APOK3, a pollen killer antidote in Arabidopsis thaliana

According to the principles of heredity, each parental allele of hybrids equally participates in the progeny. At some loci, however, it happens that one allele is favored to the expense of the other. Gamete killers are genetic systems where one allele (the killer) triggers the death of the gametes carrying the other (killed) allele. They have been found in many organisms, and are of major interest to understand mechanisms of evolution and speciation. Gamete killers are particularly prevalent in plants, where they can compromise crop breeding. Here, we deciphered a pollen killer in Arabidopsis thaliana by exploiting natural variation, de novo genomic sequencing and mutants, and analyzing segregations in crosses. We found that the killer allele carries an antidote gene flanked by two elements mandatory for the killing activity. We identified the gene encoding the antidote, a chimeric protein addressed to mitochondria. This gene appeared in the species by association of domains recruited from other genes, and it recently underwent duplications within a highly variable locus, particularly in the killer genotypes. Exploring the species diversity, we identified sequence polymorphisms correlated with the antidote activity.


33
Genetic loci that do not comply with Mendel's laws have been observed since the dawn of 34 genetics. First considered as being genetic curiosities, these loci with transmission ratio 35 distortion (TRD) are now recognized to be common in fungi, plants and animals, with a 36 particularly high incidence in plants (Fishman and McIntosh 2019). They are of major interest 37 to understand genomic evolution, adaptation and speciation (Presgraves, 2010;Lindholm et al, 38 most of the biased segregations, as expected if there is a gametophytic defect in these hybrids 114 (Table1 Source Data 1). On the other hand, among 18 accessions analyzed for a killer 115 behaviour, five induced a bias in the progeny of their hybrid with Sha, as did Mr-0 (Table 1), 116 with a distribution of genotypes that was consistent with a gametophytic defect ( Table1 Source  117 Data 2). Among the 14 accessions that were tested for both killed and killer status, five 118 accessions, including Col-0, had a neutral behaviour (neither killed nor killer). And, coherently, 119 none of these 14 accessions was found to have both a killed and a killer behaviour (Table 1). 120  Table1 We previously showed that a plant segregating Sha and Mr-0 alleles only at L3 while fixed Sha 133 in the rest of its nuclear genome (hereafter ShaL3 H ) presented a strong bias against Sha 134 homozygous progenies, which was linked to a deficit in pollen grains carrying the Sha allele 135 (Simon et al. 2016). Here, we tested whether the segregation bias was dependent on the genetic 136 background by comparing the selfed progenies of ShaL3 H with its equivalent in the Mr-0 137 nuclear background (MrL3 H ). Their progenies showed very similar deficits in Sha homozygotes 138 (Table 2), indicating that the segregation distortion was independent of the fixed parental 139 nuclear background. Accordingly, anthers from both genotypes showed similar proportions of 140 dead pollen (Figure 2). 141  We further studied the ShaL3 H genotype, which flowers earlier than MrL3 H . We observed a 145 strong distortion when ShaL3 H was used as male in a cross with Sha, but no distortion when it 146 served as female parent (Table 3). These results confirmed that the segregation distortion was 147 only due to a PK. 148  and about 35% of mature pollen grains were dead in ShaL3 H plants (Figure 4), which reflects 158 an incomplete penetrance since 50% dead pollen would be expected if the PK effect was total. 159 We concluded that, in these plants, the Sha allele at L3 is poorly transmitted because most of 160 the Sha pollen grains fail to develop properly from the binucleate stage and eventually die. 161

162
The PK at L3 contains three genetic elements 163 L3 was previously mapped in a 280 Kb interval at the bottom of chromosome 3 (Simon et al. 164 2016). We fine-mapped the PK in the genotype ShaL3 H using the presence of a segregation bias 165 in the self-descent of recombinants as a robust phenotypic trait to narrow down the L3 interval: 166 genetic markers fixed for Sha or Mr-0 alleles in recombinants with a significant bias in their 167 progeny were excluded from the candidate interval. This strategy allowed us to map all the 168 genetic elements necessary for the PK activity in an interval hereafter called PK3, 169 corresponding to the region flanked by markers M5 and M13 (68 Kb in Col-0) ( Figure 5A). 170 Out of a total of 4,717 plants genotyped, we found six recombinants between M5 and M13. 171 Recombination points of these recombinants, 27D6, 25A7, 52D12, 52D7, 8F10BH2 and 23G9 172 were finely localized. Only three of these plants, 52D7, 8F10BH2 and 23G9, were recombined 173 between M6 and M12, which are very close to M5 and M13, respectively (Figure 5 Source Data 174 1). None of them presented a bias in its progeny whereas 27D6, 25A7 and 52D12, which are 175 heterozygous between M6 and M12, did. Pollen viability of the recombinants, assessed by 176 Alexander stainings, were consistent with the presence or absence of bias in their progenies 177 ( Figure 5B). Further information on the genetic structure of the PK3 was obtained from the 178 three plants recombined between M6 and M12. First, by crossing 52D7 and 8F10BH2 with Sha, 179 we converted their fixed portion of the interval into a heterozygous region ( Figure 5A). The 180 offsprings of these new genotypes (named i-52D7 and i-8F10BH2) did not show any 181 segregation bias, whereas their siblings heterozygous along the whole interval did ( Figure 5  182 Source Data 2). This indicated that the parts of the PK3 interval that were heterozygous in 52D7 183 and in 8F10BH2 both contained elements necessary for the PK activity. We thus segmented the 184 PK3 locus into three genetic intervals, named PK3A, PK3B and PK3C ( Figure 5A), PK3A and 185 PK3C both carrying elements necessary for a functional PK. Then, in order to evaluate the 186 PK3B interval, we crossed fixed progenies of 23G9 and 52D7 that inherited the recombination 187 events from their parents, and obtain a plant (23G9#15 x 52D7#7) heterozygous at both PK3A 188 and PK3C and fixed Mr-0 at PK3B ( Figure 5A). The absence of segregation bias in the selfed 189 progeny of this plant showed that PK3B also carried an element necessary for the PK activity. 190 Therefore, each of the three parts of the PK3 interval contains at least one element required for 191 the PK activity. On one hand, Mr-0 alleles were required at PK3A and at PK3C: these two 192 intervals thus contain killer elements. On the other hand, when the Sha allele was absent at 193 PK3B while PK3A and PK3C were heterozygous, the PK was no longer active, indicating that 194 either a target element from Sha was missing, or an antidote from Mr-0 was present in all the 195 pollen grains produced. 196

197
The PK3 locus is highly variable 198 To highlight differences between Sha and Mr-0, we sequenced the entire locus in both 199 accessions. The overall structure of the PK3 locus in Sha was very similar to that of Col-0 200 ( Figure 6A), the main differences being the deletion of the transposable element (TE) 201 AT3G62455, a 1308 bp insertion in the intron of AT3G62460, and an insertion of approximately 202 1 Kb in the intergenic region between AT3G62540 and AT3G62550. In contrast, the PK3 locus 203 in Mr-0 locus was particularly complex as compared to Col-0 and Sha, showing many structural 204 variations such as large deletions, insertions, duplications and inversions ( Figure 6A). Two TEs 205 present in Col-0, AT3G62455 and AT3G62520, were missing in Mr-0. The TEs AT3G62475, 206 AT3G62480 and AT3G62490 that are located in the PK3A region of Sha, were absent from this 207 region in Mr-0, but its PK3B region presented a large insertion of over 20 Kb that contained 208 several TEs including AT3G62475, AT3G62480 and AT3G62490 homologues. Nonetheless, 209 the same protein coding genes are present in the three genotypes, even though Mr-0 has two 210 copies of the gene AT3G62510 and three copies of the genes AT3G62528, AT3G62530 and 211 AT3G62540, with one copy of AT3G62530 and AT3G62540 being inserted into the second 212 intron of AT3G62610 ( Figure 6A). 213 Because the PK3 locus is highly rearranged between these three accessions, we looked at other 214 variants of known status for the PK phenotype. The entire genomes of 10 such variants, four 215 killers, three killed and three neutral, were de novo sequenced. PK3 sequence alignments 216 revealed structural variations relative to Col-0 in all the killers ( Figure 6B). On the contrary, 217 PK3 loci of killed and neutral natural variants, excepted Bur-0, are mostly colinear with Col-0 218 ( Figure 6B). When comparing the synteny of protein coding genes at the locus between A. 219 thaliana accessions and A. lyrata (Figure 7), we observed that the locus structure in Ita-0 is 220 similar to that of A. lyrata. This locus structure is also found in other Brassicaceae, such 221 Boechera stricta (Figure 7). In contrast, Col-0, the two other neutral accessions Blh-1 and Oy-0, 222 and Sha have a duplication of the A. lyrata gene AL5G45290, that encodes a pentatricopeptide 223 repeat protein (PPR), resulting in two nearly identical genes (AT3G62470 and AT3G62540). 224 Five additional protein coding genes (AT3G62499 to AT3G62530) compared to the A. lyrata 225 and Ita-0 sequences were found between the two PPR paralogues in these accessions. We also 226 observed a great variability in the number of copies of all these genes according to the 227 accessions. Compared to Col-0, Blh-1, Oy-0 and Sha, where they were present once, some of 228 them were absent in the other killed accessions Cvi-0 and Are-10. On the contrary, in the killers, 229 they have undergone a variable number of duplications, with one to four copies depending on 230 both the genes and the accessions (Figure 7). 231 Given the complexity and diversity of the locus, no obvious candidate gene appeared for killer 232 or target/antidote elements. For the killer elements, an additional difficulty results from the fact 233 that two elements are necessary for the killing activity, one located in the PK3A interval and 234 the other in the PK3C interval, thereby preventing to draw conclusions from the comparison of 235 killer and non-killer alleles in a single interval. Moreover, the intervals PK3A, PK3B and PK3C 236 were delimited by the mapping recombinants between Sha and Mr-0, and, due to the structural 237 differences between the accessions, it is not possible to infer their frontiers in the other 238 genotypes, at least for the limit between PK3A and PK3B. However, it can be noted that the 239 locus of the neutral accession Bur-0 is nearly identical to that of the killer Jea: only the copy of 240 AT3G62530 inserted into AT3G62610 in the PK3C interval in all the killers is missing in Bur-0 241 ( Figure 7). This makes this copy, which is absent in all the non-killer alleles, a gene to be tested 242 for killer activity, although it cannot be excluded that Bur-0 lost its killer activity due to another 243 polymorphism in PK3C or in its unknown PK3A killer element. 244 We thus focused on the PK3B interval, which contains either a target element in killed alleles 245 or an antidote in killer and neutral alleles. However, comparison of the PK3B sequences did 246 not reveal a gene specific for the killed alleles that could encode a target, nor a gene specific 247 for the killer and neutral alleles that could encode an antidote. We therefore exploited mutants 248 in each of the genes present in this region. 249 250 AT3G62530, expressed in young developing pollen, encodes the antidote. 251 Col-0 had a neutral behaviour regarding the PK (Table 1). That means that, in a poison-antidote 252 model, Col-0 would carry an antidote element whose inactivation should let Col-0 pollen 253 unprotected against Mr-0 killer. A hybrid between Col-0 with an inactivated version of the 254 antidote and Mr-0 would then present dead pollen, and produce an F2 with a segregation bias 255 against the Col-0 allele at L3. In order to test the poison-antidote model, we thus used T-DNA 256 insertion mutants available in this genetic background for the eight PK3B genes. None of the 257 homozygous mutants (Col mut ) was distinguishable at the phenotypic level from Col-0 in our 258 greenhouse conditions. We crossed each Col mut with an early-flowering Mr-0 genotype carrying 259 a KO mutation in the FRIGIDA gene, hereafter named Mrfri (see Materials and Methods). All 260 the Mrfri x Col mut F1 plants exhibited only viable pollen, except the hybrid with a T-DNA 261 insertion in AT3G62530 (Col mut530 ) that presented aborted pollen ( Figure 8). Then, we 262 genotyped the F2 families at L3, which all followed the expected Mendelian proportions, with 263 the exception of the Mrfri x Col mut530 and Mrfri x Col mut499 F2s. The latter presented a slightly 264 biased segregation against the Col mut499 allele, but when the Mrfri x Col mut499 F1 was crossed as 265 female or male parent using Mrfri as a tester, no mutant allele transmission bias was detected 266  (Table 4). These results strongly suggested that AT3G62530 encoded an 274 antidote. If this is the case, the mutation in AT3G62530 should have no effect on pollen viability 275 nor on allele segregation in the absence of a killer allele. Indeed, no dead pollen was observed 276 in the anthers of homozygous nor heterozygous Col mut530 (Figure 8 Supplement 2), and no 277 segregation bias against the Col mut530 allele was detected in the self-progeny of the heterozygous 278 Col mut530 nor in the progeny of the Sha x Col mut530 hybrid (Table 4). We concluded that the 279 presence of a killer allele was necessary to trigger the death of Col mut530 pollen. This was 280 confirmed by crossing Col mut530 with another killer accession, Ct-1: while no bias was found in 281 the progeny of the Ct-1 x Col-0 hybrid, a bias against the Col mut530 allele was observed in the 282 progeny of Ct-1 x Col mut530 (Table 4), which was very similar to the bias against the Sha allele 283 observed in the progeny of the Ct-1 x Sha cross. 284
(2) genotyped with PCR primers used to characterise the mutant 286 (Figure 8 Supplement 2). (3) The progeny of the F1 (Ct-1 x Col-0) presents no TRD (Table 4  287 Supplement 1). (4) frequency of Col mut530 homozygotes (expected frequency 0.25). *** p< 288 0.001; NS, not significant. 289 290 In the frame of the poison-antidote model, the antidote must be expressed in cells that need to 291 be protected from the poison elements, in particular in developing pollen. Indeed, RT-PCR 292 assays showed that AT3G62530, in addition to being expressed in leaves, is expressed in 293 microspores (young developing pollen, before the first pollen mitosis, Figure 4) from plants 294 either Col-0, Sha or Mr-0 at the locus ( Figure 8). In addition, AT3G62530 seems one of the 295 most expressed gene in microspores amongst those of the PK3B interval, in the three genotypes. 296 All together, the above results fit perfectly with a poison-antidote system where AT3G62530 297 codes the antidote, Col-0 and Mr-0 carrying functional forms of the antidote while Sha has a 298 non-functional antidote allele. We named the gene APOK3, for ANTIDOTE OF POLLEN 299 KILLER ON CHROMOSOME 3. 300 301 APOK3 encodes a chimeric protein addressed to mitochondria 302 Because APOK3 is expressed at similar levels in microspores carrying the Sha allele, not 303 protected by an antidote, and microspores carrying Col-0 or Mr-0 alleles, which must harbor 304 an active antidote (Figure 8), we hypothesized that the allelic differences for antidote activity 305 are due to differences at the protein level. 306 In Col-0, APOK3 encodes a protein of 221 amino acids, belonging to the ARM-repeat 307 superfamily (https://www.arabidopsis.org/), defined by the presence of tandem repeats 308 generally forming alpha-helices (https://supfam.org/). Further analysis of structural domains 309 identified three HEAT repeat domains ( Figure 9A). Different parts of APOK3 are very similar 310 to parts of proteins encoded by the genes AT3G62460 (98% identity on residues 1 to 44), 311 AT3G43260 (66% identity and 73% similarity on residues 43 to 142) and AT3G58180 (57% 312 identity and 78% similarity on residues 81 to 221) ( Figure 9B). These three genes are located 313 on the chromosome 3, and it is interesting to note that AT3G62460 is in the PK3A interval. 314 AT3G43260 and AT3G58180 are both annotated as related to deoxyhypusine hydroxylases of 315 other organisms, but a close examination showed that the genuine Col-0 deoxyhypusine 316 hydroxylase is encoded by AT3G58180 (Figure 9 Supplement 1). The part of the AT3G62460 317 protein shared by APOK3 includes a mitochondria-targeting peptide ( Figure 9A), suggesting 318 that APOK3 is addressed to mitochondria. Indeed, APOK3 has been repeatedly found in A. acts in the mitochondria, the strength of the bias due to the PK could be influenced by the 322 mother plant cytoplasmic background. As a first insight in this direction, we compared the 323 segregation distortion in the progenies of plants heterozygous Sha/Mr-0 at PK3 differing only 324 by their cytoplasmic backgrounds, and we observed that the bias was stronger in the Sha than 325 in the Mr-0 cytoplasmic background ( Figure 9C). We concluded that the antidote function of 326 APOK3 is likely to be sensitive to variation in the mitochondrial genome. 327 328 APOK3 has undergone several duplication events within killer PK3 loci 329 Mr-0 has two strictly identical copies of APOK3 in the PK3B interval, which have the same 330 structure as Col-0 and Sha genes. Mr-0 also has a third copy inserted with other sequences in 331 the intron of AT3G62610 in the PK3C interval ( Figure 10A), but the N-terminal part of this 332 copy differs from that of the other two and is not homologous to AT3G62460, which suggests 333 that this copy is functionally different from the others; it was thus named APOK3-like. 334 Our analysis of the natural variation at the PK3 locus revealed that all the killers analysed have 335 multiple copies of APOK3 (in yellow on Figure 7), with Mr-0 and Ct-1 having two copies, and 336 Cant-1, Jea and Shigu-2 having three copies each. In addition, these five accessions have an 337 APOK3-like copy inserted in AT3G62610. Bur-0, which has a neutral behavior, also possesses 338 three copies of APOK3, but no APOK3-like, even if it has an insertion of AT3G62540 in the 339 intron of AT3G62610. The other neutral accessions and the killed ones have only one copy of 340 APOK3, except Ita-0, in which no copy exists at the locus ( Figure 7). No other copy of APOK3 341 was found elsewhere in the Ita-0 genome, neither by searching in the genomic sequence nor by 342 PCR amplification in Ita-0 genomic DNA. Similarly, we did not find any gene encoding a 343 protein closer to APOK3 than to AT3G58180 neither in A. lyrata nor in other Brassicaceae 344 sequences available in the databases. Altogether, these results suggest that APOK3 is specific 345 to A. thaliana and has evolved within the species. 346 347 Antidote and non-antidote forms of APOK3 differ by three amino acids 348 In order to further explore the sequence variation of APOK3 in relation with its antidote activity, 349 we amplified and sequenced APOK3 in all the accessions whose behavior for the PK3 had been 350 determined (Table 1). The APOK3 copies found in each killer accession whose genome was de 351 novo sequenced were identical. In Etna-0, the only killer accession for which no genomic 352 sequence was available, we obtained a unique Sanger sequence for APOK3, indicating that, 353 whatever the number of copies it has, they are identical. Phylogenetic analysis clustered the 354 sequences into three clades, one grouping all copies found in accessions possessing the antidote, 355 i.e. killer and neutral. A second cluster groups most of copies from the killed accessions, with 356 the exception of Are-1, Are-10 and Kas-2 which branch together as a third clade ( Figure 10B). 357 We identified 11 different APOK3 haplotypes, five from non-killed accessions and six from 358 killed accessions (Table 5). One polymorphism located 36 pb upstream of the ATG start codon 359 and three non-synonymous SNPs in the coding sequence distinguished non-killed from killed 360 alleles. Consequently, the APOK3 proteins with an antidote activity differ from those with no 361 antidote activity by the three amino acid changes C85S, V101D and C105R (Table 5). 362 Interestingly, the last two of these amino acids are located in the first HEAT-repeat domain of 363 the protein, suggesting these changes could modify the protein interactions and could make it 364 functional or non-functional as an antidote. 365 antidotes non-antidotes Amino acid change position 6 19 23 51 81 83 85 98 101 103 105 115 141

368
(1) The names of the accessions whose entire PK3 sequences are known are in bold. The Col-0 sequence (TAIR 10) is used as reference. allele, we theoretically expect 11,5% of Sha homozygotes in the progeny, which is very close 404 to the 10% observed (Table 2). Incomplete penetrance was also observed in other reported 405 gamete killers and meiotic drivers, for example in tomato (Rick 1966) (Table 4), whereas hybrids with Col-0 or Sha had fully viable pollen 433 and did not show any segregation distortion at the locus in their progenies (Table 4). These 434 results demonstrate that the Col-0 allele of APOK3 protects pollen from the effect of killer 435 alleles in hybrids, which is the definition of an antidote. In addition, three residues in the protein 436 sequence and one SNP in the 5'-UTR are strictly associated with the antidote forms of the gene 437 (Table 5), suggesting their involvement in its protective activity. 438 APOK3 molecular function still remains to be elucidated. It is annotated belonging to the 439 ARM-repeat superfamily, which mediate numerous cellular processes including signal 440 transduction, cytoskeletal regulation, nuclear import, transcriptional regulation and 441 ubiquitination (Samuel et al. 2006). The remarkable features of this protein are its interacting 442 domains, its mitochondrial location and its chimeric structure. APOK3 has three HEAT repeat 443 domains ( Figure 9A), predicted to form super-helixes (Andrade and Bork 1995)  necessary for the killer activity but they are not carried by the same allele. From our mapping 500 results, we cannot completely exclude that PK3A or PK3C also carry rescue activities not 501 redundant with PK3B, nor that PK3B also carries killer activity not redundant with PK3A and 502 PK3C. However, it is noticeable that this structure prevents the killer activity to be isolated 503 from the antidote by a recombination event, since such an event would also separate the two 504 mandatory killer elements. 505 The PK3 locus is particularly dynamic and prone to structural variations. Besides the Ita-0 allele 506 that is structurally similar to that of A. lyrata, with orthologous genes in the same order and 507 orientation (Figure 7), the other killed and the neutral alleles (excepted Bur-0) present the same 508 organization of the locus ( Figure 6B), even if some protein coding genes are missing in Are-10 509 and Cvi-0 (Figure 7). Among the neutral alleles, Bur-0 is an interesting exception: it resembles 510 a killer both in structure ( Figure 6B) and gene content (Figure 7), and we hypothesize it evolved 511 from a killer allele that lost its killing capacity. Comparison of the Bur-0 sequence with the very 512 similar allele of Jea will probably help identifying killer elements of PK3. The most complex 513 alleles are found in killers ( Figure 6B), which are highly variable, with different groups of genes 514 duplicated in one or the other orientation and in different relative positions (Figure 7). Having 515 more than one copy of APOK3 is one remarkable common feature of the killer alleles. It is 516 likely that having multiple copies is important for killer alleles not to be the victims of their 517 own activity. Gene dosage has also been reported to be critical in a poison-antidote system in 518 C. elegans (Mani and Fay 2009). In addition, the different copies of APOK3 found in each killer 519 are identical, but may be slightly different between accessions, indicating that recent 520 duplications occurred independently in the lineages of the different variants we examined. 521 These results suggest that these alleles have experienced a more intense structural evolution 522 that neutral and killed ones, and raises the questions of the mechanisms and evolutive forces 523 leading to these structures. The presence of several transposable elements at the locus, and their 524 variation in occurrence between Sha, Mr-0 and Col-0 ( Figure 6A Even if gamete killers likely exist in all plant species, none had been investigated in A. thaliana 529 until now, to our knowlege. The PK we dissected here has some general features of eukaryotic 530 poison-antidote systems, including its species-specific nature and its presence within a hyper 531 variable locus. The layout of the killer alleles is particular, with at least three mandatory 532 elements and diverse duplications of sequence blocks that contain antidote genes trapped 533 between killer elements. Continuing to exploit the natural variation of the species should help 534 identifying the killer elements, and provide clues towards the underlying mechanisms 535 responsible for the PK activity, the role of the mitochondria, and eventually the forces driving 536 the evolution of the locus. 537 538 539 Cytological analyses 566 DAPI staining of spread male meiotic chromosomes was performed according to (Ross et al. 567 1996). All stages were observed in 3 independent ShaL3 H plants and in parental controls. 568

Materials and Methods
Observations were made using a Zeiss Axio Imager2 microscope and photographs were taken 569 using an AxioCam MRm (Zeiss) camera. Propidium iodide and Alexander staining (Alexander 570 1969) of pollen were performed as described in Durand et al (2021). 571

Fine-mapping and genotyping 572
DNA extractions were conducted on leaves from seedlings as described by Loudet et al (2002). 573 Markers for the fine-mapping are described in Figure 5 Source Data 1. For CAPS markers M3 574 and M12 we used Cac8I (NEB) and Bsp1407I (ThermoFisher) restriction enzymes, 575 respectively. Other SNPs were genotyped by sequencing. For the fine-mapping of PK3, we 576 genotyped a total of 4,717 plants and identified 42 recombinants between markers M1 and M15. 577 We selected 23 informative recombinants that were tested for segregation distortion at the locus 578 by genotyping their self-descent progenies with appropriate markers (Figure 5 Source Data 1). 579 All the other primers used in this work for mutant characterization, gene expression, PCR 580 amplification and DNA sequencing are listed in Supplemental File 1. 581

RNA extractions and RT-PCR 597
We mixed on ice 500 µl of microspore suspension in TriReagent with 300 µl glass beads 598 (Sigma, G8772-100G glass beads acid washed) and two sterilized metals beads (3mm 599 diameter). The tubes were shaked 2 x 2,5 min at a 1/30 frequency in a Retsch MM400 mixer 600 mill and the solution without beads was recovered in a new RNase-free tube by making a small 601 hole at the bottom of the tube and centrifuging 2 min at 3,900 rpm at 4°C. The beads were 602 rinsed with 500 µl of TRIreagent and centrifuged again. The eluate was centrifuged 2min at 603 11,000 rpm, 4°C to remove cellular debris. RNA was extracted using the extraction RNA kit 604 (Zymo research). Leaf RNA was prepared as above except that leaves were grinded in Trizol 605 with one metal bead. In accessions with several copies of APOK3, occasional sequence errors subsist in some copies 664 in the automatic assemblies. To correct these errors, APOK3 was Sanger sequenced after gene 665 amplification with the primers AT3G62530F1 and AT3G62530R1 (Supplemental File 1). 666

Sequence availability 667
Whole genome sequences and PK3 locus sequences are available at the doi listed in Figure 6  The locus is represented as a horizontal line with the killer and killed alleles as black and white boxes, respectively. For these general models, there is no hypothesis on the number of genes present in each box. In both models, the black allele is expressed before meiosis, and the killer (K) or poison (P) persists in all the gametes. In the killer-target model, only gametes with the white allele express the target (T), which interacts with K to trigger gamete death (represented as dashed outline). In the poison-antidote model, the black allele also produces a short-lived antidote (A) that counteracts P, so only the gametes that inherit the black allele are efficiently protected. , this diagram is not to scale. Genotypes used to dissect the PK3 are presented: i-52D7 and i-8F10BH2 result from crosses of 52D7 and 8F10BH2, respectively, with Sha; the 23G9#15 x 52D7#7 genotype, with two heterozygous regions flanking a fixed central region of the PK3, was produced by crossing appropriate selfed progenies of 23G9 and 52D7 recombinants. All markers are described in Figure 5 Figure 7: Intra and Inter-specific synteny of protein coding genes at the PK3 locus. Alignment the PK3 loci from thirteen A. thaliana accessions and two related species, Arabidopsis lyrata and Boechera stricta, drawn to highlight synteny between protein coding genes. The scheme is not to scale and TEs are not represented. For A. lyrata and B. stricta the structural annotation was obtained from Phytozome (https://phytozome.jgi.doe.gov/). The names of killer, neutral and killed A. thaliana accessions are written in blue, green and red, respectively. Plain coloured arrows represent coding genes with their orientations, each colour representing orthologues and paralogues of a same gene. Gene labels correspond to the last three digits of AT3G62xxx, AL5G45xxx and Bostr.13158s0xxx gene identifiers in A. thaliana (Col-0 reference sequence TAIR10), A. lyrata (V2.1) and B. stricta (V1.2), respectively. Black arrows above the Mr-0 and Sha loci delimit the PK3A, PK3B and PK3C intervals from left to right.