Mechanism of replication fork reversal and protection by human RAD51 and RAD51 paralogs

Petr Cejka (  petr.cejka@irb.usi.ch ) Institute for Research in Biomedicine Swagata Halder University of Oxford Aurore Sanchez Institut Curie Lepakshi Ranjha Institute for Research in Biomedicine Angelo Taglialatela Columbia University Medical Center Giordano Reginato Institute for Research in Biomedicine Bellinzona Ilaria Ceppi Universita della Svizzera Italiana Ananya Acharya Universita della Svizzera Italiana Roopesh Anand Universita della Svizzera Italiana Alberto Ciccia Columbia University Medical Center https://orcid.org/0000-0003-4789-6564


INTRODUCTION
repair and fork metabolism/modulation is genetically separable 38 . Recently, the BCDX2 RAD51 pa-121 ralog complex was found to promote fork reversal alongside RAD51 in cellular assays 39 .

122
The seemingly paradoxical situation where RAD51 may both promote and prevent DNA degradation 123 highlights the difficulties explaining cellular phenotypes and thus warrants the need for a more direct 124 understanding of the underlying mechanisms 37 . Here we use reconstitution biochemistry, which allows 125 us to study elements of the reactions governing the metabolism of challenged replication forks in isola-

141
142 SMARCAL1, ZRANB3 and HLTF have unequal biochemical activities 143 SMARCAL1, ZRANB3 and HLTF have all been implicated in replication fork reversal in vitro and in 144 vivo 22,40,41 . The loss of either of these enzymes was shown to abolish nascent DNA degradation in 145 BRCA1/2-deficient cells, suggesting that these factors may act in a non-redundant manner to promote 146 fork reversal 19 . To better understand the function of these fork remodelers, we expressed and purified 147 SMARCAL1, ZRANB3 and HLTF from insect Sf9 cells (Fig. 1a). All three translocases hydrolyzed 148 ATP, as expected, with SMARCAL1 showing the highest specific activity, followed by HLTF and 149 ZRANB3 ( Supplementary Fig. 1a). We next set out to compare the relative activities of these motor 150 proteins in biochemical assays mimicking elements of fork reversal. We first used oligonucleotide-151 based DNA substrates resembling stalled replication forks with ssDNA gaps either in the leading or the 152 lagging DNA strand (Fig. 1b). We observed, as reported previously, that SMARCAL1 in the presence 153 of the ssDNA binding protein RPA was more efficient on forks with leading strand gaps, as opposed to 154 ZRANB3, which prefers RPA on lagging strand gaps 42 (Fig. 1c). Using the leading strand gap substrate, 155 SMARCAL1 was also more efficient than HLTF (Fig. 1c), while the three translocases exhibited similar 156 specific activities on the lagging strand gap substrate (Fig. 1c). In contrast to the activities of 157 SMARCAL1 and ZRANB3 that are regulated by RPA 42 , the function of HLTF was not RPA sensitive 158 (Fig. 1d). During fork reversal, the initial annealing of the nascent DNA strands leads to the formation 159 of a 4-way junction (Holliday junction, HJ), which is further branch migrated by the motor proteins, 160 leading to reversed forks of up to several kilobases in length 16 . Using a mobile HJ substrate to assay for 161 branch migration, we observed that SMARCAL1 was in contrast the least efficient enzyme, essentially 162 incapable of branch migration at physiological (150 mM) salt concentrations. However, under less re-163 strictive conditions in lower salt, the branch migration activity of SMARCAL1 was readily detected 164 (Fig. 1f,right panel). Instead, both ZRANB3 and HLTF were highly and comparably efficient in branch 165 migration at 150 mM salt (Fig. 1e, f).

166
The activity of the SMARCAL1 and ZRANB3 enzymes was first analyzed by a topoisomerase-coupled 167 assay that monitors the annealing of RPA-coated ssDNA bubbles in plasmid DNA, which can be ob-168 served as changes in DNA topology 25,43,44 . Such activity is thought to mimic the initial stages of fork 169 remodeling. Both SMARCAL1 and ZRANB3 were shown to anneal the RPA-coated DNA bubbles as 170 a result of their motor functions, as ATP hydrolysis is required for this reaction. However, the specific 171 activities of SMARCAL1 and ZRANB3 have not been compared. We observed that SMARCAL1 was 172 comparably efficient to ZRANB3, while HLTF showed much smaller capacity to anneal DNA in this 173 assay (Fig. 1g, h).
Taken together, our data indicate that SMARCAL1, ZRANB3 and HLTF possess quite different bio-175 chemical activities and substrate preferences. Our results support a model positing that fork reversal is 176 not catalyzed by a single enzyme in a processive manner, but that it is rather a dynamic process that 177 involves the sequential engagement of several factors. 178 179 SMARCAL1 specifically anneals RPA-coated ssDNA 180 Several helicases, such as members of the RecQ family, were reported to anneal two ssDNA molecules, 181 but the reactions were inhibited by RPA 45,46 . Considering that cellular RPA concentration is thought to 182 be sufficient to coat all ssDNA in most cases, the physiological relevance of these observations remains  Fig. 2a), as is the case of RAD52 48 . Therefore, the annealing of two 202 ssDNA molecules by SMARCAL1 mechanistically differs from the annealing of bubbled DNA in the 203 topoisomerase coupled assays, which largely require ATP hydrolysis 25,43,44 .

204
The N-terminal region of SMARCAL1 contains a previously defined RPA-binding site, the integrity of 205 which is required for the recruitment of SMARCAL1 to DNA damage sites, and to direct SMARCAL1 206 to substrates with RPA-coated ssDNA gaps 21 (Fig. 2e). To test for the requirement for direct interaction 207 between SMARCAL1 and RPA in ssDNA annealing, we expressed and purified SMARCAL1 (DN), 208 lacking the RPA interaction domain (Fig. 2f). The truncated SMARCAL1 was fully proficient in DNA 209 branch migration in the absence of RPA and identical to wild type SMARCAL1 as an ATPase (Sup-210 plementary Fig. 2b, c). However, the mutant was inefficient in ssDNA annealing in the presence of 211 RPA, showing that the direct interaction of SMARCAL1 with RPA is essential for this activity (Fig.   212 2g). Our results reveal that SMARCAL1 possesses a strand annealing activity similar to that of the 213 RAD52 protein family, which also rely on specific interaction with RPA 48 . We suggest that such activity 214 may be employed during the very initial steps of fork reversal, when the daughter ssDNA molecules 215 are separated from the parental strands and anneal with each other. These results further underline the 216 mechanistic differences between the fork remodeling enzymes SMARCAL1, ZRANB3 and HLTF. 217 218 RAD51 and BCDX2 paralogs promote motor-driven strand annealing activity of SMARCAL1 219 and ZRANB3 but not HLTF 220 Challenged DNA replication forks may undergo reversal, and reversed replication forks must be sub-221 sequently protected by RAD51 to prevent pathological nascent DNA degradation 9,10,15,16,49 . However, 222 RAD51, along with the RAD51 paralog BCDX2 complex, were also paradoxically implicated in pro-223 moting fork reversal, through a yet unknown mechanism 17, 39 . To elucidate whether the function of 224 RAD51 and the RAD51 paralogs in fork remodeling may be direct, we next expressed and purified 225 RAD51 and the BCDX2 complex (Fig. 3a). The BCDX2 complex was obtained upon co-expression of 226 all subunits, as the preparation of the individual proteins resulted in poor yields and solubility. The

227
BCDX2 complex did not aggregate, bound ssDNA and very weakly hydrolyzed ATP, as observed pre-

229
We next set out to test whether RAD51 and BCDX2 complex affect the strand annealing and motor 230 activities of SMARCAL1. To this point, we used the established topoisomerase-coupled assay 25,43,44 .

231
Strikingly, we observed that low concentrations of RAD51 and the BCDX2 complex promoted bubbled 232 DNA annealing by SMARCAL1, while none of these co-factors had a notable capacity to mediate DNA 233 annealing per se without SMARCAL1, even at much higher concentrations (Fig. 3b (Fig. 3b, Supplementary Fig. 3g). Similarly to SMARCAL1, we observed that RAD51 238 and the BCDX2 paralogs promoted the annealing capacity of ZRANB3. As above with SMARCAL1, 239 we observed that RAD51 and BCDX2 could both promote ZRANB3 independently of each other (Fig.   240 3c). Only limited changes in DNA topology were observed when using helicase-dead ZRANB3 and all 241 co-factors, demonstrating that the majority of the signal in the assay can be linked to the motor activity 242 of ZRANB3 leading to DNA annealing ( Supplementary Fig. 3h).

243
Differently from SMARCAL1 and ZRANB3, HLTF was not stimulated by either RAD51 or BCDX2 244 (Fig. 3d). The RAD51C-XRCC3 (CX3) complex could also promote DNA annealing with both 245 SMARCAL1 and ZRANB3 (Supplementary Fig. 3i,j,k), despite it has not been found to promote fork 246 reversal in vivo 39,51  with the co-factors. We observed that BCDX2 interacted with RAD51 (the RAD51B component was 257 detected), as described previously 52 . Importantly, we found that both SMARCAL1 and ZRANB3 also 258 interacted with RAD51 (Fig. 4a). We next immobilized the BCDX2 complex, and observed a direct 259 interaction with SMARCAL1 and ZRANB3, as detected by Western blotting and silver staining ( Fig.   260 4b, c). These results collectively suggest that the interplay of SMARCAL1 and ZRANB3 with RAD51 261 and BCDX2 likely involves direct physical interactions.

264
We next set out to define motifs in SMARCAL1 and ZRANB3 that mediate the interactions with 265 RAD51 and BCDX2. We failed to identify an interaction motif with BCDX2, but we found regions in 266 SMARCAL1 mediating the binding to RAD51. Physical and functional interactions between RAD51 267 and many of its co-factors, such as BRCA2, BARD1, MMS22L, RECQL5, SWSAP1 and FINGL1 are 268 mediated by the FXXA motif 53,54,55,56,57,58 . We identified such a motif in SMARCAL1, which is 269 conserved in evolution (Fig. 4d, Supplementary Fig. 4a). The FXXA motif is positioned ahead of the 270 conserved aI SNF2 family ATPase domain ( Supplementary Fig. 4b). The mutation of phenylalanine 271 446 into alanine (F446A) in SMARCAL1 disrupted the physical interaction with RAD51 (Fig. 4e, f).

272
In contrast, disruption of F439, which is part of a less conserved FXXA sequence in human 273 SMARCAL1 upstream of F446, did not impair the interaction (Fig. 4f , Supplementary Fig. 4a). We 274 note that SMARCAL1 F446A variant per se was very similar to wild type SMARCAL1 in its fork 275 reversal and ATPase capacities in vitro and retained its physical interaction with the BCDX2 complex 276 (Supplementary Fig. 4c,d and Fig. 4g). ZRANB3 contains a phenylalanine at the analogous position to 277 SMARCAL1 ahead of the ATPase domain. The phenylalanine however does not conform to the FXXA 278 motif, and the F47A substitution mutant retained its capacity to interact with RAD51 and was impaired 279 in its ATPase activities . We next found that mutation F736A in ZRANB3 280 disrupted interaction with RAD51, however it is likely that the mutation affected the fold of the sub-281 strate recognition domain, as it likewise abolished the biochemical activities of ZRANB3 and may thus 282 not represent a direct interaction motif (Supplementary Fig. 4i-

302
In homologous recombination, the processing of DSBs is initiated by short-range resection that involves 303 the endo-and exonuclease activities of MRN and CtIP, followed by the long-range pathways catalyzed 304 by EXO1 and/or DNA2-BLM/WRN 4, 5, 60 . The same nucleolytic pathways were shown to degrade nas-305 cent DNA at stalled replication forks, with relative contributions of the nucleases dependent on condi-306 tions and genetic background 9,27,30,32,34,35 . To investigate the function of RAD51 in DNA protection, 307 we reconstituted DNA end resection reactions without or with various concentrations of RAD51. We 308 observed that RAD51 inhibited all DNA end resection reactions tested, including the exonuclease ac-309 tivity of MRE11-RAD50, the endonuclease of MRE11-RAD50-NBS1 in conjunction with phosphory-310 lated CtIP (Fig. 5a-e), and the long-range pathways of EXO1 and DNA2-WRN (Fig 5f-k). These ex-311 periments allowed us to make several conclusions.

312
First, higher RAD51 concentrations were in general required for DNA protection compared to motor-313 driven strand annealing. Approximately 300 nM to low micromolar RAD51 was required for a robust 314 protection against nucleolytic degradation, depending on the substrate and the respective nuclease ana-315 lyzed. The RAD51 concentrations required for the inhibition of the endonuclease of MRN-CtIP were 316 similar to those required to inhibit the nonspecific NspI endonuclease, which may be used as a read-out 317 for RAD51 filament formation (Fig. 5a, b and Supplementary Fig. 5a). These results suggest that effi-318 cient inhibition of DNA end resection occurs under conditions permissive for stable RAD51 filament 319 formation.

320
Second, the DNA affinity of RAD51 directly corresponds to its efficacy in inhibiting DNA degradation.

321
To this point, we used the RAD51 variants that differ in their capacity to bind DNA 61, 62, 63 . The tightly 322 binding RAD51-KR mutant generally inhibited resection more efficiently than wild type RAD51, while 323 the poorly DNA binding RAD51 variants KA, YA and TP were largely deficient in protection h,i). Furthermore, the affinity of RAD51 to DNA depends on the presence of ATP. ATP binding 325 stimulates DNA binding by RAD51, while ATP hydrolysis leads to RAD51 dissociation from DNA 61 .

326
Correspondingly, the highest DNA binding affinity of RAD51 is observed in buffers with the non-327 hydrolysable ATP analog ATP-g-S. In accord with the modulation of DNA binding affinity of RAD51 328 by the nucleotide co-factors, we observed the strongest DNA protection by RAD51 when reactions 329 contained ATP-g-S, followed by ATP, and the weakest protection in reactions lacking ATP ( Fig. 5e-g).

330
These latter experiments could only be performed with resection nucleases that do not require ATP, 331 such as the exonuclease of MR and the exonuclease of EXO1.

332
Third, we did not find any apparent specific functional interactions between RAD51 and the DNA end 333 resection nucleases. To this point, we compared the capacity of human and yeast RAD51/Rad51 to 334 protect DNA against the human and yeast resection nucleases and nuclease complexes (Supplementary 335 . We observed that human RAD51 was generally more efficient than yeast Rad51, which may 336 suggest some degree of specificity. However, human RAD51 was also more efficient in protecting DNA 337 against non-specific nucleases Exo-III and NspI ( Supplementary Fig 5i, j). These experiments sug-338 gested that the higher efficacy of human RAD51 compared to yeast Rad51 to protect DNA is not due 339 to specific functional interactions between human RAD51 and the nuclease complexes, but may be 340 rather due to a higher DNA affinity or protection capacity of human RAD51 per se.

342
The function of RAD51 in DNA protection corresponds to its capacity to bind double-stranded 343 DNA 344 BRCA2 is required for nascent DNA protection by RAD51. Interestingly, the regions of BRCA2 re-345 quired for DNA protection are distinct from those needed for homologous recombination and RAD51 346 loading onto RPA-coated ssDNA 9 . The BRCA2 BRC repeats are needed for homologous 347 recombination, and in biochemical assays they were shown to enhance the binding of RAD51 to RPA-348 coated ssDNA, while they reduce the capacity of RAD51 to bind dsDNA 9, 64, 65 . The BRC repats are 349 however not involved in DNA protection upon replication stress. Rather, the C-terminal RAD51 bind-350 ing site is required for DNA protection 9 . The biochemistry of the C-terminal site is less understood, but 351 it was demonstrated that it facilitates RAD51 binding also to dsDNA 66 . These results prompted us to 352 investigate whether RAD51 binding to ssDNA overhangs at DNA ends, as indicated in current mod-353 els 16 , indeed explains its protection function.

354
To investigate whether RAD51 binding to ssDNA or dsDNA explains its function in DNA protection,

355
we prepared blunt-ended, 5'-overhanged and 3'-overhanged substrates. The MR exonuclease was inhib-356 ited by RAD51 to the same extent irrespectively of the presence of the overhang (Fig. 6a, b). This 357 showed that the overhang is not required to assure protection against MRE11, and infers that the pro-358 tection is instead dependent on RAD51 binding to the dsDNA part of the substrate. The same result was 359 observed with EXO1. Although EXO1 shows a clear preference to process 5'-recessed strands and is 360 least efficient on 5'-overhangs 67 , RAD51 inhibited DNA degradation to a similar extent with all struc-361 tures tested (Fig. 6c, d), suggesting that overhang is not needed for protection. Likewise, the MRN-CtIP 362 endonuclease clips dsDNA, and we showed that a variety of protein blocks promote such activity, as 363 long as the blocks are located at DNA ends or DNA overhangs. Our observation that RAD51 blocks 364 the MRN-CtIP endonuclease (Fig. 5a, b) again suggests that the inhibitory function is caused by RAD51 365 binding to dsDNA. Together, these results demonstrate that the binding of RAD51 to dsDNA is respon-366 sible for its inhibitory effect on the resection nucleases, at least in vitro. We show that the function of 367 RAD51 to protect DNA from nucleolytic degradation is structural, and directly corresponds to the af-368 finity of RAD51 to bind dsDNA. Tightly-bound RAD51 filaments then serve as a non-specific barrier 369 against DNA degradation. Our results challenge the current model suggesting that the binding of 370 RAD51 to ssDNA overhangs at the apex of the reversed fork structure explains its function in protec-371 tion 16, 27, 67 . Our data instead suggest that the capacity of RAD51 to bind dsDNA is relevant for its 372 function in DNA protection, which is ultimately responsible for the maintenance of genome stability 373 during replication stress (Fig. 7).

376
Here we used biochemistry to study the function of the replication fork remodelers SMARCAL1,

377
ZRANB3 and HLTF, their regulation by RAD51 and RAD51 paralogs, as well the interplay of RAD51

381
Depletion of either SMARCAL1, ZRANB3 or HLTF brings about a profound defect in replication fork 382 reversal, as observed by electron microscopy, or by proxy methods scoring for e.g., nascent DNA deg-383 radation in various genetic backgrounds upon stress 19,22,40,41,68,69 . To better define the function of fork 384 remodelers, we compared here the specific activities of the fork remodelers on various substrates mim-385 icking elements of fork reversal. We observed that SMARCAL1, but not ZRANB3 or HLTF, has a 386 unique capacity to anneal RPA-coated ssDNA, a function reminiscent of the RAD52 protein family.

387
The annealing function of SMARCAL1 depends on the RPA interaction motif within the N-terminus 388 of SMARCAL1. We hypothesize that such annealing function might be relevant during the initial an-389 nealing of the displaced daughter strands during the early steps of fork reversal. The annealing activity 390 of SMARCAL1, similarly to RAD52, does not involve ATPase activity. Previously, the function of 391 SMARCAL1 and ZRANB3 was monitored in assays scoring for the annealing of bubbled DNA within 392 circular plasmid. Such activity, in contrast to annealing of RPA-coated ssDNA oligonucleotides, is de-393 pendent on the motor activities of the remodelers. We show that in contrast to SMARCAL1 and 394 ZRANB3, which had similar activities, HLTF was largely inactive in this assay.

395
Using oligonucleotide-based substrates mimicking fork reversal, it has been previously reported that 396 SMARCAL1 and ZRANB3 have opposing preferences with respect to whether RPA is located on the 397 leading or the lagging DNA strand 42 . We show that in contrast, the activity of HLTF does not appear to 398 be affected by RPA. In branch migration, using 4-way junction substrates without ssDNA, HLTF and Previous cellular data suggested that RAD51 and the BCDX2 complex promote fork reversal, but the 409 underlying mechanism was not clear 17, 39 . The function of RAD51 in fork remodeling was shown to be 410 genetically separable and thus different from its canonical role in homologous recombination. Specifi-411 cally, the strand exchange function of RAD51 was dispensable, pointing at a potential structural func-412 tion 38 . We show here that RAD51 and the RAD51 BCDX2 paralog complex stimulate the strand an-413 nealing and branch migration activities of SMARCAL1 and ZRANB3, two of the key enzymes impli-414 cated in fork reversal. SMARCAL1 and ZRANB3 were stimulated when RAD51 concentration was 415 too low to support a nucleoprotein filament formation.

416
In accord with a recent cellular study that identified a function of BCDX2 in promoting fork reversal 39 , 417 we find that the paralog complex also directly stimulates SMARCAL1 and ZRANB3. Unexpectedly, 418 RAD51 and BCDX2 stimulated SMARCAL1 and ZRANB3 independently of each other, as we ob-419 served mostly additive effects when combined. The function of the RAD51 paralogs, such as the 420 BCDX2 complex, in homologous recombination remains poorly defined. Some reports suggest a joint 421 function for the paralogs and RAD51. Specifically, BCDX2 was shown to have a classical recombina-422 tion mediator function to load RAD51 on RPA-coated ssDNA 6 to remodel RAD51 filaments for acti-423 vation 71 , or to make them more resistant against disruption 72 . However, RAD51-independent function 424 of the RAD51 paralogs were also identified in cellular studies, such as in the single-strand annealing 425 pathway of DSB repair 73 , and BCDX2 was also found to physically and functionally associate with the 426 HELQ helicase 74 . The function of BCDX2 to promote SMARCAL1/ZRANB3 described here in vitro 427 also does not require RAD51.

429
The interplay of RAD51 and paralogs in promoting SMARCAL1 and ZRANB3 involves physical 430 interactions 431 RAD51 and BCDX2 did not stimulate HLTF, a third enzyme shown to catalyze fork reversal, suggest-432 ing a specificity in the interplay of SMARCAL1 and ZRANB3 with RAD51 and BCDX2. In accord, 433 we found that RAD51 and BCDX2 physically interact with SMARCAL1 and ZRANB3. We could then 434 map the RAD51 interaction site in SMARCAL1 and constructed a single point mutant (SMARCAL1-435 F446A) that disrupted the physical interaction with RAD51. The SMARCAL1 mutant was not impaired 436 in its activities per se, but was deficient in promoting nascent DNA degradation in BRCA1-deficient 437 cells, a process that requires the fork reversal activity of SMARCAL1, supporting the idea that the 438 identified interplay of SMARCAL1 and RAD51 is physiologically relevant.

440
The function of RAD51 in DNA protection is largely non-specific 441 Downstream of fork reversal, RAD51 protects nascent DNA from degradation against nucleases. We 442 reconstituted both the endo-and exonuclease activities of MRE11, as well as EXO1 and DNA2, which 443 were all implicated in nascent DNA degradation, depending on the cellular background. We then tested 444 for the effect of RAD51 on the individual nucleolytic pathways and observed that RAD51 inhibited 445 them all to a similar extent. The level of inhibition was comparable when we used other non-cognate 446 yeast or bacterial nucleases. Interestingly, human RAD51 was somewhat more efficient in DNA pro-447 tection compared to yeast Rad51, but this difference was observed in conjunction with both the cognate 448 human as well as with the non-specific nucleases, suggesting that there is no apparent functional inter-449 action between RAD51 and the resection nucleases. RAD51 therefore inhibits DNA degradation in a 450 largely non-specific manner, as a physical barrier on DNA. The concentrations of RAD51 required for 451 DNA protection were higher than those promoting fork remodeling, in accord with cellular data 75 . DNA 452 protection assays with RAD51 and non-specific nucleases are used as a proxy for RAD51 filament 453 formation. Therefore, we conclude that RAD51 filament formation is likely prerequisite for DNA pro-454 tection.
455 456 RAD51 protects nascent DNA upon binding to dsDNA 457 Contrary to current models, we provide evidence that the binding of RAD51 to dsDNA (as opposed to 458 ssDNA) may be crucial for its function in DNA protection (Fig. 7). Eukaryotic RAD51 has a similar 459 affinity to both ssDNA and dsDNA 61, 76, 77 . In homologous recombination, RAD51 needs to bind ssDNA 460 to form an active nucleoprotein filament. The binding of RAD51 to dsDNA is instead inhibitory, and 461 several recombination factors facilitate RAD51 binding specifically to ssDNA 57, 65, 77, 78 . The physiolog-462 ical relevance of the high affinity RAD51 binding to dsDNA is not known. The strong dsDNA binding 463 capacity of eukaryotic RAD51 is somewhat paradoxical, because E. coli's RecA binds preferentially to 464 ssDNA, demonstrating that stable dsDNA binding is not strictly associated with DNA strand exchange 465 activity 77, 78 . Because of the recombination paradigm, but largely without direct evidence, it has been 466 assumed that RAD51 binding to ssDNA at the apex of the reversed fork is responsible for DNA pro-467 tection 16, 27 . Our data challenge this model, and instead suggest that RAD51 binding to the dsDNA part 468 of the reversed fork is primarily responsible for its protection capacity.

469
We showed that the MRE11-RAD50 or EXO1 exonucleases were inhibited by RAD51 to a similar 470 extent when using blunt-ended dsDNA or overhanged substrates, demonstrating directly that the pres-471 ence of the overhang was not relevant for protection, at least in the reconstituted assay. RAD51 binding 472 to dsDNA also fully explains its protective function against the MRN-CtIP endonuclease ensemble.

473
MRN-CtIP only nicks dsDNA, even on overhanged substrates 3, 79 . Binding of proteins to overhangs, 474 including non-cognate factors such as streptavidin, is not inhibitory for MRN-CtIP 79 . RAD51 however 475 clearly inhibits the MRN-CtIP endonuclease when using blunt-ended dsDNA, suggesting that dsDNA 476 is the relevant substrate for RAD51 binding with respect to its protective function. We therefore propose 477 that RAD51 inhibits the resection nucleases due to its binding to dsDNA. This model better explains 478 the protective function of RAD51 in vivo, as it applies for reversed forks with any type of DNA structure 479 at the end of the reversed arm, and does not rely on the presence of an overhang to mediate DNA 480 protection. Electron microscopy analyses revealed that the regressed arms of the reversed forks are 481 mostly composed of dsDNA, irrespectively of the conditions tested 17 . Cell-based assays further indi-482 cated that fork degradation can reach multiple kilobases in length, which goes beyond the length of the 483 regressed arm of the reversed fork 17, 31 . Therefore, dsDNA is clearly being degraded at challenged rep-484 lication forks, and RAD51 is involved in its protection.

485
One of the best-known co-factors of RAD51 in DNA protection is BRCA2. BRCA2 functions as a 486 recombination mediator in homologous recombination to load RAD51 on RPA-coated ssDNA, and 487 additionally has the capacity to channel RAD51 to ssDNA away from dsDNA 64, 65 . This particular pro-488 recombination function depends on the BRC repeats of BRCA2, and could be largely recapitulated with 489 only the short BRC4 peptide 65 . However, the BRC repeats of BRCA2, despite being essential for ho-490 mologous recombination, are dispensable for fork protection 9 . Instead, the protective function of 491 BRCA2 depends on its C-terminal RAD51 binding site 9 . Strikingly, this site was previously demon-492 strated to strongly facilitate RAD51 binding to dsDNA 66 . How the site affects ssDNA versus dsDNA 493 binding is not known. The capacity of the BRCA2 C-terminus to promote RAD51 binding to dsDNA 494 thus also agrees with our direct observation that RAD51 binding to dsDNA is critical for DNA protec-495 tion in vitro. A number of additional factors including BRCA1 are also required for DNA protection 80 .

496
Whether they affect RAD51 binding to dsDNA, or facilitate DNA protection because of a different 497 mechanism, such as upon RAD51-independent DNA binding or through a direct inhibition of the nu-498 cleases, remains an interesting area of investigation for future studies. Together, our data suggest that 499 the dsDNA binding capacity of RAD51 may have evolved in conjunction with its function to prevent 500 pathological DNA degradation at the replication fork.

509
Point mutagenesis of the corresponding DNA sequences was carried out by QuikChange II site-directed 510 mutagenesis kit (Agilent Technologies), and the proteins were expressed and purified similarly as the 511 wild type counterparts. Primers used for cloning and mutagenesis are listed in Table S1.

528
The cell suspension was centrifuged for 30 minutes at 48,000 g to obtain soluble extract. The superna-529 tant was transferred to tubes containing pre-equilibrated amylose resin (New England Biolabs, 4 ml per 530 liter of Sf9 culture) and incubated for 1 hour with continuous rotation. The resin was collected by spin-531 ning at 2,000 g for 2 minutes and washed extensively batch-wise and also on a disposable 10 ml column 532 (ThermoFisher) with amylose wash buffer (50 mM Tris-HCl pH 7.5, 5 mM β-ME, 1 mM PMSF, 10% 533 glycerol, 1 M NaCl). The final wash was performed at 300 mM NaCl. Protein was eluted with amylose 534 elution buffer (50 mM Tris-HCl pH 7.5, 5 mM β-ME, 1 mM PMSF, 10% glycerol, 300 mM NaCl, 10 535 mM maltose [Sigma]) and the total protein concentration was estimated by Bradford assay. To cleave 536 off the maltose-binding protein (MBP) tag, 1/6 (w/w) of PreScission Protease, with respect to total 537 protein concentration in the eluate, was added and incubated for 1 hour at 4 °C. The sample was then 538 supplemented with 10 mM imidazole and further passed through pre-equilibrated (amylose elution 539 buffer supplemented with 10 mM imidazole) Ni-NTA agarose resin (Qiagen) on a disposable column 540 for 1 hour in flow. The Ni-NTA resin was washed 4-times with Ni-NTA wash buffer (50 mM Tris-HCl 541 pH 7.5, 5 mM β-ME, 1 M NaCl, 10% glycerol, 1 mM PMSF, 40 mM imidazole). Prior to elution, the 542 protein was washed once with the same Ni-NTA wash buffer as above but with 150 mM NaCl. Protein 543 was eluted in the same buffer supplemented with 300 mM imidazole, and subsequently dialyzed (50 544 mM Tris-HCl pH 7.5, 5 mM β-ME, 100 mM NaCl, 10% glycerol, 0.5 mM PMSF), sub-aliquoted, snap 545 frozen and stored at -80 °C for later use.

554
XRCC2co. XRCC3 was synthesized as BamHI-FLAG-XRCC3-NotI and cloned into the above vector 555 to remove the RAD51B sequence to obtain pFB-FLAG-XRCC3co-10xHis-RAD51Cco. Baculoviruses 556 expressing RAD51B-RAD51C, RAD51D-XRCC2 and XRCC3-RAD51C were prepared separately 557 and Sf9 cells were co-infected with optimized ratios for these viruses to express the BCDX2 complex 558 as a heterotetramer and the CX3 complex as a heterodimer. Both the complexes were purified in an 559 identical manner using affinity chromatography. Cells were harvested 52 hours post infection, washed 560 once with cold PBS, and the pellets were frozen in liquid nitrogen and stored at -80 °C until further use.

561
The subsequent steps were carried out on ice or at 4 °C. The cell pellet was resuspended in lysis buffer 562 (50 mM Tris-HCl pH 7.5, 2 mM β-ME, 1:400 [v/v] protease inhibitor cocktail [Sigma], 1 mM PMSF, 563 30 μg/ml leupeptin (Merck), 20 mM imidazole) for 20 minutes. Then, 50% glycerol was added to a 564 final concentration of ~16%, followed by 5 M NaCl to a final concentration of 305 mM. The suspension 565 was incubated for additional 30 minutes with gentle agitation. The total cell extract was centrifuged at 566 48,000 g for 30 minutes to obtain soluble extract. The extract was then bound to Ni-NTA resin (Qiagen) 567 for 1 hour batch wise followed by extensive washing with Ni-NTA wash buffer (50 mM Tris-HCl pH 568 7.5, 2 mM β-ME, 300 mM NaCl, 10 % glycerol, 1 mM PMSF, 10 μg/ml leupeptin, 20 mM imidazole) 569 both batch wise and on a disposable column. The protein complexes were eluted by Ni-NTA elution 570 buffer (Ni-NTA wash buffer containing 300 mM imidazole). The eluates were diluted 1:6 with a dilu-571 tion buffer (Ni-NTA elution buffer without imidazole and 0.5 mM β-ME) and bound to FLAG resin 572 (Sigma) pre-equilibrated with dilution buffer in flow with a total contact time of ~90 minutes. Protein 573 bound FLAG-resin was washed 3-times with FLAG wash buffer (50 mM Tris-HCl pH 7.5, 0.5 mM β-574 ME, 150 mM NaCl, 10 % glycerol, 1 mM PMSF) and 2 times with the same buffer with 100 mM NaCl 575 before being eluted with a FLAG elution buffer (FLAG wash buffer with 100 mM NaCl and 150 ng/µl 576 3xFLAG peptide [Sigma]). Complexes were sub-aliquoted, snap frozen and stored at -80 °C for later 577 use.

579
Drosophila Topoisomerase I. To prepare N-terminally truncated Drosophila topoisomerase I (catalytic 580 subunit) with 6xHis tag on its C-terminus, the ND423 plasmid (a kind gift from James T. Kadonaga,

612
The sample was loaded onto pre-equilibrated HiTrap Q HP column (GE Healthcare). Buffer R (20 mM 613 Tris-HCl pH 7.5, 1 mM EDTA, 0.5 mM DTT, 10% glycerol) with 150 mM NaCl was used to wash the 614 column. The same buffer R with a salt gradient up to 700 mM NaCl was used to elute the protein. Next, 615 peak fractions were pooled and dialyzed over-night in Buffer R with 100 mM NaCl and without EDTA.

682
Substrate was then stored at -20 °C until further use.

713
2% SDS [w/v], 30% glycerol, 0.1% bromophenol blue) and 1 µl Proteinase K (20 mg/ml, Roche), and 714 incubated 10 minutes at 37 °C. The mixture was resolved by 1% agarose gel electrophoresis in 1x TAE 715 buffer, and DNA was visualized by post-staining with GelRed (Biotium) according to manufacturer's 716 instructions. The gels were then imaged (InGenius3, GeneSys) and quantitated as the fraction of near 717 or fully relaxed DNA using Image J. Graphs were generated by GraphPad Prism software.

728
Samples were loaded onto 8% polyacrylamide (19:1 acrylamide:bisacrylamide) gels in 1x Tris-borate-729 EDTA (TBE) (BIO-RAD Mini-PROTEAN system, 1 mm thick) and separated for 60 minutes at 80 V 730 at room temperature. The gels were dried using a BIO-RAD gel drier on 17 CHR paper (Whatman), 731 and were exposed to storage phosphor screens and scanned using Typhoon FLA 9500 (GE Healthcare) 732 phosphor imager. The gels were quantitated with ImageJ. Graphs were generated by GraphPad Prism

805
Tween 20 (PBS-T). The mixture was incubated for 45 minutes at room temperature with gentle mixing.

811
Purified recombinant SMARCAL1 or ZRANB3 (1 µg) was added and incubated for 1 hour at 4 °C with 812 gentle mixing, and washed 4-times with wash buffer (50 mM Tris-HCl pH 7.5, 100 mM NaCl, 0.05% 813 Triton X-100). Proteins were eluted by heating the beads for 3 minutes at 95 °C in 60 µl SDS buffer 814 (50 mM Tris-HCl pH 6.8, 1.6 % SDS, 10 % Glycerol, 10% DTT, 0.01 % Bromophenol Blue) and 815 transferred to a new tube containing Avidin as a stabilizer (0.11 µg/µl). Samples were resolved by 816 polyacrylamide gel electrophoresis and protein bands were visualized either by silver staining or by Western blotting using anti-FLAG to detect SMARCAL1 and ZRANB3, and by anti-His primary anti-818 bodies to detect RAD51C of the BCDX2 complex by standard procedures.

845
Exonuclease assays with human MR were performed in the nuclease buffer as above, but with 5 nM 846 DNA substrate (in molecules, except where mentioned otherwise), which was 50 bp-long dsDNA (X12-847 3 and X12-4C) or with a 3' (X12-3SC + X12-4NC) or 5' (X12-3 and X12-4SC) overhangs. Yeast and 848 human RAD51 were then preassembled for 10 minutes on DNA at 37 °C and the assays were subse-849 quently incubated for 120 minutes at 37 °C after adding the nucleases. RAD51 assembly and subsequent 850 nuclease assays with yeast MR were performed at 30 °C.

862
Endonuclease assays with NspI (NEB) were performed with 1 nM (in molecules) 5'-end-labeled 50 bp-863 long dsDNA (X12-3 and X12-4C) in CutSmart buffer (NEB) and supplemented with 2.5 mM ATP to 864 aid RAD51 binding to DNA. Substrates were preincubated with increasing amount of human or yeast 865 Rad51 for 10 minutes at 37 °C followed by the addition of 1 unit of NspI (per 15 µl) and further incu-866 bation for 30 minutes at 37 °C.

867
EXO-III mediated Holliday junction degradation assay (0.5 nM 5'-end-labeled substrate, in molecules) 868 was performed in branch migration assay buffer containing 100 mM NaCl. However, in the assays 869 comparing the DNA protection by human and yeast RAD51/Rad51, 1 nM 5'-end-labeled 50 bp-long 870 dsDNA substrate was used in branch migration assay buffer without additional salt. RAD51 assembly 871 (10 minutes) and subsequent reactions (30 minutes) were performed at 37 °C.

952
All primary data is available in this manuscript, supplementary information or source data.

953
(b) and (c) Anti-His antibody was coupled to Protein G agarose, bound to the BCDX2 complex (bait) 1022 and tested for interaction with ZRANB3 (prey) or SMARCAL1 (prey), respectively. Samples were the labelling. Reaction products were separated by 15% denaturing polyacrylamide gel electrophoresis.

1095
A representative experiment is shown.         (b) and (c) Anti-His antibody was coupled to Protein G agarose, bound to the BCDX2 complex (bait) and tested for interaction with ZRANB3 (prey) or SMARCAL1

1096
(prey), respectively. Samples were subjected to either silver staining or Western blot analysis by anti-FLAG and anti-His antibodies. (g) SMARCAL1-F446A, as SMARCAL1-WT, interacts with the BCDX2 complex. Anti-His antibody was immobilized on protein G agarose, bound to BCDX2 complex (bait) and tested for interaction with SMARCAL1 variants (prey). Samples were subjected to silver staining.
(h) DNA fiber assay to monitor SMARCAL1-mediated nascent DNA degradation in BRCA1-deficient cells. Wild type or SMARCAL1-F446A proteins were expressed in SMARCAL1 KO MCF10A cells upon BRCA1 depletion, as indicated. SMARCAL1-deficiency renders BRCA1-depleted cells resistant to replication fork degradation upon hydroxyurea (HU) treatment, as a result of impaired fork reversal. Top: a schematic of the assay: CldU (25 minutes), IdU (35 minutes) pulse-labeling protocol to evaluate fork degradation upon HU treatment. Under wild type condition the ratio of IdU/CldU tract length will remain~1, however if there is fork degradation this ratio will be < 1. Bottom: graphical representation of IdU/CIdU tract length ratio. The median value of 100 or more IdU and CldU tracts per experimental condition is indicated. Statistical analysis was conducted using Mann-Whitney test (****p < 0.0001). Data are representative of two independent experiments.   Continued on next page.  (c) Nuclease assays with EXO1 and its inhibition by RAD51. Experiments were carried out with blunt-ended, 5'-overhanged or 3'overhanged DNA. Asterisk indicates the position of the 32 P label. Reaction products were separated by 15% denaturing polyacrylamide gel electrophoresis. Shown is a representative experiment. (d) Quantifications of experiments such as shown in panel (c), error bars indicate SEM of three replicates. The level of DNA protection by RAD51 is presented as a relative value with respect to DNA degradation for each substrate without RAD51 (lanes 2, 7 and 12 in panel c).

Figure 7. Model for replication fork reversal and protection.
Fork remodelers have unequal biochemical functions. SMARCAL1 anneals RPA-coated ssDNA and may promote initial steps in fork reversal.
ZRANB3 and HLTF are more proficient in branch migration. RAD51 and BCDX2 interact with SMARCAL1 and ZRANB3 and promote their activities. Reversed replication forks are prone to pathological degradation, in certain genetic backgrounds, unless protected by RAD51. We show that unexpectedly the dsDNA-binding capacity of RAD51 promotes DNA protection against nucleases.