A structural framework for mutation-guided immunotherapy sequencing of GPRC5D-targeted T cell engagers in multiple myeloma Authors: Ricardo Ahrens, Kyra Ahrens, Maximilian Al-Bazaz, Lisa Leypoldt, Carsten Bokemeyer, Winfried Alsdorf, Katja Weisel Published: 2026 Dear Editor, G-protein–coupled receptor family C group 5 member D (GPRC5D) is a leading immunotherapy target in multiple myeloma (MM), with diverse GPRC5D-targeting agents—including T-cell engagers (TCE), antibody–drug conjugates, and chimeric antigen receptor (CAR) T cells—being evaluated clinically for relapsed/refractory MM. In a recently published relapse cohort, 13 of 19 patients treated with anti-GPRC5D therapy had acquired GPRC5D alterations, with coding single-nucleotide variants or indels in 7 of these 13 patients [1]. Because GPRC5D-directed agents can bind distinct receptor regions, a single coding variant may have different implications for different drug formats, and the same relapse biopsy can encode either a target-class exit event or a mutation that disables only one antibody footprint. Current guidance supports BCMA- and GPRC5D-directed immunotherapy sequencing but remains largely target-based rather than mutation-based [2, 3]. Clinically, relapse must be separated into surface loss, which argues against further GPRC5D targeting, and surface-expressed epitope escape, which may preserve alternate-epitope sensitivity. This raises a clinically urgent question: when a specific GPRC5D mutation is detected at relapse, which agent should the patient receive next? Talquetamab has established clinical activity as an approved GPRC5D TCE [4]; forimtamig, although no longer an active clinical sequencing option, provided the key clinically measured alternate-epitope comparator in the published post-GPRC5D relapse dataset [1] and in early clinical development [5]. Its value here is mechanistic: it is an experimentally anchored N-terminal GPRC5D binder that tests whether talquetamab epitope escape is intrinsically rescuable by an alternate-epitope binder rather than by leaving the target class. We built a two-stage structural model that separates antigen loss from epitope loss (Fig. 1). First, a surface-expression module classifies truncating or frameshift variants as absent and uses auxiliary stability estimates to flag missense mutations likely to destabilise the receptor fold. Second, surface-expressed variants are scored for agent-specific binding disruption using the talquetamab–GPRC5D cryo-EM structure (PDB 9IMA) [6]. This contact-energy binding model is the central method; the stability predictors are used to distinguish folded, targetable receptors from destabilised receptors that should exit the GPRC5D class. D239N removes the negative charge from the talquetamab CDR H3 salt-bridge anchor, whereas Y257S lies outside either antibody interface and instead disrupts the GPCR transmembrane core. Generic variant-effect scores do not make the receptor-loss/epitope-escape distinction needed for same-target sequencing. Because no forimtamig co-structure is available, the GPRC5D–scFv150_18 structure (PDB 8YZK), which maps to the same N-terminal region, was used as a structural proxy [7]. The full mathematical specification, tool implementation, parameter perturbation analysis, cross-complex validation, species cross-reactivity analysis and SKEMPI 2.0 benchmark are shown in the Supplementary Information [8]. Mutation-specific triage of GPRC5D escape. a GPRC5D receptor topology showing the talquetamab extracellular-loop/transmembrane (ECL/TM) pocket, the N-terminal forimtamig-region proxy epitope (residues 5–16), D239 at the ECL3/TM7 edge of the talquetamab pocket and the buried Y257 position. b Interface schematic derived from the talquetamab–GPRC5D cryo-EM structure (PDB 9IMA). GPRC5D D239 forms a three-salt-bridge clamp with R102 in talquetamab complementarity-determining region H3; D239N removes the negative charge and is predicted to disrupt talquetamab binding selectively, while sparing the N-terminal forimtamig-region epitope. c Clinical triage algorithm separating truncating or frameshift variants, destabilising surface-expression loss and surface-expressed epitope mutations. d Predicted versus observed binding mean fluorescence intensity across the 10 published mutation-by-agent validation conditions. e, f D239N dose-response data showing loss of talquetamab binding and retained forimtamig binding; in (f), D239 lies outside the forimtamig-region epitope. The model reproduced the clinically relevant binding hierarchy across all five experimentally characterised GPRC5D variants in the post-GPRC5D relapse validation cohort (10 mutation-by-agent conditions) [1]. D239N, a charge-neutralising substitution at the talquetamab CDR H3 salt-bridge anchor, was correctly classified as talquetamab escape with retained forimtamig binding. Although D239 accounts for only a small fraction of interface contacts, D239N produces an approximately 275-fold talquetamab cytotoxicity shift; in contrast, the cytotoxicity of forimtamig was retained [1]. The D239N–forimtamig condition is therefore the key independent test because the model was calibrated on D239N–talquetamab but still had to recognise that the same mutation lies outside the N-terminal epitope. In contrast, Y257S and the truncating variants R233Ter, E146Ter and L174Wfs were classified as target-class escape because they disrupt surface expression or receptor integrity. Categorical agreement of the model was 8 of 10, with both discrepancies involving Y257S, where residual suprapharmacological signal did not alter the treatment decision. All three clinical actions were concordant: preserve GPRC5D targeting only for D239N-like epitope escape, avoid GPRC5D agents for surface-loss or truncating lesions, and do not overinterpret weak residual binding near assay background. The same energy terms also transferred without refitting to the independent 8YZK antibody complex, cross-species reactivity data and 115 SKEMPI interface mutations, supporting the model as a constrained biophysical classifier rather than a curve fit to the validation variants. The saturation scan defines the prospective escape landscape. Across all 6 555 possible single-amino-acid substitutions in GPRC5D, only 21 variants selectively abolished talquetamab while retaining the N-terminal binder proxy, concentrated at D239, N160 and Q86 (Fig. 2). No single-nucleotide-accessible variant showed the reciprocal pattern. Intersecting this map with codon accessibility identified seven single-nucleotide accessible talquetamab-selective escape variants, including the clinically observed D239N. This small number is clinically useful: a targeted relapse assay does not need to solve the entire GPRC5D mutational landscape to distinguish the main actionable class from truncation or antigen loss. Prospective escape landscape and dual-target hypothesis. a Per-position vulnerability score from the saturation scan of all 6 555 possible GPRC5D single-amino-acid substitutions, with receptor topology shown above the plot. D239 is the dominant talquetamab-selective vulnerability; N160 is a secondary hotspot. b Differential sensitivity map comparing predicted talquetamab and N-terminal binder proxy effects for each substitution. Red indicates talquetamab-selective loss, blue indicates greater N-terminal proxy sensitivity and white indicates no differential effect. c AlphaFold 3 cross-agent chain-pair ipTM changes for D239N, Y257S and D239N + Y257S across five GPRC5D-binding arms, expressed relative to matched wild-type baselines. Negative values indicate lower interface confidence and should be interpreted as structural hypotheses, not measured affinity. d Ramantamig avidity matrix across 88 GPRC5D-by-BCMA resistance combinations. Under the dual-target avidity model, single-arm escape with the alternate tumour antigen intact reduces rather than abolishes engagement, whereas concurrent disruption of both arms is predicted to abolish binding. It also turns the model into a practical panel-design problem: a few hotspot codons can be prioritised for clinical reporting while preserving the broader rule that truncation and antigen loss override epitope-level interpretation. The asymmetry reflects binding architecture: talquetamab depends on a small number of high-energy contacts, especially D239, whereas the N-terminal binder distributes binding energy across a compact peptide segment. The practical implication is that a future clinically available N-terminal GPRC5D binder would be rational after D239/N160/Q86-like talquetamab escape. Conversely, biallelic deletion, epigenetic silencing, truncation or severe destabilisation should prompt a shift away from the GPRC5D class. This is why mutation detection, not interpretation, becomes the limiting step. GPRC5D coverage is not guaranteed across targeted next-generation sequencing panels, yet the actionable coding space is small enough for targeted incorporation. A relapse report that only states “GPRC5D altered” is insufficient: D239N and R233Ter imply opposite decisions. Serial marrow sequencing or circulating tumour DNA could, in principle, identify an emerging epitope-specific clone before it dominates the relapse, whereas complete antigen loss would prompt a different therapeutic branch. This fits recent sequencing guidance that increasingly emphasises molecular profiling, but moves the decision one level deeper: from target choice to residue-level escape mechanism. The same principle is emerging across MM immunotherapy. BCMA exhibits both epitope-specific escape and complete antigen loss under sequential immune pressure [9], while genetic or epigenetic GPRC5D inactivation can remove the target altogether [10]. Thus, antigen-class sequencing should move from a target-name framework to a genotype-and-epitope framework. The broader cross-agent screen of our proposed model also included prospective GPRC5D formats outside the TCE class, including AZD0305 as an antibody–drug conjugate candidate and the arlocabtagene autoleucel (arlo-cel) clone 109/MCARH109 CAR binder, to test whether epitope vulnerability tracks with drug class or with binding footprint (Supplementary Note 3) [11]. Ramantamig extends the same logic to multi-antigen targeting. This BCMA×GPRC5D×CD3 trispecific antibody is designed to preserve tumour-cell engagement when one antigen arm is weakened [12], with early phase 1 activity reported [13]. We therefore asked whether single-arm escape would be sufficient to abolish ramantamig binding. This analysis should be read as hypothesis-generating: the GPRC5D-arm prediction assumes that GC5B680 shares the talquetamab-region footprint, the BCMA arm uses a PDB 4ZFO-derived structural proxy, and AlphaFold 3 chain-pair ipTM is a confidence metric rather than an affinity measurement [14]. Subject to these assumptions, no single GPRC5D or BCMA mutation alone abolished dual-target engagement; complete resistance required concurrent disruption of both arms (4 retained, 36 reduced and 48 abolished among 88 GPRC5D-by-BCMA variant/condition combinations; Fig. 2). That caveat is clinically important because concurrent loss of CD38, BCMA and GPRC5D has been observed in advanced hepta-refractory myeloma [15]. Multi-targeting agents may therefore be most useful before repeated single-antigen pressure selects diversified multi-antigen loss, rather than only as late salvage after several antigen-directed therapies have failed. Several limitations should be considered. The forimtamig and ramantamig analyses depend on structural proxies: 8YZK for N-terminal GPRC5D binding, an assumed GC5B680 footprint for ramantamig and a PDB 4ZFO-derived BCMA-arm proxy. AF3 chain-pair ipTM provides structural confidence rather than affinity and, in the talquetamab complex, did not itself detect D239N disruption (Supplementary Note 3). The model captures coding mutations only and does not detect promoter methylation, biallelic deletion without a coding lesion, low target expression, antigen shedding, or T-cell exhaustion. The surface-expression threshold is calibrated by one destabilising mutation, so variants with ΔΔGstab near 2–4kcal/mol should be treated cautiously. Stability predictors are trained largely on soluble proteins, and performance for seven-transmembrane GPCRs remains less established. Finally, no prospective clinical trial has yet assigned post-talquetamab therapy by GPRC5D genotype. Despite these constraints, the output of the model is straightforward: a truncating, silencing, or destabilising event argues against further GPRC5D targeting, whereas a surface-expressed mutation in the talquetamab epitope argues for an alternate-epitope binder if one is available. A dual-target agent may raise the resistance barrier, but should be used before serial immune pressure creates multi-antigen loss. Our proposed framework is target-agnostic, requiring only an antigen–antibody complex structure and one calibration mutation. With the GPRC5D therapeutic landscape now spanning bispecific antibodies, trispecific antibodies, CAR T-cell therapies, and antibody–drug conjugates, and as clonal evolution under sequential immune pressure creates increasingly complex resistance patterns [15], prospective resistance prediction is becoming essential for rational treatment sequencing. This approach may extend beyond MM to any targeted immunotherapy facing antigen escape. Supplementary information Supplementary Note Acknowledgements Structure predictions were performed using AlphaFold Server (Google DeepMind and Isomorphic Labs). The authors received no external funding for this study. Author contributions RA conceived and designed the study, interpreted results, provided clinical expertise and wrote the manuscript. KA developed and optimised the computational framework, performed computational analyses and co-wrote the manuscript. MAB and LL provided clinical context and reviewed the manuscript. CB provided departmental supervision. WA and KW interpreted results, provided clinical expertise, supervised the study and critically revised the manuscript. All authors reviewed and approved the final manuscript. Data availability Structures used in this study are publicly available from the Protein Data Bank: 9IMA, 8YZK, 4ZFO, 1VFB, 3HFM, 1DQJ, 1JRH, 1DVF, 3NGB and 1AHW. Validation data are from published flow-cytometry and dose-response assays [1]. SKEMPI 2.0 data are available at https://life.bsc.es/pid/skempi2/. Code availability Custom code for the contact-energy binding model, surface-expression prediction, avidity framework and AlphaFold 3 analysis pipelines is available from the corresponding author upon request. Model parameters, energy terms and classification thresholds are reported in the Supplementary Information. Competing interests RA reports travel support from Amgen, Sanofi, BeOne and Pfizer and served as a consultant for Stemline and Pfizer. KA declares no competing interests. MA-B reports travel support from Stemline. LL reports travel support from Sanofi, Oncopeptides, J&J and Amgen, served as a consultant for GSK, Sanofi, J&J and Pfizer, received honoraria from Adaptive, Amgen, AbbVie, J&J, Pfizer, Sanofi, Regeneron, Takeda and AstraZeneca and received research funding from AbbVie, GSK (both to the institution) and Amgen. CB received honoraria from AOK Germany, AstraZeneca, Bayer, Berlin Chemie, Bristol Myers Squibb, GSO Research Organisation, J&J, med update, Merck Serono, MSD, Novartis, Roche and Sanofi. WA has received travel grants from J&J, Immatics and BioNTech, honoraria from GSK, Astellas, Janssen and AstraZeneca and research funding (to the institution) from Affimed and BioNTech. KW served as a consultant for AbbVie, Adaptive Biotech, Amgen, BeOne, GSK, J&J, Karyopharm, Stemline, Oncopeptides, Pfizer, Regeneron, Roche, Sanofi, and Takeda and received honoraria from AbbVie, Amgen, AstraZeneca, BeOne, Bristol Myers Squibb, Celgene, GSK, J&J, Novartis, Oncopeptides, Sanofi, Stemline, and Takeda and served on the Board of Directors or advisory committees for AbbVie, Adaptive Biotech, Amgen, BeOne, Bristol Myers Squibb, Celgene, GSK, J&J, Karyopharm, Stemline, Novartis, Oncopeptides, Pfizer, Regeneron, Roche, Sanofi, and Takeda and received research funding from AbbVie, Amgen, GSK, J&J, Pfizer, Bristol Myers Squibb, Novartis, Sanofi, and Takeda. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41408-026-01616-6. References Associated Data Supplementary Materials Supplementary Note Data Availability Statement Structures used in this study are publicly available from the Protein Data Bank: 9IMA, 8YZK, 4ZFO, 1VFB, 3HFM, 1DQJ, 1JRH, 1DVF, 3NGB and 1AHW. Validation data are from published flow-cytometry and dose-response assays [1]. SKEMPI 2.0 data are available at https://life.bsc.es/pid/skempi2/. Custom code for the contact-energy binding model, surface-expression prediction, avidity framework and AlphaFold 3 analysis pipelines is available from the corresponding author upon request. Model parameters, energy terms and classification thresholds are reported in the Supplementary Information.