Can saturation mutagenesis classify RAD51D variants of uncertain significance?

A survival readout turns variants into function scores

The study asks whether a laboratory measurement can tell a harmful RAD51D missense variant from a harmless one. RAD51D is a tumor suppressor needed for homologous recombination, the DNA repair pathway that fixes double-strand breaks, and pathogenic variants in it are linked to breast and ovarian cancer. Most missense variants found in patients remain variants of uncertain significance, so clinicians lack a functional basis for classifying them.

The authors used a multiplex assay of variant effect. They built a library covering 97.6% of the 6,888 possible single-amino-acid RAD51D variants, with each cell carrying at most one stably integrated variant copy in a U2OS cell line whose endogenous RAD51D had been knocked out by CRISPR. Cells were treated with 250 nM olaparib, a PARP inhibitor that kills cells unable to repair damage through homologous recombination, for two passages. Variant frequencies were read by deep amplicon sequencing before and after selection, and each variant received a function score; negative scores mean the variant was depleted, so it failed to restore repair.

The scale matters because it converts a per-variant biochemical question into a ranked measurement across nearly the whole protein. The score is not a clinical verdict by itself: it is a functional readout in one cell line under one drug selection, and the authors used it as evidence within a classification framework rather than as a stand-alone diagnosis.

Score cutoffs split variants into three functional classes

To make the continuous scores usable, the authors set thresholds from control variants. Synonymous variants, which change no amino acid, defined the neutral end, and nonsense variants, which truncate the protein, defined the loss-of-function end. Variants with scores between −0.170 and −0.446 were called intermediate, and those below −0.446 were called loss-of-function, provided they also passed a significance cutoff on the local false sign rate.

The controls behaved as intended. Stop-gain variants before residue 305 had a mean score of −0.989, while synonymous variants averaged 0.023, and the two groups were perfectly separated with a precision-recall area of 1.00 across 494 variants. Truncations from residue 306 onward scored like synonymous variants, which the authors interpret as the mostly unstructured C-terminal region being dispensable in this assay. Among missense variants, 74.5% (4,593) were neutral with a mean score of −0.07, 16.8% (1,038) were loss-of-function with a mean of −0.78, and 8.6% (536) were intermediate with a mean of −0.31.

These class boundaries are the unit that later clinical comparisons rely on. A variant called intermediate is not the same as one called loss-of-function, and the paper treats them differently when counting how many uncertain variants can be reclassified.

Known pathogenic and benign variants land on opposite sides

The strongest check on the map is agreement with variants that already have clinical classifications. The authors curated 211 unique protein-level variants from ClinVar, including 155 benign or likely benign and 36 pathogenic or likely pathogenic entries, plus 28 entries with conflicting submissions that they resolved as pathogenic in one case and benign in 27. All 37 pathogenic or likely pathogenic variants scored as loss-of-function, and all 174 benign or likely benign variants scored as neutral.

Individual cases support the same picture. The common polymorphism R165Q scored neutral, consistent with its 14.1% minor allele frequency in gnomAD. The two missense variants previously classified as pathogenic or likely pathogenic, S207L and S207W, were both called loss-of-function, with S207L milder at −0.503 and S207W more severe at −1.703.

The authors also translated the scores into ClinGen OddsPath ratios, which quantify how well a functional assay separates known pathogenic from known benign variants. The ratio was 174 for functionally abnormal variants and 0.027, or 1/37, for functionally normal ones, corresponding to strong evidence toward pathogenicity and toward benignity respectively. That framing is what allows the assay to contribute points in a clinical classification rather than only describing biochemistry.

Most uncertain variants score functionally normal

The clinically consequential test is the 789 RAD51D missense variants listed in ClinVar as uncertain or conflicting. The authors excluded 49 predicted to affect splicing and 68 that were intermediate or failed the significance cutoff, leaving 672 variants that could be assigned a direction. Of these, 87% scored functionally normal, which the authors note would be enough, absent conflicting evidence, to move them toward likely benign. The remaining 85 variants, 12.6%, scored loss-of-function and could receive points toward pathogenic classification.

The distribution is the paper's practical answer to the classification question: the assay does not reclassify every uncertain variant, but it gives most of them a strong push in one direction. The authors also note that nearly two-thirds of the 85 loss-of-function variants were absent from gnomAD, and that among those present the cumulative minor allele frequency was 2.03 × 10−4. They present that figure as a lower-bound estimate that RAD51D missense loss-of-function variants may contribute to hereditary breast and ovarian cancer risk for roughly 1 in 20,000 individuals.

That risk figure is an inference from population frequency, not a measured penetrance, and it applies to the missense loss-of-function subset rather than to all RAD51D variants. The reclassification language also depends on the ClinGen point system, so the assay supplies evidence whose weight is set by that framework.

Damage concentrates in the DNA-binding and ATPase core

Mapping scores onto structure shows where substitutions hurt most. Variants at buried residues were on average more deleterious than those at surface-exposed residues, with mean scores of −0.272 versus −0.113. But the screen also found surface-exposed residues near the RAD51C or XRCC2 interfaces that were highly deleterious, which a simple buried-versus-exposed rule would miss.

Several residues were almost substitution-intolerant. At the Walker A motif, the conserved threonine T114 tolerated only serine. At the Walker B aspartate D206, 18 of 19 substitutions produced strong loss of function with a mean score of −1.22, with the chemically similar glutamate less severe at −0.52. Residues near the single-stranded DNA backbone at positions 263–267 strongly rejected glutamate and aspartate substitutions, which the authors attribute to repulsion from the negatively charged backbone, and glycine 265 tolerated no substitution at all.

These patterns identify hotspots of deleterious variants and give the map a mechanistic reading: the DNA-binding loops and ATPase core are the regions where amino acid changes most severely compromise homologous recombination in this assay.

RAD51D appears to restrain BCDX2 ATP hydrolysis

The biochemical work addresses what RAD51D actually does inside its complexes. RAD51D forms a heterotetramer, BCDX2, with RAD51B, RAD51C, and XRCC2, and it also belongs to a second complex, X3CDX2, with XRCC3, RAD51C, and XRCC2. Only the RAD51B–RAD51C pair hydrolyzes ATP; the RAD51D–XRCC2 pair regulates that activity through a mechanism the paper set out to define.

Using purified subcomplexes, the authors found that adding the RAD51D–XRCC2 pair to RAD51B–RAD51C reduced ATP hydrolysis. Variants defective in single-stranded DNA binding failed to produce that reduction, leaving hydrolysis comparable to the RAD51B–RAD51C pair alone. The RAD51D–RAD51C interface within BCDX2 therefore appears essential for regulating ATPase activity, and the authors hypothesize that RAD51D's primary function is paradoxically to slow BCDX2 ATP hydrolysis, allowing time and space for RAD51 filament assembly.

This is presented as a hypothesis rather than a demonstrated in vivo causal chain. It does, however, explain why interface variants that look harmless on a structural surface map can still be deleterious: they break the regulation rather than the fold.

One allele uncouples BCDX2 from X3CDX2 activity

The screen also turned up a variant that separates two functions. RAD51D-L4H showed no discernible biochemical defect within BCDX2 and still formed the X3CDX2 complex, yet it bound DNA more weakly within X3CDX2 and failed to promote RAD51-mediated strand exchange on RPA-coated single-stranded DNA. The authors describe it as a separation-of-function allele that may uncouple BCDX2 and X3CDX2 activities.

In strand exchange assays, wild-type X3CDX2 promoted RAD51-mediated exchange on RPA-coated DNA, while the tested X3CDX2 mutant complexes did not, consistent with the homologous recombination defects seen in the cellular assay. Six of the seven deleterious cancer-identified variants examined altered regulation of RAD51B–RAD51C ATPase activity, either by weakening RAD51D interaction with RAD51C or by reducing BCDX2 DNA binding, and six also affected X3CDX2 formation.

The takeaway is that RAD51D supports homologous recombination through at least two routes: regulating BCDX2 ATPase activity and supporting X3CDX2 strand-exchange activity. A variant's functional score can therefore reflect either route, which is why the biochemical follow-up matters alongside the pooled survival readout.

Use the map as strong evidence, not a final verdict

The paper's most decision-relevant comparison is between the assay's two directions of evidence. A loss-of-function score carries 4 points toward pathogenicity, and a neutral score carries 4 points toward benignity, but intermediate scores and variants failing the significance cutoff were excluded from the uncertain-variant count rather than forced into a class. That exclusion is the boundary a reader should keep in view: the 87% and 12.6% figures describe the 672 interpretable variants, not all 789 uncertain entries.

The comparison with computational prediction sharpens the priority. Among variants that AlphaMissense called pathogenic with high confidence, 35% were functionally normal in this screen, which the authors read as a potentially high false-positive rate for that predictor. For RAD51D, a measured function score therefore carries information that a sequence-based prediction does not, and the paper's own framing places the strongest weight on the experimentally scored loss-of-function and neutral calls.

The practical priority is thus to use the map where it gives a strong directional call, to treat intermediate and non-significant variants as unresolved, and to read the biochemical mechanism as the explanation for why interface and ATPase-core variants behave as they do rather than as a separate clinical test.

Sources used in this answer