Saturation mutagenesis enables RAD51D clinical variant interpretation and identifies distinct biochemical functions

Abstract

The tumor suppressor RAD51D is essential for homologous recombination (HR). Pathogenic variants in RAD51D are associated with breast and ovarian cancers. However, most clinical missense variants are of uncertain significance. Here, we report a multiplex assay of variant effect to test 6,888 RAD51D coding variants for loss-of-function. The resulting variant-to-function map perfectly separates known pathogenic and benign variants and is validated by orthogonal HR and biochemical assays across 70 clinical variants. Our screen shows that variants in the DNA-binding or ATPase core most severely compromise HR, and we identify the RAD51D-RAD51C interface within the BCDX2 complex as essential for regulating its ATPase activity. We hypothesize that, paradoxically, the primary function of RAD51D is to slow the ATPase activity of BCDX2, thereby allowing sufficient time and space for RAD51 filament assembly. Together, we identify hotspots of deleterious RAD51D variants and uncover the mechanisms by which variants compromise its biochemical functions.

RAD51D is a tumor suppressor gene where pathogenic variants are associated with increased risk of breast and ovarian cancer. This study reports a high throughput functional screen of 6,888 RAD51D coding variants along with analysis of 70 clinically-identified breast and ovarian cancer variants.

Introduction

Homologous recombination (HR) is a central DNA repair pathway that repairs double-strand breaks. A central step in HR is the formation of RAD51 nucleoprotein filaments at single-stranded DNA ends. This HR step is tightly regulated. Defects in HR genes that regulate RAD51 are closely associated with breast and ovarian cancers, as well as Fanconi anemia, a rare disease characterized by cancer and bone marrow failure¹. As a result, numerous HR genes are routinely included on cancer screening panels². Among these, RAD51D is a paralog of the recombinase RAD51, and plays an essential role in the repair of DNA double-strand breaks through the HR pathway and in the repair of replicative damage. Loss of RAD51D function often occurs through germline inheritance of a single deleterious variant, with the eventual somatic loss of the second wild-type RAD51D copy contributing to tumor development^3,4.

Although disruption of the HR pathway is well known as a cancer risk factor, defining which variants in HR genes are pathogenic poses a significant barrier for patients and their family members who may inherit the same alleles. Family history of RAD51D-associated cancers is often challenging to recognize, given the lower penetrance compared with other hereditary breast and ovarian cancer genes such as BRCA1 and BRCA2⁵. While nonsense variants are often classified as pathogenic, missense variants are routinely identified in clinical practice, and most remain variants of uncertain significance (VUS). In RAD51D alone, >850 missense variants are reported in ClinVar, with >96% having VUS or conflicting interpretations (Fig. 1a). The absence of functional characterization for these VUS variants poses a significant barrier in clinical decision-making, hindering the ability to identify which patients require increased surveillance and/or specialized treatment strategies.

Fig. 1. Overview of pooled variant-to-function mapping in RAD51D.

a ClinVar classifications of RAD51D missense variants each year. P/LP = Pathogenic/Likely Pathogenic, B/LB = Benign/Likely Benign, VUS/CON = uncertain significance/conflicting reports. b, c RAD51D forms an obligate heterodimer with XRCC2 to form the (b) BCDX2 (PDB: 8GBJ) or (c) XRCC3 (X3CDX2) complexes (PDB: 9SVX) through its interaction with RAD51C. d Schematic of pooled RAD51D variant function assay. RAD51D mutant library cell population is treated with 250 nM olaparib over several passages, selecting for cells with normal RAD51D function. RAD51D variants are quantified within the starting (P0) and final (P2) cell populations by sequencing and scored by their relative enrichment in P2 versus P0.

Multiplex assays of variant effect (MAVE) have recently enabled high-throughput measurement of variant functional consequences across clinically relevant and actionable human disease-associated genes^6,7. Following genome editing or stable expression of one variant per cell, a functional selection is applied based upon cell fitness or resistance to genotoxic drugs, after which deep sequencing identifies variants that become enriched or depleted. This approach has been successfully used to identify functionally disruptive variants in HR genes, including BRCA2, and the RAD51 paralog RAD51C^8–10.

Central to understanding the clinical impact of RAD51D is understanding its HR function. RAD51D forms a heterotetrameric complex with the RAD51 paralogs RAD51B, RAD51C, and XRCC2 (herein BCDX2; Fig. 1b). Additionally, RAD51D is a key member of another RAD51 paralog-containing complex, the XRCC3 complex (herein X3CDX2), which comprises XRCC3, RAD51C, RAD51D, and XRCC2¹¹ (Fig. 1c). In ssDNA-bound structures of both BCDX2 and X3CDX2 complexes, the DNA-binding loops L1 and L2 of RAD51D play a critical role, as mutations in these regions impair BCDX2’s ability to bind ssDNA^11,12. Although all of the paralogs share 20-30% structural similarity to RAD51^13,14, particularly in conserved ATPase core domains and motifs, only the RAD51B-RAD51C (herein BC) heterodimer hydrolyzes ATP, whereas the RAD51D-XRCC2 (herein DX2) heterodimer regulates BC ATPase activity through an unknown mechanism¹². Not surprisingly, previous functional analysis of the related RAD51C protein has identified the ATPase domains as central for its HR function, and deleterious variants in these regions have been identified in breast and/or ovarian cancers^15,16. Although RAD51C variants have been identified in approximately 3% of HR deficient ovarian cancers, RAD51D is mutated more frequently, occurring in 5% of such cases¹⁷. In contrast to BRCA2 and RAD51C, only a handful of variants in RAD51D have had any prior functional analysis to date, and the impact of clinically identified breast and ovarian cancer variants on HR is largely unknown (summarized in Supplementary Data 1). Furthermore, the functional regions and domains essential for RAD51D activity, and the manner in which RAD51D, together with XRCC2, regulates BC ATPase activity, have yet to be elucidated.

Here, we used a MAVE to analyze 6,888 single-codon RAD51D variants for loss of viability using sensitivity to the poly (ADP-ribose) polymerase inhibitor (PARPi), olaparib, which is currently used to treat HR-deficient ovarian cancers¹⁸. The resulting variant-to-function map provides measurements for 97% of all single-residue variants, offering a global view of missense constraint concentrated in the ATPase-like and DNA-binding domains. Next, 70 clinically relevant variants identified in breast and/or ovarian cancers were selected for subsequent functional studies of HR activity and complex interactions involving BCDX2 and X3CDX2. Having found that the results of the MAVE analysis were strongly correlated with mechanistic studies of individual variants, we identified seven residues in RAD51D that are critical for HR function and performed biochemical analyses to investigate their unique mechanistic contributions to BCDX2 complex formation, DNA binding, and ATPase regulation. We found that six of the seven deleterious RAD51D cancer-identified variants alter the regulation of BC ATPase activity in the BCDX2 complex, either by attenuating RAD51D interaction with RAD51C or by reducing BCDX2 DNA binding affinity. Notably, six of these variants also impact the formation of the X3CDX2 complex. Lastly, we identified a separation-of-function allele in the N-terminus of RAD51D that alters its HR function, not through modulating BC ATPase activity but rather by inhibiting RAD51-mediated strand exchange as part of the newly identified X3CDX2 complex^11,19,20. These results suggest that the strand exchange activity of X3CDX2 can occur independently of BCDX2. Together, our work is informative in identifying and reclassifying deleterious variants in RAD51D, providing mechanistic insight into how the disruption of these functional hotspots contributes to RAD51D function as part of the BCDX2 and X3CDX2 complexes.

Results

A variant-to-function map of all possible RAD51D single-codon variants

Since RAD51D is essential in non-transformed cells, we used the well-characterized osteosarcoma cell line U2OS in which RAD51D has been knocked out by CRISPR/Cas9 (herein U2OS RAD51D KO)²¹. It was previously shown that restoring RAD51D expression in these knockout cells rescues HR function and resistance to PARP inhibition through olaparib treatment^21,22. We leveraged this assay to systematically test missense variants in the RAD51D gene (Fig. 1d). Mutational libraries were stably integrated into U2OS RAD51D KO cells such that each resulting cell expressed a single RAD51D variant. We performed saturation mutagenesis of the RAD51D cDNA and obtained 97.6% coverage of the 6,888 possible single-amino acid variants (at each of the 328 residues, 21 mutations, including 19 substitutions, plus nonsense and synonymous mutations). Variants were deemed present in the library if their starting frequency was ≥1/10,000 (Supplementary Fig. 1). These mixed cellular populations were then treated with 250 nM olaparib for two passages to enrich for cells expressing functionally normal RAD51D mutants. Variant frequencies were quantified by deep amplicon sequencing before and after selection, and used to calculate a function score for each variant, with negative scores indicating variants that became depleted during olaparib treatment.

As expected, premature truncation variants showed strong depletion, with stop-gain variants prior to residue 305 having a mean score of −0.989. These were perfectly separated from the scores for synonymous variants (mean: 0.023; area under the precision-recall curve, prAUC=1.00, n = 494 variants). Notably, C-terminal truncations from residue 306 onward scored comparably to synonymous variants, consistent with this mostly unstructured region being dispensable in our assay. Variants were classified as neutral (NEU), intermediate (INT), or loss-of-function (LOF), using function score cutoffs set by the 2.5% and 97.5% percentiles of synonymous and nonsense variants, respectively (Methods). To be classified as INT (function scores between −0.170 and −0.446) or LOF (function score below −0.446), variants also had to pass a significance cutoff (local false sign rate, lfsr ≤0.01). As a result, we classified 74.5% of measured RAD51D missense variants (n = 4593) as NEU (mean score: −0.07), 16.8% as LOF (n = 1038; mean score: −0.78), and 8.6% as INT (n = 536; mean score: −0.31) (Fig. 2a). Supporting these classifications’ accuracy, all scored synonymous variants were classified as NEU, while nearly every nonsense variant prior to codon 305 (304/305, 99.7%) scored as INT or LOF.

Fig. 2. A variant-to-function heatmap of RAD51D.

a Distributions of function scores by variant class; nonsense variants in codons 1–304 are perfectly separated from synonymous variants. Vertical lines denote cutoffs for function classifications (LOF, loss-of-function; INT, intermediate; NEU, neutral). Source data are provided in Supplementary Data 3. b Per-residue mean missense function scores, grouped by residue surface area exposure status and secondary structure. Unique LOF exposed residues are labeled. c Variant-to-function heatmap across RAD51D, shaded by function score (white, WT-like; red, null-like) for each mutant amino acid (rows) at each codon position (columns). Variants that do not pass multiple testing correction (lfsr>0.01) are shaded from white to gray, WT residues are boxed, and dark gray denotes missing data. Tracks above heatmaps, from top to bottom: conservation score, protein secondary structure, key domains. Asp (D) or Glu (E) are shown in boldface to highlight stronger effects of these substitutions at some sites. d Function scores separate pathogenic from benign variants among single-residue variants reported in ClinVar, and provide evidence for missense VUS and those with conflicting reports (all variants plotted SpliceAI score <0.2). Point color denotes variant type, and filled/open points denote statistical significance (i.e., filled: lfsr ≤0.01). e Precision-recall curve showing classification performance of function scores to distinguish between clinical variants with consensus interpretations of P/LP versus B/LB (n = 211 distinct variants at the protein level). f Enrichment of RAD51D LOF mutations among tumor variants in Project GENIE. Each point represents one cancer type, with P-value plotted vs odds-ratio for Fisher’s exact tests of LOF enrichment among MAVE-scored RAD51D SNVs found in that respective cancer type. Orange points: all variants tested; blue: missense (MIS) only. Dashed line denotes Bonferroni-corrected threshold for testing 11 cancers at P < 0.05. (OVCA: ovarian cancer).

Missense intolerance was enriched in structured, annotated domains of RAD51D (Figs. 2b, c, 3). As expected, variants at buried residues were on average more deleterious than those at surface-exposed residues (mean function scores: −0.272 vs −0.113, Welch’s t: P < 2.2×10⁻¹⁶; Fig. 2b). However, we uncovered surface exposed residues near RAD51C or XRCC2 interfaces at which variants were highly deleterious (Fig. 3a–c, e). Structural analysis identified several residues that make contact with DNA, ATP, or other proteins via hydrogen bonding or salt bridges that were largely intolerant to substitution (mean function score < −0.446, Fig. 3d, f, g, Supplementary Fig. 2a). We found six highly constrained residues at which two or fewer variants were tolerated as NEU or INT. These appeared at well-conserved sites, including the Walker A motif, where the conserved terminal threonine (T114) could be replaced only by serine, notable for being the other final amino acid of the classical Walker A motif²³. Similarly, the Walker B motif ends in a conserved aspartic acid (D206) at which 18 of 19 substitutions exhibited strong LOF (mean function score: −1.22), except the chemically similar glutamic acid, which was less severe (score: −0.52). Residues near the ssDNA backbone (263-267) showed strong intolerance to Glu and Asp variants, which would be expected to repel the negatively charged backbone (Fig. 3f, Supplementary Fig. 2b). Glycine 265 was completely substitution-intolerant, suggesting it may provide functionally critical flexibility for fitting around the ssDNA backbone.

Fig. 3. High-throughput screening reveals disrupted RAD51D interaction surfaces.

a Surface representation of RAD51D within the BCDX2 complex (PDB: 8GBJ). Surface residues are colored according to either red (50% pathogenicity) or yellow (20% pathogenicity). Cartoon representations of XRCC2 (purple), RAD51D (light blue), RAD51C (green), and RAD51B (red) are shown consistently throughout. Bound ATP molecules are represented as yellow-orange sticks. b Specific loss-of-function variants (red, 50% pathogenicity) listed above or (yellow, 20% pathogenicity) listed below the RAD51D schematic are shown with the indicated interaction interfaces indicated. c–g) Binding sites for RAD51C (c), ATP site 1 (d), XRCC2 (e), ssDNA (f), and ATP site 2 (g).

MAVE scores provide strong evidence for RAD51D variants

To compare our function scores to existing clinical variant classifications, we curated a set of control variants from ClinVar. We identified 211 unique variants at the protein level, with B/LB (n = 155) or P/LP (n = 36) classification, along with an additional 28 “soft” conflicts, meaning that there were two of more submissions with the same definitive classification (pathogenic or likely pathogenic (P/LP) or benign or likely benign (B/LB), along with VUS. To gain calibration power, we resolved these conflicts as P/LP (n = 1) or B/LB (n = 27; Supplementary Data 2). We observed perfect concordance with the clinical classifications; all 37 pathogenic or likely pathogenic (P/LP) variants scored as LOF and all 174 benign or likely benign (B/LB) variants scored as NEU in our MAVE (Fig. 2d, e). The benign polymorphism R165Q scored as NEU, consistent with its high population allele frequency (gnomAD MAF = 14.1%). Only two missense variants, S207L and S207W, have been classified as pathogenic or likely pathogenic (P/LP and VUS in ClinVar)²⁴. Both were correctly identified as LOF, with S207L having comparatively mild effects (score: −0.503) and S207W showing more severe impairment (score: −1.703). Additionally, we compared our function scores to AlphaMissense, a computational tool developed to predict missense variant pathogenicity from AlphaFold structural predictions and sequence conservation²⁵. Notably, 35% of the variants called by AlphaMissense as pathogenic with high confidence (probability ≥0.9) were classified as functionally normal in our screen, suggesting a potentially high rate of false positives among these computational predictions (Supplementary Fig. 2c).

To estimate the strength of evidence provided by our MAVE map, we translated the functional scores into evidence for clinical variant classification. We calculated OddsPath odds ratios as recommended by ClinGen²⁶. These reflect the ability of MAVE scores to correctly identify previously classified pathogenic as functionally abnormal and benign variants as functionally normal²⁶, among the 211 unique pathogenic or benign RAD51D protein variants with existing clinical interpretations The resulting OddsPath scores were 174 for functionally abnormal variants in our assay, and 0.027 (1/37) for normal variants, corresponding to strong level of evidence towards pathogenicity (4 points) and benignity (−4 points), respectively²⁷ (Table 1).

Table 1.

OddsPath-based assay evidence calibration using RAD51D clinical variants

Pathogenic variants					Benign variants
n	Abnormal	Normal	Odds Path	Evidence points	n	Abnormal	Normal	Odds Path	Evidence points
37	37	0	174	+4	174	0	174	1/37	−4

Known pathogenic or benign variants were obtained from ClinVar, requiring two stars (multiple submitters, without discordance), and resolving conflicts when at least two non-VUS interpretations were available (Methods). OddsPath scores were determined as described²⁶. Variants predicted to be splice disruptive by SpliceAI (score≥0.2) were excluded and treated as ‘not determined’.

We next examined the 789 RAD51D missense variants listed in ClinVar as VUS or having conflicting interpretations. We excluded 49 variants predicted to have splicing effects (SpliceAI²⁸ scores ≥ 0.2), and the 68 variants that scored as INT and variants that scored as LOF but failed to pass the significance threshold. Of the remaining 672 variants, 87% were functionally normal (NEU); absent any conflicting evidence, the addition of the strong evidence (−4 points) would be enough to reclassify these as likely benign²⁹. The other 85 missense variants (12.6%) scored as LOF and could receive 4 points toward pathogenic classification. Nearly two-thirds (n = 56) of these were absent from gnomAD, but among those present in that population dataset, the cumulative minor allele frequency (MAF) is 2.03 × 10⁻⁴, indicating that, as a lower bound estimate, RAD51D missense LOF variants may contribute to hereditary breast and ovarian cancer risk for approximately 1 in 20,000 individuals. As for missense SNVs not yet reported in ClinVar (n = 1,264), a similar proportion ( ~ 13%) exhibit abnormal function, and our function scores will provide evidence to classify them as they are newly clinically observed.

To explore the potential for RAD51D variant reclassification in a real-world clinical setting, we queried the University of Michigan Breast and Ovarian Cancer Risk Evaluation Center registry. Out of 24 RAD51D VUS carriers identified by multi-gene panel testing between 2017 and 2025, we identified five patients with potentially abnormal VUS (based on the MAVE score, Supplementary Data 3), a proportion much higher than in unselected population databases. Of these five individuals, two carry c.202 G > A, which showed abnormal splicing in a previous minigene study³⁰ and was flagged by SpliceAI as disruptive (deltaMax score=0.42). The remaining three individuals each carry a different missense variant showing strong loss of function in our map (Table 2). Two of these three individuals had a personal history of breast or ovarian cancer, along with a family history of cancers (Supplementary Fig. 3).

Table 2.

Functionally abnormal RAD51D clinical VUS

Patient Info				RAD51D VUS		Assay results
Current Age	Personal History of Cancer	Age at diagnosis	Family History of Cancer	cDNA	AF	MAVE score	Classification
60–64	Ovarian cancer	60-64	Unilateral breast, dx 40 s	c.802 T > A (p.Trp268Arg)	1.4 × 10-6	−1.410	Abnormal
55–59	Triple-negative breast cancer	55-59	Colon, dx 70 s	c.823 C > T (p.Arg275Trp)	9.6 × 0-6	−1.278	Abnormal
30–34	None	N/A	N/A	c.11 T > C (p.Leu4Pro)	N/A	−0.648	Abnormal

RAD51D follows the classical ‘two-hit’ model of tumor suppressor genes³¹, with biallelic mutations in LOF carriers’ tumors⁴. We reasoned that somatic mutation databases would therefore be enriched for LOF variants. We queried the AACR Project GENIE database³² and identified 854 tumors with RAD51D somatic variants scored by our variant-function map. Across all cancers, these variants’ function scores were significantly lower than the >2,300 assayed SNVs not found in the Project GENIE dataset (P = 2.4 × 10⁻⁷, t-test; Supplementary Fig. 4), though this difference disappeared when we focused only on missense variants (n = 793 vs 1,689; P = 0.96). However, stratifying by cancer type revealed that ovarian cancer alone was significantly enriched for functionally abnormal variants overall (8.2-fold, P = 4.1 × 10⁻⁶, Fisher’s exact test) and for abnormal missense mutations specifically (4.9-fold, P = 1.9 × 10⁻³; Fig. 2f). Of the 10 other cancer types tested, several including breast cancer showed nominal enrichment but did not survive multiple testing correction.

Characterization of RAD51D variants identified in breast and/or ovarian cancers for altered DNA repair function through HR and for altered BCDX2 complex formation

To determine how well our variant-function map can predict the ability of RAD51D variants to perform HR, we performed a sister chromatid recombination (SCR) assay on 70 clinically-identified breast and/or ovarian cancer RAD51D missense variants, of which 14 were determined to have LOF or intermediate scores in our map (Supplementary Data 1). These variants were identified through clinical collaborations, in cancer databases such as Catalog of Somatic Mutations in Cancer (COSMIC) or The Cancer Genome Atlas (TCGA), or in the academic literature (Supplementary Data 1). Using the U2OS RAD51D KO cells transiently transfected with pcDNA3 RAD51D-WT or a series of synonymous or truncation alleles and the resulting HR activity was used to benchmark and threshold RAD51D VUS HR activity (Supplementary Fig. 5; Supplementary Data 4). To measure HR function, we utilized a GFP-based SCR reporter construct that has been stably integrated into these cells²¹ (Supplementary Fig. 5a). This assay utilizes a non-functional GFP, which is disrupted by the inclusion of a unique I-SceI endonuclease cleavage site cloned downstream of a truncated GFP fragment. Upon transfection of a plasmid that expresses the I-SceI enzyme, a DNA double-strand break is induced in the non-functional GFP fragment, which then undergoes DNA repair using the upstream, truncated GFP fragment as a template, thereby restoring GFP expression and subsequent fluorescence. We compared these 70 clinically identified RAD51D missense variants relative to known synonymous (benign) or nonsense (pathogenic) variants from the ClinVar database (Fig. 4; Supplementary Fig. 5).

Fig. 4. Breast and ovarian cancer-identified variants in the ATPase and DNA-binding regions of RAD51D exhibit reduced HR proficiency.

a HR proficiency of 70 RAD51D missense variants of unknown significance. HR function was classified in comparison to the range of synonymous (green, HR > 0.75) and truncation variants (red, HR < 0.6). Missense variants are colored corresponding to function (HR > 0.7 in grey; intermediate yellow, and HR < 0.6 in light red). See Supplementary Fig. 5b, c for all synonymous and truncation variants tested, and note that a subset of those variants analyzed are replotted here as representative variants. The experiment was performed independently 4–9 times per variant (see Source Data for exact n of each variant) and plotted as mean values ± s.d. b, e Olaparib and cisplatin sensitivity of breast/ovarian cancer identified RAD51D variants with reduced HR. Representative images of U2OS RAD51D KO cell lines stably expressing pcDNA3 RAD51D WT or the indicated RAD51D variant that were treated with increasing concentrations of Olaparib (b) or cisplatin (d). c, e Clonogenic survival assays were quantified by percent colony area and normalized to the area of untreated or vehicle control. Means of 4–12 trials (see Source Data for precise n for each variant) are plotted ± s.d., and drug concentrations with colony area <0.001 are omitted. f RAD51D variants with reduced HR are expressed. Western blot analysis of U2OS RAD51D KO cell lines stably expressing WT or the indicated RAD51D variants. RAD51D protein expression was assessed using an anti-RAD51D antibody, and equal protein loading was assessed using an anti-Tubulin antibody. Note that protein expression can be variable, but notably L4H consistently exhibits reduced protein expression. Experiment performed in triplicate. g Structures of the RAD51 paralog pentamer X3CDX2 (PDB: 9SVX) and BCDX2 (PDB: 8GBJ) complexes. Cartoon representations of XRCC2 (purple), RAD51D (light blue), RAD51C (green), XRCC3 (yellow), RAD51 (orange), RAD51B (red), bound ATP (yellow-orange), and ssDNA (orange) are shown consistently throughout. h, i Variants causing complete (red, h) or intermediate (yellow, i) deficiency in homologous recombination are highlighted on RAD51D. The interacting surfaces of RAD51C and XRCC2 are outlined with black dashed lines. The contact interfaces are identical between the XRCC3 and BCDX2 complex. j, k Variants that disrupt RAD51C-RAD51D interactions are shown as red sticks, mapped onto the BCDX2 complex structure. Source data are provided as a Source Data file.

We observed a strong correlation with MAVE results (Pearson’s r: 0.75, P = 1.27×10⁻¹⁹; Fig. 5d) and identified four RAD51D HR-deficient variants (L4H, L164P, M178R, S207L) and three alleles with attenuated HR (G96C, G265R, R275W) (Fig. 4a, Supplementary Data 1). We verified transient expression of the pcDNA3 RAD51D missense variants, and all variants are expressed, although to varying degrees (Supplementary Fig. 6a). Since these HR-disrupted variants all had LOF scores in the MAVE olaparib selection assay, using a clonogenic survival assay, we sought to validate their sensitivity to olaparib and test whether they exhibited sensitivity to cisplatin, which is a platinum-based compound used as a front-line treatment for ovarian cancer³³ (Fig. 4b–e). To do this, we created stable cell lines in the U2OS RAD51D KO cell lines solely expressing the variant of interest and validated both expression and loss of HR function compared to our transiently transfected cell lines (Fig. 4f; Supplementary Fig. 6b). Using these stably complemented cell lines, we found that cells expressing the HR deficient variants exhibited variable sensitivity to olaparib and/or cisplatin, but none were as sensitive as the RAD51D knockout cell line (Fig. 4b–e).

Fig. 5. A subset of RAD51D breast and ovarian cancer identified variants alter interaction with XRCC2 or RAD51C by Y2H or Y3H, respectively.

a, b Interaction proficiency of RAD51D VUS was measured via (a) yeast 2-hybrid (Y2H) of pGAD-RAD51D or its variant, expressed in a GAL4 DNA activating domain expressing plasmid with pGBD-XRCC2 expressed in the GAL4 DNA binding domain expressing plasmid and (b) pGAD-RAD51D or its variant, expressed in the pGAD (GAL4 DNA activating domain) plasmid with pGBD-RAD51C (GAL4 DNA binding domain) plasmid with pADH1-RAD51B via yeast 3-hybrid (Y3H). RAD51B serves to stabilize RAD51C protein levels. Empty vectors are used as a negative control. Quantification of yeast growth from three experiments is plotted as mean ± s.d relative to the WT control. Representative images for variants and controls are shown. Variants with <50% interaction are classified as deficient and plotted in red. c RAD51D variants with reduced interaction with XRCC2 or RAD51C are expressed. Western blot analysis of protein extract from yeast cells in (a) expressing WT RAD51D or the indicated RAD51D variants was assessed using an anti-RAD51D antibody, and equal protein loading was assessed by measuring the amount of Kar2 using an anti-Kar2 antibody. Note that a subset of variants has reduced protein expression relative to WT RAD51D. Experiment performed in triplicate. See Supplementary Fig. 7 for Western blots of all variants. d–f MAVE function scores vs individual variant assay results of (d) HR activity, p-value determined by simple linear regression, (e) XRCC2 binding, and (f) RAD51C binding. Missense variant point shape denotes functional status (LOF/INT/NEU) based upon MAVE result. Source data are provided as a Source Data file.

Upon mapping the identified HR deficient variants on the solved cryo-EM structures of RAD51D within the BCDX2 and X3CDX2 complexes, we observed no clear localization of the critical residues: two of the HR deficient residues are located in the ATP-binding core (L164 and S207), two are found in the DNA-binding loop (G265, R275), while M178 is located in the RAD51D-RAD51C binding interface, and L4 in the N-terminus (Fig. 4h–i). Together, these results suggest that variants that disrupt the ATPase core domain or DNA binding interfaces may be critical for HR function.

RAD51D is a member of the BCDX2 and X3CDX2 complexes, where it directly interacts with XRCC2 as part of the subcomplex DX2 and with RAD51C, which is part of the BC or X3C subcomplex^11,34. Our previous work on RAD51C demonstrated that variants that disrupt BCDX2 complex formation are indicative of loss-of-function^15,16. To determine whether any of the 70 breast/ovarian cancer identified variants alter BCDX2 complex formation, we examined the RAD51D variants interaction with XRCC2 by yeast-2-hybrid analysis (Y2H). Using this approach, we found that three of the variants had reduced interaction with XRCC2 (Fig. 5a; L164P, M178R, R275W), and all exhibited LOF (Fig. 5e). We next examined whether the RAD51D variants would alter their interaction with RAD51C using a yeast-3-hybrid system (Y3H; Fig. 5b), where RAD51B is used as a third vector to help stabilize RAD51C. Importantly, 26 of the 70 RAD51D variants exhibited defects in interaction with RAD51C and had an overall intermediate MAVE score, with 10 functionally defective variants (mean = −0.940), including the three variants that also exhibited loss of XRCC2 interaction (Fig. 5e, f). Expression of the variants from the yeast cells was confirmed by Western blot (Fig. 5c; Supplementary Fig. 7). In some cases, the defect in protein interaction could be due to reduced expression (Supplementary Fig. 7; S46C). Together, these findings suggest that the interaction between RAD51D and its binding partners, XRCC2 and RAD51C, is critical for function.

To determine if loss of HR observed in the RAD51D cancer-identified variants results in alteration in complex formation or RAD51D biochemical activity, we purified recombinant DX2 subcomplex with or without one of the seven different RAD51D variants (Fig. 6a) with a FLAG-tagged XRCC2. We then incubated the BC subcomplex with the DX2 subcomplex and performed pulldown assays using DX2-FLAG. We found that four of the RAD51D variant-containing complexes exhibited defects in BCDX2 complex formation (Fig. 6b; G96C, L164P, M178R, S207L, as indicated by the comparison of enrichment in the eluate to the flow-through). To further investigate how these RAD51D variants affect BCDX2 biochemical function, we performed biochemical analyses using reconstituted wild-type or mutant BCDX2 complexes. We created BCDX2 complexes by mixing wild-type BC subcomplex with either wild-type DX2 subcomplexes, or DX2 containing the mutation of interest.

Fig. 6. RAD51D HR deficient variants differentially impact ssDNA binding, BCDX2 ATPase activity, and/or RAD51-mediated strand exchange.

a Coomassie-stained SDS–PAGE gel analysis of the purified CX3, BC, and DX2 sub-complexes carrying either WT or RAD51D variants, one representative gel shown. b Pull-down analysis of WT BC (RAD51B-His/RAD51C) and DX2 (RAD51D/XRCC2-FLAG) sub-complexes with indicated RAD51D variants using anti-FLAG resin. Mean values ± s.d. from two independent are indicated below. c ssDNA binding of BCDX2 paralog complexes reconstituted by mixing WT BC and DX2 sub-complexes with indicated RAD51D variants. Results from three independent experiments were plotted as mean values ± s.d. The p-values for the significance of differences were calculated using two-way ANOVA with DX2 variants and Protein concentration as fixed factors, followed by Dunnett’s post-hoc test for multiple comparisons against the Wild-Type (WT) group. d ATPase analysis of 0.5 μM WT BC and DX2 sub-complexes with indicated RAD51D variants in the presence of ssDNA after 60 min incubation. Pi indicates released inorganic phosphate after hydrolysis. The histograms represent mean values ± s.d. of three independent experiments. Statistical significance between mean values was assessed using an unpaired two-sided t-test. e Pull-down analysis of WT CX3 (RAD51C-His/XRCC3-STREP) and DX2 (RAD51D/XRCC2-FLAG) sub-complexes with indicated RAD51D variants using anti-FLAG resin. Mean values ± s.d. of the pulled-down from two independent experiments are indicated below. f ssDNA binding of X3CDX2 paralog complexes reconstituted by mixing WT CX3 and DX2 sub-complexes with indicated RAD51D variants. Results from three independent experiments were plotted as mean values ± s.d. The p-values were calculated using two-way ANOVA with DX2 variants and protein concentration as fixed factors, followed by Dunnett’s post-hoc test for multiple comparisons against the Wild-Type (WT) group. g Reaction schematic of DNA strand exchange in the presence of RPA. (h) DNA strand exchange in the absence or presence of 30 nM RPA, with 100 nM RAD51 alone or a mixture of RAD51 and indicated RAD51 paralog complexes. Reactions were performed with AMP-PNP as the nucleotide cofactor. The positions of the Cy5‑labeled (indicated by asterisk) 40‑bp dsDNA and the strand‑exchange product are annotated alongside the gel images. Mean values from four independent experiments are plotted ± s.d. Statistical significance between mean values was assessed using an unpaired two-sided t-test. Source data are provided as a Source Data file.

Six of the seven RAD51D variants with reduced HR exhibit altered DNA binding and BCDX2 ATPase activity

The association of BC and DX2 sub-complexes into the BCDX2 complex synergistically enhances ssDNA binding at DNA break sites to facilitate RAD51 filament formation in an ATPase-dependent manner^12,16. To assess whether RAD51D variant can affect ssDNA binding of BCDX2 complex, we conducted electrophoretic mobility shift assays (EMSA) in the presence of Mg²⁺/ATP using an 80-nucleotide single-stranded DNA substrate at increasing concentrations BCDX2 complexes (Fig. 6c). BCDX2 complex containing RAD51D-L4H is DNA binding proficient, whereas six of the seven RAD51D variants displayed attenuated DNA binding compared to WT BCDX2 (Fig. 6c). We note that defective DNA binding of BC with DX2 sub-complexes containing RAD51D variants G96C, L164P, M178R, S207L was primarily attributable to defective BCDX2 complex assembly (Fig. 6b, c). In contrast, RAD51D variants G265R and R275W formed intact BCDX2 complexes but displayed severe DNA-binding deficiencies, consistent with their mutations residing within the second DNA-binding loop of RAD51D (Figs. 6c, d and  4j, k).

Our previous work demonstrated that the BC sub-complex exhibits robust ssDNA-dependent ATPase activity, whereas the DX2 sub-complex does not. Interestingly, addition of DX2 to BC reduced ATP hydrolysis, suggesting that DX2 modulates BC ATPase activity at DNA break sites¹². Next, we examined whether the regulation of BC ATPase activity by DX2 would be altered in complexes containing the various RAD51D variants. Similar to wild type, DX2 with RAD51D-L4H dampened BC ATP hydrolysis (Fig. 6d). In contrast, six variants defective in ssDNA binding failed to influence BC ATPase activity, as ATP hydrolysis remained comparable to BC alone (Fig. 6d). These findings indicate that the synergistic stimulation of ssDNA binding by BC-DX2 association is essential for ATPase regulation within the BCDX2 complex, and that RAD51D variants impaired in complex formation and/or ssDNA binding cannot modulate BC ATPase activity.

Identification of RAD51D-L4H as a separation-of-function mutant proficient for BCDX2 activity with reduced X3CDX2 function

It was recently discovered that in addition to RAD51D’s central function in the BCDX2 complex, the DX2 subunit is a member of another RAD51 paralog-containing complex with RAD51C and XRCC3, X3CDX2¹¹. Notably, interfaces of RAD51C and RAD51D interaction are similar in both BCDX2 and X3CDX2 complexes^11,12,35. Therefore, we asked whether RAD51D variants with altered HR would also exhibit defects in X3CDX2 formation. To address this question, we examined DX2 subcomplexes containing WT RAD51D, or the variant of interest, for interaction with CX3 (Fig. 6e). Similar to BC–DX2 interactions, CX3 associates with DX2 when RAD51D contains L4H or G265R, but not with variants G96C, L164P, M178R, or S207L (Fig. 6e). Additionally, unlike BC–DX2, CX3 fails to interact with DX2 carrying RAD51D-R275W (Fig. 6e). We then assayed whether ssDNA binding of the X3CDX2 complex would be impacted, and indeed, all the tested variants exhibit reduced ssDNA binding or form some lower molecular weight complexes (Fig. 6f). It is important to note that while RAD51D-L4H exhibits no discernible biochemical defect within the BCDX2 complex and still forms the X3CDX2 complex, it shows attenuated DNA binding within the latter (Fig. 6f). These results suggest that RAD51D-L4H is a separation-of-function allele that may uncouple the activities of the BCDX2 and X3CDX2 complexes.

The X3CDX2 complex interacts with RAD51 to synergistically bind ssDNA to facilitate RAD51-mediated strand exchange. To determine whether the RAD51D variants impact RAD51 function within the X3CDX2 complex, we performed strand exchange assays using a single-stranded DNA (ssDNA) oligo that is homologous to a double-stranded DNA (dsDNA) radioactively labeled template (Fig. 6g). RAD51 can readily promote strand exchange in the absence of the ssDNA binding complex, RPA (Fig. 6g, lane 2). However, when the ssDNA is RPA coated, then RAD51 requires its mediators overcome the inhibitory effects of RPA (Fig. 6g, compare lane 3 to 4). While wild-type X3CDX2 complex promotes RAD51-mediated strand exchange on RPA-coated ssDNA, the X3CDX2 mutant complexes were unable to promote strand exchange (Fig. 6g, lanes 5-11). These results are consistent with the HR defects observed in our cellular HR assays (Fig. 4). Importantly, RAD51D-L4H mutant complexes are also deficient in RAD51-mediated strand exchange, suggesting that the X3CDX2 complex plays a critical role in this activity that can be uncoupled from BCDX2. Together, our results demonstrate that RAD51D promotes HR by regulating BCDX2 ATPase activity and X3CDX2 strand-exchange activity.

Discussion

To identify critical regions for RAD51D function, we undertook a high-throughput screen of RAD51D’s response to olaparib, examining 6,888 synonymous, nonsense, and missense changes across the full-length RAD51D cDNA. We identified critical regions near ATPase-like domains, as well as the binding interfaces between RAD51C-RAD51D and RAD51D-XRCC2, and key structural elements critical for olaparib response. Functional screening of 70 clinically identified variants for HR and/or BCDX2 complex interactions showed a high correlation with our MAVE results, with HR-deficient variants and variants that disrupt RAD51D interaction with XRCC2 also significantly depleted. As RAD51D variants are routinely uncovered in high-risk patients for breast and ovarian cancer, these findings validate and enforce the utility of using high-throughput variant analysis and classification.

While RAD51D’s role as a tumor suppressor is well established, surprisingly little is known about its biochemical function during HR. Our groups recently found that RAD51D, when partnered with XRCC2, attenuates BC ATPase activity¹². The significance of this attenuation has been largely overlooked. Here, we provide evidence that RAD51D DNA-binding and ATPase-like domains are critical for this regulation (summarized in Fig. 7a, b). These results are unexpected, given that the BCDX2 complex is an ATPase, and one might have expected that DX2 would stimulate ATP hydrolysis of BC, rather than decrease it. Clues to this regulatory mechanism come from RAD51 itself. RAD51 is an ATPase that forms nucleoprotein filaments on the DSB ends, and its affinity for DNA is strongly enhanced upon ATP binding, whereas ATP hydrolysis results in RAD51 removal from the ssDNA³⁶. Several key regulatory factors rely on this mechanism to regulate RAD51 loading at sites of DNA damage, such as Srs2 in yeast^37–39 and RECQL5 and FIGNL1 in humans^40–42. Though BC subcomplex displays weak ssDNA binding, like RAD51, its ATPase activity is stimulated upon ssDNA [Fig. 7c, BC (ATP hydrolysis)]. In contrast, despite having ATPase core domains and motifs, DX2 is not an ATPase; rather, its interaction with BC attenuates this ATPase activity and is required for stable ssDNA binding of the BCDX2 complex¹²; Fig. 7c, DX2 (ssDNA binding with BC). Given the high degree of conservation between RAD51 and its paralogs, we hypothesize that slowing the hydrolysis of ATP would increase BCDX2 DNA binding dwell times and therefore facilitate RAD51 filament elongation (Fig. 7c). Consistent with this hypothesis, loss of DX2 or its DNA-binding activity increases ATP hydrolysis of BC, thereby reducing RAD51-mediated HR (Fig. 6). Indeed, our most severe HR-deficient variants affect ATPase regulation, and all decrease DNA binding to varying degrees (Fig. 6b,e). CX3 directly interacts with RAD51, thus facilitating its binding to ssDNA [Fig. 7c; CX3 (RAD51 ssDNA targeting]. Both BCDX2 and X3CDX2 are important for RAD51 filament formation, while X3CDX2 plays an additional role in RAD51-mediated strand exchange (Fig. 7c). Notably, we identified a separation-of-function variant, L4H, that is proficient in BCDX2-mediated ATPase regulation but compromises X3CDX2-mediated RAD51 strand exchange. This suggests that RAD51D plays a critical role in both RAD51 loading and strand exchange activities required for HR.

Fig. 7. Summary of findings and model for RAD51D function as part of the BCDX2 and XRCC3 (X3CDX2) complexes.

a Summary of cellular findings. Schematic of RAD51D (1-328 aa) is shown with the Walker A and B motifs (green), and DNA binding loops (blue) indicated. The variants analyzed are shown above with the results from the cellular studies (MAVE variant-function map; Fig. 2), RAD51D Y2H interaction with XRCC2 and its Y3H interaction with RAD51C (Fig. 5), SCR recombination (HR), as well as olaparib and cisplatin clonogenic survival assays (Fig. 4) are summarized based on functional score. Loss-of-function is shown in red, an intermediate function is shown in yellow, and wild-type function is indicated in green. b Summary of biochemical findings. Variants included in biochemical analysis for BCDX2 or XRCC3 (X3CDX2) complex formation, DNA binding, ATPase restraint, and RAD51 strand exchange from Fig. 6 are shown. c Model for BCDX2 and X3CDX2 function in DSB repair. Upon DSB formation, the 3’ ssDNA end is resected and coated by RPA (a heterotrimeric protein complex, blue circles). The BCDX2 and X3CDX2 complexes facilitate the displacement of RPA and the loading of RAD51 (yellow circle) onto ssDNA. This is achieved by DX2 regulation of BC ATP hydrolysis, which facilitates its binding to ssDNA. CX3 in complex with DX2 then targets RAD51 to ssDNA. The combined functions of the BCDX2 and X3CDX2 complexes are necessary to facilitate RAD51 filament assembly. Subsequently, the X3CDX2 complex aids in the RAD51-mediated homology search and strand exchange activities needed for homologous recombination. RAD51D plays a critical role in both of these BCDX2 and X3CDX2 functions.

MAVE analysis is useful not only for ascertaining the novel variants’ impact, but also its global view of functional constraint has enabled key mechanistic insights into the structure and function of RAD51D. When we mapped the depleted variants of RAD51D onto the recently solved cryo-EM structure of BCDX2 complexed with DNA, we found five beta sheets that are aligned within the central structure of RAD51D, which, when mutated, result in loss of function (Fig. 3). Our high-resolution functional mapping of this structure showed high intolerance to mutation, pairing well with structural data, and together suggests that this core RAD51D beta sheet is required to promote this modulatory effect—possibly by propagating conformational changes based on the RAD51D-DNA interface (Fig. 3f). Furthermore, our Y3H experiments indicate that the RAD51C-RAD51D interaction is much more fragile than the RAD51D-XRCC2 interaction, with many protein-coding changes fully disrupting the BC-DX2 complex. Further biochemical and structural studies into critical structural regions of RAD51D and mechanisms driving its modulatory functions are warranted.

High-throughput functional assays can assist clinical variant interpretation, provided their measurements are carefully evaluated and calibrated^6,43. Supporting our map’s suitability for clinical use, it perfectly separated synonymous variants from nonsense variants (excluding truncations after codon 305). Likewise, classifications based upon our function scores and cutoffs were completely concordant with the standing classifications across the 31 missense variants with consensus interpretations in ClinVar. Calibrated to these clinical variant sets, our map meets the ‘strong’ level of evidence in both directions of effect (i.e., for or against pathogenicity) under accepted clinical variant interpretation guidelines²⁶. Thus, our map will assist the resolution of 789 existing missense VUS in RAD51D and inform the interpretation of the other 1,264 such variants as they are detected clinically in the future.

To demonstrate these functional measurements’ utility, we applied them to interpret clinical variants. In a high-risk breast and ovarian cancer cohort, we identified three carriers with functionally abnormal missense variants. Notably, one of these was nearly two decades younger than the typical age of onset for RAD51D-associated cancers⁴⁴, demonstrating an example of the potential for prevention in VUS carriers. Somatic RAD51D mutations impose a similar interpretation burden. Although the two-hit pattern is not universal among all HR tumor suppressors⁴⁵ (e.g, CHEK2), RAD51D is often biallelically inactivated in tumors⁴. We leveraged a large collection of tumor somatic mutations and observed that functionally abnormal SNVs in RAD51D were significantly enriched over functionally normal ones. Notably, this enrichment was not solely driven by truncating variants, but extended to missense loss-of-function, specifically in ovarian cancer. As with germline missense variants, these often remain as VUS, but their resolution could inform the use of treatments such as PARP inhibitors.

Our MAVE analyses are also corroborated by orthogonal functional assays, including HR and protein interaction assays for 70 individually tested variants. We further compared our function scores to a saturation genome editing (SGE) map of RAD51D⁴⁶. These two maps are highly significantly correlated (Pearson’s r = 0.60, P < 10⁻²³⁸) and yielded consistent classifications for 92.6% of the variants scored by both (Supplementary Fig. 8). Among these, for the 266 of these listed in ClinVar with two or more non-VUS interpretations, the two maps agreed with each other and with the ClinVar consensus interpretation in 260 cases (97.7%). This level of concordance is remarkable for maps produced by separate groups using different cell lines (U2OS vs HAP1) and distinct functional selections (olaparib selection vs cell fitness).

Nevertheless, a potential limitation of our cDNA-based MAVE is that it cannot capture variant effects upon mRNA splicing and cannot distinguish LOF caused by biochemical vs expression level changes. We used SpliceAI to flag 120 variants with potential to disrupt mRNA splicing despite having normal protein activity in the context of an unspliced cDNA; we removed these from our comparison to clinical variants and did not apply functional evidence to them for clinical variant classification. This filter captures nearly half of the variants (41/92) that appeared neutral in our assay but were abnormal in SGE, which unlike our approach, can detect loss of function due to splicing disruption. Conversely, our cDNA-based approach detected a broad range of intermediate effects, which may, on their own, be too subtle to perturb cell fitness in SGE and related assays. As it remains unclear whether these hypomorphic variants are pathogenic, we did not use them to apply clinical evidence. Even so, the functional constraint they reveal enabled our new insights into RAD51D function. Likewise, by interrogating the full missense space, our map highlights the constraint patterns such as the complete constraint of glycine 265, and the strong preference for compact hydrophobic residues at leucines 4, 12, and 20.

One important finding from our work is that most clinically identified variants have little to no impact on HR. While it is possible that the majority of these variants are passenger mutations or benign, it suggests that other additional functions of RAD51D may also be critical for tumor suppression. For example, RAD51D has an HR-independent function in replication fork progression; however, the significance of this function in cancer is unknown. The modulatory role of RAD51D may serve critical additional functions to promote RAD51 filament formation, depending on the DNA substrate and cell cycle contexts, thereby facilitating high-fidelity DNA repair. Furthermore, our screening methodologies may not detect variants with replication defects. Future studies examining this important aspect of DNA repair will shed light on the impact of BCDX2 and X3CDX2 replicative functions in promoting genomic integrity and cancer prevention.

Methods

RAD51D Multiplexed Assay of Variant Effect (MAVE)

Tissue culture, cell lines, and reagents

Human osteosarcoma U2OS SCR (sister chromatid recombination) #18 wild-type (gifted from Mauro Modesti;²¹) and RAD51D CRISPR knockout (KO; clone #4, purchased from DSMZ, no. ACC835) were cultured in Corning DMEM supplemented with 10% fetal bovine serum (FBS) and penicillin-streptomycin antibiotic (50 U/mL). Cells were incubated at 37 °C in a 5% CO₂ atmosphere. Cells were transfected using Lipofectamine 2000 (Invitrogen) diluted with OptiMEM serum-free media following the manufacturer’s instructions. Cisplatin and olaparib (Selleck Chem; AZD2281).

Library synthesis and transduction

Saturation mutagenesis libraries were prepared as described^47,48; briefly, a wild-type RAD51D cDNA plasmid (GenBank: NM_002878.4) was synthesized (Twist Bioscience). The cDNA was divided into 6 tiles (Supplementary Table 1), and for each, a mutant oligonucleotide library was designed in which each codon was sequentially replaced by an ‘NNN’ randomer. Mutant oligonucleotides were synthesized together as a single oPool oligo pool (IDT), then PCR-amplified using tile-specific primers to add sequences for cloning into the starting cDNA plasmid. Mutant cDNA libraries were cloned into a ‘Tet-On’ inducible lentiviral expression construct (pCW57.1; Addgene #80922), transformed into Endura electrocompetent cells (Lucigen). Plasmids were prepared with ZymoPure II kits, and pooled lentivirus libraries were prepared by the University of Michigan Vector Core. U2OS RAD51D knockout cells were transduced with WT RAD51D lentivirus or pooled mutant lentivirus, at low multiplicity of infection (MOI < 0.1) such that each cell carries zero or one stably integrated copies of RAD51D cDNA. Per replicate, three million cells were infected with 200 µL lentiviral supernatant in the presence of 8 μg/mL polybrene, sufficient for each mutation to be integrated into 200-300 unique cells per replicate. At 48 h post-infection, transduced cells were selected with 50 µg/mL hygromycin, which was maintained in all subsequent steps.

Olaparib functional selection assay

Mutant library-expressing cells were cultured in experimental triplicates in T175 flasks in DMEM media (supplemented with 10% fetal bovine serum and 1% penicillin-streptomycin) with 1 mg/mL doxycycline added to induce RAD51D expression. From each baseline passage (“P0”) culture, one million cells were collected for gDNA extraction and three million cells were seeded for olaparib selection, with media containing 50 µg/ml hygromycin, 1 mg/mL doxycycline, 1 µg/mL blasticidin (to enforce RAD51D expression), and 250 nM olaparib (as used in ref. ⁴⁹). Cells were grown to 90% confluence for two rounds. At each round, three million cells were harvested for gDNA extraction and then seeded with three million cells to avoid population bottlenecks.

Mutant library sequencing

Genomic DNA from P0 and P2 cell populations was isolated using CleanNA Blood and Tissue magnetic bead kit (Bulldog Bio). To amplify integrated RAD51D mutant libraries, “outer PCR” was performed on a 500 ng gDNA template with primers flanking the RAD51D cDNA using the PrimeSTAR GXL kit (Takara), with 16 replicate PCR reactions for each gDNA sample to maintain library complexity. Gel electrophoresis and shallow shotgun sequencing of the full-cDNA PCR products were performed to confirm the absence of large rearrangements, insertions, or deletions across cell passages.

To quantify mutation frequency in sorted cells, mutant libraries were subjected to deep amplicon tile sequencing⁵⁰ as described⁴⁷. Briefly, 1 µL of outer PCR library product was amplified in a limited ( ≤ 10) cycle PCR, using primers directed at constant bases flanking the targeted tile within the RAD51D cDNA. Primer 5’ tails contained constant sequencing adaptor stub sequences. Each resulting product was then diluted and further amplified to add sequencing adaptors and sample-specific unique dual 10 bp index sequences. Amplicon size was verified by electrophoresis on 6% TBE native PAGE gel stained with SYBR Gold. Indexed libraries were pooled, purified with SPRI beads, and sequenced with paired-end 150 bp reads on an Aviti24 sequencer (Element Biosciences) at the University of Michigan Advanced Genomics Core. Libraries were sequenced to a median depth of 12 million read pairs per replicate.

Function score calculation

Paired-end reads were overlapped, and an error-corrected consensus was taken by PEAR (v0.9.6)⁵¹, then aligned to a RAD51D cDNA reference with bwa mem (v0.7.17)⁵². Variant counts were tabulated by a pipeline implemented in Snakemake (https://github.com/kitzmanlab/tileseq/). Counts were summed across synonymous codons and the resulting amino acid-level counts for each replicate were combined by Rosace (v1.1)⁵³, yielding a single ‘fitness’ score for each variant and a local false sign rate (lfsr) score, a measure of confidence that the sign of the fitness score is correct, taking into account variation across replicates and due to sampling at low-frequency variants⁵⁴ (Supplementary Data 2, 3). To normalize between tiles, within each tile, each variant score s_i was scaled and shifted by a constant factor to center synonymous scores at 0, and nonsense scores at −1: r_i’=(r_i -med_syn)/(med_syn-med_non), where med_syn and med_non are per-tile medians of r_i for synonymous and nonsense variants, respectively. Variants with low initial abundance ( ≤ 100 counts per million at timepoint P0; n = 154, 2.2%) were excluded from further analysis. For DE intolerance analysis, the mean score of DE variants was subtracted from the mean score of all non-DE variants.

Variant classification

We determined functional cutoffs by fitting normal distributions to the scores of nonsense variants (residue ≤304) and synonymous variants, and taking the 97.5^th percentile and 2.5^th percentile of the respective distributions. Variants scoring below the lower cutoff (−0.446) and having an lfsr ≤0.01 were classified as loss-of-function (LOF). Variants scoring above the upper cutoff (score ≥−0.170) were considered as neutral (NEU). Variants with function scores between the two cutoffs and with lfsr ≤0.01 were classified as intermediate (INT), and these were not used for clinical variant analyses. Variants in the intermediate scoring range but with lfsr >0.01 were considered as neutral. Variants with LOF scoring but lfsr >0.01 were excluded from analysis (n = 38, 0.57 %). ClinVar interpretations were downloaded on 9/28/2025. Variants with ‘soft conflicts’ were included only if there were two or more non-VUS interpretations, all in the same direction (e.g., all B/LB or all P/LP), and variants were filtered to retain only those with assertion criteria provided. Splicing effects were predicted by SpliceAI²⁸, and variants with a deltaMax score ≥0.2 were considered potentially splicing-disruptive. Missense variant stability was predicted by ThermoMPNN⁵⁵.

Clinical and population datasets

Population variants were obtained from gnomAD v4.1. Somatic variants in cancer were obtained from AACR Project GENIE (v19.0). Carriers of RAD51D VUS were ascertained from the University of Michigan Breast and Ovarian Cancer Risk Evaluation Center study. Binned ages and family pedigrees with uninformative members disguised to preserve anonymity are presented here. Data collection and sharing procedures were reviewed by the University of Michigan Medical School IRB (studies HUM00112234 and HUM00288570), and the use in this study was declared exempt from human subject regulations.

RAD51D missense VUS selection

RAD51D missense variants were identified in databases (COSMIC, TCGA, ClinVar), as well as through our clinical collaborators and the literature (See Supplementary Data 1). Of the missense RAD51D variants in our list, 70 variants were previously reported in ClinVar as somatic or germline variants identified in breast and ovarian cancers and either CON/P or VUS (with the exception of p.Ser207Leu (S207L) classified as P/LP and p.Arg165Gln (R165Q), p.Arg232Gln (R232Q), p.Glu233Gly (E233G) classified as B or B/LB).

Molecular cloning

All plasmid constructs were prepared through either site-directed mutagenesis of the corresponding RAD51D-WT cDNA template or synthesized by Gene Universal (Gene synthesis and mutagenesis services, Newark, DE). For yeast assays, RAD51D-WT was cloned into the pGAD-C1 backbone at the 5’-EcoRI and 3’-SalI restriction cut sites. Breast and ovarian cancer-identified VUS were created by site-directed mutagenesis using primers (synthesized by Integrated DNA Technologies) that contained the corresponding nucleotide mutation. For yeast assays, RAD51C and XRCC2 were cloned into the pGBD-C1 backbone at the 5’-EcoRI and 3’-SalI restriction cut sites. For cellular HR assays and stable cell line creation, RAD51D-WT and corresponding mutated VUS were subcloned into the 5’-HindIII and 3’-EcoRI restriction sites of the cFlag pcDNA3 backbone. cFlag pcDNA was a gift from Stephen Smale (Addgene plasmid #20011). All mutations were validated through Sanger sequencing (Azenta).

Yeast-2-hybrid and yeast-3-hybrid analysis

Yeast-2-hybrid (Y2H) and yeast-3-hybrid (Y3H) assays were performed as previously described^22,56, with the exception that RAD51D WT and corresponding VUS were subcloned into the pGAD-C1 backbone (as described above) and expressed as C-terminal fusions to the GAL4 activating domain. Wild-type XRCC2 and RAD51C were subcloned into the pGBD-C1 backbone (as described above) and expressed as a C-terminal fusion to the GAL4 DNA binding domain for Y2H and Y3H, respectively. Following co-transformation in the PJ694-alpha yeast strain, 3-4 single colonies were cultured in liquid SC-Leu^–Trp^– (Y2H) or SC-Leu^–Trp^–Ura^– (Y3H) media overnight. Plates were incubated for 72 h at 30 °C before imaging. Colony growth was quantified in ImageJ and normalized to the respective wild-type (positive) and empty vector (negative) controls. Images were adjusted for uniform brightness and contrast using Adobe Photoshop.

Homology-directed repair reporter assay

Plasmid constructs of each RAD51D VUS were subcloned into the cFlag pcDNA3 backbone (Addgene #20011) as described above. pCBA-SceI was provided as a gift from Maria Jasin (Addgene plasmid #26477). U2OS-SCR18-WT and U2OS-SCR18-RAD51D-KO cells were seeded in a single well of a 6-well plate at densities of two and four million cells per well, respectively (due to the slower growth rate of RAD51D knockout cells). Cells were incubated for 12–24 h before liposomal-mediated transfection with pCBA-SceI (2 µg) alone or co-transfected with empty vector (cFlag pcDNA3, negative control) or vector expressing individual RAD51D VUS (2 µg, 1:1 ratio) for 4 h. Transfection efficiency was assessed by transfecting a single well with pEGFP-N1. Following a 48 hr incubation, cells were collected by trypsinization and fixed in an aqueous 2% paraformaldehyde solution. Fixed cells were re-suspended in phosphate-buffered saline (PBS) and analyzed on the BD Accuri C6 Plus flow cytometer. For each sample, 30,000 events were recorded, and results were normalized to the complemented RAD51D WT control. Each variant was analyzed in at least three biological replicates.

Western blot

For analysis of protein levels in yeast, 3-4 single colonies of PJ69-4α co-transformed with pGAD-RAD51D-WT or VUS and pGBD-XRCC2 were cultured overnight. The culture density was measured by spectrophotometry (0.75 OD₆₀₀) and yeast were collected through centrifugation and proteins isolated through TCA precipitation. For protein level analysis in mammalian cells, cell lysates were prepared by lysing at least 4 million cells in RIPA buffer. Protein concentration in cell lysates was measured against a BSA standard curve (0-2.0 mg/mL) using the BCA protein quantification kit. Yeast (25 µL) and cell lysates (10 µg) were resolved on 10% SDS-PAGE gels with 4% stacking gels. Proteins were transferred to PVDF membranes and blocked with either 5% [w/v] milk or 5% [w/v] BSA diluted in Tris-buffered saline supplemented with 1% Tween-20 (TBST, blocking buffer is antibody specific). Antibodies against RAD51D (Abcam, 202063, 1:1000), KAR2 (Santa Cruz, sc-33630, 1:1000), and α-tubulin (Cell Signaling Technologies, 2144, 1:1000) were diluted in their respective blocking buffers and incubated with the membrane at 4 °C overnight. The next day, membranes were rinsed with TBST and then incubated with secondary antibody (Jackson anti-rabbit #111-165-003, 1:10,000) diluted in TBST for 1 h. Yeast blots were developed on film. The chemiluminescence reagent was imaged on a Bio-Rad Gel Doc with exposure times ranging from 5 to 300 seconds.

Clonogenic survival assays

U2OS-SCR18-WT and U2OS-SCR18-RAD51D-KO cells stably complemented with RAD51D WT or RAD51D VUS were seeded in a 6-well plate in DMEM supplemented with FBS (10%) and penicillin-streptomycin antibiotic (50 U/mL) at cell densities ranging from 1000 to 5000 cells per well (for cisplatin treatment) and 10,000 to 20,000 cells per well (for olaparib treatment). For cisplatin treatment, cells were incubated in media supplemented with cisplatin (at least nine unique treatment concentrations ranging from 0.05 nM to 25 µM) for 48 h, at which point cisplatin-supplemented media was removed and replaced with fresh, non-treatment-supplemented media. Cells were incubated for 14 days post-cisplatin treatment, and the media was refreshed 7 days post-treatment. For PARPi treatment, cells were incubated in either DMSO- (0.1%) or olaparib- (at least nine unique treatment concentrations ranging from 0.01 nM to 25 µM) supplemented media. Cells were incubated for a total of 15–21 days following initial treatment, and DMSO or PARPi-supplemented media were refreshed every 3–4 days. At the end of incubation, the media was aspirated, and the cells were rinsed with PBS. They were then fixed in 100% methanol and stained with crystal violet. Cells were imaged on a BioRad ChemiDoc Imager, and colony area was quantified using the colony area plug-in tool in ImageJ. All variants and controls were analyzed in at least three biological replicates.

Protein expression and purification

Purifications of DX2 (RAD51D and XRCC2-FLAG) and BC (RAD51B-His and RAD51C) sub-complexes were carried as previously described¹². For purification of the CX3 (RAD51C-His and XRCC3-Strep), Hi5 insect cells were infected for 48 hrs with baculoviruses expressing RAD51C-His and XRCC3-Step, cells were collected and kept frozen until further used. All the purification steps were carried out at 0–4 ^oC. The crude cell lysate was prepared from 6 g cell pellet (from 800 mL insect cell culture) by sonication in 50 mL T buffer (25 mM Tris-HCl, pH 7.5, 10% glycerol, 0.5 mM EDTA, 1 mM DTT, 0.05% IGEPAL, 1 mM PMSF, and protease inhibitors) containing 300 mM KCl, 5 mM ATP, and 2 mM MgCl₂. Lysates were clarified by centrifugation at 100,000 × g for 1 h. The supernatant was incubated with 1 mL Ni-NTA affinity resin for 1 h in the presence of 25 mM imidazole. The resin was sequentially washed with 100 mL T buffer containing 1 M KCl and then 50 mL T buffer containing 300 mM KCl, with all wash buffers containing 25 mM imidazole, 2 mM ATP and 2 mM MgCl₂. Bound proteins were eluted with 10 mL T buffer containing 200 mM imidazole, 300 mM KCl, 2 mM ATP, and 2 mM MgCl₂. The partially purified CX3 complex was passed through 1 mL HiTrap StepXT column, the column was further washed with 15 mL T buffer containing 300 mM KCl, 2 mM ATP, and 2 mM MgCl₂ (T300-ATP), and further eluted in same buffer also containing 50 mM biotin. Peak fractions were pooled and concentrated to 0.4 mL in an Amicon 30 concentrator and subjected to size exclusion chromatography in a Superdex200 Increase 10/300 GL column in (T300-ATP). Peak fractions containing CX3 complex were concentrated and snap frozen in liquid nitrogen and stored in −80 °C.

DNA binding

For DNA binding, 2 nM of 5’ Cy5 labeled 80-nt ssDNA¹⁵ was incubated with the indicated concentration of purified RAD51 paralog complexes in 10 μl reaction buffer (50 mM Tris-HCl, pH 7.5, 120 mM KCl, 1 mM DTT, 1 mM ATP, 1 mM MgCl₂ and 100 ng/μl BSA) for 10 min at 37^oC. DNA-protein complexes were resolved on the 5% polyacrylamide gels in Tris-borate buffer (45 mM each) and gels were visualized using the ChemiDoc imaging system (Bio-Rad) and proportion of bound versus unbound DNA was quantified using the Fiji ImageJ software (v2.90/1.53t). The mean values were calculated from the data obtained from at least three independent experiments and plotted using the R (3.6.1) package ggpubr. Error bars indicate standard deviation. A two-way Analysis of Variance (ANOVA) was conducted to evaluate the effects of RAD51D variants and protein concentration on DNA binding shift. Post-hoc comparisons stratified by protein concentration were performed using Dunnett’s adjustment, comparing each variant back to the WT control baseline.

ATPase assays

The indicated combinations of BC and DX2 subcomplexes (0.5 μM) were incubated in the presence of ssDNA (20 μM nucleotides, phiX174 virion) in 10 μl of reaction buffer (20 mM HEPES, pH 7.5, 1 mM DTT, 1 mM MgCl₂, and 30 mM KCl) containing 0.05 mM ATP with 0.25 μCi [γ-³²P] ATP at 37 °C. After 1 hour incubation reaction was stopped by adding 10 μl of 0.5 M EDTA and the level of ATP hydrolysis was determined by layer chromatography on PEI cellulose F sheets (Millipore, 105579) in 0.375 M potassium phosphate (pH 3.5), followed by phosphorimaging analysis. The mean values were calculated from the data obtained from at least three independent experiments and plotted using the R (3.6.1) package ggpubr. Error bars indicate standard deviation. Significance of differences in mean values was calculated using the unpaired t-test.

Pull-down assay

Pulldown assays were performed by incubating 1 μg of the bait complex DX2 (RAD51D and XRCC2-FLAG) with 1 μg of either prey complex BC (RAD51B-His and RAD51C) or CX3 (RAD51C-His and XRCC3-Strep) in the presence of 2 μL α-FLAG resin (Pierce, A36804). Incubation was carried out in T buffer supplemented with 100 mM KCl, 2 mM ATP, and 2 mM MgCl₂ (T100-ATP) at 4 °C for 1 h with gentle agitation. After incubation, reactions were centrifuged to collect the flow-through (FT), and the resin was washed twice with T100-ATP. The resin-bound proteins were analyzed by SDS-PAGE followed by Coomassie blue staining.

Homologous DNA pairing and strand exchange assay

To analyze strand exchange by RAD51 in the presence of RPA, the 80 nt ssDNA (8 nM)¹⁹ was premixed with RPA (30 nM) in reaction buffer (35 mM Tris pH 7.5, 1 mM DTT, 0.1 μg/μl BSA, 2 mM AMP-PNP and 1 mM MgCl₂). This is followed by addition of RAD51 or RAD51 and indicated CX3 and DX2 combinations and incubation at 37 °C for 10 min. The reaction was initiated by adding homologous 40 bp 5’ Cy5-labeled dsDNA (8 nM)¹⁹ and 4 mM spermidine hydrochloride (pH 7.5) and incubated for 30 min at 37 °C. The reaction was stopped by adding stop mix to final concentration 0.1% SDS and 0.5 mg/mL proteinase K and 1 X loading buffer (7% glycerol, 10 mM Tris pH 8.0, 1 mM EDTA and Orange G) for 10 min at 37 °C. The samples were resolved in 8% native polyacrylamide gels in TGS (160 mM Tris pH 8.3 / 1.2 M Glycine / 0.22 mM SDS) buffer at room temperature. Gels were visualized using the ChemiDoc imaging system (Bio-Rad) and quantified using the Fiji ImageJ software (v2.90/1.53t). Mean values were calculated from the data obtained from four independent experiments. Data were processed and plotted using the R (3.6.1) package ggpubr. Standard deviations were calculated and presented as error bars together with the mean values. Significance of differences in mean values was calculated using the unpaired t test.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Author contributions have been credited using the Contribution Roles Taxonomy (CRediT) system. Conceptualization (K.E.D., S.L.H., Y.R., N.J.G., S.K.O., P.S., J.O.K., K.A.B.), Data Curation (S.L.H., J.O.K.), Formal Analysis (K.E.D., S.L.H., Y.R., N.J.G., J.O.K.), Funding Acquisition (S.K.O., P.S., J.O.K., K.A.B.), Investigation (K.E.D., S.L.H., Y.R., N.J.G., P.S.P, G.S.G., C.M.S., K.G.O., E.R., E.G., N.B., A.K.K., A.M.D.), Resources (M.W.S., K.J.M., S.D.M., K.L.N., S.M.D., L.M.S.), Software (J.O.K.), Supervision (S.K.O., P.S., J.O.K., K.A.B.), Validation (S.C., M.W.S., L.M.S., J.O.K., K.A.B.), Visualization (K.E.D., S.L.H., Y.R., N.J.G., G.S.G., C.M.S., S.K.O., J.O.K., K.A.B.), Writing – Original Draft Preparation (K.E.D., S.L.H., Y.R., N.J.G., J.O.K., K.A.B.), Writing – Review and Editing (S.L.H., Y.R., N.J.G., S.K.O., P.S., J.O.K., K.A.B.)

Peer review

Peer review information

Nature Communications thanks Andrew Hong and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This work was supported by the National Institutes of Health grants [R01 ES031796 (K.A.B.), R01 ES030335 (K.A.B.), GM115568 (S.K.O), GM128731 (S.K.O), R35 CA241801 (P.S.), P01 CA275717 (P.S.), R01 CA293655 (S.K.O.), R35 GM153286 (J.O.K.), the American Cancer Society (DBG-1243615 to K.A.B), the Penn Center for Genome Integrity, the Basser Center for BRCA), the Department of Defense grants (BC201356 to K.A.B.). N.G. is supported by NIEHS (5T32ES019851-10). S.K.O. is the recipient of a CPRIT Rising Star Award (RR200030). S.L.H. is supported by an American Cancer Society Postdoctoral Fellowship.

Data availability

Variant function scores and associated annotations are provided in Supplementary Data 2 and 3. Processed variant counts are available at NCBI’s Gene Expression Omnibus through GEO Series accession number GSE312720 [https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc= GSE312720]. Raw sequencing reads are available at NCBI’s Bioproject PRJNA1375770. Functional scores (for HR, Y2H/Y3H, and biochemical data) are available in the extended data files. Images of raw blots and data points for graphs are available in the Source Data File. Source data are provided with this paper.

Code availability

The TileSeq pipeline is available on GitHub (https://doi.org/10.5281/zenodo.21707670).

Competing interests

J.O.K. serves on the scientific advisory board of Myome, Inc. The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77941-y.