutions 82 from beneficial to detrimental) exists as the substitutions that contribute to this 83 dual phenotype are combined. 84 85 The ability to push an evolved TEM genotype back to the wild type state would 86 limit the range of antibiotics to which it could confer resistance. To embark upon 87 our effort of determining the best way to do this, we decided to create a model 88 system based upon the TEM-50 genotype, which differs from TEM-1 by four 89 amino acid substitutions. All four substitutions by themselves confer clearly 90 defined resistance advantages in the presence of certain antibiotics. Additionally, 91 TEM-50 is one of the few genotypes that simultaneously confers resistance to 92 cephalosporins and inhibitor combined therapies. 93 94 Results 95 From experimental data to mathematical models 96 97 We created all 16 variant genotypes of the four amino acid substitutions found in 98 TEM-50 using site directed mutagenesis (Table 2) and measured the growth 99 rates of 12 replicates of E.coli DH5α-E expressing each genotype in the 100 presence of one of fifteen β-lactam antibiotics (Table 3). Each genotype was 101 grown in each antibiotic in 12 replicates. We computed the mean growth rate of 102 those replicates (Table 4) and the variance of each sample, as well as the 103 significance between adjacent genotypes that differed by one amino acid 104 substitution. This was done using one-way ANOVA analysis. 105 106 The results are summarized in Figures 1-15, where the arrows in the landscape 107 maps connect pairs of adjacent genotypes. For each comparison of adjacent 108 genotypes, we indicate the one whose expression resulted in the faster growth by 109 directing the arrowhead towards that genotype, and implying that evolution would 110 proceed in that direction if the two genotypes occurred simultaneously in a 111 population. In other words, the one indicated by the arrowhead would increase in 112
4
frequency and reach fixation in the population, while the other would be lost. Red 113 arrows indicate significance, and black arrows indicate differences that were not 114 statistically significant by ANOVA, but that may still exist if a more sensitive 115 assay was used. 116 117 We rank ordered the genotypes (Table 5) in each landscape diagram with a 118 score from 1 to 16, with the genotype promoting the fastest growth receiving a 119 score of “1” and the genotype with the slowest growth a score of “16”. This 120 analysis shows that all genotypes have a score of 5 or better and a score of 13 or 121 worse, in at least one landscape, indicating that there is abundant pleiotropy as 122 antibiotic selective pressures change. That pleiotropy provides a basis for 123 effectively alternating antibiotic to restore the wild type. 124 125 Based on the strong patterns of pleiotropy we observed, we reasoned that the 126 choice and the succession of antibiotics were at least as important as other 127 cycling considerations. We formalized our approach to optimal cycling as follows. 128 129 We start by considering the 15 antibiotics previously mentioned in Table 3: AMP, 130 AM, CEC, CTX, ZOX, CXM, CRO, AMC, CAZ, CTT, SAM, CPR, CPD, TZP, and 131 FEP. For each of these 15 antibiotics, we select exactly one TEM fitness 132 landscape that exists at a specific concentration of the antibiotic. That landscape 133 is a real 2 × 2 × 2 × 2 tensor f = ( fijkl )whose entries are the growth rates we 134 measured. Those growth rates depend upon the states of the four functionally 135 important amino acid residues involved in the evolution of TEM-50. The indices 136 i, j,k,l correspond to four possible amino acid substitutions and exist in either 137 state 0, corresponding to no substitution at that site, or 1, which corresponds to 138 an amino acid substitution that is involved in the resistance phenotype. We can 139 identify f with a vector of length 16 whose coordinates are indexed by {0,1}4 . 140 The resulting 15 vectors, one for each antibiotic, are the rows in Table 4. f ai( ) 141 contains the growth rates of each genotype as a function of antibiotic ai . M ( f ) 142 is a transition matrix whose rows contains the fixation probabilities for all possible 143 transitions in a single antibiotic (eg M f( )u ,v is the probability that genotype u is 144 replaced by genotype v). For this reason, a transition matrix has nonnegative 145 entries and its rows sum to 1. The rows and columns of M ( f ) are labeled by 146 {0,1}4 , in lexicographical order that is fixed throughout. We require that our 147 transition matrices respect the adjacency structure of the 4-cube, that is, 148 M ( f )u ,v = 0 unless u and v are vectors in {0,1}4 that differ in at most one 149 coordinate. For that reason our transition matrices can have at most four entries 150 corresponding to transitions to the immediately adjacent genotypes that add to 151 one. In other words, we reasoned that resistant strains are most likely to be in 152 competition with those that express resistance genotypes that are immediately 153 adjacent (vary by a single amino acid substitution). In our model, each TEM has 154
5
four genotypes that are adjacent, since there are four amino acid residues under 155 consideration. 156 157 Our substitution model is a function M :16 → 16×16 that assigns a transition 158 matrix M ( f (ai )) to the fitness landscape for antibiotic ai . An entry of that matrix 159 is denoted M ( f (ai ))u ,v .This represents the fixation probability for genotype u 160
transitioning to genotype v in the presence of antibiotic ai . If (a1,a2,....,ak ) is a 161 sequence of k antibiotics, then the matrix product is 162 M ( f (a1))∗M ( f (a2 ))∗...∗M ( f (ak )) . Our goal is to maximize the matrix entry 163 M ( f (a1))∗M ( f (a2 ))∗...∗M ( f (ak )){u ,0000} for all 15 genotypes u other than 0000. 164
For each u this requires searching over all antibiotic sequences of length k . 165 166 Finding optimal sequences of antibiotics 167 We used two substitution models to determine the optimal (most probable) 168 sequences of β-lactams for returning TEM genotypes back to their wild type 169 state. Briefly, the Correlated Probability Model (CPM) allows probabilities to be 170 based upon the actual growth rates. It is given by applying formula (7) to the 171 growth rates in Table 4. The Equal Probability Model (EPM) assumes that 172 beneficial mutations are equally likely and that only the direction of the arrows in 173 Figures 1-15 is important. This means that the matrix entry M ( f )u ,v is 1/ N if 174 genotype u has N outgoing arrows and there is an arrow from u to v . 175 176 A visual summary of the highest probabilities seen in the 15 CPM transition 177 matrices is provided in Figure 16. The CPM provides good estimates 178 if fitness differences between genotypes are small [14,16,17,18]. The EPM has 179 been used in settings where only rank order (as in Table 5) is available [19]. 180 181 For all sequences of antibiotics of a fixed length (2, 3, 4, 5, and 6), we examined 182 the probability that a given genotype is returned to the wild type state. For every 183 starting genotype, we found we were able to return to the wildtype genotype with 184 a probability between 0.6 and 1.0 when using the CPM model and a probability of 185 0.375 and 1.0 when using the EPM model. These results are summarized in 186 Tables 6-9 and Figure 17. These results show the number of paths and their 187 probabilities (Tables 6 and 7) and the substitutions of the most probable paths 188 (Tables 8 and 9) for returning to the wild type state from various starting points. 189 190 Once returned to the wild type state, we identified cycles that would allow for 191 alternation of antibiotics, and allow for some variation through amino acid 192 substitution, but then rapidly return bacteria to the wild type state (Figure 18). 193 Such cycles were possible for path length of 2, 4, and 6 and the probabilities of 194 those paths were respectively 0.704, 0.617, 0.617. We found that in the most 195 probable cases, the genotype varied by only one amino acid substitution before 196 reverting back to the wild-type state. However, when treatment plans with lower 197
15k
6
probabilities are considered, we find that more amino acid substitutions in the 198 genotype are allowed. 199 200 Discussion 201 In this study, we have developed an experimental approach for measuring 202 pleiotropy and a computational approach for optimizing antibiotic treatment paths. 203 The experimental approach we developed is rapid and high throughput, and 204 should be applicable to many species of resistant bacteria. The mathematical 205 model we created expresses the problem of antibiotic resistance in general 206 terms, and can therefore be applied to other resistance phenotypes where 207 pleiotropy occurs to identify the antibiotic treatment plans that have the highest 208 probability of reversing the evolution of resistance. 209 210 The purpose of this study was to determine whether it is possible to use selective 211 pressures to return TEM-genotypes to the wild-type state, as observed in 1963. 212 The methods may also be used to select for any particular genotype within our 213 data set, and can therefore be used generally to select, with reasonable 214 precision, for resistance genotypes that may have existed at any time point up to 215 the present. To emphasize the potential of this approach, we have named our 216 computational software package “Time Machine”. 217 218 Once given growth rates of adjacent genotypes, Time Machine returned 219 treatment plans that restored the wildtype state as observed in 1963 with 220 probabilities >0.6 when using the CPM model and greater than 3/8 (>0.375) 221 when using EPM. These results suggest that when possible a CPM model 222 including actual growth rates rather than rough ranking data is desirable. 223 224 Tables 6 and 7 suggest that the maximum probabilities in each row no longer 225 increase after a limited number of steps. This is not always the case. We have 226 constructed a particular example (see supplemental information) of two 227 substitution matrices on a 3-locus system where the maximum probabilities 228 increase by the number of steps indefinitely. 229 230 These results show that great potential exists for remediation of antibiotic 231 resistance through antibiotic treatment plans when pleiotropic fitness costs are 232 known for an appropriate set of antibiotics. While developed using a model of 233 Gram-negative antibacterial resistance, this approach could also be used for 234 Gram-positive bacteria and HIV treatment plans. 235 236 Methods 237 Experimental methods 238 Strains and Cultures 239 We expressed 16 mutant constructs of the blaTEM gene in plasmid pBR322 from 240 strain DH5-αE. The 16 genotypes differ at all combinations of four amino acid 241
7
residues and have been previously described [14]. We grew them overnight (16 242 hours) in standing cultures and diluted them to a concentration of 1.9X105 as 243 described elsewhere [14]. 244 245 We transferred 80 µl of each culture to a 384-well plate with one genotype 246 present in each of the 16 rows. The first 12 wells of each row were antibiotic free 247 (controls) and the last 12 wells contained a single antibiotic at an inhibitory, 248 sublethal concentration 249 250 After plating, a membrane is placed over the plate and simultaneously 251 incubated/measured in the Eon Microplate Spectrophotometer at a temperature 252 of 25.1°C for 22 hours. This relatively cool (<37º) temperature is used because 253 degradation of the antibiotics is much slower, while the growth rate of the 254 bacteria is still sufficient to capture the complete exponential period of growth 255 over the duration of the experiment. Overall, we have found that a temperature 256 ~25ºC yields more reliable and consistent measurement of growth rates in the 257 presence of antibiotics. 258 259 Measurements of cell density (light scattering) at a wavelength of 600 260 nanometers were automatically collected every 20 minutes after brief agitation to 261 homogenize and oxygenate the culture. 262 263 Growth Rates 264 The data obtained from the microplate spectrophotometer is exported to the 265 GrowthRates program to derive the growth rates. In essence, by measuring the 266 optical density at frequent intervals the GrowthRates program can estimate the 267 growth rate, a, through a linear regression algorithm fitting the data from the 268 exponential growth phase. Details can be found in [20] in the section entitled 269 “The Growth Curve” located on pages 233-4. The output of this program for the 270 data we collected was a list f (a1), f (a2 ),..., f (ak ) of 15 tensors, each of format271 2 × 2 × 2 × 2 . These are the rows in Table 4. So if u ∈{0,1}4 is a genotype, then 272 f (ai )u is the fitness of genotype u in the presence of antibiotic ai . This fitness is 273
a growth rate, so we are here using the letter f for a quantity often denoted byα . 274 275 One-Way Analysis of Variance (ANOVA) was then used to compare the means 276 of the growth rates obtained, and to determine if there were significant 277 differences between the growth rates of adjacent genotypes. 278 279 Time Machine Programs 280 -Derivation of Correlated Probability Model (CPM) 281 Once the growth rates have been determined under various experimental 282 conditions, the next step is to use them to compute fixation probabilities. 283 284 If the (multiplicative) absolute fitnesses Wu and Wv of two neighboring genotypes 285
8
u and v, differ by a small quantity then the (additive) relative fitness lnWu
Wv
⎛⎝⎜
⎞⎠⎟
286
can be approximated by 287 288
289
ln Wv
Wu
⎛⎝⎜
⎞⎠⎟= T fv − fu( ) 290
where T is the generation time. After a Taylor series approximation 291 292
ln Wv
Wu
⎛⎝⎜
⎞⎠⎟≈ Wv
Wu
−1 . 293
If Wv >Wu , then 294
pu ,v =fv − fufuj − fu( )∑
295
is the probability for v to substitute u , where uj are the neighbors of u 296 with higher fitness than u [17]. 297 298 -Derivation of Equal Probability Model (EPM): 299 According to the EPM model, the probabilities are equal for all beneficial 300 mutations, so that one needs the fitness graphs only for computing the 301 probabilities. The matrix entry M ( f )u ,v is 1/ N if genotype u has N outgoing 302 arrows and there is an arrow from u to v . 303 304 CPM is accurate if fitness differences between genotypes are small. EPM 305 may provide better estimates if fitness differences are substantial. Indeed, if the 306 fitness effects of all available beneficial mutants exceed some threshold, then 307 fixation probabilities are independent of fitness values [21]. We applied both 308 CPM and EPM, since no complete theory for substitution probabilities exists. 309 Additionally, comparison of two models is useful in learning how sensitive our 310 results are for variation in substitution probabilities. 311 312 -Optimal antibiotic sequences and pathways of genotypes 313 Let M[d] denote the 16 ×16 transition matrix we derived for the antibiotic labeled 314 a . For any sequence a1,a2,...ak of k antibiotics, we consider the matrix product 315 M[a1]M[a2 ]...M[ak ] . This product is also a 16 ×16 transition matrix. Its entry in 316 row a and column b is the fixation probability of genotype u mutating to genotype 317 v under the antibiotic sequence a1,a2,...ak . That probability is a sum of products 318 of entries in the individual matrices M f ai( )( ) , with one sum for each possible 319 pathway of genotypes from u to v . Our optimization algorithm enumerates all 320 15k antibiotic sequences of length k , and it selects all sequences that maximize 321
9
the entry in row a and column b of the matrix product. In a subsequent step we 322 then analyze these optimal antibiotic sequences, and for each such sequence, 323 we extract the full list of genotype pathways that contribute. 324 325 We implemented this algorithm in the computer algebra software Maple, and we 326 ran it for k = 2,3,4,5,6 . The running time of the program is slow because of the 327 exponential growth in the number of sequences. At present we do not know 328 whether an efficient algorithm exists for solving our optimization problem for 329 larger values of k . 330 331 Figure Legends 332 Figures 1-15 333 These figures present a visual summary of the adaptive landscape 2x2x2x2 334 tensors in which each resistance phenotype conferred by each TEM genotype is 335 enumerated. Arrows pointing upward represent addition of a mutation. Arrows 336 pointing downward represent reversions. Red arrows indicate significance 337 between adjacent growth rates as determined by one way ANOVA. Genotypes 338 that confer the most resistance to each antibiotic are shown in red. 339 340 Figure 16 341 Summary of CPM Substitutions with the Highest Probabilities. Each arrow is 342 labeled by the drug or drugs corresponding to the maximal transition probability, 343 taken over all 15 drugs. Each arrow is also labeled by the maximal probability. 344 345 From the graph, it is possible to find candidate, un-optomized treatment plans. 346 For example, when starting at genotype 1010 the graph shows that the 347 probability for ending at 0000 is 0.71for the sequence ZOX-TZP (0.71 is the 348 product of the arrow labels). Similarly, when starting at 1111 the probability for 349 ending at 0000 is 0.62 for the sequence CEC-CAZ-TZP-AM. When starting at 350 0001 the graphs shows that a single drug gives probability at most 0.29, whereas 351 the probability for ending at 0000 for the sequence AMC-CRO-AM (one arrow up, 352 two arrows down) is at least. 353 354 This graph can also be used to generate circular paths. For example, from a 355 starting point of 0000, the probability for ending at 0000 is 0.62 for the sequence 356 CEC-SAM-AMP-FEP-CPR-CAZ-TZP-AM (4 substitutions and 4 reversions). 357 358 Figure 17 359 Summary of Optimal 6 Step CPM and EPM Treatment Paths. Black arrows show 360 transitions present in six step paths computed using both the CPM and the EPM. 361 Red arrows signify transitions found only in optimum paths computed using the 362 CPM whereas blue signify transitions only found using the EPM. 363 364 Figure 18 365
10
Summary of Optimal CPM 2, 4, and 6 Step Antibiotic Cycles. Two step cycles 366 are shown in red. Four and six step cycles are shown in blue. Four and Six Step 367 cycles differ only in the number of steps, but the substitutions used within them 368 and the probabilities are identical. 369
11
Table 1 370 Number of amino acid substitutions
Number of TEM genotypes
1 53 2 53 3 37 4 31 5 10 6 2 7 2 8 0 9 0
10 1 11 1
371
12
Table 2 Variant Genotypes Created, Binary Codes and Names of 372 Genotypes Identified in Clinical Isolates 373 Number of Substitutions