Host

To test if the changes in λ’s fitness landscape facilitated its evolution to infect via OmpF, we simulated λ’s evolution on the two landscapes we measured and recorded whether OmpF+ genotypes arose to a high enough frequency that we would have detected them in the original laboratory evolution. We initiated λ populations with no mutations and allowed them to evolve and mutate at any of the 10 focal sites used to construct the landscapes (Supplementary file 1a). For each treatment (Figure 2), we simulated 300 separate trials using a modified Wright-Fisher model with discrete generations and a fixed population size.

For each simulation, a new fitness landscape was constructed by assigning fitness values to all the genotypes (210=1024 genotypes). This was done in a way to account for error in estimating fitness and the error associated with imputing the values of missing data points. For the genotypes that had empirical fitness data available (Figure 1a and b), we did not simply average the values of the replicate fitness measurements but instead we performed a bootstrapping protocol in order to account for the error in the fitness estimate. We randomly resampled from the four replicate measurements (with replacement) and then computed the mean of the four. Some genotypes were not present in all four replicates, in these cases we resampled as many times as there were replicates. Genotypes with only one replicate data point were, thus, assigned the same corresponding fitness values in all the runs of the simulations. Next, we imputed fitness of the genotypes that were missing from the empirical landscape. A genotype with a missing fitness value was randomly chosen and assigned the mean of the fitness values of its nearest neighbors (one-mutation away genotypes) present in the landscape. This was iterated until the full genotypic space was complete. Note that the order in which genotypes are chosen can affect the value that is estimated for a given genotype. This is because as the landscape is filled in, each genotype will have more neighbors to draw values from. This means that if a genotype is randomly chosen early, its fitness will be based on fewer neighbors than if it were chosen later, and its value will be slightly different. Since missing genotypes are randomly chosen in each iteration, the order will vary from one simulation to the next. This method introduces an extra source of variation in the simulation runs and captures uncertainty associated with the imputation of fitness values.

After constructing a complete fitness landscape, we evolved a λ population through repeated cycles of reproduction, selection, and mutation. For each generation, reproduction in the population was simulated by a multinomial sampling where the number of trials was equal to the population size N (set to ~6.3 × 109 based on Figure 4—figure supplement 2) and the success probability associated with a genotype i was given by pi=niwi/∑iniwi , where ni is the abundance of the genotype i, and wi is defined as the exponential of selection rate used in fitness landscape. Thus, the probability of (k1,k2,k3,…km) offsprings for genotypes 1,2,3,..., m would be:

To incorporate mutations, all genotypes whose frequencies increased were mutated as per λ’s mutation rate (7.7 × 10–8 per base per replication Drake, 1991). Consider a genotype (say i) that increased in abundance and let the number of additional individuals produced by this genotype be denoted by zi (with zi=ki-ni). Each of these zi individuals retained its parent’s genotype with a probability e-μ (assuming a Poisson distribution for mutations). Otherwise, the individual was assigned to a random neighboring (i.e. mutant) genotype. Given λ’s mutation rate, the probability of multiple mutations is very small and was ignored here. We simulated this modified Wright-Fisher cycle of reproduction and mutation for 960 generations. It is difficult to know how many generations phages undergo because the evolved phages have a high spontaneous death rate (Petrie et al., 2018) and can also experience other sources of mortality. Given this, we decided to run the simulation for a somewhat arbitrary amount of time, 960 generations which corresponds to two doublings per hour for the 20-day experiment we are trying to replicate Meyer et al., 2012. This is likely an overestimate, which is unintuitively conservative for our purposes because more cycles will cause additional evolution and exploration of the fitness landscape, enhancing the possibility of OmpF+ evolution in the negative controls, and reducing our ability to detect treatment differences.

We evaluated whether λ evolved OmpF+ by examining the population for genotypes that had the necessary mutations to be OmpF+. Previous studies revealed that OmpF genotypes must have four mutations (Maddamsetti et al., 2018). They all possess two specific changes (A3034G) and (G3319A) and a third change that can occur at positions 3320 or 3321. For the 10 mutations we studied, T3321A is the only third mutation that satisfies this requirement, meaning all OmpF genotypes must have three specific changes (A3034G, G3319A, and T3321A). Many J mutations satisfy the last requirement (Maddamsetti et al., 2018), so we implemented what we call the ‘3+1’ rule where genotypes are designated as OmpF+ if they have the three necessary J mutations plus any additional mutation (see Supplementary file 1i). If any such genotype crossed the threshold of 5000 λ particles during the course of a simulation, the λ population in that run was marked to have evolved OmpF-function. We based the threshold value on the detection limit of OmpF+ genotypes in the original laboratory coevolution experiments (~500 pfu/ml, see ‘Coevolutionary Replay Experiments’ section).

We implemented host switching by first evolving λ on the landscape measured with ancestral E. coli, stopping the simulation early, and quantifying the frequency of each λ genotype. Next, we would initiate evolution on the malT— for the remainder of the time but starting with genotypes at the frequencies recorded on the ancestral landscape. A total of 11 different coevolution treatments were run by varying how long λ evolved on the ancestral host landscape and malT— landscape (Figure 2). 300 simulation trials were run for each treatment. We calculated error in the simulations in order to detect significant treatment differences by batching the runs into 30 and estimating the frequency of OmpF+ for each subset. This allowed us to calculate a 95% confidence interval using a Student’s t distribution for the frequency of OmpF+ evolution in each treatment, and then to use an ANOVA coupled with Tukey’s multiple comparison test to test for treatment differences.