Evolution of satellite plasmids can stabilize the maintenance of newly acquired accessory genes in bacteria

Plasmids play a principal role in the spread of antibiotic resistance and other traits by horizontal gene transfer in bacteria. However, newly acquired plasmids generally impose a fitness burden on a cell, and they are lost from a population rapidly if there is not selection to maintain a unique function encoded on the plasmid. Mutations that ameliorate this fitness cost can sometimes eventually stabilize a plasmid in a new host, but they typically do so by inactivating some of its novel accessory genes. In this study, we identified an additional evolutionary pathway that can prolong the maintenance of newly acquired genes encoded on a plasmid. We discovered that propagation of an RSF1010-based IncQ plasmid in Escherichia coli often generated ‘satellite plasmids’ with spontaneous deletions of accessory genes and genes required for plasmid replication. These smaller plasmid variants are nonautonomous genetic parasites. Their presence in a cell drives down the copy number of full-length plasmids, which reduces the burden from the accessory genes without eliminating them entirely. The evolution of satellite plasmids may be favored relative to other plasmid fates because they give a more immediate fitness advantage to a cell’s progeny and because the organization of IncQ plasmids makes them particularly prone to certain deletions during replication. Satellite plasmids also evolved in Snodgrassella alvi colonizing the honey bee gut, suggesting that this mechanism may broadly contribute to the importance of IncQ plasmids as agents of bacterial gene transfer in nature. Significance Statement Plasmids are multicopy DNA elements found in bacteria that replicate independently of a cell’s chromosome. The spread of plasmids carrying antibiotic-resistance genes to new bacterial pathogens is a challenge for treating life-threatening infections. Often plasmids or their accessory genes encoding unique functions are lost soon after transfer into a new cell because they impose a fitness burden. We report that a family of transmissible plasmids can rapidly evolve ‘satellite plasmids’ that replicate as genetic parasites of the original plasmid. Satellite plasmid formation reduces the burden from the newly acquired genes, which may enable them to survive intact for longer after transfer into a new cell and thereby contribute to the spread of antibiotic resistance and other traits within bacterial populations.

in a cell drives down the copy number of full-length plasmids, which reduces the burden from 23 the accessory genes without eliminating them entirely. The evolution of satellite plasmids may 24 be favored relative to other plasmid fates because they give a more immediate fitness advantage 25 to a cell's progeny and because the organization of IncQ plasmids makes them particularly prone 26 to certain deletions during replication. Satellite plasmids also evolved in Snodgrassella alvi 27 colonizing the honey bee gut, suggesting that this mechanism may broadly contribute to the 28 importance of IncQ plasmids as agents of bacterial gene transfer in nature. 29 Introduction 9 plasmid were missing in these strains. For the ancestral plasmid, AvaI digestion of the linearized 153 plasmid yields four fragments, one of which (F1a) is too small to visualize (Fig. 1A). Compared 154 to the ancestral plasmid, all three of these evolved plasmids kept the oriV bearing fragment 155 (F1b), but the other two visible fragments (F2, F3) were either truncated or lost (Fig. 1E). 156 Sequencing these PCR products allowed us to precisely localize the missing plasmid pieces. 157 A region beginning within repA or repC, including the gfp and aadA genes, and ending before 158 oriV was deleted from each plasmid (Fig. 1A). The repA and repC genes encode the DNA 159 helicase and iteron-binding protein, respectively, which are required for initiation of plasmid 160 replication (27). Loss of either function should lead to a plasmid that is unable to replicate on its 161 own. As expected, attempts to transform the shorter plasmid variants into new E. coli cells were 162 unsuccessful. The PCR assay will preferentially amplify these smaller plasmids when they are 163 present in a mixture, but weak bands corresponding to the linearized full-length pQGS plasmid 164 are visible in the same sample in some cases (Fig. 1D). Since the qPCR results also show that 165 copies of the gfp gene are still present in all three of these strains (Fig. 1C), we conclude that the 166 smaller plasmids are satellite plasmids that co-exist with and require copies of the full-length 167 plasmid in the same cell. From the perspective of the original plasmid, these satellite plasmids 168 are parasites. They take advantage of its replication machinery without encoding all of the 169 components needed for autonomous replication or any of the accessory genes. 170 We next examined how prevalent satellite plasmid (SP) evolution was in our experiment. 171 Using the PCR assay, we detected SPs in all 6 populations on day 5 (Fig. S1). SP formation was 172 so widespread that multiple different SPs were even observed in the same population in two 173 cases (B1 and B4). Sequencing the linearized PCR products from clonal isolates carrying each of 174 these new SPs revealed that they were similar to the initially characterized plasmids. All had 10 deletions that completely eliminated the gfp and aadA genes, overlapped one or more rep genes, 176 and left the oriV sequence intact (Fig. S2). The deletions leading to SP formation were nearly 177 always flanked by short, near-perfect sequence homologies with lengths of 7 to 15 base pairs 178 (Table S1), suggesting that SPs may form through RecA-independent processes (38). 179 180 Accessory Genes are Eventually Lost or Integrated into the Chromosome 181 To understand the role of satellite plasmids in the maintenance of accessory genes, we further 182 examined the time courses of plasmid evolution in the experimental populations. We found 183 plasmids with deletions leading to 3.5-6.0 kb PCR amplicons had evolved in all six populations 184 by day 5, persisted through day 9 in all but one population, and were still present on day 15 in 185 three of the six populations ( Fig. 2A, B and Fig. S3A). Sequencing these PCR products showed 186 that they were all nonautonomous SPs with deletions overlapping replication and accessory 187 genes (Fig. S1). In four populations (B3, B4, B5, B6), larger PCR amplicons with sizes of 7.0-188 8.0 kb emerged at later time points as satellite plasmids disappeared. These amplicons 189 corresponded to various plasmids with accessory gene deletions that included loss of the entirety 190 of lacI gene and a portion of gfp in every case (Fig. S2). Complete sets of replication proteins 191 and the aadA gene were retained in these plasmids, meaning that they remained fully 192 autonomous. In the other two populations (B1 and B2), SPs disappeared, and no plasmid 193 amplicon of any type was detectable on day 21 and day 25 ( Fig. 2A and Fig. S3A). 194 For two representative populations of each type (B2 and B6), we correlated flow cytometry 195 data (Fig. 2C, D) with whole-genome sequencing (WGS) data (Fig. 2E, F) to characterize 196 plasmid evolution in a way that was not subject to PCR amplification biases. In addition to better 197 defining evolutionary dynamics involving satellite and deletion plasmids, these data revealed an 198 11 additional plasmid fate (Fig. 2G). In both populations, WGS reads consistent with an insertion of 199 an IS186 element into the pQGS plasmid repF gene were detected by day 6 (Fig. 2E, F). IS186 200 is one of the most active insertion sequences in E. coli. It inserted into pQGS at a site matching 201 its preferred target site of 5′-GGGG(N6/N7)CCCC (39-41). This IS insertion likely creates a 202 second class of satellite plasmid that is unable to replicate autonomously because it separates the 203 repA and repC genes from their promoter which is located upstream of repE gene (27). 204 At the same time we detected the IS186 insertions, there was a precipitous drop in the copy 205 number of the plasmid origin of replication (oriV) in both populations (Fig. 2E, F), and 206 subpopulations of cells with greatly reduced fluorescence began to dominate (Fig. 2C, D). These 207 results are consistent with integration of the IS186-disrupted plasmid sequences into the E. coli 208 chromosome, either concomitant with IS186 insertion or shortly thereafter, through single-209 crossover recombination with one of the three original copies of IS186 (Fig. 2G). We were able 210 to detect hybrid PCR products when using one chromosomal and one plasmid primer, indicating 211 that this was the case. The presence of multiple peaks with intermediate fluorescence 212 intensities-at the same time in population B6 (Fig. 2D) and later and to a lesser extent in 213 population B2 (Fig. 2C)-might indicate that the integrated plasmid sequence is sometimes 214 duplicated within the chromosome. This could occur through homologous recombination during 215 chromosomal replication, facilitated by the flanking IS186 copies (Fig. 2G). 216 We also tracked the prevalence of satellite plasmids from the numbers of WGS reads that 217 spanned deleted regions. SPs first appeared by day 4 in both populations (Fig. 2E, F). In 218 population B2, we could detect three different sizes of SPs. SP incidence as a percentage of all 219 plasmids in the population peaked at ~50% in B2 and ~20% in B6. In population B2, the copy 220 number estimated from the read-depth coverage of the accessory gene completely deleted in all 12 SPs (lacI) dropped noticeably more quickly than that of the origin (oriV), as expected if satellite 222 plasmids and ancestral plasmids coexisted in a sizable number of cells in the population. SPs 223 dropped below the detection limit by WGS at day 9 and 7 in populations B2 and B6, 224 respectively. At this point the IS-mediated integrants dominated. Completely nonfluorescent 225 cells became dominant in population B6 near the end of the experiment (Fig. 2D). A new 226 plasmid variant with a deletion of lacI and a portion of gfp was detected at day 19 by WGS in 227 this population (Fig. 2F). According to the flow cytometry data, the prevalence of cells 228 containing only the deletion plasmid increased to 47% at day 21 and 66% at day 25. 229 Overall, these evolutionary trajectories suggest that satellite plasmid evolution was beneficial 230 to the fitness of a cell because it curtailed the expression of gfp and/or lacI. These results further 231 suggest that IS186-mediated integration of the plasmid into the chromosome or deletion of these 232 accessory genes from an evolved plasmid that remained capable of self-replication conferred 233 greater fitness benefits than SP formation. Flow cytometry time courses show that these two 234 fates eventually dominated in the other four populations of the evolution experiment (Fig. S3B). 235 To directly test the fitness consequences of evolving the different plasmid fates, we 236 reconstructed cells of each type and competed them against a reference strain of E. coli 237 BW25113 with the spectinomycin resistance gene (aadA) integrated into its chromosome and a 238 restored ability to utilize arabinose as a neutral genetic marker. First, we isolated satellite 239 plasmids from each population (Sat2 and Sat6) and retransformed them into cells harboring the 240 ancestral plasmid. Next, we isolated the deletion plasmid from population B6 (Del6) and 241 transformed it into the ancestral strain with no plasmid. Finally, we used transduction to move an 242 integrated copy of the plasmid from an evolved cell in population B2 to the ancestral strain (Int2) 243 so that its fitness could be measured without interference from any other evolved alleles. 244 13 As a control, we competed the plasmid-free ancestor strain BW25113 against the reference 245 strain in media that did not contain any antibiotic. The fitness of the ancestral strain (No plasmid) 246 was not significantly different from that of the reference strain (p = 0.768, two-tailed t-test). The 247 ancestral strain/plasmid combination of BW25113/pQGS (Anc) had a greatly reduced fitness of 248 0.431 relative to the reference strain with no plasmid. The addition of a satellite plasmid to a cell 249 significantly alleviated this cost, with a relative fitness of 0.781 and 0.840 for Sat2 and Sat6, 250 respectively (Fig. 2H). The cases of integration into the chromosome (Int2) or deletion of the 251 accessory genes from the plasmid that we tested (Del6) restored fitness even more toward that of 252 cells with no plasmid, to relative fitnesses of 0.923 and 0.910, respectively. As predicted by the 253 coexistence of cells with deletion plasmids and cells with chromosomally integrated plasmids in 254 population B6, the fitness values of these evolved plasmid fates were not significantly different 255 (p = 0.474, two tailed t-test). All strains with evolved plasmid variants were significantly more fit 256 than the ancestor with pQGS (p ≤ 3.6 ´ 10 -14 , Bonferroni-corrected two-tailed t-tests) and less fit 257 than the reference strain with no plasmid (p ≤ 0.011, Bonferroni corrected two-tailed t-tests). 258 259

Satellite Plasmids Evolve in the Honey Bee Gut 260
We next used Snodgrassella alvi wkB2, a core gut symbiont of honeybees (Apis mellifera) to 261 test whether SP evolution from plasmid pQGS would occur in a different bacterial species and in 262 a complex environment associated with an animal host. S. alvi is a β-proteobacterium that has 263 only recently been cultured (42). We first used the PCR assay to check for the evolution of SPs 264 from pQGS within five populations of S. alvi wkB2 (designated S1-S5), during serial passage in 265 vitro (Fig. 3A). By sequencing the PCR products, we determined that two identical bands 266 detected at day 8 in all populations (pS621 and pS521) corresponded to SPs (Fig. S2). Bands that 267 14 likely correspond to smaller SPs (e.g., pS631) became dominant by day 12 in all populations. 268 Larger amplicons that may be accessory-gene deletion plasmids were also observed in two 269 populations at this point. 270 Next, we examined whether SPs would evolve and persist in S. alvi wkB2 carrying pQGS 271 when they colonized microbiota-free bees (43, 44). As expected from previous studies (45), S. 272 alvi with the pQGS plasmid began to colonize the inner wall of the ileum by 24 h post-273 inoculation as visualized by GFP fluorescence in the bee gut (Fig. 3B). We further confirmed 274 colonization by measuring colony-forming units in guts isolated from sacrificed bees. On day 3 275 and day 4, there were approximately 10 7 and 10 8 CFUs per bee gut, respectively. Previous 276 studies have showed that S. alvi fully colonizes microbiota-free bees by 4 to 6 days post-277 inoculation (46). Therefore, we used PCR to check for SP formation in bees sacrificed every 24 h 278 during the first 4 days of colonization from ten different enclosures (Fig. 3C, Fig. S4). Three SPs 279 (pC562, pC765, pC1025) were identified from bees reared in different enclosures 3 to 4 days 280 after they were inoculated. No SP was detected in bees sacrificed 8 and 12 days after inoculation, 281 suggesting that SPs may be a transient evolutionary intermediate in S. alvi colonizing the honey 282 bee gut as they were in E. coli and S. alvi passaged in vitro. 283 284

Multiple Factors Favor Satellite Plasmid Evolution Over Accessory Gene Deletion 285
Mutations that eliminate different regions of the pQGS plasmid can result in nonautonomous 286 satellite plasmids or in autonomous plasmids that have deleted just the accessory genes. We 287 found that E. coli cells with accessory-gene deletion plasmids (DPs) have a much higher fitness 288 than cells containing a mixture of satellite and ancestor plasmids. If both types of cells arise at 289 similar rates in a bacterial population, then cells with DPs are expected to dominate to such an 290 extent that SPs should never reach an observable frequency, but this was not the case in our 291 experiments. To investigate this discrepancy, we examined ways in which the genetic 292 organization of the pQGS plasmid and its multicopy nature might favor satellite plasmid 293 evolution and thereby maintenance of the newly acquired accessory genes in a population. 294 First, there might be a mutational bias favoring the evolution of satellite plasmids. DPs must 295 maintain the entire plasmid from oriV to aadA intact so that they remain autonomous and still 296 encode antibiotic resistance. If we also assume that they must delete at least a portion of the both 297 the lacI and gfp coding sequences to realize their full fitness advantage, then this constrains these 298 deletions to relatively few combinations of starting and ending base positions. On the other hand, 299 SPs must only maintain oriV while deleting the accessory genes plus aadA and least one of the 300 replication genes. Given just the raw numbers of start-end coordinate combinations fitting each 301 of these scenarios (Fig. 4A), SPs would be expected to arise at 4.0 times the rate of DPs. 302 All ten unique SPs we characterized resulted from deletions that had DNA sequence 303 microhomologies consisting of near-exact 7 to 15 base-pair repeats overlapping their endpoints 304 in the ancestral plasmid (Table S1). We investigated whether there was any bias in the 305 abundance of similar microhomologies in different regions of the plasmid that might additionally 306 favor deletions leading to SPs versus DPs (see Materials and Methods). Over a range of 307 different microhomology lengths, the number of near-repeats that could mediate the formation of 308 each type of evolved plasmid did not noticeably deviate from the ratio of 4:1 expected if they 309 were randomly distributed in the plasmid (Fig. 4B). Surprisingly, all four DPs did not have any 310 microhomology near their endpoints (Table S1), despite the fact that near-exact repeats exist in 311 these regions. This striking contrast may indicate that the mechanisms of deletions in different 312 regions of the pQGS plasmid may vary due to how IncQ plasmids replicate (see Discussion). 313 Mutant plasmids start as a single copy within one cell, so the dynamics of plasmid replication 314 and segregation affect how much that cell and its progeny benefit from the mutant plasmid and 315 the chances that copies of the mutant plasmid will survive in the cell population (47). The lag 316 between when a mutant plasmid first appears in a cell and when that cell's descendants realize 317 the full fitness benefit possible from harboring multiple copies of that plasmid is known as its 318 phenotypic delay (48). We examined phenotypic delay using multilevel stochastic simulations 319 with parameters fit to be consistent with our E. coli evolution experiment (see Materials and 320 Methods). If SPs replicated more quickly than other plasmid types due to their smaller size, then 321 they would be expected to have a much shorter phenotypic delay. However, models in which we 322 gave SPs an advantage for within-cell replication were not consistent with the experimentally 323 observed numbers of ancestral and satellite plasmids co-existing within cells in our experiment 324 and/or the relative fitness of different cell types. Therefore, we gave all plasmids equal rates 325 (chances) of replication in our simulations, which is consistent with a model of plasmid 326 replication in which initiation, rather than elongation or other steps, is rate-limiting. 327 These simulations predict that SP cells will initially fare better in terms of their fitness before 328 DP cells surpass them (Fig. 4C). Because a one-to-one trade of a copy of the ancestral plasmid 329 for a SP is more beneficial than replacing it with a DP, the model predicts that if a cell with a SP 330 and a cell with a DP evolved at the same time, the descendants of the SP cell would be more fit, 331 on average, than those of the DP cell for the first 12 generations. This reduced phenotypic delay 332 is expected to give a newly evolved SP a better chance of avoiding stochastic loss. Rep) that is needed to initiate replication outside of the origin sequences to which it binds. Thus, 359 these plasmids could theoretically evolve or be engineered into satellite plasmids, as long as SPs 360 and intact full-length plasmids are able to stably co-exist within cells and their progeny. 361 IncQ plasmids replicate by a strand displacement mechanism and do not encode a 362 partitioning system, so they randomly segregate into daughter cells (24). RepC has been reported 363 to positively regulate copy number (49), while MobC and MobA suppress it (50). Despite these 364 regulatory feedbacks, we found that total plasmid copy number was largely unchanged in cells 365 that evolved SPs that inactivated repC in the E. coli experiments. We also observed SPs with a 366 large range of sizes that deleted a variety of combinations of replication and mobilization genes, 367 indicating that many different deletions are compatible with stable maintenance of these satellite 368 plasmids. Our modeling indicates that smaller pQGS-derived SPs do not have a significant 369 advantage over full-length plasmids in terms of their replication rate within cells. This could be 370 another reason that it is possible for a cell to maintain enough full-length plasmids such that there 371 is not too great a fitness cost for SP evolution due to some of a cell's progeny inheriting only 372 nonautonomous plasmids and thereby losing plasmid function entirely. It is possible that 373 plasmids from other families, which commonly utilize theta or rolling-circle replication 374 mechanisms, are subject to regulatory controls that make the presence of derived satellite 375 plasmids more disruptive to their continued replication and segregation. 376 The rate at which new satellite plasmids are generated during plasmid replication is another 377 important factor in whether they will be observed relative to other plasmid fates. We investigated 378 whether there were aspects of SP organization and population dynamics that could favor their 379 evolution in our experiments, especially relative to plasmids that delete burdensome accessory 380 genes but maintain the ability to self-replicate. We found that there are more opportunities for 381 deletions that generate SPs (i.e., mutations that generate them are expected to be more common) 382 and that they are expected to initially give greater fitness benefits to the host cell and its progeny 383 (i.e., they have a reduced phenotypic delay). However, their rather slight advantages in these 384 areas are not enough on their own to explain the prevalence of SPs in our experiments. 385 Our observation that the deletions which create satellite plasmids are flanked by sequence 386 microhomologies but those that generate accessory-gene deletion plasmids are not, suggests that 387 the unusual strand displacement mechanism of IncQ plasmid replication may be a key reason 388 that SPs evolve in our experiments. This mechanism begins with RepC binding to the iteron in 389 RecA-independent recombination mediated by annealing of short sequence homologies (54-56). 397 We hypothesize that the presence of large amounts of ssDNA in cells with the pQGS plasmid 398 may explain why SPs evolve so quickly. Deletions that remove only the accessory genes may 399 occur at a much lower rate and through alternative mechanisms because both of their endpoints 400 are closer together and located on the same arm of the plasmid relative to oriV. Similarly, 401 plasmids with theta or rolling-circle replication mechanisms may only rarely evolve SPs. 402 Formation of SPs is one mechanism that can lead to persistence of accessory genes after they 403 are transferred to a new bacterial host. Another is that the plasmid can become temporarily or 404 permanently single-copy and nonautonomous by integrating into the bacterial chromosome. 405 20 Integration was a common endpoint in our E. coli evolution experiment. In the case that we fully 406 characterized, this process occurred via insertion of an IS186 copy into the plasmid backbone 407 and recombination with a copy of this transposable element in the genome. Similar integration 408 events have also been observed in evolution experiments with a different family of conjugative 409 plasmids in Pseudomonas putida (10). In another instance, reversible chromosomal integration 410 mediated by IS elements was reported for the virulence plasmid pINV in Shigella (57). 411 We have shown that a transmissible plasmid possesses an organization and replication 412 mechanism that can favor the rapid evolution of satellite plasmids that reduce the copy number 413 of any costly cargo genes that it has spread to a new cell. Awareness of the propensity of IncQ 414 plasmids to evolve SPs is an important consideration when studying the ecological range of this 415 broad-host-range plasmid and when using it to engineer cells. In this latter context, SP evolution 416 could be interpreted as a fortuitous "safety valve" for titrating down toxic or burdensome gene 417 expression or as an unwanted complication that leads to a different effective copy number for 418 engineered DNA sequences than was intended (58). In summary, our results suggest that satellite 419 plasmids may be a common evolutionary intermediate that can stabilize the spread of plasmid-420 borne accessory genes, including those responsible for antibiotic resistance. replicate competition assays for a pair of strains as previously described (63). 511

Bee gut symbiont experiments 512
Plasmid pQGS was conjugated into bee gut symbiont S. alvi strain wkB2 as described 513 previously (58). For the in vitro evolution experiment, five populations were started from single 514 colonies that were inoculated into 6 ml of Insectagro DS2 medium (Corning) amended with 30 515 µg/ml Spec (DS2-Spec). S. alvi was cultivated in 6 ml of medium in culture tubes under 5% CO2 516 at 35°C without shaking. At 4-day intervals, 6 µl of each culture was transferred into 6 ml of 517 fresh media and checked for the presence of SPs by PCR and Sanger sequencing. 518 For the in vivo evolution experiment, microbiota-free bees were obtained as described by 519 Zheng et al. (44) with the following modifications. Late-stage pupae were removed manually 520 25 from brood frames and placed in sterile plastic bins. The pupae emerged in an incubator at 35°C 521 and humidity of 75%. Newly emerged bees were kept in cup cages with sterilized sucrose syrup 522 (0.5 M) and bee bread. For inoculating these bees, S. alvi wkB2 cells containing the pQGS 523 plasmid cultured on DS2-Spec agar were first suspended in 1×PBS to an OD600 of 1.0. Then, 524 batches of 20-25 bees were placed into a 50-ml conical tube, and 50 µl sucrose syrup was added. 525 The tube was rotated gently to coat the syrup on the surface of bees. Finally, 50 µl of the S. alvi 526 wkB2/pQGS suspension (~5×10 7 cells) was added to the tube, and it was rotated again. S. alvi 527 from the body surface colonizes the guts of these bees as a result of auto-and allogrooming. 528 Inoculated bees were reared in cup cages. Five of these enclosures were set up in each of two 529 different experiments with more than 12 bees in each cage. The inoculated bees were fed sucrose 530 syrup with spectinomycin (60 µg/ml) throughout the experiment. Bee guts were dissected 24 h 531 after the inoculation, and wkB2/pQGS colonization was examined by fluorescence microscopy. 532 To characterize SP evolution, one bee gut from each cup was dissected every day until 4 days 533 after inoculation. Colonization levels were determined at day 3 and day 4 as colony-forming 534 units (CFUs) present in dissected guts, as described by Kwong et al. (64). Bee guts were 535 homogenized in 728 µl of cetyltrimethylammonium bromide (CTAB) buffer with 20 µl 536 Proteinase K solution (20 mg/ml). DNA was then extracted using a bead-beating method as 537 previously described (65). SPs were detected by PCR and characterized by Sanger sequencing. 538

Repeat analysis 539
We created a Python script to enumerate microhomologies (short, near-perfect repeats) in 540 pQGS that could mediate deletions leading to satellite plasmid formation or accessory-gene loss. 541 We ran this code with settings that identified repeats that were at least 7 bases long, included at 542 most one inserted or deleted base (indel), had at most five total mismatches (including indels), 543 26 and with at least 75% sequence identity. When multiple valid alignments overlapped, only one 544 was counted, giving precedence to the longer one, and then to the one with fewer mismatches. 545

Plasmid evolution simulations 546
We implemented multilevel stochastic simulations that replicate cells and plasmids using 547 Monte Carlo methods in Python. We established parameters that fit the observed numbers and 548 types of plasmids typical for each cell type to its fitness by first assuming a linear cost for 549 expression of each protein from the plasmid (66, 67). With 18 total plasmids per cell, an additive 550 fitness model with a cost of 3.06% per full-length plasmid and 0.50% per deletion plasmid in a 551 cell matches the experimentally observed fitness values for cells containing solely ancestral or 552 solely deletion plasmids. A model in which satellite plasmids have no fitness cost and the 553 collection of plasmids in each cell is exactly doubled and randomly but evenly assorted to 554 daughter cells leads to an equilibrium level of 5.2 ancestral plasmids and 12.8 satellite plasmids 555 per cell, on average. This is close to the ratio of the two types observed for several evolved 556 strains. Note that instead of the expected relative fitness value of 0.841 for the cost of those 5.2 557 full-length plasmids, loss of all ancestral plasmids from some offspring due to segregation causes 558 this population of cells to realize a relative fitness of just 0.77, which also closely matches the 559 experiments. Alternative plasmid replication and segregation models that introduced more 560 complexity and randomness gave lower fitness values and/or prevalence of satellite plasmids in 561 cells than we observed. For example, we tested allowing plasmids to replicate via their own 562 stochastic growth process and unevenly segregating plasmids between daughter cells. 563 The stochastic simulations track a collection of cells of interest that may contain different 564 numbers of each plasmid type as they compete against a background of cells with a fixed fitness 565 value and the cell population is transferred via serial dilution. We used them in three ways: (1) 566

27
To determine the equilibrium fitness value of each cell type and equilibrium number of satellite 567 plasmids per cell, we tuned the fitness of the background population in the model until a 568 population seeded with 10,000 initial cells of the type of interest were able to maintain a stable 569 population size over 500 generations of growth with 2-fold serial dilutions. chance that a new cell with a satellite plasmid or deletion plasmid would successfully escape 577 dilution and establish under the conditions of our evolution experiment, we started 1,000,000 578 simulations, each with one cell containing one mutant plasmid that arose at a random cell 579 division within a daily growth cycle with a 2,000-fold transfer dilution. Then, we followed the 580 cell number of this population competing against a background of cells with only the ancestor 581 plasmid until cells with the mutant plasmid either went extinct or reached a population size that 582 was 20-fold higher than the dilution factor at the end of a growth cycle. The chance that a new 583 mutant will establish is number of events of the latter type divided by the total number of trials.