Mutualism-enhancing mutations dominate early adaptation in a microbial community

Species interactions drive evolution while evolution shapes these interactions. The resulting eco-evolutionary dynamics, their outcomes and their repeatability depend on how adaptive mutations available to community members affect fitness and ecologically relevant traits. However, the diversity of adaptive mutations is not well characterized, and we do not know how this diversity is affected by the ecological milieu. Here we use barcode lineage tracking to address this gap in a competitive mutualism between the yeast Saccharomyces cerevisiae and the alga Chlamydomonas reinhardtii. We find that yeast has access to many adaptive mutations with diverse ecological consequences, in particular, those that increase and reduce the yields of both species. The presence of the alga does not change which mutations are adaptive in yeast (i.e., there is no fitness trade-off for yeast between growing alone or with alga), but rather shifts selection to favor yeast mutants that increase the yields of both species and make the mutualism stronger. Thus, in the presence of the alga, adaptations contending for fixation in yeast are more likely to enhance the mutualism, even though cooperativity is not directly favored by natural selection in our system. Our results demonstrate that ecological interactions not only alter the trajectory of evolution but also dictate its repeatability; in particular, weak mutualisms can repeatably evolve to become stronger.


29
Ecological communities are often perturbed by environmental shifts 1,2 , demographic noise 3 and 30 species turnover 4,5 . Such perturbations can not only displace communities from their ecological 31 equilibria but also precipitate adaptive evolution [6][7][8][9] . Evolutionary changes within one species can 32 be rapid and can alter its ecological interactions with other community members, which can 33 cause further evolution [8][9][10] . Although such eco-evolutionary feedbacks appear to be 34 widespread 7,[9][10][11][12][13][14][15][16][17][18][19][20][21][22][23][24][25] , the population genetic mechanisms that underlie them are not well 35 understood 26 . For example, how does the spectrum of adaptive mutations available to a species 36 (i.e. the genomic locations and fitness effects of mutations that provide fitness benefits) depend 37 on the composition of the surrounding community? In particular, how does it change when a 38 species is lost from the community or a new species invades it? How many and which of the 39 adaptive mutations available to a species affect its interactions with the rest of the community? 40 Which of these mutations are likely to spread and fix? And thus, how diverse and repeatable are 41 the ecological outcomes of evolution? 42 Empirical data supporting answers to these questions would help us develop a better theoretical 43 understanding of eco-evolutionary dynamics. For example, many existing models assume that 44 any combination of traits can be produced by mutations so that the eco-evolutionary trajectories 45 and outcome are determined exclusively by natural selection 27 . However, recent evidence 46 suggests that the availability of mutations can significantly impact evolution 28-32 . Yet, we know 47 very little about the distributions of ecological and fitness effects of new mutations in multi-48 species communities and how these distributions shift when the ecological milieu changes-for 49 example, due to the addition or extinction of community members. 50 Here, we address this gap in one of the simplest experimentally tractable microbial communities. 51 Our community consists of two species, the alga Chlamydomonas reinhardtii and the yeast 52 Saccharomyces cerevisiae, that interact in our environment via competition and mutualism 33 . 53 Although communities in nature often contain more members, understanding eco-evolutionary 54 dynamics in simple model communities is helpful for developing an intuition and expectations 55 for the behaviors of more complex ecosystems 34-36 . We measure how adaptive mutations arising 56 in one member of our community, the yeast, affect its competitive fitness (a metric that 57 determines the evolutionary success of a mutant lineage), the absolute abundances of both 58 species in the community (a metric that informs us about the type of interactions between species 59 and the stability of the community) as well as basic life-history traits of yeast (growth rates and 60 carrying capacities) that contribute to both fitness and abundances. We specifically ask whether 61 and how the statistical distribution of effects of adaptive mutations in yeast are altered by the 62 presence/absence of the alga. To this end, we use the barcode lineage tracking (BLT) 63 technology 37,38 to isolate hundreds of adaptive mutations arising in yeast when it evolves alone or 64 in community with the alga. Our data offer us a detailed view on how inter-species interactions 65 affect the evolutionary dynamics of new mutations, and how these mutations in turn alter the 66 ecology of our community. 67

68
Yeast and alga form a facultative competitive mutualism 69 In a previous study, Hom and Murray showed that in a sealed environment in which nitrite is 70 provided as a sole source of nitrogen and glucose is provided as a sole source of organic carbon, 71 the yeast Saccharomyces cerevisiae and the alga Chlamydomonas reinhardtii spontaneously 72 form an obligate mutualism 33 . Under such conditions, C. reinhardtii consumes nitrite and 73 produces ammonium that is secreted and utilized by S. cerevisiae, which consumes glucose and 74 produces CO2 that is in turn utilized by the alga. When the environment is opened to ambient gas 75 exchange and ammonium is supplied in the medium, both the yeast and the alga can survive 76 without each other. In this study, we grow the yeast and the alga alone and together over multiple 77 5-day growth and dilution cycles in well-mixed and well-lit conditions, open to gas exchange, in 78 a medium supplemented with 0.5 mM of ammonium (see Methods for details). 79 Time-course cell-density measurements of the wild-type yeast and alga over a single cycle 80 confirm that both species grow significantly differently in each other's presence than alone 81 ( Figure 1, Extended Data Figure 1, Data S1, repeated-measures ANOVA P = 10 -4 for the yeast 82 and P = 5×10 -8 for the alga), indicating that they ecologically interact in our experimental 83 environment. Specifically, the alga achieves higher densities over the entire growth cycle in the 84 community compared to growth alone. The presence of the alga alters yeast growth dynamics in 85 a more complex way. When yeast grows alone, it reaches peak cell density at 48 hours, after 86 which point its density gradually declines ( Figure 1A), suggesting that it exhausts the initial 87 supply of ammonium after about 48 hours. When yeast grows in community with the alga, the 88 two species initially compete, likely for ammonium. This can be seen by a reduction of the peak 89 yeast cell density at 48 hours ( Figure 1A). Upon the depletion of supplemented ammonium, the 90 alga subsequently reduces nitrite to ammonium that it then secretes 33 . This nitrogen provisioning 91 by the alga reduces the yeast's rate of population decline between days 3 and 4 (t-test P = 5×10 -92 4 , Extended Data Figure 1, Table S1). As a result, yeast reaches approximately the same density 93 by the end of the cycle in the community as it does alone, despite having a lower peak density on 94 day 2. Thus, in the latter portion of the growth cycle, the yeast experiences a benefit from its 95 interaction with the alga. Since yeast and alga initially compete and later cooperate in our 96 conditions, we refer to our system as a competitive mutualism 39 . 97 Although the ecological interactions in our system are quite complex, it would be convenient to 98 quantify them with some simple summary statistics. We can do so, for example, by comparing 99 the final densities that each species achieves at the end of the 5-day cycle when growing in 100 community versus alone ( Figure 1). We refer to these final densities as "yields". Specifically, we 101 compute the ratio of yeast yield in community (YYC) to its yield alone (YYA) and the ratio of 102 alga yield in community (AYC) to its yield alone (AYA). Both ratios exceeding unity indicate 103 that cooperation is on the whole more important than competition. Conversely, when both ratios 104 are less than one, competition is more important on the whole than cooperation. For our wildtype 105 community, we find that YYC to YYA ratio is not significantly different from one, while the 106 AYC to AYA ratio equals 3.00 (95% Confidence Interval (CI) [2.64, 3.36], Student's t-test t = 107 13.24, df=7.12, P = 3×10 -6 ; Figure 1). This indicates that, on the whole, the alga benefits from its 108 interactions with the yeast, while the yeast, on the whole, neither benefits nor suffers from its 109 interactions with the alga. Thus, according to this metric, yeast and alga form a net commensal 110 relationship. However, we emphasize that this net commensalism is a result of a balance between 111 the underlying competitive and cooperative interactions. 112 The fact that the yeast and the alga can grow in our conditions either alone or together as a 113 community allows us to inquire how evolution of one species is affected by its ecological 114 interactions with the other, under otherwise identical environmental conditions. We focus on the 115 initial phase of adaptive evolution. Since the yeast is likely to adapt faster than the alga (see 116 Supplementary Information), we characterize the distribution of ecological and fitness effects of 117 adaptive mutations arising in yeast and examine how these distributions depend on the 118 presence/absence of the alga. 119 The presence of the alga alters the fitness effects of beneficial 120 mutations in yeast without apparent trade-offs 121 We first asked whether and how ecological interactions with the alga change the distribution of 122 fitness effects of adaptive mutations arising in yeast. To this end, we carried out five replicate 123 BLT experiments in yeast evolving alone (the "A-condition") and in a community with the alga 124 (the "C-condition"; Methods). In each population, we tracked the frequencies of ~5×10 5 neutral 125 DNA barcodes integrated into the yeast genome 40 for 17 growth and dilution cycles. We 126 identified on average 2,820 and 2,905 adapted barcode lineages per culture in A-and C-127 conditions, respectively (Methods; Extended Data Figures 2, 3, Figures S1-S8). The similarity of 128 these numbers suggests that the presence of the alga does not dramatically change the rate at 129 which beneficial mutations arise in yeast. 130 Each adapted lineage is expected to initially carry a single beneficial driver mutation 37,38 . Thus, 131 by tracking barcode frequencies, we can estimate the competitive fitness benefits of many 132 simultaneously segregating driver mutations relative to the ancestral yeast strain (Methods). The 133 estimated distributions of fitness effects of beneficial mutations (bDFEs) are much broader than 134 expected from measurement noise alone in both A-and C-conditions (Supplementary 135 Information), indicating that yeast has access to multiple mutations with different fitness 136 benefits, consistent with previous work 37, 41 . The bDFEs in A-and C-conditions are different in 137 both median effect and breadth ( Figure 2A, S6 and S7). Specifically, the presence of the alga 138 reduced the bDFE median (1.60 in A vs. 1.50 in C; P < 10 -4 , two-sided permutation test, see 139 Methods) and increased its width (interquartile range (IQR) = 0.31 in A vs. IQR = 0.37 in C; P < 140 10 -4 , two-sided permutation test). This increase in width is associated with the appearance of two 141 peaks with higher relative fitness values around 2.0 and 2.5 (Figures 2A and S6). Since the 142 dynamics of adaptation depend on the shape of the bDFE 42,43 , these results indicate that the 143 presence of the alga alters evolutionary dynamics in the yeast population. 144 The presence of the alga can alter the yeast bDFE by changing which mutations are beneficial 145 (i.e., by imposing a fitness trade-off relative to the A-condition) or by changing the fitness 146 benefits provided by adaptive mutations, or both. To discriminate between these possibilities, we 147 randomly sampled 221 yeast clones from distinct adapted lineages in the A-condition ("A-148 mutants") and 189 yeast clones from distinct adapted lineages in the C-condition ("C-mutants"). 149 Clones were sampled at cycle nine, a time point at which most adapted lineages are expected to 150 still be driven by a single beneficial mutation (Methods). We then used competition assays to 151 measure the fitness of all A-and C-mutants relative to their ancestor in both A-and C-conditions 152 (Methods; Figures S9-S11; Data S2). These direct measurements of competitive fitness are 153 concordant with our estimates from the BLT experiment (see Supplementary Information; Figure  154 S12). We found that the C-mutants had significantly higher fitness in their "home" C-condition 155 than A-mutants, by on average 10.8% (95% CI [7.7%, 14.0%], P = 2×10 -11 , ANOVA model: 156 fitness ~ environment , F = 45.83, df = 1) and that A-mutants were more fit in their home A-157 condition than C-mutants by on average 14.4% (95% CI [10.2%, 18.5%], P = 2×10 -11 ; F = 158 45.42, df = 1), consistent with a signature of local adaptation. However, the fitness distributions 159 of both A-and C-mutants are wide and overlapping in both conditions, so that some A-mutants 160 are more fit than some C-mutants in the C-condition and vice versa. 161 Interactions with the alga significantly alter the fitness of 88% (362/410) of all sampled mutants 162 (false discovery rate (FDR) = 11%, obtained by permutation). However, fitness is positively 163 correlated between the two conditions across all mutants ( Figure 2B; Pearson R = 0.70, 95% CI 164 [0.65, 0.75], two-sided P = 10 -63 , t = 19.9, df = 408). Importantly, none of the A-or C-mutations 165 are deleterious in their "non-home" condition. Thus, the presence of the alga changes the fitness 166 benefits provided by adaptive mutations in yeast but does not impose a measurable fitness trade-167 off, in the sense that it does not alter which mutations are beneficial. 168 The presence of the alga alters the distribution of mutations 169 contending for fixation in yeast 170 Given that all sampled yeast mutants are beneficial both in the presence and in the absence of the 171 alga and that mutant fitness is correlated between the two conditions, we expected that the sets of 172 A-and C-mutants would be genetically indistinguishable. We quantified the diversity of adaptive mutations carried by A-and C-mutants with the 191 probability of genetic parallelism, Pg, which is the probability that two random clones share a 192 mutation at the same driver locus (lower Pg values imply higher diversity). We found that Pg is 193 slightly higher among A-mutants than among C-mutants (Pg = 6.0 ± 0.6% and 8.5 ± 0.8%, 194 respectively), although this difference was not statistically significant (P = 0.06, two-sided 195 permutation test; see Methods). Nevertheless, there were large differences in the frequency 196 distribution of driver mutations among A-and C-mutants (Figures 2C and Extended Data Figure  197 4, Data S4; P = 5×10 -8 , two-sided χ 2 -test, χ 2 = 160.51, df = 76), suggesting that the chance for 198 any given beneficial mutation to rise to a high enough frequency and be sampled varied 199 dramatically between A-and C-conditions. The starkest difference was observed for the 200 amplification of chromosome XIV (mutation chrXIV-3n): 10% (18/181) of C-mutants carried it 201 but none of the 215 A-mutants did (Figures 2C and Extended Data Figure 4), despite chrXIV-3n 202 mutations being beneficial in both conditions (Figure 2A  explanation, we find that a typical chrXIV-3n mutant is 2.05 ± 0.04 times more fit per cycle than 209 the ancestral wild-type and ranks in the top 11 ± 1.3% of most fit mutants in the C-condition, 210 while being only 1.59 ± 0.03 times more fit than the ancestor and ranking at 53 ± 5.3% in the A-211 condition ( Figure 2A). Other mutations with strong discrepancies in their representation among 212 A-and C-mutants show similar shifts in their fitness effects and rank order ( Figure 2A, Data S4). 213 In general, interactions with the alga shifted the fitness ranks of mutants between conditions by 214 14.1 ± 0.8% on average. We used simulations to confirm that such differences in fitness effect 215 and rank are sufficient to explain the observed genetic differences between A-and C-mutants 216 (see Supplementary Information, Extended Data Figure 5). 217 The fact that the sets of A-and C-mutants are genetically different implies that the yeast 218 populations in these two conditions are about to embark on distinct evolutionary trajectories. 219 Indeed, A-and C-mutants primarily represent high-frequency lineages in the respective 220 populations. Therefore, the mutations that they carry are more likely to win the clonal 221 competition towards fixation in the condition from which they were sampled. Then, the fact that 222 A-and C-mutants carry statistically distinct sets of mutations implies that mutations contending 223 for fixation in yeast in the A-versus C-conditions are also different. In other words, by altering 224 the fitness benefits of mutations, the ecological interactions with the alga change the 225 evolutionary trajectory of yeast, at least over the short-term. 226 Adaptive mutations in yeast have diverse ecological 227 consequences for the community 228 A mutation that spreads and fixes in the yeast population could subsequently alter the ecological 229 dynamics of the yeast-alga community; and adaptive mutations at different loci may have 230 different ecological consequences. For example, some mutations could increase yeast's 231 competitive ability and ultimately lead to the exclusion of the alga from the community. Others 232 could increase yeast's cooperativity and thereby strengthen the mutualism. To assess the 233 prevalence and the magnitude of different ecological effects of adaptive mutations in yeast, we 234 selected 28 C-mutants and 31 A-mutants that are representative of the genetic diversity of 235 contending mutations (Methods). We formed 59 "mutant communities" by culturing each of 236 these yeast mutants with the ancestral strain of the alga. As all A-mutants and C-mutants are 237 adaptive in both the A-and C-conditions, we pool all of them to increase our power to identify 238 their ecological effects. We quantified the ecological effect of each mutation by measuring YYC 239 (yeast yield in community) and AYC (algal yield in community), as well as YYA (yeast yield  240 alone; Methods, Figure S18, S19). As before, we define yield as the species density at the end of 241 the 5-day growth cycle. We focused on yields for two reasons. Yield is a measure of absolute 242 species abundance which determines the robustness of our community to demographic 243 fluctuations: communities with higher yields of both species are more ecologically stable 3 . In 244 addition, as discussed above, the ratio of YYC to YYA as well as the ratio of AYC to AYA (alga  245  yield alone) inform us about the net balance between competition versus cooperation in our  246 communities. 247 We found that many of the adaptive mutations significantly affected both YYA and YYC ( Figure  248 3A, S19A,B). The majority of tested mutations, 53% (31/59), significantly decrease YYA while 249 15% (9/59) of them significantly increase it (two-sided FDR = 28%, see Methods). At the same 250 time, we found no mutations that significantly decreased YYC, but 34% (20/59) of them 251 significantly increased it (FDR = 26%). Mutations have uncorrelated effects on YYA and YYC 252 ( Figure 3A), suggesting that yeast yield in the two conditions is determined by different 253 underlying traits. Mutations in yeast also significantly alter AYC, with 24% (14/59) of mutations 254 increasing it and 8% (5/59) decreasing it (FDR = 32%), with the effects on AYC and YYC being 255 positively correlated ( Figures 3B and S19C). 256 The fact that mutations alter the yields of both species suggests that some of them may also tip 257 the balance between cooperation and competition in one or the other direction. Indeed, we found 258 that 24% (14/59) of mutants have significantly increased both AYC/AYA and YYC/YYA ratios, 259 39% (23/59) have a significantly increased the YYC/YYA ratio only, and 12% (7/59) have 260 significantly decreased one or both ratios (two-tailed FDR = 5%, Figure 3C). Thus, seven 261 mutants acquired mutations that weaken the yeast-algal mutualism, at least for one of the 262 partners, and 37 mutants acquired mutations that enhance the mutualism. 263 These results show that yeast has access to beneficial mutations with ecologically diverse 264 consequences and suggest that our community has the potential to embark on a variety of eco-265 evolutionary trajectories with possibly different ecological outcomes. 266 Mutations favored by selection in the presence of the alga 267 strengthen the mutualism 268 Given that adaptive mutations in yeast have a variety of ecological consequences and that yeast 269 populations in the absence or presence of the alga are likely destined to fix different adaptive 270 mutations, we next asked whether mutations that contend for fixation in the A-or C-conditions 271 might have systematically different ecological effects. 272 We first noticed that the C-mutants clustered in the top right corner of the YYC versus AYC plot 273 ( Figure 3B) 122%]; P = 4×10 -3 ; n1 = 28 C-mutants and n2 = 31 A-mutants). These differences remained large 277 and significant even after accounting for the frequencies with which different driver mutations 278 were observed in our yeast populations (YYC: P = 10 -5 ; AYC: P = 3×10 -3 , permutation test, 279 Extended Data Figure 6, see Methods), indicating that this trend is not an accidental byproduct of 280 our choice of A-and C-mutants. Instead, the observed differences in yield must be caused by 281 systematic genetic differences between the A-and C-mutants. In other words, in the presence of 282 the alga, natural selection favors yeast mutants that produce higher yields in the community. This 283 conclusion is further corroborated by the fact that both YYC and AYC are correlated strongly 284 and significantly with competitive fitness in the C-condition, but only weakly with fitness in the 285 A-condition (Extended Data Figure 7). 286 A mutation in yeast that increases YYC may concomitantly lead to either a gain or a loss of 287 YYA ( Figure 3A). Therefore, such a mutation may either increase or decrease the net benefit that 288 yeast derives from its interactions with the alga over the growth cycle. We found that 86% 289 (24/28) of the C-mutants had a significantly higher YYC/YYA ratio than the ancestor (FDR = 290 21%), of which many, 32% (9/28), had also a significantly higher AYC/AYA ratio (FDR = 6%; 291 Figure 3C) but only 1/28 had a significantly lower AYC/AYA ratio (FDR = 58%). In contrast, a 292 smaller fraction, only 52% (16/31), of the A-mutants had a higher YYC/YYA ratio (FDR = 293 35%), of which 16% (5/31) also had a higher AYC/AYA (FDR = 12%) and 2/31 had lower 294 AYC/AYA (FDR = 29%). Thus, C-mutants both benefit more often from the presence of the 295 alga and reciprocally more often provide benefits to the alga compared to the A-mutants. In other 296 words, adaptive mutations that dominate yeast adaptation in the presence of the alga are more 297 likely to make yeast more cooperative and/or less competitive and thereby strengthen the 298 mutualism, compared to mutations that dominate adaptation in the absence of the alga. 299 An interesting potential consequence of this shift in selection on yeast precipitated by the 300 presence of the alga is that it can change the repeatability of yeast evolution along the 301 competition-mutualism continuum. We quantified such ecological repeatability by the 302 probability that two randomly drawn yeast mutants that contend for fixation in a given condition 303 both increase or both decrease the YYC/YYA ratio and simultaneously both increase or both 304 decrease the AYC/AYA ratio. The probability of ecological parallelism would be 25% under a 305 uniform null model. For the A-mutants, this probability is 33 ± 2.8%, indistinguishable from the 306 null expectation (P = 0.21, two-sided χ 2 -test, χ 2 = 1.55, df = 1). In contrast, it is 68 ± 5.5% for the 307 C-mutants, which is significantly higher than expected (P = 6×10 -8 , χ 2 = 29.3, df = 1) and also 308 significantly higher than for the A-mutants (P = 0.031, two-sided permutation test; see Methods). 309 Thus, yeast evolves more repeatably (towards stronger mutualism) in the presence of the alga 310 than in its absence. 311 In summary, our results show that mutations contending for fixation in yeast populations 312 evolving alone have relatively diverse effects on the ecology of the yeast-algal community, with 313 some strengthening and some weakening the mutualism. In contrast, mutations contending for 314 fixation in yeast evolving in the presence of the alga predominantly lead to higher yields of both 315 species, which strengthens the yeast-alga mutualism and makes evolution more repeatable at the 316 ecological level. 317 Mutualism enhancement is not selected directly but is likely a 318 byproduct of selection for other yeast life-history traits 319 We next asked how stronger mutualism could possibly evolve in our community. Specifically, 320 does natural selection in the presence of the alga favor mutations that increase cooperativity 321 and/or decrease competitiveness in yeast directly or is this bias a byproduct of selection for other 322 traits 25,52 ? Natural selection can directly favor rare mutualism-enhancing (i.e., more cooperative 323 and/or less competitive) yeast mutants only if such mutants preferentially receive fitness benefits 324 from their algal partners, that is, if there is a partner-fidelity feedback 53 . Our system is well-325 mixed, so that all diffusible benefits are shared by the entire culture, eliminating any potential 326 fitness advantage of rare mutualism-enhancing mutants 25,54 . The only way to prevent such 327 diffusion and ensure preferential benefit exchange with an algal partner is for such a mutant to 328 form a physical association with the partner 21 . However, we found no evidence for such 329 associations in any of the sampled mutants (Extended Data Figure 8; Methods). Given the 330 absence of a plausible partner-fidelity feedback, the increased cooperativity and/or decreased 331 competitiveness of the C-mutants must be a byproduct (pleiotropic effect) of selection for one or 332 more other traits. 333 We sought to identify traits under selection in the C-condition that could cause yields to increase. 334 Both competitive fitness and yield depend on fundamental physiological and life-history traits 335 embodied by yeast and alga, such as their growth rates, mortalities, nutrient consumption 336 efficiencies, etc 55 . Since measuring all potentially relevant traits was not feasible in this study, 337 we focused on two key traits that are known to be under selection in environments with variable 338 nutrient availability. The maximum population growth rate, r, is important for competitive 339 fitness when resources are abundant 56,57 , a condition that takes place at the beginning of each 340 growth cycle in our cultures. The carrying capacity, K, is an indicator of nutrient utilization 341 efficiency, which is important for competitive fitness when resources are scarce 56,57 , a condition 342 that takes place at the later phase of each growth cycle. We estimate r and K in the A-condition, 343 reasoning that these intrinsic traits would be relevant for fitness and yield in both A-and C-344 conditions. We estimate r by regressing the natural logarithm of the yeast cell density against 345 time during the initial phase of the growth cycle (Methods). We estimate K as the maximum 346 yeast cell density during the growth cycle, which is usually achieved on day 2. 347 We estimated r and K for all 59 sampled mutants (Figures S20, S21; Methods) and found that 348 many mutations significantly increased and decreased either one or both traits ( Figure 4A). We 349 found a negative correlation between the effects of mutations on r and K (Pearson R = -0.38, 350 95% CI: [-0.59, -0.13], two-sided permutation P = 0.004, Figure 4A), indicating a trade-off 351 between growth rate and nutrient utilization efficiency, which is often observed in other 352 systems 58-65 . More specifically, we found 16 C-mutants and 4 A-mutants have a significantly 353 higher K and a significantly lower r than the ancestor (FDR = 18%), an observation that is rare in 354 experimental evolution studies 60 where selection usually favors higher r 55,59,66-69 . However, 355 theory suggests that high-K/low-r mutations can be favored in the presence of an r-K trade-off in 356 populations near starvation 56,57 . We confirmed that 60% (12/20) of our significant high-K/low-r 357 mutants can in fact invade the ancestral yeast population in simulations of a logistic growth 358 model (see Supplementary Information and Figure S22). While this model demonstrates the 359 plausibility of selection favoring high-K/low-r mutants, it does not capture all the important 360 complexities of our system. Thus, we next explicitly tested whether r and K are under selection 361 in our A-and C-conditions. To this end, we examined the correlation between these traits and 362 competitive fitness among all 59 assayed mutants. 363 We found that neither r nor K are significantly correlated with fitness in the A-condition 364 ( suggest that other traits that we have not measured must be more important for fitness in the A-368 condition than either r or K. In contrast, fitness in the C-condition is positively correlated with K 369 ( Figure 4B, R = 0.51, 95% CI [0.29, 0.68], P < 10 -4 ) and negatively correlated with r (Extended 370 Data Figure 9A, Pearson's R = -0.51, 95% CI [-0.68, -0.29], P < 10 -4 ), consistent with the 371 observed r-K trade-off, and both of these traits together explain 37% of variation in competitive 372 fitness in the C-condition. Interestingly, a negative correlation between fitness and r persists 373 even after controlling for K (Table S4), suggesting that other unmeasured traits must also be 374 important for fitness in the C-condition. Regardless, C-mutants reach K values on average 12.7% 375 higher than the ancestor (P = 4×10 -4 , permutation test) and 9.5% higher than A-mutants (95% CI 376 [2.6%, 16.3%], P = 0.03). A typical C-mutant also has a significantly lower r than both the 377 ancestor (average ∆r = -8.7 ± 2.3%, P = 0.001) and a typical A-mutant (∆r = -11%, 95% CI [-378 17.0%, -5.1%], P = 7×10 -4 ). These observations suggest that nutrient efficiency is an important 379 component of fitness in the C-condition, and that yeast high-K/low-r mutants are favored by 380 selection in the presence of the alga. 381 We next asked whether higher-K mutants achieve higher yields. We might expect a strong 382 positive correlation between these quantities because, all else being equal, mutants that reach 383 higher density in the middle of the growth cycle due to their higher carrying capacity are more 384 likely to maintain higher density at the end of the growth cycle. However, we found no 385 correlation between K and YYA (Extended Data Figure 10C). This lack of correlation further 386 confirms that adaptive mutations that we sampled must affect other unmeasured traits which are 387 more important for yield than K. In contrast, we found that K and YYC were positively 388 correlated ( Figure 4C), suggesting that higher nutrient efficiency is important for achieving 389 higher yields in the community with the alga. 390 Our observations suggest a plausible model for how adaptive evolution can favor mutualism 391 enhancement in the absence of partner-fidelity feedbacks: ecological interactions with the alga 392 intensify selection for yeast mutants that use resources more efficiently (i.e., those that reach 393 higher K even at the expense of reduced r); once these mutants spread in the yeast population, 394 they support higher yields of both members of the community. Whether mutualistic partners 395 generally induce selection for lower r and/or higher K, and whether such selection consistently 396 leads to increased yields of both species remains an open question. 397 Similar to our analysis of ecological parallelism, we asked whether the presence of the alga alters 398 the probability of parallelism at the level of life-history traits, which we define as the probability 399 that r would be affected in the same direction in two randomly sampled mutants and that K 400 would also be affected in the same direction in these mutants (see Methods). We find that the 401 probability of trait parallelism is 30.7 ± 1.7% for the A-mutants, which is not significantly 402 different from 25% expected under the uniform null model (P = 0.47, two-sided χ 2 -test, χ 2 = 403 0.53, df = 1). In contrast, the probability of trait parallelism is 46.3 ± 4.8% for the C-mutants, 404 which significantly exceeds 25% (P = 0.009, χ 2 = 6.7, df = 1), suggesting that evolution in the 405 presence of the alga becomes more repeatable not only at the ecological level, as shown in the 406 previous section, but also at the level of underlying life-history traits. 407 In summary, our results show that interactions with alga shift natural selection on yeast to favor 408 mutants that increase K and decrease r, which in turn leads to increasing yields of both species in 409 the community. The shift in selection imposed by the alga makes evolution more repeatable both 410 at the level of life-history traits and even more so at the ecological level. 411

412
We characterized early adaptation in the experimental yeast-alga community and made three 413 main observations. First, we found that yeast have access to adaptive mutations that are not only 414 genetically diverse but also have diverse ecological effects. Second, even though there are no 415 measurable fitness trade-offs for yeast between growing alone or with the alga, the presence of 416 the alga modifies the fitness benefits provided by many mutations. This shift in selection 417 pressures is sufficient to change the set of mutations that contend for fixation in yeast and 418 thereby to alter the course of its evolution. Third, mutations that are strongly favored by selection 419 in the presence versus absence of the alga have different ecological consequences. Specifically, 420 the presence of the alga shifts selection on yeast to favor mutations that enhance the yeast-alga 421 mutualism (as measured by the yield of both species at the end of the growth cycle), making 422 evolution at the ecological level more repeatable. 423 Insofar as our yeast-alga community is representative of other ecological communities, our 424 results suggest that (i) organisms have access to a variety of adaptive mutations with diverse 425 ecological consequences and (ii) ecological perturbations, such as removal or addition of species, 426 can change the fitness effects of many of these mutations, thereby altering future outcomes of 427 evolution not only at the genetic but also at the ecological level. while yeast has access to a set of adaptive mutations that are quite diverse in terms of their 441 ecological effects, natural selection acting on yeast growing in the community strongly favors a 442 biased subset of these mutations, namely those that produce higher yields of both yeast and alga. 443 When viewed in the context of these prior observations, our findings suggest that ecological 444 interactions may limit the space of the most likely evolutionary trajectories. In our system in 445 particular, the presence of the alga modifies the effects of mutations in yeast in such a way that 446 yeast evolution becomes more repeatable at the ecological level, at least over the short-term. In 447 other words, ecological interactions may canalize evolution. Whether such canalization is a 448 general feature of evolution in a community context remains to be determined. 449 In our competitive mutualistic community, canalization appears to occur in the direction of 450 enhanced mutualism in the sense that the presence of the alga shifts selection on yeast in favor of 451 mutations that benefit both species. There are no demonstrated mechanisms that would favor 452 such enhanced mutualism in our community directly, but our results suggest another plausible 453 scenario for how it can evolve. Mutations in yeast favored in the presence of the alga tend to 454 increase yeast's carrying capacity in our medium and reduce its growth rate. Increased carrying 455 capacity could provide the competitive advantage necessary for such mutants to spread. Once 456 these mutations dominate, increased K and/or decreased r could enhance cooperation or reduce 457 competition with the alga. Specifically, increased K implies that there are more yeast cells to 458 generate CO2, which stimulates algal growth. Reduction in r could also benefit the alga via the 459 "competitive restraint" mechanism 77 in which slower growing yeast compete less for the initial 460 supply of ammonium and thereby offer the alga an opportunity to grow more and supply more 461 ammonium during the latter portions of the growth cycle. However, competition for the initial 462 ammonium can probably not be reduced to zero solely by mutations in yeast because yeast lacks 463 the molecular machinery for metabolizing the only other nitrogen source, nitrite. Therefore, a 464 single mutation or even a few mutations cannot alleviate yeast's basic requirement for 465 ammonium. Furthemore, traits other than r and K most certainly contribute to both fitness and 466 yield. Thus, additional experiments will be needed to determine how adaptive mutations in yeast 467 modify the competitive and cooperative phases of the growth cycle to provide an evolutionary 468 advantage and increase the yields of both species. 469 How the presence of the alga amplifies the fitness advantage of high-K mutants is currently 470 unclear. An analysis of the genetic and biochemical basis of yeast adaptation may help us answer 471 this question and assess how general the ecological mechanisms of mutualism enhancement 472 might be. However, one challenge is that many mutations driving adaptation in yeast are large 473 chromosomal amplifications and deletions, and it is unclear which amplified/deleted genes 474 actually cause the fitness gains and changes in the ecologically relevant traits. At this point, we 475 can only speculate on this subject. For example, it is known that ChrXIV-3n amplifications are 476 adaptive under ammonium limitation, possibly driven by the copy number of the gene MEP2 that 477 encodes a high affinity ammonium transporter 78 . We suspect that these adaptations are 478 particularly beneficial to yeast in the C-condition because the alga provides a continuous but low 479 flux of ammonia. Another interesting example are mutations in genes HEM1, HEM2 and HEM3, 480 which provide much larger fitness benefits in the C-condition compared to the A-condition (Data 481 S4) possibly because they shift the metabolic balance towards fermentation at higher 482 concentrations of dissolved oxygen produced by the alga (see Supplementary Information). 483 Elucidating these and other mechanisms of physiological adaptation in our competitive 484 mutualistic systems is the subject of future work. 485 To conclude, our results suggest that microbial adaptation in the community context is driven by 486 many mutations that are genetically and phenotypically diverse and have diverse ecological 487 consequences. Changes in the ecological milieu, such as loss of some species or invasions by 488 others, may not necessarily alter which mutations are beneficial to community members. 489 Nevertheless, such ecological changes can quantitatively alter the benefits of mutations, so that 490 evolutionary trajectories become canalized towards certain ecological outcomes. 491

492
Barcode lineage tracking (BLT) experiment and data analysis 493 Strains. We used the strain CC1690 of the alga Chlamydomonas reinhardtii, which can also be 494 obtained from the Chlamydomonas Resource Center. The barcoded library of the diploid yeast 495 Saccharomyces cerevisiae strain GSY6699 40 was kindly provided by Prof. Gavin Sherlock. This 496 is a diploid, prototrophic strain derived from the BY genetic background, homozygous 497 throughout the genome, except for locus YBR209W, where one copy of a DNA barcode was 498 integrated 37 . Our starting library consists of about 5 × 10 5 clones, each of which carries a unique 499 DNA barcode at this locus. In principle, the genomes of all clones should be identical 500 everywhere else prior to our barcode lineage tracking (BLT) experiment. However, as discussed 501 in Supplementary Information (Sections 1.3 and 3), we found that our initial population already 502 contains some pre-existing polymorphisms, which arose prior to our BLT experiment. 503 Growth conditions. Both yeast monocultures and yeast-alga communities were cultured in a 504 defined minimal medium 33 ("CYM medium") supplemented with 2% dextrose, 10mM KNO2 and 505 0.5mM NH4Cl, which we thereafter refer to as the "growth medium". All cultures were grown in 506 10mL of the growth medium in 50mL flasks (FisherSci #FS2650050) capped with 50mL plastic 507 beakers (VWR #414004-145) at room temperature (21°C) on a platform shaker with 70 foot-508 candles of constant light (three Feit Electric #73985 suspended approximately 24 inches above 509 the platform shaker) shaking at 125 RPM, unless noted otherwise. 510 BLT pre-cultures. Prior to the BLT experiment, yeast and alga were pre-cultured in 50mL of 511 growth medium in 250mL delong baffled flasks (PYREX #C4446250) for two and 10 days 512 respectively. Alga pre-cultures were started from colonies. To start yeast pre-cultures, the 513 barcoded yeast library was thawed from frozen stock at room temperature, then 500µL were 514 transferred into 50mL of the growth medium. 515 BLT initiation and propagation. We conducted five replicate BLT experiments for each of two 516 treatments, yeast monoculture (the A-condition) and yeast + algae community (the C-condition). 517 Each monoculture BLT experiment was initiated from 100µL of the yeast pre-culture. Each 518 community BLT experiment was initiated from 100µL of the 1:1 (v/v) yeast and alga mixture. 519 Cultures were grown for 5 days before being diluted 1:100 for the next growth cycle (100µL into 520 10mL fresh media). A total of 17 growth/dilution cycles were completed. A detailed discussion 521 on the number of generations per growth cycle is provided in Supplementary Information 522 (Section 1.1). Throughout this work, we ignore adaptation in the alga, as discussed in 523 Supplementary Information (Section 1.2). 524 Culture preservation. Glycerol stocks were taken of the yeast pre-culture and yeast + algae 525 inoculum mixture, as well as at the end of every odd growth cycle. Separate stocks were stored 526 for DNA extraction and cell isolation purposes with two replicates each, for a total of 4 stocks 527 per culture per time point. Cell isolation stocks were created by aliquoting 1.5mL of culture into 528 500µL of 80% glycerol, mixing by vortex and storing at −80°C. DNA stocks were created by 529 removing the supernatant of the remaining 7mL of culture via centrifugation and resuspending in 530 2mL of 20% glycerol (80% glycerol diluted with 1x PBS), which was then stored as two separate 531 1mL stocks at −80°C. 532 DNA isolation. DNA stocks were thawed and DNA isolated using a "salting out" method, based 533 on established protocols 79,80 . The thawed stocks were first centrifuged, and the supernatant 534 removed. The pellet was resuspended in 300µL 3% SET buffer (3% SDS, 10mM EDTA, 30mM 535 Tris) and incubated at 65°C for 15 minutes. The tube was cooled to room temperature by 536 immersing it in room temperature water, then 2.5µg of RNAse A was added. After vortexing, the 537 mixture was incubated at 37°C for 1 hour, after which it was cooled on ice. 150µL of 3M Sodium 538 Acetate was then added and mixed by inversion, after which it was cooled on ice for a further 5 539 minutes before centrifuging at maximum speed for 10 minutes on a tabletop centrifuge. The 540 supernatant was transferred to a new tube and DNA was precipitated by the addition of 500µL 541 isopropanol, which was mixed by inversion and then pelleted by centrifugation for 1 minute at 542 maximum speed. The supernatant was removed and the pellet was washed with 200µL cold 70% 543 ethanol without vortexing before being allowed to dry inverted for 30 minutes at room 544 temperature before the DNA was resuspended in 50µL molecular biology grade water. 545 Sequencing library preparation. The barcode locus was amplified through a 2-step PCR 546 protocol slightly modified from Ref. 38 . The first amplification added inline indices for sample 547 multiplexing, universal molecular identifiers for removing PCR duplicates during analysis and 548 Illumina-compatible adapter sequences for a second round of amplification with standard 549 Illumina Nextera XT primers. For the first reaction, 10µL of template was mixed with 25µL of 550 OneTaq 2x Master Mix, 1µL of 25mM MgCl2, 1µL each of the forward and reverse primers (at 551 10mM concentration) and 12µL of molecular biology grade water. Primer sequences are as 552 described 38 . This mixture was amplified using the following conditions: (1) 94°C for 10 min; (2) 553 94°C for 3 min; (3) 55°C for 1 min; (4) 68°C for 1 min; (5) Repeat steps 2-4 for a total of 8 554 cycles; (6) 68°C for 1 min; (7) Hold at 4°C. The amplified product was purified using Ampure 555 XP magnetic beads using established protocols (with 50µL of beads used per sample). Then, 556 10µL of the purified product was used as the template for a second reaction along with Illumina 557 Nextera XT primers (1µL of 10mM stock for each primer), 1µL of 25mM MgCl2, 25µL of 558 OneTaq 2x Master Mix and 12µL of water with the following reaction conditions: (1) 94°C for 5 559 min; (2) 94°C for 30 sec; (3) 62°C for 30 sec; (4) 68°C for 30 sec; (5) Repeat steps 2-4 for a 560 total of 25 cycles; (6) 68°C for 5 min; (7) Hold at 4°C. The PCR products were purified using 561 Ampure XP beads as before. 5µL of each sample was mixed to form a pool for Illumina 562 sequencing, concentrated using Ampure beads as before with equal volume of beads as a pooled 563 sample and size-selected via agarose gel extraction to isolate the correct amplicon before 564 submitting for sequencing. 565 Sequencing. Populations A1 and C1 were initially sequenced on a MiSeq platform. All 566 populations (including A1 and C1) were then also sequenced on a HiSeq platform. Data from 567 both runs were combined for all downstream analysis of frequency trajectories. We obtained an 568 average of 2.6 million paired-end reads per time point. 569 To identify and count DNA barcodes, we used a custom python pipeline BarcodeCounter2 570 available at https://github.com/sandeepvenkataram/BarcodeCounter2. The package first uses the 571 BLASTn tool to identify sequences known to flank the barcode region within each read pair. If 572 reads contain inline indices, samples can be demultiplexed. Universal molecular identifier (UMI) 573 sequences can be extracted if present within the reads. If a sequence contains multiple barcode 574 regions, these extracted regions are concatenated together. To account for sequencing errors, 575 DNAClust 81 is then used to cluster the concatenated barcodes into clusters of nearly identical 576 sequences which presumably originated from the same DNA molecule. The output of DNAClust 577 is a FASTA database of all unique barcode sequences present in the library. Then, BWA 82 is 578 used to map the barcode sequences from each sample onto the clustered FASTA database. 579 Mapping of reads is necessary because the clustering process removes identifying information 580 associating barcode sequences with samples from which they came. PCR duplicates are removed 581 based on the UMI sequences, and the total number of unique reads corresponding to every 582 barcode in the FASTA database is counted. The final output is a tab-delimited table of the read 583 counts for every barcode in every sample. The software is designed to be user-friendly and 584 highly customizable, with simple text files describing the input files, multiplexed samples and a 585 sequence template describing the structure of the sequenced reads. The package is built using 586 python3, and uses the popular BioPython package. The package has multithreading support, and 587 can be run on both personal computers and supercomputing clusters. 588 When generating the database of all unique barcode sequences, we clustered sequences at 95% 589 similarity, so that, given that the length of our barcode is 52bp, sequences with 3 or more base-590 pair differences were merged into the same cluster. This set of clustered sequences was 591 generated only once using all of the time points from all sequenced populations. 592 We developed an iterative heuristic procedure to identify adaptive lineages from lineage tracking 593 data and estimate their fitness. where ̅ ( ) is the mean selection coefficient (per cycle) of the population at cycle t. Thus, we 602 estimate the mean selection coefficient of the population at cycle t at iteration k as 603 . 604 The mean fitness at cycle t is then defined as 1 + ̅ (&) ( ).  CV of the lineage exceeds the median CV across all lineages or if si (k) < 0 (see Figure S1A). 622 6. Termination. The procedure is terminated when the set of neutral lineages at iteration k 623 differs by less than 5% compared to the set at iteration k − 1. 624 Our procedure converges to a stable set of neutral lineages within 5 iterations. We have carried 625 out sensitivity analyses of our procedure with respect to the choice of various parameters, as well 626 as other sanity checks, as described in the Supplementary Information (Section 1.3). Note that 627 we report fitness values relative to the ancestral strain on a per-cycle basis, rather than the per-628 generation basis typically used in the literature because our cultures experience growth phases 629 other than exponential growth 38,83 ( Figure 1, Supplementary Information (Section 1.1)). 630 Permutation tests for differences in bDFE across conditions. We randomly relabel adapted 631 lineages as being from either the A-or C-condition, and calculate the difference in bDFE median 632 and IQR from this permuted data. This procedure was conducted 10,000 times, to generate a 633 random distribution of median and IQR difference values. 634 Competitive fitness assays Isolation of random clones. To isolate adapted clones, frozen stocks of the monoculture and 643 community populations from cycle 9 were thawed, plated onto standard 100mm Petri dishes with 644 CYM + 1% agarose at a dilution of approximately 100 cells per dish, and incubated at 30°C for 645 three days (algae do not grow at 30°C). 88 random colonies were isolated from each population, 646 i.e., a total of 440 clones from the A-and C-condition each. Eight additional clones from each 647 population were harvested at cycle 17 and are present in the pools described below, but they are 648 not included in any of the analyses presented in this study. Each colony was transferred into a 649 well of a 96-well plate (Corning 3370) with 200µL of CYM media and incubated for two days at 650 30°C. 10µLwas used for each of the "population", "row" and "column" pools described below. 651 Then, 50µL of 80% glycerol was added to each well, and the plate was stored at −70°C. 652 Barcode genotyping. The DNA barcodes of all isolated clones were identified by sequencing, 653 using the Sudoku method 84,85 . Specifically, 10µL of each clone was pooled into 10 "population" 654 pools (one pool for each source population), eight "row" pools and 12 "column" pools. DNA 655 barcodes in each pool were amplified and sequenced as described above We expect that a given 656 combination of row, column and population pools would have a single barcode in common, 657 defining the isolate in the corresponding well of the appropriate plate. We determined all 658 barcodes present in the intersection of each combinations of row, column population pool. If a 659 single barcode is identified in the intersection, the corresponding well is assigned that barcode 660 identity. If multiple barcodes or no barcode are identified, the associated clone is removed from 661 further analysis. 662 Generation of A, C and N pools. We used the heuristic procedure described above to classify 663 clones with identified barcodes into three groups described below: non-adapted, adapted in the 664 C-condition or adapted in the A-condition. Clones were pooled into three libraries defined by 665 their membership in these three groups as follows. To construct each pool, we transferred 20µL 666 of thawed frozen stock of each isolate into a 2mL 96-deep-well plate (Corning P-2ML-SQ-C) 667 filled with 1.8mL of growth media supplemented to 10mM ammonia. After incubating these 668 plates at 30°C for three days, we formed each pool by combining 200µL of individual saturated 669 cultures. Three pools were stored in 20% glycerol at −70°C. 670 For the sake of efficiency, clone pooling was based on an earlier version of the heuristic 671 procedure used for the classification of lineages. The current version of the procedure (as 672 described above) classifies lineages slightly differently. As a result, pools do not perfectly 673 correspond to the classification of clones according to the current heuristic procedure, which is 674 provided below. This minor discrepancy has no bearing on our results because the final 675 classification of clones is based on fitness estimates from the competition assays (see below). 676 The A pool contains 214 clones that are classified as adapted in A populations. The C pool 677 contains 223 clones that are classified as adapted in C populations. The N pool contains 144 678 clones, 84 of which are classified as neutral from the BLT analysis (i.e., not adapted in either A 679 or C populations), 29 clones that are classified as adapted in the A-condition and 31 clones that 680 are classified as adapted in the C-condition. Thus, we measured competitive fitness for a total of 681 581 clones. 682 Competition assay experiment. To conduct the competitive fitness assays, we pre-cultured each 683 of the three pools (N, A and C; see above) separately in the growth media for two days. We also 684 pre-cultured algae for 10 days, starting from colonies. We then combined A, C and N pools in 685 the 1:1:18 ratio. We carried out three replicate competitions in the A-condition and three 686 replicate competitions in the C-condition. To this end, we inoculated each of the six replicates 687 with 100µL of the combined A/C/N pool. In addition, the three C-condition replicates were 688 inoculated with 100µL of the algae preculture (∼ 10 6 cells / mL). 689 All replicates were propagated in conditions identical to the BLT experiment for a total of five 690 growth cycles. Glycerol stocks were made at the end of each growth cycle after the dilution step 691 by centrifuging the full culture volume, removing the spent media and resuspending the pellet in 692 2mL of 20% glycerol + PBS. Two 1mL aliquots of this glycerol suspension were stored at 693 −70°C. One of these aliquots was harvested for DNA extraction and barcode sequencing using 694 protocols described in the section on the heuristic BLT analysis procedure. 695 Competition assay data analysis. Barcodes were identified and counted as described above. 696 The resulting barcode count data were analyzed as described previously 38 using software 697 available at https://github.com/barcoding-bfa/fitness-assay-python. Briefly, the 84 non-adapted 698 barcodes (as defined from the BLT analysis described above) from the N pool were used to 699 estimate the mean fitness trajectories and the additive and multiplicative noise parameters for 700 each pair of time points in each assay 38 . These estimates were used to estimate the fitness of 701 every lineage for each pair of neighboring time points along with the error in the estimate. The 702 variance of an estimate for a given pair of time points was calculated as the inverse of the read 703 depth at the earlier of the two timepoints + the estimated multiplicative noise parameter. Inverse 704 variance weighting was then used to combine estimates across all time point pairs to generate a 705 single fitness and error estimate for each lineage in each replicate. Replicate estimates were 706 combined using further inverse variance weighting to generate the final fitness estimate for each 707 isolate in the A-and C-conditions. For each mutant in each condition, we also calculated the 708 95% confidence interval around the fitness estimate based on the variability in fitness 709 measurements between replicates (assuming that measurement errors are distributed normally). 710 Fitness estimates are provided on a per growth cycle basis, as discussed above. Validation of this 711 analysis procedure and additional statistics are described in Supplementary Information (Section 712 2). 713 Genome sequencing and analysis Reads were sorted, duplicates marked and short variants were called and filtered using GATK (v. Variants were annotated with ENSEMBL Variant Effect Predictor using their command-line 737 tool. As many variants had multiple possible annotations, coding sequence annotations 738 ("missense variant", "frameshift variant", "stop gained" and "stop lost") were prioritized over 739 synonymous annotations, which were prioritized over upstream noncoding annotations (within 740 2kb of a gene) and finally downstream noncoding annotations (again within 2kb of a gene). 741 Variants further of 2kb of any gene or those within 2kb but with no annotations were removed as 742 likely nonfunctional. Finally, variants with less than 3 reads of support for the derived allele 743 were removed as putative false positives. 744 Additional filtering to remove erroneous and ancestral variants. The procedure described 745 above identified 34,720 small variants across 428 sequenced isolates. We expect that many of 746 these variants are sequencing and/or mapping errors that our procedure failed to remove, as well 747 as fixed differences from the reference genome present in the ancestor of our experiment. To 748 further filter out such spurious variants, we estimate the ancestral allele frequency spectrum from 749 24 sequenced ancestral clones and compare it with a typical allele frequency spectrum of 24 750 adapted mutants (averaged over 1000 random draws of 24 adapted mutants). As Figure S15  751 shows, the latter has an excess of variants that are present only in one clone, as expected for de 752 novo mutations. The fact that there is no excess of mutations present in two or more adapted 753 clones suggests that all or most mutations observed in two or more adapted clones are not 754 adaptive. coverage plots for each sequenced clone by averaging read depth into 1kb windows with 759 bedtools genomecov. An example plot is shown in Figure S13A. As coverage negatively 760 correlates with the distance to telomeres ( Figure S14), we re-calculate coverage after correcting 761 for this variation ( Figure S13B). We then manually identify CNVs from these corrected coverage 762 plots by visualizing the coverage distribution at higher resolution ( Figure S13C). 763 We identified 176 CNVs across 167 strains (Data S3), of which 85 are whole-chromosome 764 aneuploidies. We found no CNVs in 32 sequenced ancestral and neutral isolates ( Figure S16C, 765 "driver" mutations among residual ancestral and erroneous variants as well as nonadaptive 771 "passenger" mutations, we rely on the idea of genetic parallelism, i.e., the fact that loci under 772 selection gain mutations in independent lineages more often than expected by chance 87,88 . For 773 each gene, we define its multiplicity as the number of clones that carry a mutation in this gene. 774 Since shorter genes require lower multiplicity to be called adaptive, we bin genes by their length 775 into six 1kb bins plus one bin for genes with length ≥ 6kb. For each length bin l = 1,2,...,7, we 776 count the number of genes whose multiplicity is m = 1,2,..., denoting these counts by klm. We 777 obtain the number of such genes 〈 01 〉 expected in the absence of selection as follows. 778 We randomly and independently redistribute N = 1718 mutations (small mutations in adaptive 779 clones after removing multiple mutations in the same gene in the same clone) across 7226 yeast 780 genes 1000 times, with the probability for each mutation landing in a given gene being 781 proportional to its length + 2kb. Then, the observed excess number of mutations with 782 multiplicity m in length bin l is 01 = max{0, 01 − 〈 01 〉}. These excess mutations are 783 assumed to be adaptive. We redistribute the remaining ⌊ − ∑ ∑ 01 1 2 03( ⌋ potentially non-784 adaptive mutations as before, and identify additional excess mutations as adaptive. We repeat 785 this procedure iteratively until the total number of excess mutations is ≤ 1. We achieve 786 convergence after 8 iterations and thereby obtain the final expected counts 〈 01 〉. 787 Next, we would like to identify specific loci that carry adaptive mutations as those with 788 multiplicities above some threshold. Specifically, we would like to determine multiplicity 789 thresholds Ml for each length bin l = 1,...,7, so that all loci in that bin with multiplicities ≥ Ml are 790 called adaptive. To do so, for all l we calculate the expected FDR at the given multiplicity 791 threshold Ml as 792 We choose the multiplicity thresholds Ml, so that β(Ml) ≤ β * for all l and for some desired β * . We 794 use β * = 10%. We estimate the overall FDR as 795 We conduct this analysis considering all 396 sequenced adaptive isolates together, as well as A-797 and C-mutants separately. Loci identified in any one of these three analyses are defined as 798 putative adaptive loci. After identifying putative 63 adaptive loci this way, we assume that all 799 discovered 185 mutations at these loci are adaptive (Data S4). An extended discussion of our 800 analysis of adaptive mutations can be found in the Supplementary Information (Section 3). 801 Probability of genetic parallelism. We calculate the probability of genetic parallelism Pg for a 802 set of mutants as follows. We consider every pair of mutants, and calculate the proportion that 803 have a mutation in at least one common adaptive locus (including both small variants and 804 CNVs Devices Spectramax i3x, excitation at 435nm and observation at 670nm). Chlorophyll 826 fluorescence intensity measurements were converted into cell density estimates by using a 827 calibration curve as described in the Supplementary Information (Section 5.1). 828

Selection of A-and C-mutants for phenotyping experiments.
We selected 31 C-mutants and 829 28 A-mutants to cover a diversity of mutations and fitness values represented among all sampled 830 adaptive mutants (see Data S4 for the number of mutants selected from each mutation class). Our 831 reasoning for this non-random sampling was that mutants carrying a driver mutation at the same 832 locus would have similar phenotypic values. If we selected clones for phenotyping randomly, we 833 would have likely not observed more rare phenotypes. To account for this over-dispersion in the 834 selection of mutants we apply the mutation weighting procedure described below. 835

Measurement of YYA.
To estimate YYA, we inoculated all 60 yeast strains (including the 836 ancestor) individually into the standard conditions (10 mL of media in 50 mL flasks) from frozen 837 stocks and propagated them for two cycles (10 days). During the third cycle, we estimated yeast 838 densities after 5 days of growth by plating and colony counting as described below. Correlations 839 between replicate measurements are shown in Figure S19A. µL of alga were used instead of 100 µL because the density of algae culture was approximately 848 50% of that at the initiation of the BLT experiment). These mutant communities are grown for 849 one cycle in our standard conditions. On day 30, we transfer 100 µL of each mutant community 850 into 10 mL fresh media, as in the BLT experiment. We estimate both yeast and alga density on 851 day 35, as described previously. Yield estimates can be found in Data S2. Correlations between 852 replicate measurements are shown in Figure S19B,C. 853 Microscopy. To detect potential physical associations between algae and beneficial yeast 854 mutants, we created mutant communities as described above. After 5 days of growth, 855 communities were then mounted on glass microscope slides (Fisher Scientific 12550143), sealed 856 with Dow Corning high vacuum grease (Amazon B001UHMNW0) and imaged on a light 857 microscope using a 20x objective with DIC. Extended Data Figure 8 shows one representative 858 community; the remaining 17 imaged mutant communities along with WT controls can be found 859 in the Dryad data repository. 860 Mutant growth curve measurements. To estimate the growth parameters r and K for individual 861 beneficial mutants, we carried out growth curve measurement experiments of individual yeast 862 mutants and the ancestor in the A-condition. To this end, we inoculated all 60 yeast strains 863 (including the ancestor) individually into the BLT condition (10mL of media in 50mL flasks) 864 from frozen stocks and propagated them for two cycles (10 days). During the third cycle, we 865 estimated yeast densities on days 10, 10.5, 11, 11.5, 12, 13, 14 and 15 by plating and colony 866 counting as described in the section "Measurements of YYC and AYC." The growth curves are 867 shown in Figure S20 and the data are provided in Data S2. 868 We estimate r as the slope of the relationship between log(CFU/mL) and time (in hours) for the 869 three measurements between 12 and 36 hours of growth. We estimate K as the maximum 870 observed density (in CFU/mL). r and K estimates can be found in Data S2. Correlations between 871 replicate measurements are shown in Figure S21. 872 Mutation weighting. Even though not all A-or C-mutants were phenotyped, we would like to 873 make certain statistical statements about the distribution of phenotypes among all sampled A-or 874 C-mutants. To this end, we associate each of the 59 phenotyped mutants with a single driver 875 mutation. Mutants with multiple driver mutations are associated only with the most common 876 driver mutation. Mutants with no identified driver mutations are associated with a unique 877 unknown mutation. To obtain the prevalence of a given phenotypic value among all A-or C-878 mutants, we weight each measured phenotypic value by the number of sequenced A-or C-879 mutants with the same driver locus as the phenotyped mutant and divide by the total number of 880 phenotyped A-or C-mutants. 881 Kernel density estimation. We use kernel density estimate (KDE) to determine how likely 882 certain phenotypic trait values would occur among all A-and/or C-mutants. Specifically, we 883 obtain the kernel density estimates DA(y,a) and DC(y,a) for the probabilities that a community 884 formed by the ancestral alga and a random A-or C-mutant, respectively, would produce yeast 885 yield y and alga yield a. To estimate DA, we apply the kde2d function in R with bandwidth 1 886 along the x-axis and 4/3 along the y-axis to the mutation-weighted yield data for the A-mutants. 887 We analogously obtain DC(y,a). 888 Phenotypic parallelism analysis. We quantify the degree of parallelism among a set of mutants 904 with respect to a pair of quantitative traits X and Y by estimating the probability that, for two 905 randomly selected mutants, trait X changes in the same direction in both mutants and trait Y 906 changes in the same direction in both mutants. Mathematically, if one randomly selected mutant 907 has trait increments ∆Xi and ∆Yi relative to the ancestor and the other mutant has trait increments 908 ∆Xj and ∆Yj, we estimate the probability that both (∆Xi)(∆Xj) ≥ 0 and (∆Yi)(∆Yj) ≥ 0. This 909 measure of phenotypic parallelism emphasizes the direction of change rather than the magnitude. 910

Accounting for measurement errors in statistical tests
When we compute this measure for the pair of yields of mutant communities, in which case X 911 and Y are yeast and alga yields, we refer to it as the probability of ecological parallelism, Pe. 912 When we compute this measure for the growth phenotypes, in which case X and Y are r and K, 913 we refer to it as the probability of phenotypic parallelism, Pph. 914 We obtain these probabilities of parallelism for the A-and C-mutants. We test the significance of 915 the deviation of these probabilities from the expectation of 25% parallelism via a χ 2 -test. To 916 determine the statistical significance of the difference between the probabilities of parallelism for 917 the A-and C-mutants, we resample the phenotypic values of each of the 59 mutants from the 918 normal distribution (see above) and permute mutant genotype and home-environment labels, so 919 that the genotype and label are always associated with each other but dissociated from the 920 phenotypic values. We then calculate the parallelism probabilities for these permuted and 921 resampled data. We estimate the P-value by carrying out this permutation and resampling 922 procedure 1000 times. 923    probability of sampling at least one clone with a beneficial mutation that arises at a certain rate 1223 (x-axis) and provide a certain fitness benefit in the A-condition (y-axis). The most common 1224 driver loci are shown by points (colors are the same as in Figure 2 in the main text). The 1225 estimated beneficial mutation rate and the selection coefficient for each mutation class are given 1226

Data Availability
in Table S3. B. Same as A but for the C-condition. The mutation rate for each locus is assumed 1227 to be the same in both conditions, but the selection coefficients vary.   Table S1. P-values for comparison of absolute abundances and net population change rates 1288 across conditions. 1289 Table S2. Statistics of mutational counts among sequenced mutants. 1290 Table S3. Numbers of isolates carrying the most common adaptive mutations and the estimated 1291 mutation rates. 1292 The rate at which new mutations arise in a yeast population and the amount of physical time it takes for these mutations to survive drift and establish depends on the number generations that yeast go through during our BLT experiment. Thus, if yeast go through substantially different numbers of generations per cycle between the A-and the C-conditions, it could cause differences in our power to detect adapted lineages, which in turn could confound our ability to compare measured bDFEs. In fact, it is possible that yeast go through different number of generations because they reach different peak densities and because they likely do not reproduce after Day 2 in the A-condition, but they likely do reproduce in the C-condition (since the alga supplies the limiting ammonium).
We can estimate the number of generations in the A-condition if we assume that there is no death during Days 1-2 and there is no growth during Days 2-5. Given that the yeast starts the growth cycle at 1.37 × 10 4 cells per mL and reaches peak density of 4.72 × 10 6 cells per mL on Day 2 (Figure 1, Data S1), the estimated number of generations per cycle in the A-condition is 8.4. Knowing that yeast yield in the A-condition is 9.83 × 10 5 cells per mL, we can also estimate the average per capita death rate in the A-condition in Days 3-5 as 0.52 per day.
We can estimate the number of generations per cycle in the C-condition as follows. We again assume that there is no death in Days 1-2. Thus, given that the yeast starts the growth cycle at 2.39 × 10 4 cells per mL and reaches peak density of 2.83 × 10 6 cells per mL on Day 2 ( Figure 1, Data S1), we estimate the number of generations in this part of the cycle to be 6.9. To estimate the number of generations in Days 2-5, we assume that the per capita death rate during this phase of the cycle is the same in the C-condition as in the A-condition, i.e., 0.52 per day. Thus, in the absence of new births, yeast would have reached density of 5.95 × 10 5 cells per mL by the end of the cycle. Instead, we observe that yeast yield is 8.78 × 10 5 cells per mL, which implies that 2.83 × 10 5 new yeast cells were produced during this period, corresponding to 0.13 generations. Thus, we estimate that yeast go through on average 7.0 generations per cycle in the C-condition.
Given that yeast go through fewer generations and generally have lower population sizes in the C-condition than in the A-condition, we would expect that fewer adaptive mutations would arise in the C-condition and it would take them longer to spread. However, we actually detect slightly more adapted lineages in the C-condition than in the A-condition, suggesting that these differences do not substantially undermine our ability to identify adapted lineages.

Alga.
Since the alga grow continuously during the entire cycle both alone and in community with yeast, we can estimate their number of generations by assuming that there is no death. Given that the alga growing alone starts at density 9.3 × 10 3 cells per mL and reaches density 2.6 × 10 6 cells per mL (Figure 1), we estimate that the alga goes through approximately 8.1 generations when growing alone. Given that the alga growing in community with yeast starts at density 4.83 × 10 4 cells per mL and reaches density 7.90 × 10 6 cells per mL, we estimate that the alga goes through approximately 7.4 generations when growing in community.

Justification for ignoring adaptation in alga
In the C-condition, that is, when yeast and alga are co-cultured together, yeast and alga reach final yields of 8.78 × 10 6 and 7.90 × 10 7 cells, respectively, with the bottleneck sizes being by a factor of 100 smaller. Numerous evolution experiments with yeast populations of comparable or smaller size across various environmental conditions have shown that adaptive mutations arise and reach high frequencies within 250 generations or less [1,2,3,4,5]. Although much less is known about the rates of adaptation in Chlamydomonas reinhardtii, one study reports a failure of the alga (strain CC2344) to adapt within 1000 generations of evolution at the bottleneck size of 10 5 cells [6]. Another study reports 35% growth rate gain after 1880 generations of evolution in alga strain CC-503 cw92 mt+ [7] with a comparable population size to ours and detectable gains appeared only after about 300 generations. Our community BLT experiments last for about 66 algal generations until adapted yeast mutants are sampled, which is likely insufficient for new mutations to arise in the alga population. Furthermore, since the environment in our experiment is well-mixed and there is no evidence for physical associations between the two species (see Extended Data Figure 8), the only way the alga can modify the environment for yeast is through the medium. Thus, even if some alga mutants arose during the BLT experiment, they would not be able to reach high enough frequencies in the population to substantially alter the medium.

BLT data analysis Sanity checks.
We performed two "sanity" checks to ensure that our analysis gives reasonable results. First, mean fitness is expected to monotonically increase over time in each population. We find that mean fitness does indeed increase monotonically during the first 11 cycles in all but one populations ( Figure S1B). Second, we visually inspected how well individual lineage trajectories are fitted by equation and found that these fits are generally very good ( Figure S1C,D shows typical fits).
Sensitivity of the procedure for detecting adapted lineages with respect to the choices of parameters. The heuristic procedure for identifying adapted lineages has three key parameters: (1) the minimum lineage frequency threshold (set to 10 −4 ); (2) the expansion factor used to identify neutral lineages (set to 100); and (3) the "SE parameter" which is the number of standard errors separating the lineage's fitness from zero required to call it adapted (set to 2). We tested the sensitivity of our procedure with respect to the choice of these parameters.
We varied the minimum lineage frequency threshold and the expansion factor by an order of magnitude in either direction. We find that the number of adapted lineages that we detect varies primarily with the minimum lineage frequency threshold, as expected ( Figure S2C). However, our claims regarding the differences between the bDFEs in the A-and C-condition remain reasonably robust ( Figure S2A,B). The most sensitive result is the difference in the width of the bDFEs which disappears when we lower the minimum lineage frequency threshold, presumably because too many non-adapted lineages are being falsely called adapted.
We then tested how the SE parameter affects the bDFE median in each condition, while keeping the other two parameters at their chosen values. We find that the bDFE median as a function of the SE parameter has a plateau between values 1 and 2 ( Figure  S3). When the SE parameter declines below 1, we see a precipitous decline in the bDFE median, as would be expected from the associated growth in the number of false positives. When the SE parameter is increased beyond 2, we observe a gradual increase in the bDFE mean, as would be expected from the associated growth in false negatives. This result suggests that any choice of the SE parameter within the plateau would be appropriate, as it would strike a balance between keeping down both false positives and false negatives.
Precision and accuracy of the procedure for detecting adapted lineages. We developed a simulation framework to quantify the precision and accuracy of our procedure. We simulated evolution using a population initialized with 100,000 barcoded lineages, of which 2,000 had already acquired an adaptive mutation prior to the beginning of the evolutionary simulation. The fitness effects of these mutations per growth cycle were drawn from a normal distribution, with µ = 0.4 and σ = 0.2. These parameters were selected to allow for relatively weak adaptation, as a strong adaptation regime would lead to extinction of many adapted mutants. Observing adaptation in this regime allows us to get better statistics on the precision and accuracy of our method.
We simulated the evolution of our population iteratively as follows. Given that the population size of lineage i at the beginning of cycle t is N i t , by the end of the cycle it deterministically grows or shrinks toÑ i t+1 = N i t e s i −ωt . Here, s i is the fitness of lineage i, ω t = i N i t s i / i N i t is the mean fitness of the population at time t. After growth, we simulate dilution, so that the size of lineage i at the beginning of cycle t + 1 is drawn from the Poisson distribution with parameterÑ i t+1 . The initial size N 0 of each lineage was drawn from an exponential distribution with mean 100.

5
The population was propagated for total 18 growth cycles. Every other cycle, barcodes were sampled and "sequenced" . We simulate this stochastic process by drawing r i t reads of barcode i from the negative binomial distribution with mean u i = R N i t / i N i t , and variance u i (1 + ϵ u i ). The total read depth was set as R = 10 6 . The error parameter ϵ = 0.01 allows for variation greater than the mean and represents systematic sources of experimental error (gDNA extraction, PRCs, sequencing). The choice of the negative binomial distribution and parameter ϵ was based on Ref. [8]. We then applied our lineage calling procedure to these simulated barcode data.
Our procedure identified 2306 lineages as adapted, of which 1413 were true positives and 893 were false positives ( Figure S4). We thus estimate the false positive rate to be 71%, the false positive rate to be 1% and the false discovery rate to be 39%. For the true positives, the estimated fitness correlates very well with their true fitness ( Figure S5A). The median of the estimated bDFE is lower compared to that of the underlying true bDFE (0.314 versus 0.406, Figure S5B), presumably due the presence of false positives. Indeed, 93% (829 out of 893) of the false positives have estimated fitness effects below the median. As a result, for lineages above the median the FDR is 4.3%, indicating, as expected, that our procedure captures the higher-fitness lineages very well.
Pre-existing mutations. Some mutations could have arisen prior to the beginning of the experiment, specifically, prior to splitting the barcoded yeast library between five A and five C replicates. Such pre-existing mutations cause two issues with the downstream analysis and interpretation. First, they make it more difficult to identify causal mutations that increase fitness of adapted lineages. We discuss this problem and how we address it in Section 3. The second issue is that pre-existing beneficial mutations confound the inference of bDFEs and their comparison across treatments. Specifically, if multiple lineages (within the same population or in different replicate populations) carry the same pre-exisiting beneficial mutation and are identified as such, this mutation will be counted multiple times and will therefore inflate the size of a fitness class in the inferred bDFE.
To diagnose this problem, we note that there are two types of pre-existing mutations. Those that arose after the transformation that introduced barcodes into our strain (type I) and those that arose before this transformation (type II). All pre-existing mutations of type I are linked to different barcodes. Therefore, these mutations can be identified by comparing barcodes of lineages identified as adapted in different populations. We found that pairs of populations have on average 416 (15%) adapted lineages with a common barcode, whereas only 23 (0.8%) would be expected by chance ( Figure S8). Barcodes are more often shared between A populations than other types of population pairs (22% vs 13%, t-test P = 10 −10 ). We suspect that this is because the majority of adapted mutants have higher relative fitness in the A-condition than the C-condition ( Figure 2B in the main text), such that pre-existing mutations are more likely to reach high frequencies and be detected in the A-condition. We eliminate pre-existing beneficial mutations of type I from our bDFE analysis by constructing these distributions from lineages with unique barcodes.
The same pre-existing mutation of type II can be linked to multiple barcodes and is indistinguishable in the barcode data from multiple independently arising mutations. However, pre-existing mutations of type II can be identified in the genome sequencing data. As discussed in the methods section on genome sequencing and analysis, we find no evidence of such mutations, suggesting that they must be rare in our populations.
Additional analyses. The median standard error in our estimates of fitness of adaptive mutants was 0.02, and median CV was 0.08 ( Figure S1A). The standard deviation of the bDFEs is much larger, 0.22 for A populations and 0.33 for C populations, suggesting that there are multiple classes of adaptive mutations available to yeast in both conditions. We confirm that the observation of a multimodal bDFE is not an artefact of combining the bDFEs of all replicate C populations ( Figure S6). Furthermore, the fitness estimates have fairly consistent error distributions across all replicate populations ( Figure S7).

Competitive fitness assays
We obtained barcode frequency data at cycles 1, 2, 3, 4 and 5 ( Figure S9). Fitness estimates of individual lineages were concordant between replicates (Figure S10), with relatively low estimation errors ( Figure S11). Estimated fitness of sampled clones in both conditions along with errors and confidence intervals are provided in Data S2.
Determining valid clones for further analysis. Of the 581 clones that were pooled in the competition assays, we filtered out many clones from downstream analysis via a number of filters. First, clones that were sampled at cycle 17 were excluded from further analysis, as were clones that lacked a valid fitness measurement in either environment or had variance in fitness measurement in either environment ≥ 0.5. These filters removed 151 of 581 from consideration, leaving us with 430 clones for further analysis.

Calling adapted clones.
We use the competition assay data to call adapted clones. A clone is called as adapted in a given environment if its estimated competitive fitness in that environment is more than two standard errors greater than 1 (2.3% FDR based on normal distribution). We identified 401 clones as adapted in the A-condition (see Data S2). 221 of them were sampled from the A-condition and we refer to them as the Amutants. We identified 402 clones as adapted in the C-condition (see Data S2). 189 of them were sampled from the C-condition and we refer to them as the C-mutants. There are 16 clones that are not adapted in either A or C condition, 13 clones that are adapted in the C-condition but not in the A-condition and 12 clones that are adapted in the Acondition but not in the C-condition. Of the 16 clones not adapted in either the A or C conditions, 12 come from lineages determined not to be adaptive in the BLT analysis, 7 which we define as our neutral clones (Data S2). The genomes of 215 A-mutants, 181 C-mutants and eight neutral mutants were later sequenced, as described in the Methods and Section 3.
Relationship between fitness estimated in competition assays and in BLT experiments. The correlation between fitness of adapted mutants (i.e., A-and C-mutants) estimated in the competition assays and those estimated from the BLT data is reasonably good ( Figure S12, R = 0.57, P = 8.77 × 10 −37 ), but there are also some systematic differences. Specifically, competition assays over-estimate BLT fitness in the A-condition and under-estimate it in the C-condition ( Figure S12). We discuss possible reasons for these discrepancies below.
When we place mutants onto the bDFE in the non-home environment (as in Figure  2A in the main text) and when we estimate the probability of sampling a mutation in the non-home environment (see Section 4), we need to know the BLT fitness of mutants in their non-home environment. We have no direct measurements of BLT fitness of A-or Cmutants in their non-home environment, but we do have the non-home fitness estimates of all mutants in competition assays. However, directly substituting BLT fitness for competitive fitness would lead to biases due to the aforementioned discrepancies between the two estimates. To correct for these discrepancies, we linearly regress BLT fitness of A-mutants against their competitive fitness in the A-condition and we regress BLT fitness of C-mutants against their competitive fitness in the C-condition (see Figure S12). We then use these regressions to estimate the BLT fitness of C-mutants in the A-condition and A-mutants in the C-condition.
Difference in adaptive mutant fitness between conditions. Mutant's fitness were called as significantly different between A and C-conditions if the two 95% confidence intervals did not overlap. We discovered 178 such clones (108 A-mutants and 70 Cmutants), all of which were adaptive in both environments.
Possible reasons for the slight discrepancy between BLT and competitive fitness estimates. We can think of at least three possible reasons for this discrepancy. One possible reason is that we use a different set of reference lineages in the BLT experiments and competition assays. It is possible, for example, that some of the lineages that we use as reference in the competition assay are not in fact neutral in one or both of the conditions.
Another possible reason is that fitness is weakly frequency-dependent (this is expected in batch culture experiments [9]). Frequency dependence can manifest itself in a discrepancy between BLT and competition assays because lineages are present at much higher frequencies in the competition assays than in the BLT experiments. Specifically, the median frequency of an adapted clone (as defined from the competition analysis) in the competition assay is 4 × 10 −4 at the initial time point, whereas the median frequency of an adapted lineage is 3 × 10 −6 at the initial time point (the frequency of the adapted mutant driving the lineage frequency must be even lower since many if not most adapted lineage also initially contain non-adapted individuals).
A third possible reason is that the media composition in the BLT experiments and in competition assays could be somewhat different at later stages of the growth cycle because population compositions are different and the media composition is determined by the collective metabolism of all variants present in the population.
Given the overall concordance of fitness estimates, we decided that dissecting these relatively minor effects was not particularly important within the scope of the present work.

Genome sequencing and analysis
Small mutations. The distributions of derived small variants per clone is shown in Figure S16A and their summary statistics are given in Table S2. These data show that the majority of small variants detected in the evolved clones are still likely ancestral or erroneous. We can estimate the expected number of adaptive small variants as follows. A typical adapted isolate carries 1.18 more mutations compared to an ancestral isolate (Figures S15, S16 , Table S2). However, not all of these mutations may be adaptive. Indeed, given that small indels and single-nucleotide mutations occur at rate 3 × 10 −3 per genome per generation [10], we expect 0.18 of such mutations to have occurred on the line of descent of any isolate sampled at cycle 9 (∼ 60 generations). Based on this estimate, a typical adapted clone is expected to carry 1.18 − 0.18 = 1.00 adaptive mutation.
One potential problem with this estimate is that our yeast strain may have a somewhat different mutation rate than the strain used in Ref. [10]. We can obtain an alternative estimate for the number of adaptive mutations per clone by comparing adapted clones with the neutral ones. We find that a typical neutral clone carries 1.96 ± 0.88 more mutations than a typical ancestral isolate 1 (Table S2), which is greater than 1.18 extra mutations carried by a typical adapted clone. Thus, it is possible that a typical adapted clone carries no beneficial small mutations.
The problem with the second estimate is that some of the clones identified as neutral may in fact carry adaptive mutations 2 . Therefore, we use both methods to obtain bounds on the number of adaptive mutations carried by a typical adapted clone and conclude that a typical adapted clone is expected to carry between 0 and 1 small adaptive mutations. Carrying out the same calculations for the A-and C-mutants separately, we estimate that a typical A-mutant carries between 0 and 1.34 small adaptive mutations (so that between 0 and 288 out of 1008 small mutations found in A-mutants are expected to be adaptive), and a typical C-mutant carries between 0 and 0.61 small adaptive mutations (so that between 0 and 110 out of 717 small mutations found in C-mutants are expected to be adaptive; Table S2).
Using genetic parallelism, we identify 185 mutations at 63 loci as adaptive (on average 0.47 mutations per clone), consistent with our expectation. 76 of these mutations at 39 loci are found in 60 A-mutants (on average 0.35 mutations per clone) and 109 mutations at 56 loci are found in 81 C-mutants (on average 0.60 mutations per clone; Table S2).
Total number of identified adaptive mutations per clone. After combining small mutations and CNVs together, we find that a typical A-mutant carries 0.74 identified adaptive mutations with 100 A-mutants having no identified adaptive mutations (Data S3). A typical C-mutant carries 1.11 identified adaptive mutations with 46 C-mutants having no identified adaptive mutations (Data S3). Since all A-and C-mutants gained fitness in their home environment, each of them must have at least one adaptive mutation. Therefore, the "unknown" sectors in Figure 2C in the main text refers to the numbers of A-or C-mutants without any identified adaptive mutations.
Pre-existing mutations. We use genetic information to test for the prevalence of preexisting mutations of type II, i.e., those that arose prior to the integration of the DNA barcodes, so that multiple lineages may carry an adaptive mutation identical by descent. If such mutations were prevalent, we would expect to observe an excess of adapted clones carrying identical genetic mutations, compared to ancestral isolates, but this is not the case ( Figure S15). In fact, an identical small mutation is found less frequently in multiple adapted clones than in multiple ancestral clones, suggesting that some of the pre-existing genetic variation may have been deleterious.
CNV events are not suitable for this analysis because they occur at very high rates, and thus we have no way of ascertaining whether or not two identical CNV events are identical by descent. Indeed, many CNVs are aneuploidies, which occur at rate 9.7 × 10 −4 per generation [10]. The only consistent segmental CNVs are ChrIV-1n and ChrIV-3n. ChrIV-1n mutations are localized to identical breakpoints across 64 mutants; these breakpoints are concordant with highly similar repetitive regions, YDRWTy2-2/YDRCTy1-2 and YDRWTy2-3/YDRCTy1-3. There are three different classes of ChrIV-3n amplifications across 10 mutants; at least some of these breakpoints also appear to be localized close to known repetitive regions, including YDRCTy2-1, YDRCTy1-1 and YDRWTy2-2. Thus, we suspect that these recombination-driven segmental events also occur at high rates and are not necessarily indicative of pre-existing variation.

Simulations of evolutionary dynamics and the estimation of rates of adaptive mutations
We found that the sets of A-and C-mutants are genetically distinct (Figures 2C in the main text, Extended Data Figure 4 and Data S4), despite all of them being more fit than the ancestor in both environments. In this section, we show that this somewhat puzzling observation can be explained by the differences in the evolutionary dynamics in the concurrent mutation regime [11]. There are two key differences between the A-and C-conditions. First, the bDFEs are different (Figure 2A in the main text), resulting in different increases of population's mean fitness over time. Second, the fitness rank orders of adaptive mutations are also different ( Figure 2A in the main text). These two facts imply that in the concurrent mutation regime the chances for a given mutation to escape drift while rare and reach a certain frequency and be sampled can be substantially different in the two conditions. Next, we develop a quantitative version of this argument.
As described in Section 3, we classify all discovered adaptive mutations into mutation classes by the type of CNV or the gene in or near which the mutation occurred (Data S4). Let k A im and k C im be the number of sequenced A-or C-mutants that carry a mutation from class m sampled from the replicate population i = 1, 2, 3, 4, 5 (see Table S3). We would like to know whether the differences between k A im and k C im can be explained by the observed differences in the bDFEs and by differences in the fitness benefits provided by mutations of class m in the A-and C-conditions.
The challenge is that the probability Pr {k; s, U } of observing k mutants of a certain type in a sample from a given population depends not only on the selection coefficients s of adaptive mutations of this type (which we have measured, as described in Section 2) but also on the unknown rate U at which such mutations arise. Therefore, we first find the mutation rates that fit our data and confirm that these rates are biologically plausible. We then ask how well the expected numbers of A-and C-mutants carrying specific mutations match the observed numbers k A im and k C im , given the estimated mutation rates.

Data
Exclusion of pre-existing mutations.
A number of lineages are identified as adapted in multiple populations, indicative of pre-existing mutations of type I (see Methods for more details). Since such mutations do not carry information about the mutation rate, we exclude them from this analysis. That is, numbers k A im , k C im given in Table S3 and in Extended Data Figure 5 are the numbers of sampled and sequenced clones with unique barcodes.

Selection coefficients.
Fitness of individual mutants in the BLT experiments are estimated based on the competition assay data, as described above (see Section 2), and selection coefficients s A m and s C m are taken as averages over all mutants carrying a mutation from a given mutation class (Table S3).

Mean fitness trajectories.
Our model described below takes into account the fact that the probability of a newly arisen beneficial mutation to survive genetic drift depends on how the mean fitness of the population changes over time. Thus, when simulating the evolutionary dynamics, we use the empirical mean selection coefficient trajectoriess A (t) ands C (t) in the A-and C-conditions, respectively.s A (t) ands C (t) are computed by first averaging the values of the mean selection coefficient (estimated in as described in the Methods) over all five replicates and then fitting the logistic function to these data points (see Figure S1B).

Model
To keep the model as simple as possible, we assume that our population has the constant size N = 2 × 10 6 in both A-and C-conditions, mutations from the mutation class m arise at (an unknown) rate U m per division in both conditions. We assume that all mutation from class m confer the same (known) fitness benefits s A m and s B m relative to the ancestor in the A-and C-condition, respectively (Table S3). We also assume that all mutants are sampled at time T = 60 generations after the beginning of the experiments in both A-and C-conditions, which roughly corresponds to cycle 9, assuming log 2 100 = 6.64 doublings per cycle (see Methods).
To estimate the mutation rate U m for mutation class m we define the log-likelihood function The sampling probability Pr {k; s, U } depends on the number of lineages that independently arose in the population and that carry a mutation of the focal type, as well as on the frequencies of these lineages at the sampling time point T . It is difficult to obtain an analytical expression for this probability because it requires integrating over all these nuisance parameters. We therefore compute the sampling probability numerically, using population dynamic simulations described below.

Evolutionary dynamics simulations and the estimation of the sampling probability
As mentioned above, difference in the bDFE between the A-and C-conditions may contribute to the observed differences in the genetic composition of sampled mutants. The bDFE indirectly affects the survival probability of a newly arisen beneficial mutation by altering how the mean fitness of the population changes over time. Thus, we design our simulations so that they match the average empirical mean selection coefficient trajectories in the A-or C-condition (see Methods and Figure S1B). Matching these trajectories in simulations of the full Wright-Fisher where all mutation classes segregate simultaneously is difficult. Instead, we simulate the arrival and spread of mutations of each mutation class separately, while accounting for changes in mean fitness using our logistic fit of the observed mean fitness trajectories ( Figure S1B). This approach also allows us to use branching process approximations described below in Section 4.2.2, which greatly speed up the calculations.
Since in this section we are concerned with mutants of one mutational class in one environment, we omit the subscript m and superscripts A/C. In other words, we assume that the mutations arrive at rate U per individual per generation and have selection coefficient s, and we sample 88 mutants from this population at time point T = 60.
In our simulations, we allow for new mutations to arise between 1 cycle prior to the beginning of the experiment up to cycle 9. Thus, we divide the time interval between −6.6 and T = 60 generations into 260 to 1300 segments of length ∆t = 0.01/s generations. For each time segment (t i , t i + ∆t), we draw the number of new mutants arising in the population in that segment from the Poisson distribution with rate N U ∆t. Each of these mutants survives until the sampling time point T with probability P surv (T ; s, t i ), given by equation (2) below. For each mutant that survives, we draw its establishment time τ from the distribution given by equation (3) below. We then set the mutant's frequency at the sampling time point to n(T ; s m , τ )/N , where n is given by equation (4) below. At the end of each simulation run, we obtain a list of frequencies of all independently arisen mutants carrying mutation of type m. We assume that each independently arisen mutants is linked to a distinct barcode. We then randomly sample 88 clones from the whole population and discard those that do not carry the focal adaptive mutation. If multiple clones are sampled from the same lineage, only one is retained. We further randomly sub-sample these clones with 97.5% success probability, simulating the small whole-genome sequencing failure rate, which results in the final number k of sampled clones that carry a mutation from the focal mutation class. We estimate the probability of sampling k clones, Pr {k; s, U }, by running 10 4 simulations and recording the fractions of simulations where we observe k sampled clones.

Branching process approximations for mutant growth dynamics
Consider a mutant that arises at some point t 0 < T on the ancestral background. The early population dynamics of such mutant lineage can be modeled as a branching process [11]. In particular, Desai and Fisher [11] used the branching process approximation to derive the probability that a mutant with selection coefficient s that arose at time zero in the ancestral population (i.e., whose mean selection coefficient is zero) has not gone extinct by time t and the related probability that the mutant "establishes" at time τ (see equations (11) and (17) in Ref. [11]). However, this model ignores the fact that the mean fitness of the population is changing over time while the focal mutant is still at low frequency. In our experiment, mean fitness changes very rapidly, as can be seen in Figure S1B, which can significantly alter mutant's survival probability and its establishment time. Assuming that the mean selection coefficient trajectorys(t) is known (see Section 4.1), we can model mutant dynamics analogous to Ref. [11] but with a generalized birth-death model with growth rate 1 + s −s(t) and death rate 1. In this model, the probability that a mutant that arose at time t 0 survives until time t is given by where provided that ω(t 0 ) > 0. Conditional on surviving, the probability that this mutant establishes at time τ is given by Once the mutant establishes at time τ and provided that it is still beneficial, i.e., ω(τ ) > 0, its subsequent dynamics are essentially deterministic, so that its size n(t; τ ) at time t is approximately given by which is analogous to equation (14) in Ref. [11] in the limit t ≫ 1/s.

Estimation of mutation rates by maximum likelihood
For each mutation class m, we compute the likelihood function L(U m ) using equation (1) for 40 discrete values of u m = log 10 U m between −7 and −3 and interpolate between them using a polynomial of degree 4. The likelihood function for chrIII-3n mutations is shown as an example in Figure S17. We find the maximum of this function as a point where the polynomial approximation of L has a zero first derivative with respect to u m . We estimate the standard error (SE) as the inverse of the square root of Fisher information [12].

Results
We estimated the rates of driver mutations at 9 loci that were mutated in at least five adapted clones, after excluding gene NUM1 where all mutations are likely pre-existing (see Extended Data Figure 4, Table S3 and Data S4). Our estimates are generally consistent with those derived from a published mutation accumulation experiment by Zhu et al [10]. In particular, our estimate of the rate of small adaptive mutations in genes HEM1 and HEM2 is ∼ 3 × 10 −7 per generation, consistent with per basepair mutation rate of 1.67×10 −10 estimated by Zhu et al [10]. Gene YIL169C is an exception, with an estimated rate of adaptive point mutations of 7.76 × 10 −6 per generation. We estimate the rates of large CNV events to be ∼ 10 −5 per generation, again consistent with the genome-wide rate of aneuploidies of 1.04 × 10 −4 estimated by Zhu et al [10]. Our estimate of the segmental duplications ChrIV-3n is much lower (3.24 × 10 −7 per generation), also consistent with Ref. [10], although they do not provide a quantitative estimate of such events.
The probability that at least one sequenced mutant from all five replicate populations has a mutation with a given selection coefficient s and mutation rate U is shown in Extended Data Figure 5A,B, for both A-and C-conditions. As expected, this function is different between conditions because the mean fitness dynamics are different ( Figure S1B). Finally, we directly compare the observed numbers of mutations of each mutation class with the expected numbers in the A-and C-conditions and find a reasonable match (Extended Data Figure 5C).

Calibration curve for the measurement of alga cell density
We generated a calibration curve for interpreting fluorescence measurements in terms of alga cell number as follows. We grew alga in our standard growth conditions both alone and in a community with the ancestral yeast. On each day of the growth cycle, we counted alga cells using a haemocytometer and we took a fluorescence measurement using the a plate reader, as described in Methods. In addition, we generated a 2-fold dilution series starting with the saturated alga cultures at the end of the growth cycle both alone and in the community. We also obtained haemocytometer counts and fluorescence measurements for the alga grown in a 59 mutant communities (see Methods in the main text). We combined all these data to obtain a single the calibration curve by plotting the fluorescence against haemocytometer counts, both log-transformed ( Figure S18). We found a good positive correlation (R = 0.96). We used the following equation to convert fluorescence measurements into cell density, log(Cell density) = 1.236 × log(Fluorescence) − 2.843.

Analysis of plausibility of natural selection favoring high-K /low-r mutants
Our measurements of r and K for the adapted yeast mutants suggest that the increases in K at the expense of r and increases in r at the expense of K can both be adaptive in both A-and C-conditions. Theory suggests that high-K mutants can be favored by natural selection when the population is close to starvation [13,14], but mutants with higher K, and especially those with lower r, are rarely found in evolution experiments. We therefore wanted to test whether the mutants with our measured r and K values (especially those with higher K and lower r than the wildtype) can plausibly invade the wildtype yeast population.
To this end, we constructed a simple coupled logistic growth model. We consider two strains, strain 1 (wildtype) and strain 2 (mutant) whose per capita growth rates are r i and carrying capacities are K i , i = 1, 2. The dynamics of the population sizes N i (i = 1, 2) are then described by equations We set r 1 = 0.168 h −1 and K 1 = 5.5×10 6 for the wildtype and we set the mutant values r 2 and K 2 based on our measurements given in Data S2. We set the initial population sizes of the wildtype and mutants to be 10 4 and 10 2 individuals, respectively, and simulate the growth of these strains for 120 hours (the length of our standard growth cycle). We determine that a mutant can successfully invade if its frequency increases by the end of the growth cycle.
If our model perfectly captured our experimental conditions, all of the mutants would be able to invade, since all of them are experimentally found to be more fit than the wildtype. In fact, we find that only 37 out of 59 (63%) mutants are able to invade ( Figure S22). As expected, five mutants that have both r and K lower than the wildtype could not invade in our model. Similarly, eight mutants with higher K but low r and nine mutants with higher r and lower K could not invade, suggesting that our model does not capture some processes and traits that are likely important in our experiment (e.g., it does not capture the dependence of growth and death rates on resource concentrations). Nevertheless, 16 out of 24 mutants with higher K and lower r are able to invade the wildtype in this simple model, which supports our argument that selection can favor high K in our conditions even if it comes at the expense of decreasing r.

Relationship between r, K and fitness
As described in the main text, we found that each r or K individually explain about 26% of variation in competitive fitness in the C-condition but do not explain any statistically significant variation in the A-condition. However, it is possible that a linear combination of both variables would improve our ability to predict fitness in both conditions. To this end, we consider a multiple regression model Fitness ∼ r+K. We report the results of this regression analysis in Table S4. We find that r and K jointly do not explain any variation in fitness in the A-condition, but they jointly explain 37% of variation in fitness in the C-conditions, which is significantly more than each of them explains individually. Both r and K contribute approximately equally to fitness. Furthermore, we find that decreasing r increases fitness even if K is being held constant. This surprising observation could be explained if one or more unobserved traits (other than r and K) were important for competitive fitness and if there was a trade-off between r and such unobserved trait.
6 Possible function effects of mutations in HEM1, HEM2 and HEM3 genes Three critical components of the heme biosynthesis pathway, HEM1, HEM2 and HEM3 are putative targets of adaptation in this study, and provide a substantially larger fitness benefit in the C-condition than in the A-condition (see Data S3 and Data S4). Most mutations found in these genes probably compromise the function of the encoded enzymes and lead to a decreased "siphoning" of succinyl-coenzyme A (sCoA) from the TCA cycle and the production of less heme and/or decoupling of aerobic-driven regulation from aspects of central metabolism. Although yeast is already capable of performing fermentation under ambient oxygen concentration, we speculate that these mutations enables yeast to ferment under even higher oxygen concentrations that results from alga photosynthesis.

HEM1
HEM1 encodes the mitochondrial 5-aminolevulinate synthase (ALAS) that catalyzes the first committed step of porphyrin through the condensation of glycine and sCoA using the cofactor pyridoxal 5'-phosphate (PLP): sCoA + glycine + PLPcofactor 5-aminolevulinate + CoA + CO 2 + PLPcofactor ALAS links heme/cytochrome production with the TCA cycle and aerobic respiration via sCoA, and thus plays a key role in cellular energetics [15]. ALAS functions as a homodimer (with subunits referred to as A and B below). A crystal structure of S. cerevisiae HEM1 is available (https://www.rcsb.org/3d-view/5TXR/1).
We found 5 adaptive mutations in the HEM1 gene, four of which occurred in the C-mutants (Data S3 and Data S4). A nonsense mutation at residue 166 (out of 548) in HEM1/ALAS leads to a truncated coding sequence and presumably loss of function. All other mutations alter amino acids that are conserved in all sequenced S. cerevisiae strains, which we speculate may also significantly compromise normal HEM1/ALAS function, as described below.
His107Pro mutation does not seem to be involved in any substrate or co-factor binding directly, but a change from histidine to a proline, with significant backbone confirmational constraints may have serious consequences for the folding (and thus function) of the enzyme. Bracketing this position (107) nearby is Arg91, which plays a critical role in sCoA binding, and Asn121, which forms an alpha-carboxylate hydrogen bond with the glycine substrate [16]. Moreover, His107 (of subunit B) forms important structurally stabilizing side-chain interactions with Glu111 (subunit B) and Lys142 (subunit A) (see Figure S3 of [16]), which may very well be disrupted with a proline substitution (lacking a side chain).
Asn152Lys and Asn157Lys mutations occur in a region of the ALAS protein that becomes ordered upon PLP cofactor binding. N152 plays a direct role in coordinating sCoA [16], hydrogen bonding with the carboxylate (COO-) moiety of the sCoA succinyl group. Mutation from asparagine to a positively charged lysine may both disrupt the formation of this hydrogen bond and prevent a key side-chain interaction between Asn152 (subunit A) and Arg91 (subunit B) that stabilizes key structural elements needed for PLP cofactor binding (we expect the side chain of mutation Lys152 to repel that of Arg91 since both are positively charged). Asn157Lys is a few amino acids downstream from Ile153, which also plays a role in carboxylate/sCoA coordination like Asn121 [16]. Asn157 is adjacent to Ile158 (subunit A) that forms stabilizing main-chain and side-chain interactions upon PLP cofactor binding with Asn95 and Asn 97 (both on subunit B); mutation to Lys may disrupt these interactions, along with that of nearby stabilizing interactions between Arg98 (subunit B) and Ala147 (subunit A), due to positive charge repulsion between Arg98 and Asn157Lys (see Figure S3 of [16]).
Gly344Cys mutation occurs at a site that does not directly bind substrate or co-factor but is each bracketed by amino acids that do play key active site roles and could also have structural consequences as conformational flexibility is likely lost with the change away from glycine. Gly344 is flanked (although several amino acids away) by K337 that forms a critical covalent pyridoxyl-lysine bond, and F365 which delineates the sCoA substratebinding pocket [16]. Gly344 resides on the juncture between a beta-sheet and alpha-helix motif in the protein structure, with little space for a side chain. Mutation of this site to a cysteine may alter backbone flexibility and the folding of key structural elements. Gly344 is also in close proximity to Cys182 that forms stabilizing hydrogen bonds with amino acids Asn129, Thr275, and Gly 276; having another cysteine nearby may very likely interfere with that hydrogen bonding network.

5-aminolevulinate porphobilinogen + 2 H 2 O
A crystal structure of S. cerevisiae HEM2 is also available (https://www.rcsb.org/ 3d-view/1AW5/1). We found 5 adaptive mutations in the HEM2 gene, four of which also occurred in the C-mutants (Data S3 and Data S4). Two frame-shift mutations lead to premature stop codons ∼ 100 out of 342 amino acids. The three remaining mutations Ile129Phe, Val132Phe, and Pro264Arg are predicted to have moderate impact.
Ile129 resides on a beta-strand with a side chain embedded in a relatively hydrophobic environment. It is unclear what the consequences of mutation Ile129Phe might be, as phenylalanine is a comparably bulky hydrophobic side chain to isoleucine that would at first glance seem to be a conservative substitution [17]. However, several aromatic amino acids are in the vicinity, including Trp30 and Tyr127; Ile129Phe may form aromatic ringstacking interactions with these residues to disrupt fold structure.
Val132 is similarly on a beta-strand with a side chain embedded in another nearby hydrophobic pocket and interface with other non-polar amino acids on alpha-helix. The Val132Phe mutation could disrupt the tight packing at this interface as phenylalanine is substantially bulkier than valine. Moreover, Tyr168 is nearby which hydrogen bonds a key water molecule that is hydrogen-bonded to several other backbone atoms flanking Val132; the Val132Phe mutation may also disrupt this by interacting with Tyr168 through aromatic stacking interactions. Pro264 resides in a sharply kinked beta-turn between beta-strand and alpha-helix elements, which may be important for constraining and enabling key interactions of flanking residues with amino acids distributed across different structural elements: Val262/Ser265/Tyr287, Ser265/Glu292, and Lys263/Tyr287 (see https://www.rcsb.org/3d-view/1AW5/1).
Pro264Arg mutation is pronounced not just for the loss of backbone constraint provided by the imino acid proline [17], but it introduces a large positively charged side change that may form "inappropriate" interactions to disrupt monomer fold and subsequently oligomerization; these include interactions with nearby: negatively charged amino acids: Glu292 and Glu313 aromatic amino acids (through cation-pi interactions) Phe211 Figure S1. Estimation of fitness of adapted lineages in the BLT data. A. The mean lineage fitness (x-axis) is plotted against the coefficient of variation (CV) of the estimate (y-axis). Blue (red) points indicate lineages identified as adapted (non-adapted). B. Mean fitness of each population over time. Individual populations are shown in grey, the average mean fitness trajectory over all populations is shown in black, and a logistic fit to this average trajectory is shown in red. C, D. Examples of selection coefficient estimation for one lineage in one A population (C) and one C population (D).       Figure S8. Evidence for pre-existing mutations in the BLT experiments. For each population on the x-axis, we plot the fraction of lineages identified as adapted in that population which are also identified as adapted in every other population (y-axis, black points) as well as the expectation for this fraction (red points). The expected overlap between populations i and j is calculated by comparing the observed adaptive lineages in population i to a random set of lineages from population j that reach detectable frequency. All lineages are considered equally for sampling.               Table S1. P -values for comparison of absolute abundances and net population change rates across conditions. P -values are obtained from t-tests and corrected for multiple testing using the Benjamini-Hochberg procedure.

Mutation class m/Locus
Culture isolates chrIV-1n chrIII-3n chrXIV-3n chrIV-3n chrIX-3n chrXIII-3n YIL169C HEM1 HEM2 A1  Table S3. Numbers of isolates carrying the most common adaptive mutations and the estimated mutation rates. Only clones with unique barcodes are shown. Column "Seq. isolates" shows the total number of sequenced isolates with unique barcodes sampled from each culture. Last row shows the estimated u m = log 10 U m .

46
Condition R 2 P β r P r β K P K A 0.067 0.14 0.15 0.72 0.73 0.05 C 0.374 2 × 10 −6 −0.81 0.002 0.72 0.002 Table S4. Multiple regression analysis of fitness against life-history traits. β r and β K are standardized regression coefficients for r and K, respectively. P r and P K are P -values indicating whether adding r or K to the model significantly improves the fit above and beyond the single-variable model.