HDX-MS-Guided computational modeling for rapid and accurate antibody-antigen interface mapping Authors: Will McMahon, Hanzhi Zhang, Xiuling Li, Vaheh Oganesyan, Sterling Payne, Nurjahan Mehzabeen, Nydia van Dyk, Wenhai Liu, Wesley Howard, Pengyi Zhao, Yeying Zhang, Mulin Fang, Xiaoyu Chen, Yariv Mazor, Yu Shi Published: 2026 ABSTRACT ABSTRACTAccurate mapping of paratopes and epitopes at antibody-antigen (Ab-Ag) interfaces is essential for the rational design of therapeutic antibodies in early drug development. Traditional methods for determining high-resolution Ab-Ag complex structures, such as X-ray crystallography and cryo-electron microscopy, are labor-intensive and require substantial time and resources. Although recent computational approaches have increased throughput, prediction accuracy and reliability remain limited, particularly without experimental structural constraints. Here, we present a hydrogen-deuterium exchange mass spectrometry (HDX-MS)-guided structural modeling workflow for efficient and accurate epitope and paratope mapping across three prospective case studies in therapeutic protein discovery and engineering. Across these systems, peptide-level HDX-MS data were sufficient to guide docking and structural modeling, whereas targeted electron transfer dissociation (ETD) HDX-MS was applied in one system as a higher-resolution, orthogonal approach to further localize critical binding residues. The workflow was applied to diverse Ab-Ag architectures spanning Fab and VHH binders and homodimeric and heterodimeric antigen ectodomains, and HDX-MS data were integrated with HADDOCK, AlphaFold2, and Protenix to generate Ab-Ag complex models. Comparison with subsequently obtained in-house X-ray crystal structures, used here as orthogonal high-resolution reference structures, showed strong agreement in overall binding orientation and interface placement. These results demonstrate the robustness and practical utility of this integrative approach for structure-informed antibody engineering in therapeutic protein discovery. Introduction Identification of antibody paratopes and antigen epitopes is fundamental to the development of antibody-based therapeutics. The epitope is a distinct antigenic region recognized and bound by a particular antibody, often dictating the antibody’s specific function and its mechanism of action.1 Epitopes located at ligand-binding or allosteric sites can determine agonistic or antagonistic properties,2 while the spatial context of the epitope, such as concave, flexible, or membrane-proximal regions, affects cellular accessibility and rates of internalization.3 Evolutionarily conserved epitopes are particularly important for achieving cross-reactivity across multiple species.4 By contrast, the paratope comprises the discrete antibody residues, typically within the complementarity-determining regions (CDRs), that directly interact with the epitope and are therefore crucial for antibody engineering in early drug development. Accurate paratope characterization enables identification of key binding residues and facilitates rational approaches for affinity maturation and optimization of therapeutic antibody candidates.5 Traditional experimental techniques for elucidating antibody-antigen (Ab-Ag) complex structures are often time- and resource-intensive. A commonly used approach involves co-crystallizing fragment antigen-binding (Fab) domains with their respective antigens followed by X-ray crystallography, but this can be limited by difficulties in producing suitable crystals and diffraction-quality samples.6 Single-particle cryo-electron microscopy (cryo-EM) offers a relatively faster alternative for some systems, yet it requires specialized instrumentation and rigorous sample preparation that may be difficult to justify in early-stage discovery.7 Alanine scanning mutagenesis provides another indirect strategy for mapping epitopes and paratopes,8 but such mutations may compromise protein stability or alter the native binding interface, thereby limiting reliability. Various in silico tools have been developed for predicting epitopes, paratopes, and Ab-Ag complex structures, but predictive accuracy remains a major challenge. For example, Paragraph, a leading paratope prediction tool, has a precision-recall area under the curve (PR-AUC) of 0.696,9 while the widely used docking algorithm HADDOCK has a success rate of 25%.10 Moreover, both Paragraph and HADDOCK rely on high-resolution crystal or cryo-EM structures of Fab and antigen. Recent advances in this space, such as AlphaFold3, report a 60% success rate for Ab-Ag complex prediction.11 Despite these advances, the moderate predictive rates and dependency on structural inputs limit the effectiveness of in silico tools as stand-alone solutions for antibody engineering and drug development. To address these limitations, hybrid experimental-computational approaches, integrating relatively high-throughput experimental analyses with computational predictions, have emerged for the structural characterization of Ab-Ag complexes.12 In this arena, structural mass spectrometry (MS) techniques, including chemical footprinting, crosslinking, ion mobility, and hydrogen-deuterium exchange (HDX), are increasingly used for high-throughput interrogation of protein-protein interfaces.13,14 Among these, HDX-MS is particularly attractive because it probes proteins in near-native solution conditions and can report on changes in solvent accessibility, hydrogen bonding, and conformational dynamics associated with binding. Advances in gas-phase fragmentation methods, such as electron transfer dissociation (ETD) and electron capture dissociation (ECD), have further expanded the potential of HDX-MS by enabling residue-level localization in selected cases, albeit with increased experimental complexity.15–18 In parallel, advances in liquid-handling automation, instrument control, and data processing have increased the practicality and throughput of HDX-MS workflows across laboratories.19 Nevertheless, HDX-MS-based interface mapping remains subject to several limitations, including incomplete sequence coverage, deuterium-hydrogen back exchange, and the possibility of allosteric or binding-induced conformational effects that can complicate interpretation. Building on recent efforts to integrate structural MS data with computational docking,12 we present an HDX-MS-guided structural modeling workflow that integrates experimental constraints with modern deep learning-based structure prediction. Applied prospectively to three Ab-Ag complexes relevant to therapeutic protein discovery and engineering, the workflow yielded interface root-mean-square deviation (I-RMSD) values of 1.05–1.57 Å for the selected Protenix models. The workflow integrates three strategies: (1) systematic use of peptide-level HDX-MS differential uptake data to define ambiguous interaction restraints for HADDOCK3 and residue distance constraints for Protenix in an Ab-Ag modeling context; (2) application of Protenix’s constraint-esm mode to HDX-MS-guided modeling by translating peptide-level HDX-MS protection hotspots into approximate residue-pair distance constraints; and (3) HDX-MS-based filtering criteria for selecting among computationally generated docked models, improving identification of biologically correct binding orientations beyond what energy-based scoring alone could achieve. Across the systems studied, peptide-level HDX-MS data were sufficient to guide docking and structural modeling, whereas targeted ETD HDX-MS was applied in one case as a higher-resolution, orthogonal approach to further localize critical binding residues. To preserve confidentiality, the systems are described using target-agnostic structural features, yet collectively span Fab and VHH binders and homodimeric and heterodimeric antigen ectodomains, representing diverse Ab-Ag architectures relevant to therapeutic design. Comparison with subsequently obtained in-house X-ray structures, used here as orthogonal high-resolution reference structures, showed strong agreement in overall binding orientation and interface placement, demonstrating the robustness and practical utility of this integrative approach. Materials and methods Protein expression and purification To maintain confidentiality, the antigens studied here are described using target-agnostic structural features rather than target identity. Relevant contextual information, including oligomeric state, approximate molecular size, and overall architecture, is provided to support evaluation of the HDX-MS-guided modeling workflow. Recombinant IgG and Fc-fused variable domains of heavy-chain-only antibodies (VHHs) were cloned and expressed in CHO cells using established protocols.20 The Fab and VHH domains were enzymatically cleaved from the Fc with Fabulous (Genovis), followed by sequential purification via MabSelect SuRe (Cytiva) and Superdex 75 (Cytiva) in phosphate-buffered saline. Antigen 1 (Ag1, homodimeric) was purchased from ACROBiosystems (Newark, DE) as a lyophilized powder produced in HEK293 cells and reconstituted according to the manufacturer’s recommendations. Ectodomains of antigen 2 (Ag2, heterodimeric) were recombinantly expressed in E. coli, purified by refolding protocols previously described,21,22 and diluted in phosphate buffer to final working concentrations. Size-exclusion chromatography (SEC) Three types of samples were prepared in 10 mM phosphate buffer (pH 7.0), with final concentrations ranging from 0.5 to 2.6 mg/mL: Ab alone, Ag alone, and Ab-Ag mixtures at defined molar ratios. For SEC, 3–8 μL of each sample was injected onto an ACQUITY BEH SEC column (200 Å, 1.7 µm, 2.1 mm × 150 mm, Waters, Milford, MA) using 100 mM ammonium acetate as the mobile phase at a flow rate of 100 μL/min and column temperature of 25°C. Separation was performed on either a Vanquish Horizon UHPLC (Thermo Fisher Scientific, Waltham, MA) with a variable wavelength detector, or an ACQUITY UPLC (Waters) with a tunable UV detector. UV absorbance was monitored at 280 nm, and chromatograms were processed using Freestyle (Thermo Fisher Scientific) or MassLynx v4.1 (Waters) software. HDX and in-line pepsin digestion Temperature-controlled HDX reactions for Ab, Ag, and Ab-Ag mixtures were prepared and analyzed in triplicate using either an HDx-3 PAL platform (Trajan Scientific and Medical, Morrisville, NC) or a Trajan liquid-handling system integrated with an ACQUITY HDX Manager (Waters). The use of the two platforms reflected the chronological progression of the studies and a subsequent upgrade of the laboratory’s HDX automation instrumentation; platform selection was not based on Ab format or complex type. The Fab1-Ag1 study was conducted first using the liquid-handling system integrated with the ACQUITY HDX Manager. Following the instrumentation upgrade, the VHH1-Ag1 and VHH2-Ag2 studies were conducted using the HDx-3 PAL platform. For each complex, all free and bound samples, including all replicate measurements, were analyzed using the same platform and experimental workflow. Chronos software (Trajan Scientific and Medical) was used to automate reaction timing and sample handling for all experiments. Stock protein solutions (0.3–3.4 mg/mL) were diluted approximately 10-fold in 10 mM phosphate buffer (pH 7.0) prepared in deuterium oxide (D2O) and incubated at 23°C for exchange times ranging from 30 to 7200 s; the time points used for each experiment are listed in Table S1. Exchange reactions were quenched by cooling to 1.5°C and performing a 1:1 dilution with buffer containing 0.1 M glycine, 4 M guanidine-HCl, and 0.25 M tris(2-carboxyethyl)phosphine, adjusted to approximately pH 2.5. Quenched samples were maintained at 1.5°C for 3–5 minutes to ensure protein denaturation, then diluted up to 3-fold with 0.1% formic acid in water. A total of 100 μL of each sample was immediately injected for online pepsin digestion using an Enzymate BEH column (2.1 × 30 mm, Waters) at 20°C, with 0.1% formic acid in water as the mobile phase at flow rates between 100 μL/min (ACQUITY HDX Manager) and 150 μL/min (HDx-3 PAL). Resulting peptides were collected on an ACQUITY BEH C18 VanGuard pre-column (130 Å, 1.7 μm, 2.1 × 5 mm; Waters) for 3 minutes at 1.5°C prior to LC separation. Control (no H/D exchange) samples were prepared and processed in parallel. Liquid chromatography-mass spectrometry (LC-MS) Digested peptide mixtures (2–9 μg) were eluted from the trap column using a 6.8-minute gradient from 15 to 28% mobile phase B (mobile phase A: 0.1% formic acid in water; mobile phase B: 0.1% formic acid in acetonitrile) at 40 μL/min. Peptides were separated using a reversed phase column (ACQUITY BEH C18, 300 Å, 1.7 μm, 1.0 × 10 mm; Waters) maintained at 2°C and analyzed in positive ion mode by an Orbitrap Eclipse Tribrid Mass Spectrometer (Thermo Fisher Scientific) equipped with a HESI-II source. Data used for peptide identifications were collected under data-dependent acquisition (DDA) using control samples. MS1 data for precursor ions were acquired at 120,000 resolution over 300–2000 m/z. For each scan cycle, the 10 most abundant precursor ions were isolated and fragmented by HCD (CE = 30%) for MS/MS analysis in the linear ion trap; dynamic exclusion was set for 7 s during each scan cycle. For deuterated samples, only MS1 data were collected over the same mass range and source settings but at a higher resolution of 240,000. Peptide identification was performed using Byos v5.0 (Protein Metrics, LLC, Boston, MA), and all assignments were manually validated from control data. Sequence coverage for each Ab and Ag component is summarized in Table S3. Spectra were analyzed with the peptide workflow and nonspecific digestion parameters, employing mass error tolerance of 10 ppm for precursor ions and 50 ppm for product ions. Curated lists of validated peptides were subsequently imported into HDExaminer v3.0 (Sierra Analytics, Modesto, CA) for quantitation of absolute deuterium uptake (D) and percentage deuterium uptake (%D) at each time point. The %D value was calculated as the number of incorporated deuterium atoms divided by the theoretical maximum number of exchangeable amides, which is estimated as the peptide length minus the number of proline residues and minus one for the N-terminal amide. The significance threshold for absolute deuterium uptake differences was determined by HDExaminer using the variance across replicate experiments, with all time points analyzed collectively. Relative differences in D (∆D) and %D (∆%D) between bound and unbound states were used to assign paratope and epitope regions. ETD HDX-MS For Fab1-Ag1, after paratope regions were localized to the peptide level, additional experiments were conducted using ETD HDX-MS. Prior to sample analysis, a selectively labeled peptide standard (Peptide P1, HHHHHHIIKIIK, Genscript Corp, Piscataway, NJ) was infused to tune source parameters and limit H/D scrambling.23,24 ETD HDX-MS was conducted on peptides localized to the paratope by initial HDX-MS, using targeted inclusion mass lists for precursors (charge z > 2, intensity > 1 × 105). MS1 acquisition for HDX samples selectively triggered ETD fragmentation for precursors meeting criteria for mass error (10 ppm) and retention time window (30 s). Product ions were analyzed by Orbitrap at 60,000 resolution. ETD data were processed using HDExaminer ETD mode, assigning the deuterium uptake for each identified fragment. Residues showing the greatest differences in deuterium uptake ∆D between bound and unbound forms were defined as binding sites. HDX-MS-guided HADDOCK3 modeling Crystal structures of antigens were obtained from the Protein Data Bank,21,22 processed by isolating ectodomains and relaxing via Rosetta FastRelax.25 The variable fragment (Fv) regions of antibodies were modeled using AlphaFold2 (multimer option).26 HDX-MS data were mapped onto structures using the Molecular Operating Environment (MOE) HDX-MS plugin, visualizing deuterium uptake difference between the bound and unbound states. Ab-Ag complex docking was performed using HADDOCK3.10 Hydrogen atoms of Fv and antigens were removed and chain identifiers were renamed using PyMOL. Surface residues with high ∆%D were selected as ambiguous interaction restraints defining the putative epitope and paratope regions. Center residues were defined as active restraints, while the edge residues were defined as passive. The preprocessed structures and the HDX-MS-derived restraints were used as HADDOCK3 inputs. A total of 10,000 rigid-body docking models were generated and clustered by fraction of common contacts (FCC). Up to 20 models per cluster were retained for flexible refinement and energy minimization. HADDOCK scores and cluster statistics for the three complexes are provided in Table S5. Clusters were ranked by HADDOCK score, and the 10 highest-ranked clusters were subsequently evaluated for consistency with the HDX-MS-defined epitope and paratope hotspots and for CDR-centered engagement. The model that best satisfied these criteria without predominant nonspecific contacts between the Ab framework and Ag was selected as the final docking model. HDX-MS-guided Protenix modeling Together with the antigen and Fab/VHH sequences, structural modeling was performed using the constraint inference with multiple sequence alignment (MSA) workflow implemented in Protenix (protenix_base_default_v0.5.0, ByteDance)27 on a workstation equipped with an NVIDIA A100 GPU. No modification of the Protenix source code or model architecture was made. Instead, peptide-level HDX-MS protection hotspots were translated into inputs compatible with the existing constraint-esm workflow. For each complex, one representative residue near the center of the HDX-MS-defined antigen hotspot and one aromatic residue within the antibody CDR-H3 hotspot were selected as the constraint residues. A maximum distance of 8 Å was specified between these residues. Because no specific atom names are defined in the constraint input, Protenix applies the distance threshold to the closest atom pair between the two residues, which may correspond to side-chain atoms, Cα atoms, or Cβ atoms depending on residue geometry. These constraints are implemented as soft restraints rather than hard distance constraints, allowing the model to be guided toward, but not strictly forced to satisfy, the specified distance and thereby providing tolerance to moderate variation in constraint parameters. Multiple sequence alignments were generated using LocalColabFold. Template use was disabled. For each complex, a single seed (101) was used, and five models were generated from that seed. Candidate structures were ranked according to the predicted interface template modeling score (ipTM), and the highest-scoring model for each complex was selected. The selected model was subsequently inspected to confirm consistency with the HDX-MS-defined epitope and paratope hotspots; this inspection did not alter the ipTM-based selection. Model selection was performed without reference to the subsequently obtained crystal structures, which were used exclusively for retrospective evaluation of model accuracy. Results Paratope and epitope mapping of Ab-Ag complexes by HDX-MS The paratopes and epitopes of three Ab-Ag complexes, VHH1-Ag1, VHH2-Ag2, and Fab1-Ag1, were identified, with HDX-MS serving as the primary experimental input for downstream computational modeling and without prior knowledge of the bound complex structures. These studies were performed during the discovery and optimization of therapeutic proteins incorporating the engineered binding modules VHH1, VHH2, and Fab1, rather than retrospectively on previously established benchmark systems. Collectively, the three cases span two antibody formats, VHH and Fab, and two antigen assembly types. Ag1 is a homodimeric antigen composed of two identical protomers, each containing a 161-residue target sequence of approximately 15 kDa with a poly-His tag. Each protomer contains six predicted O-glycosylation sites within residues 117–137 and one predicted N-glycosylation site at position 28. Ag2 is a heterodimeric antigen comprising an α subunit (204 residues, ~22.3 kDa) and a β subunit (244 residues, ~27.8 kDa), with a combined molecular mass of approximately 50 kDa. Thus, the workflow was evaluated on both symmetric and asymmetric antigen ectodomains and on both single-domain and conventional antibody binding modes. At the time of analysis, neither the sequences nor structural information for VHH1, VHH2, and Fab1 were available in public structure databases, ensuring that the workflow was independent of existing antibody structural references. Subsequently, high-resolution in-house X-ray structures became available and were used as orthogonal reference structures to assess the modeled bound-state geometries. SEC was used to confirm the formation of Ab-Ag complexes and to optimize their molar ratios for subsequent HDX-MS analysis. Pure Ab and Ag samples each exhibited a single predominant peak, whereas Ab-Ag mixtures produced up to three distinct peaks, depending on the mixing ratio, signifying the presence of free Ab, free Ag, and complexed species. The optimal Ab-Ag mixing ratios were determined through iterative SEC experiments (Table S2). For paratope-focused studies, an excess of Ag was used, while an excess of Ab was used for epitope mapping experiments. This strategy maximized the sensitivity of the HDX-MS readout for the binding partner of interest. To map the Ag1 binding site, we compared deuterium uptake (D) between free Ag1 and the Fab1-Ag1 complex (Figure 1). The residual plot in Figure 1A summarizes the differences in deuterium uptake (ΔD) for each Ag1 peptide, where positive ΔD values indicate reduced uptake in the bound state and therefore increased structural protection upon complex formation. Significant protection, defined as ΔD greater than the 1.56 Da threshold, was observed in peptide regions of residues 26–48, 49–78, and 98–103, with the strongest protection detected in residues 49–78. Representative uptake kinetics for protected peptides are shown in Figure 1B, whereas representative peptides without significant differences are provided in Figure S1i-iii. Together, these peptide-level HDX-MS data defined the Ag1 epitope regions used for subsequent structural modeling. Figure 1.HDX-MS epitope analysis of the Fab1-Ag1 complex.Two-panel figure showing the HDX-MS epitope analysis of Ag1 in the Fab1-Ag1 complex. Panel A is a residual plot comparing deuterium uptake by unbound Ag1 with uptake by Ag1 in the Fab1-Ag1 complex. Unique Ag1 peptides are arranged along the horizontal axis, and the vertical axis shows the difference in deuterium uptake, ΔD, between the unbound and bound states. Gray bars represent cumulative ΔD across five exchange time points, while overlaid colored points and lines represent values at the individual time points of 30, 90, 300, 900, and 3600 seconds. A red dashed horizontal line marks the global significance threshold of 1.56 daltons. Positive values above this threshold identify Ag1 peptide regions that become protected upon Fab1 binding. Significant protection occurs in peptide regions spanning residues 26 to 48, 49 to 78, and 98 to 103, with the strongest protection in residues 49 to 78. Panel B shows deuterium uptake as a function of exchange time for two representative protected Ag1 peptides, comparing free Ag1 with Ag1 in the Fab1-Ag1 complex.HDX-MS analysis identifying Ag1 peptide regions protected upon Fab1 binding through differential uptake measurements and representative uptake kinetics.(A) Residual plot displaying ΔD values between unbound (Ag1 alone) and bound (Fab1-Ag1 complex) states. The x-axis represents unique Ag1 peptides, while the y-axis shows the corresponding ΔD values. Each bar indicates the cumulative ΔD across five time points, with overlaid points denoting the ΔD at each individual time point. The red dashed line marks the global significance threshold considering all five time points together (0.312 Da × 5 = 1.56 Da), determined by HDExaminer. (B) HDX kinetic profiles for two representative peptides from epitope regions, illustrating deuterium uptake dynamics over time. HDX kinetics of representative peptides with no significant ΔD are provided in Figure S1i-iii. HDX-MS epitope analysis of the Fab1-Ag1 complex. Consistent with the epitope analysis, peptide-level HDX-MS was also used to localize the Fab1 paratope. The major protection hotspots were mapped to the CDR-H2 and CDR-H3, spanning ten residues in CDR-H2 and fourteen residues in CDR-H3 (Figure 2A), whereas no significant protected regions were observed in the light chain (Figure S2). Because these peptide-level data were sufficient to define the paratope regions used for downstream modeling, targeted ETD HDX-MS was applied only as a higher-resolution, orthogonal analysis to further localize critical binding residues within these protected segments. Within CDR-H2, the two peptides exhibiting the largest ΔD values (L50-F59 and I51-F59) were selected for ETD fragmentation to generate c- and z-ion series (Figure 2B). In this analysis, ΔD denotes the difference in deuterium uptake between free Fab1 and the Fab1-Ag1 complex, and the color scale reflects the magnitude of ΔD for each precursor and fragment ion. Using the combined c/z fragment information across these peptides, the largest ΔD shift was localized to Y52-D54, identifying these residues as key contributors for Ag1 binding. A similar analysis of peptide Y94-F107 in CDR-H3 (Figure 2C) localized the principal protected segment to D102-Y105, further refining the paratope assignment within this region. Thus, ETD provided higher-resolution localization within peptide-level HDX-MS hotspots, whereas the peptide-level HDX-MS data themselves defined the regions used for downstream modeling.Figure 2.HDX-MS paratope analysis of the Fab1-Ag1 complex.Multi-panel figure showing the HDX-MS paratope analysis of Fab1 in the Fab1-Ag1 complex. Panel A is a residual plot comparing deuterium uptake by free Fab1 with uptake by Fab1 bound to Ag1. Unique Fab1 heavy-chain peptides are arranged along the horizontal axis, and the vertical axis shows ΔD between the free and bound states. Gray bars represent cumulative ΔD across five exchange time points, overlaid colored points and lines represent individual time-point values, and red dashed horizontal lines indicate the positive and negative global significance thresholds. The approximate locations of CDR-H1, CDR-H2, and CDR-H3 are labeled above the plot. The strongest positive ΔD values occur in CDR-H2 and CDR-H3, indicating protection upon Ag1 binding. Panel B contains an uptake-versus-time plot for a representative CDR-H2 peptide and electron transfer dissociation (ETD) fragment-ion maps for two overlapping CDR-H2 peptides spanning residues 50 to 59 and 51 to 59. The fragment-level differences localize the significantly protected segment within CDR-H2 to residues Y52 to D54. Panel C contains an uptake-versus-time plot and an ETD fragment-ion map for a CDR-H3 peptide spanning residues 94 to 107, localizing the principal protected segment to residues D102 to Y105. Red and blue shading in the fragment-ion maps represents the magnitude and direction of ΔD for detected precursor and fragment ions at the 90-second exchange time, while gray cells indicate fragment ions that were not detected.Peptide-level and ETD HDX-MS analyses localizing Ag1-induced protection within the Fab1 CDR-H2 and CDR-H3 regions.(A) Residual plot showing ΔD values between free Fab1 (unbound) and the Fab1-Ag1 complex (bound). The x-axis represents unique Fab1 heavy-chain peptides, and the y-axis indicates ΔD. Each bar displays the cumulative ΔD across five time points, with overlaid points representing ΔD at individual time points. The red dashed line marks the global significance threshold considering all five time points together (0.640 Da × 5 = 3.20 Da), determined by HDExaminer. (B) (i) HDX kinetics of a representative CDR-H2 peptide (residues 50–59). (ii) ETD-generated c/z fragment ions for CDR-H2 peptides (residues 50–59 and 51–59), with ΔD indicated for each detected fragment ion. (C) (i) HDX kinetics of a representative CDR-H3 peptide (residues 94–107). HDX kinetics of representative peptides with no significant ΔD are provided in Figure S1iv-vi. (ii) ETD-generated c/z fragment ions for CDR-H3 peptide (residues 94–107), with ΔD reported for each detected fragment ion. Scale bars in B (ii) and C (ii) correspond to the ∆D of precursor and c/z fragment ions measured at the 90-second time point. HDX-MS paratope analysis of the Fab1-Ag1 complex. A similar peptide-level HDX-MS workflow was applied to define the paratope and epitope regions in the VHH1-Ag1 and VHH2-Ag2 complexes. For VHH1, comparison of deuterium uptake between free and bound states identified protection within all three CDRs, with relatively minor protection in CDR1 and larger ΔD values in CDR2 and CDR3 (Figure S3). Epitope mapping of Ag1 revealed multiple overlapping protected peptides consistent with a broad protected surface on the antigen (Figure S4). Together, these peptide-level HDX-MS data defined the paratope and epitope regions used for downstream structural modeling of the VHH1-Ag1 complex. For VHH2, the primary protected regions were localized to CDR1 and CDR3, whereas relatively minor protection was observed in CDR2 (Figure S5). In the corresponding Ag2 epitope analysis, multiple protection hotspots were detected in both the α and β protomers of the heterodimer (Figure S6). As with Fab1-Ag1, these results show that peptide-level HDX-MS data were sufficient to identify the key paratope and epitope regions used for downstream docking and structural modeling. Exact residue ranges for all three complexes are summarized in Table S4. Mapping peptide-level HDX-MS data onto structural hotspots using MOE To translate the peptide-level HDX-MS assignments into structure-based inputs for modeling, we first used residual plots of absolute deuterium uptake differences (ΔD) between bound and unbound states to identify peptides exceeding the global significance threshold and thereby define the initial paratope and epitope regions. For structural visualization and downstream modeling, these peptide-level uptake values were then converted to percentage deuterium uptake (%D), defined as the number of incorporated deuterium atoms divided by the theoretical maximum number of exchangeable backbone amides. This conversion reduced the influence of peptide length and differential exchange capacity. Relative differences in %D between bound and unbound states (Δ%D) were subsequently mapped onto the antibody and antigen structures in MOE, which integrates overlapping peptide information into localized surface heat maps (Figure S7).28 Accordingly, the orange-to-red regions in Figure S7 represent localized hotspots of differential deuterium exchange that are consistent with, but more spatially refined than, the broader peptide-level regions summarized in Table S4. Generating HDX-MS-guided Ab-Ag complex models Using the HDX-MS-defined paratope and epitope hotspots as experimental constraints, we next generated Ab-Ag complex models using two complementary workflows. Initial unguided trials with protein docking tools (HADDOCK3, LightDock, RosettaDock) and modeling platforms (AlphaFold2, Protenix) failed to generate models with acceptable Ab-Ag relative orientation. To overcome this limitation, we developed two HDX-MS-guided modeling strategies (Figure 3) that used the MOE-derived Δ%D hotspot maps to constrain and prioritize candidate binding modes. Figure 3.HDX-MS-guided Ab-Ag complex modeling workflow.Two-panel workflow diagram showing complementary strategies for incorporating HDX-MS information into antibody-antigen complex modeling. Panel A shows the HADDOCK3 workflow. Antibody sequence information is used to generate an antibody variable-region model with AlphaFold2, while an antigen crystal structure is processed with Rosetta FastRelax. In parallel, HDX-MS identifies antibody paratope and antigen epitope regions. The antibody and antigen structures and the HDX-MS-derived ambiguous interaction restraints are supplied to HADDOCK3 for docking. Generated models are clustered, the 10 highest-ranked clusters are retained, and agreement with the HDX-MS-defined paratope and epitope regions is used to select the final model. Panel B shows the Protenix workflow. Antibody and antigen sequences are combined with HDX-MS results to select one representative antibody paratope residue and one representative antigen epitope residue. The selected residue pair defines a distance constraint that is supplied to the Protenix constraint-esm workflow for antibody-antigen complex modeling. Connecting arrows indicate the flow of sequence, structural, and experimental information through each workflow.Schematic of two HDX-MS-guided workflows using HADDOCK3 or Protenix to generate antibody-antigen complex models.(A) Workflow combining AlphaFold2 and HADDOCK3 using ambiguous interaction restraints, guided by HDX-MS. (B) Workflow utilizing the constraint-esm module of Protenix using the residue distance constraint from HDX-MS. HDX-MS-guided Ab-Ag complex modeling workflow. In the first workflow (Figure 3A), we combined AlphaFold2‘s protein modeling capabilities, antigen crystal structures, and the docking tool HADDOCK3. We used the relevant antigen ectodomains crystal structure and performed a Rosetta FastRelax to mitigate conformational effects on surface residues caused by crystal contacts or binding partners. Subsequently, we modeled the variable region of Fab and VHH structures using AlphaFold2. Recognizing AlphaFold2‘s limitations in Ab-Ag complex modeling, we then employed HADDOCK3.10 Guided by the ambiguous interaction restraints extracted from HDX-MS data, we generated 10,000 docking models, clustering them by binding orientation. The top 10 clusters were selected based on the HADDOCK score. These top clusters were further filtered to maximize alignment with HDX-MS-identified epitopes and paratopes, favoring models with optimal CDR involvement and excluding extremely angled binding orientations, such as those where one side of the framework region is adjacent to the antigen (Figure S7A). The second workflow (Figure 3B) utilized Protenix, a new constraint-guided protein modeling tool (constraint-esm mode). Protenix uses residue distances between antibody and antigen as constraints to model the complex structure. For each of the three complexes, we selected a large aromatic residue on CDR-H3 within the highest ∆%D (paratope) and a residue in the central high ∆%D region (epitope) of the antigen. We set a distance constraint of 8 Å between these amino acids and ran Protenix in constraint mode to model the structures (Figure S7B). Assessing modeled complexes against orthogonal X-ray reference structures To evaluate the accuracy of the HDX-MS-guided models, we compared them with subsequently obtained in-house X-ray structures for VHH1-Ag1, Fab1-Ag1, and VHH2-Ag2, which served as orthogonal high-resolution reference structures for the bound-state geometries. Although these structures do not capture the full solution-phase ensemble sampled by HDX-MS, they provide a valuable benchmark for assessing overall binding orientation and interface-level agreement. As shown in Figure S7, models generated independently with HADDOCK3 and Protenix closely reproduced the relative arrangement of antibody and antigen observed in the corresponding reference structures. Moreover, the HDX-MS protection patterns mapped onto the modeled complexes aligned well with the crystallographic interfaces, supporting the reliability of the experimentally defined paratope and epitope regions. We first compared the crystallographic complexes with structures modeled using AlphaFold2 and HADDOCK3. The top-ranked candidates showed binding orientations consistent with the crystal structures, with I-RMSD values (DockQ v2-calculated) ranging from 2.92 to 6.82 Å across complexes (Figure 4).29 Our refined analysis yielded higher ligand root-mean-square deviation (L-RMSD) than the top models prioritized by HADDOCK scores. This discrepancy is expected since HADDOCK3 primarily performs rigid-body docking, whereas antibody CDRs are intrinsically flexible, making precise interface geometry difficult to recapitulate with this approach (Figure 4B). These backbone conformational differences result in inaccuracies in specific interactions, including hydrogen bonds and hydrophobic contacts. Nevertheless, the correct global binding orientation is valuable for lead selection, particularly in the development of bispecific and multispecific antibodies. However, for precision design tasks such as affinity maturation, higher interface-level accuracy is required. Figure 4.Comparison of HADDOCK3 and Protenix models with orthogonal X-ray reference structures.Two-panel figure comparing HDX-MS-guided HADDOCK3 and Protenix models with orthogonal X-ray crystal structures for VHH1-Ag1, Fab1-Ag1, and VHH2-Ag2. Panel A contains six structural superpositions arranged in two rows and three columns. The columns, from left to right, correspond to VHH1-Ag1, Fab1-Ag1, and VHH2-Ag2. The upper row shows HADDOCK3 models in purple superimposed on the corresponding crystal structures in cyan. The lower row shows Protenix models in gold superimposed on the crystal structures in cyan. Panel B reports the interface root-mean-square deviation, I-RMSD, and ligand root-mean-square deviation, L-RMSD, for each selected model relative to the corresponding crystal structure.Structural superpositions and RMSD values comparing HADDOCK3 and Protenix models with X-ray structures for three antibody-antigen complexes.(A) Top: HADDOCK3 models (purple) superimposed with crystal structures (cyan); Bottom: Protenix models (gold) superimposed with crystal structures (cyan). (B) I-RMSD and L-RMSD values were calculated using DockQ v2.29 Comparison of HADDOCK3 and Protenix models with orthogonal X-ray reference structures. We next compared the crystallographic complexes with models generated using Protenix. Across all three test cases, HDX-MS-guided Protenix modeling reproduced the crystallographic binding orientations with high accuracy, yielding DockQ scores of 0.62–0.82, I-RMSD of 1.05–1.57 Å, and L-RMSD of 2.74–2.94 Å (Figure 4 and Table S6). These results show that, in the three complexes examined, a single HDX-MS-derived distance constraint was sufficient to guide Protenix toward models with improved interface accuracy relative to the corresponding HADDOCK3 models. A detailed inspection of the binding interfaces revealed excellent agreement between the Protenix models and the orthogonal X-ray structures, particularly for the placement of the epitope–paratope interactions outside the CDR-H3 region (Figure 5). Although the CDR-H3 loop, which is typically conformationally heterogeneous and often represents the most difficult region to model accurately, remained more challenging, the individual CDR-H3 RMSDs were 1.52 Å for VHH1-Ag1, 2.43 Å for Fab1-Ag1, and 2.21 Å for VHH2-Ag2 (Table S6). Importantly, despite these local deviations, key interface features and residue placements were largely preserved relative to the crystal structures. Figure 5.Comparison of the Ab-Ag binding interface in Protenix models and orthogonal X-ray reference structures. The interface details of Protenix models (left) and crystal structures (right) with interface residue side chains shown as sticks. Antigen interface residues are shown in green. Protenix paratope residues are shown in gold, and crystallographic paratope residues are shown in pink. The paratope residues on CDR-H3 are circled in red. The CDR-H3 RMSDs for the Protenix models were 1.52 Å for VHH1-Ag1, 2.43 Å for Fab1-Ag1, and 2.21 Å for VHH2-Ag2. Complete values for both modeling methods are provided in Table S6.Close-up comparison of Ab-Ag binding interface in Protenix models and orthogonal X-ray reference structures, highlighting interface residues and CDR-H3 placement. Comparison of the Ab-Ag binding interface in Protenix models and orthogonal X-ray reference structures. The interface details of Protenix models (left) and crystal structures (right) with interface residue side chains shown as sticks. Antigen interface residues are shown in green. Protenix paratope residues are shown in gold, and crystallographic paratope residues are shown in pink. The paratope residues on CDR-H3 are circled in red. The CDR-H3 RMSDs for the Protenix models were 1.52 Å for VHH1-Ag1, 2.43 Å for Fab1-Ag1, and 2.21 Å for VHH2-Ag2. Complete values for both modeling methods are provided in Table S6. Because model quality can be influenced by the selection of HDX-MS-derived residue pairs and constraint definitions, future studies will evaluate additional antibody-antigen complexes and alternative constraint strategies to further assess the robustness and general applicability of the HDX-MS-guided Protenix workflow. Discussion This study establishes and validates HDX-MS-guided computational workflows for high-throughput mapping of Ab-Ag interfaces. Using standard peptide-level HDX-MS, we reliably identified paratope and epitope regions within days to weeks, and these data were sufficient to guide docking and structural modeling across the three systems studied. Mapping resolution can, however, be limited in regions where pepsin digestion yields longer peptides, increasing ambiguity in local interface assignment. In this study, targeted ETD HDX-MS was applied in the Fab1-Ag1 case as a higher-resolution, orthogonal approach to further localize critical binding residues within peptide-level protection regions. Accordingly, ETD HDX-MS is best viewed here as an optional refinement strategy rather than a required component of the modeling workflow, with potential value in cases where peptide-level localization remains ambiguous. One recurring challenge in interpreting HDX-MS data is that changes in deuterium uptake do not necessarily correspond exclusively to direct intermolecular contacts. Because HDX-MS reports changes in solvent accessibility, hydrogen bonding, and protein dynamics, Ab-Ag binding may produce protection through direct interface burial, conformational stabilization, or allosteric effects. For Fab1-Ag1 and VHH1-Ag1, the HDX-MS-defined epitope and paratope hotspots were generally consistent with the crystallographic interfaces, although the protected regions were broader, as expected from the peptide-level resolution of HDX-MS (Table S7). In contrast, several protected regions in VHH2-Ag2, including α-subunit residues 107–133 and 167–181 and β-subunit residues 66–90, were not part of the crystallographic interface and may reflect indirect binding-associated structural effects. The crystallographic epitopes in all three complexes comprised multiple noncontiguous sequence segments that converge spatially in the folded structures and were therefore predominantly conformational rather than strictly linear. Because such interfaces can produce protection across separated peptides, a single residue-pair constraint may not capture the complete interaction surface. The HDX-MS-derived constraints were therefore used to provide approximate spatial guidance rather than comprehensive residue-level definitions of the Ab-Ag interfaces. Integration with structural modeling and orthogonal high-resolution structures helped distinguish direct contacts from broader binding-associated protection patterns. We also emphasize that HDX-MS and X-ray crystallography provide complementary rather than hierarchical views of Ab-Ag structure. HDX-MS is performed under near-native solution conditions and reports ensemble-averaged deuterium uptake behavior that reflects solvent accessibility, hydrogen bonding, conformational dynamics, and, in some cases, binding-induced allosteric changes. By contrast, X-ray crystallography provides a high-resolution but static snapshot of a crystallizable bound state and may not capture the full range of conformations populated in solution. Accordingly, the crystallographic structures in this study were not treated as absolute representations of the native solution ensemble, but rather as orthogonal high-resolution reference structures for evaluating whether the HDX-MS-guided models recovered plausible bound-state geometries. Agreement between the HDX-MS protection patterns and the crystallographic interfaces therefore supports mutual consistency between the methods rather than strict one-way validation. Integration of HDX-MS data with HADDOCK3 yielded reliable predictions of the overall binding orientation for all three Ab-Ag complexes. Throughout rigid-body docking with HADDOCK, we observed that reliance on HADDOCK scores alone did not always prioritize the correct orientation. Filtering candidate models for agreement with the HDX-MS-defined epitope and paratope regions, together with appropriate CDR engagement, substantially improved the selection of biologically reasonable poses. For discovery-stage applications such as lead selection, epitope positioning, and multispecific format design, recovery of the correct relative orientation is highly informative. However, interface-level precision becomes increasingly important for downstream applications such as affinity maturation and fine-grained molecular engineering. To improve local interface accuracy, we incorporated residue-to-residue constraint-guided modeling with Protenix. By selecting two proximal amino acids, a large aromatic residue within the antibody CDR-H3 at the center of the highest Δ%D paratope region and a residue at the center of the highest Δ%D epitope region on the antigen, we guided Protenix’s modeling and achieved near-atomic interface accuracy, particularly in non-CDR-H3 regions. The 8 Å distance constraint represents an intermediate estimate between typical interface Cα-Cα distances (<12 Å) and direct interatomic contact distances (<4 Å). As described in Methods, Protenix applies this threshold to the closest atom pair between the two specified residues. To assess constraint robustness, we tested alternative residue selections within the same HDX-MS-defined hotspot regions and found that randomly selected high-Δ%D residues yielded comparable binding orientations. Additionally, increasing the distance constraint from 8 Å to 15 Å produced similar results. This tolerance to moderate variation in both residue selection and distance parameter is consistent with Protenix’s soft-constraint implementation, in which the generative diffusion process is steered toward, but not rigidly locked to, the specified distance. Together, these observations indicate that only approximate spatial guidance from HDX-MS data is required to achieve correct interface placement. As a recently developed tool, Protenix requires further benchmarking to validate its reliability across diverse systems. Because the present Protenix analysis used a single seed and generated five models per complex, broader evaluation across additional seeds and replicate runs will be needed to assess sampling robustness. Nevertheless, integration of ipTM and pLDDT scores, as well as systematic exploration across additional complexes, will be important for establishing broader applicability. In the three systems examined, unconstrained Protenix predictions did not recover the crystallographically observed binding orientations. Although the present dataset does not identify a single mechanistic cause, plausible contributors include limited paired evolutionary information for Ab-Ag interfaces, uncertainty in CDR-H3 conformations, the large number of potential Ag binding surfaces, and Ag oligomeric or conformational complexity. The HDX-MS-derived residue-pair constraint reduced the interfacial search space by providing approximate Ab-Ag proximity information. These results support the practical value of experimental spatial guidance but do not establish the specific cause of failure for each unconstrained prediction. More broadly, despite rapid advances in deep learning-based structure prediction, unguided modeling of Ab-Ag complexes remains challenging. Next-generation tools such as Protenix, Chai-1, and Boltz offer promising approaches for incorporating experimental constraints. Our initial results with Protenix emphasize the value of HDX-MS-derived spatial information for guiding interface modeling. Continued progress will likely depend on modeling frameworks that can more directly incorporate HDX-MS and complementary structural MS data. Rapid and confident elucidation of Ab-Ag complex structures is increasingly important for antibody engineering and early-stage drug discovery. Although traditional methods such as X-ray crystallography and cryo-EM remain gold-standard structural methods, their time and resource demands can limit routine deployment across large numbers of discovery-stage complexes. In the case studies examined here, HDX-MS provided experimentally informed epitope and paratope regions on a timescale compatible with discovery-stage decision-making. This workflow may therefore complement, rather than replace, higher-resolution structural methods when rapid interface localization is needed. Current in silico docking approaches also face challenges in model ranking and interface-level accuracy. The hybrid workflows demonstrated here may help accelerate interface localization while improving confidence in predicted binding orientation and interface placement. In summary, our HDX-MS-guided modeling strategy provided accurate and practically useful structural insight across three prospective Ab-Ag case studies relevant to therapeutic protein discovery. Peptide-level HDX-MS was sufficient to guide docking and structural modeling in all three systems, while targeted ETD HDX-MS served as an optional higher-resolution refinement in one case. The strong agreement between the resulting models and orthogonal in-house X-ray reference structures supports the robustness of this integrative approach and highlights its value for structure-informed antibody engineering. We acknowledge that nondisclosure of antigen identities limits precise external replication of the specific case studies presented here. However, the workflow and constraint strategies are described in sufficient detail for adaptation to other Ab-Ag systems. Continued benchmarking, refinement of constraint strategies, and expansion to more complex targets will be important next steps for broadening the applicability of these workflows. Supplementary Material Supplementary Information_2nd Revision.docx