Integrating deep learning with physics based modeling enables high precision antibody antigen interface prediction

Abstract

Accurate antibody-antigen (Ab-Ag) docking is hindered by CDR flexibility, discontinuous epitopes, and the absence of reliable binding-site restraints. This study presents a deep learning-augmented docking framework that integrates ParaDeep, a sequence-based paratope predictor, with the PyDockWEB scoring engine to provides a practical and interpretable framework for guiding docking using sequence-derived paratopes. ParaDeep predicts binding residues directly from concatenated VH/VL sequences, and these residues are used as spatial restraints within the PyDockWEB pipeline. Across 50 Ab-Ag complexes from AACDB, DL-guided targeted docking improved performance for the majority of cases relative to blind docking. Interface RMSD decreased overall (median 10.171 Å to 1.193 Å, p = 0.0016), and TM-score proximity showed a significant shift toward native folds (p = 0.0256). DockQ distributions exhibited a clear rightward shift, with median scores increasing from 0.0523 to 0.6799 and 46% of targeted models reaching high-quality classification. Structural analysis indicated that high-DockQ interfaces were more hydrophilic (–1.33 ± 0.52 vs. − 0.56 ± 0.88, p = 0.037) and enriched in coil regions, suggesting that moderate flexibility and polar complementarity may be associated with near-native docking convergence. Cross-metric analysis evaluated strong agreement between TM-score and DockQ (ρ = 0.854 for targeted vs. 0.782 for blind), indicating concurrent improvements in interface and global accuracy. Importantly, paratope-size correlation analyses showed no association with docking accuracy, whereas re-analysis of initially misclassified models using AppA-derived paratopes recovered most models, suggesting that the spatial precision of predicted restraint placement is a major contributor to docking outcomes in this rigid-body setting, while restraint count alone is not informative. In summary, ParaDeep-guided docking provides a practical and interpretable framework for integrating DL-derived paratope information into a physics-based docking framework. Rather than introducing a new docking paradigm, this work suggests that DL-derived residue-level priors can improve the efficiency and accuracy of rigid-body Ab-Ag docking on average, while retaining physical transparency and mechanistic interpretability. The framework offers a scalable and biologically informed complement to blind docking, with potential for integration into iterative antibody design and structure-guided immune-engineering workflows.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-39466-8.

Introduction

The molecular foundation of adaptive immunity is built upon antibody (Ab)-antigen (Ag) interactions, which enable the immune system’s specific recognition and neutralization of diverse targets^1–3. The remarkable specificity of this binding hinges on the hypervariable complementarity-determining regions (CDRs), flexible loops situated within the antibody’s variable domains (VH and VL)^4,5. These loops fuse to form the paratope at the binding surface, where their amino-acid composition and conformational diversity govern affinity, selectivity, and structural alignment^5–8. The resulting complex’s strength and thermodynamic stability are dictated by the intricate interplay between CDR side chains and antigenic epitopes⁵. Consequently, the rational use of computational modeling and mutational engineering to optimize CDR configurations holds enormous potential for developing high-affinity therapeutic and diagnostic applications^9–11. Despite a surge in high-resolution antibody structures, experimental determination via X-ray crystallography or cryo-EM remains resource-intensive, time-consuming, and limited in throughput^12,13. Hence, accurate computational modeling has become a crucial tool for closing structural knowledge gaps, guiding design efforts, and providing atomic-resolution insight into immune recognition^14–17.

Protein-protein docking aims to accurately predict the native complex structure from its unbound components, yet it remains a persistent challenge due to the extension of the conformational search space and the inherent complexity of biomolecular energy landscapes^18–21. This difficulty is complicated in Ab-Ag systems by significant factors like the pronounced flexibility of CDR loops, subtle interdomain movement between VH and VL, and the often shallow, discontinuous nature of antigenic epitopes^6,22,23. Conventional rigid-body docking tools, such as ZDOCK and ClusPro, rely predominantly on shape complementarity and electrostatics, but their sampling is blind to crucial biological context, frequently resulting in an excess of non-native solutions^24,25. Moreover, without contextual constraints from experimental data or learned binding patterns, scoring functions often struggle to reliably distinguish correct, near-native poses from incorrect decoys²⁶. The strategic integration of spatial restraints or contact information, whether derived from mutagenesis, NMR, or predictive computation, effectively resolves this ambiguity, rapidly collapsing the search space toward plausible interfaces^27,28. This restraint-based methodology has been successfully implemented in docking tools, including PyDockWEB, HADDOCK, and RosettaDock, establishing a necessary foundation for integrating data-driven machine learning (ML) guidance into established physics-based workflows^29–31.

Recent developments in ML have dramatically advanced our capability to predict antibody binding determinants, offering data-driven insights into how underlying sequence and structure define paratope specificity^32–37. Early ML platforms, including Parapred and ParaSurf, employed convolutional–recurrent and surface graph-based neural networks to classify interface residues using sequence- or structure-driven features, thereby generating coarse-grained maps of CDR regions^3,35. Subsequent methods, such as ACME, Paragraph, CSM-AB, and proABC-2, incorporated attention mechanisms, graph convolutional networks, and physicochemical embeddings to capture long-range dependencies across VH and VL domains and improve prediction accuracy for conformational paratopes^36–40. Numerous ML frameworks, spanning sequence-based, structure-based, and hybrid representations, thus already provide diverse strategies for inferring antibody contact residues, highlighting that paratope prediction is an active and well-established field with multiple validated approaches. Despite this progress, many contemporary predictors still emphasize local sequence context independently, often underrepresenting inter-chain complementarity and co-evolutionary signals that help define the true antigen-contact surface. It is important to note that the present study does not seek to benchmark ParaDeep against other predictors; instead, our objective is to evaluate whether the incorporation of sequence-based paratope priors can enhance Ab-Ag docking within the PyDockWEB framework, regardless of the specific ML model used.

To address the limitations of existing paratope predictors, this study makes use of ParaDeep, a hybrid deep-learning (DL) architecture designed to infer antibody contact residues directly from VH/VL sequence information. ParaDeep integrates bidirectional long short-term memory (BiLSTM) encoding with multilayer perceptron (MLP) classifiers to capture both contextual sequence dependencies and residue-level physicochemical properties, including hydrophobicity, charge, and side-chain accessibility⁴¹. These representations are further refined through a self-attention mechanism, which highlights inter-residue relationships important for distinguishing true antigen-contact sites, particularly within the structurally diverse CDR-H3 loop. The model outputs residue-level probability scores that can be readily interpreted or converted into putative spatial restraints. Because ParaDeep is a sequence-only predictor, it does not require experimentally resolved structures, making it suitable for early-stage antibody screening and for antibodies lacking crystal or cryo-EM data. Importantly, ParaDeep is used here as a representative example of a sequence-based paratope predictor rather than as a replacement for or competitor to existing approaches. The aim of this work is therefore not to suggest that ParaDeep is superior to other ML models, but to evaluate how sequence-derived paratope probabilities, regardless of the predictor used, can serve as practical spatial priors for guiding Ab-Ag docking. ParaDeep therefore provides a scalable, structure-independent source of residue-level probabilities that enables a controlled assessment of how DL-derived paratope information influences PyDockWEB’s rigid-body sampling and scoring. Accordingly, improvements observed in downstream docking should be interpreted as reflecting the general utility of restraint-guided sampling, rather than performance attributable uniquely to ParaDeep.

PyDockWEB is a widely used rigid-body docking platform that employs a Fast Fourier Transform (FFT)-based algorithm to efficiently sample large numbers of rotational and translational configurations between partners^29,42,43. The resulting poses are ranked using an empirical energy function dominated by electrostatics and desolvation, with a smaller van der Waals contribution, providing a computationally efficient and reproducible framework for high-throughput docking^44,45. Nevertheless, when applied to Ab-Ag systems characterized by flexible CDR loops and discontinuous epitopes, fully blind docking often yields many non-native solutions because the search and scoring procedures lack biological context⁴⁶. Consequently, the use of residue-level spatial restraints to guide docking represents a well-established strategy rather than a new conceptual advance. In this study, we systematically evaluate how ParaDeep-derived paratope probabilities, when converted into soft spatial restraints, influence PyDockWEB’s performance. These restraints do not modify the underlying physical model or scoring function; instead, they act as biologically informed priors that bias rigid-body sampling toward predicted interface regions, potentially increasing the probability of recovering near-native orientations. Although recent diffusion-based generative models can produce Ab-Ag complexes in an end-to-end manner, they often operate as computationally intensive black-box predictors and rely on structural inputs. In contrast, the ParaDeep-PyDockWEB framework is a lightweight, modular, and interpretable alternative that links sequence-level DL-derived paratope hypotheses to physics-based docking outcomes, and is therefore complementary rather than competitive with generative approaches. Because PyDockWEB assumes rigid-body interactions, it cannot explicitly model induced fit or CDR loop rearrangements; accordingly, this study evaluates DL-guided restraints under bound-state conditions to provide a controlled assessment of paratope-guided sampling. Finally, while information-driven docking is known to benefit from true interface knowledge⁴⁷, the objective here is not to introduce a new docking paradigm, but to test whether sequence-only DL paratope predictions can serve as a practical and scalable source of interface priors in the absence of experimental epitope data.

In this study, we systematically evaluate how sequence-based, DL-derived paratope predictions influence the performance of a conventional rigid-body docking framework when used as spatial restraints. ParaDeep-predicted residue probabilities were integrated into the PyDockWEB pipeline and benchmarked against a restraint-free blind docking baseline across 50 curated Ab-Ag complexes from the Antigen-Antibody Complex Database (AACDB)^48,49. Docking performance was assessed using CAPRI-compliant metrics, including interface RMSD (iRMSD), TM-score for global fold accuracy, and DockQ as a composite quality measure^50–52. Importantly, the goal of this work is not to introduce a new docking paradigm, but to rigorously assess the practical utility of DL-derived paratope information as a scalable, sequence-based source of spatial priors for augmenting a widely used physics-based docking platform. We therefore intentionally restrict the present evaluation to bound-state complexes to isolate the effect of DL-derived paratope restraints under controlled rigid-body conditions, acknowledge that these likely overestimates real-world unbound performance, and identify full unbound docking benchmarks as a key direction for future work.

Materials and methods

Model sets Preparation

For this initial evaluation, all antibody and antigen structures were taken from their bound co-crystal conformations in the Antigen-Antibody Complex Database (AACDB), enabling a controlled assessment of restraint-guided docking without additional variability from unbound conformational changes. From AACDB, a set of 50 Ab-Ag complexes was selected using systematic criteria designed to maximize structural diversity and minimize selection bias (Table S1)⁴⁸. Although AACDB contains substantially more crystallographic Ab-Ag complexes, this representative subset was chosen to balance realistic interface variability with the substantial computational cost associated with performing paired blind and targeted docking runs, consistent with standard practice in large-scale docking benchmarks.

A total of 50 Ab-Ag complexes were obtained by systematic sampling (sorted chronologically by PDB ID) and chosen through a randomized but quality-controlled procedure to ensure representative coverage across antibody classes and epitope types. The inclusion criteria required that (i) both the antibody and antigen structures were experimentally determined by X-ray crystallography with a resolution ≤ 3.0 Å (mean ± SD: 2.50 ± 0.51 Å); (ii) each complex contained complete variable domains (VH/VL) with unambiguous chain annotation; (iii) the antigen partner represented either a protein or peptide epitope with a defined continuous interface (≥ 400 Ų buried surface area); (iv) the complex was free from nonstandard ligands, glycans, or fusion tags near the binding interface; and (v) redundant entries sharing > 95% sequence identity were excluded to avoid overrepresentation. No complexes meeting these criteria were excluded, and no performance-based filtering was applied at any stage.

Because this benchmark set is drawn exclusively from AACDB, it may encode structural, antigen-type, or species-specific biases that limit generalizability beyond the represented epitopes and antibody classes. These dataset characteristics should be considered when interpreting the scope of performance gains and when extending the workflow to broader Ab-Ag landscapes.

Although the primary cutoff for resolution was ≤ 3.0 Å, a few complexes with slightly lower crystallographic quality (for example, structures resolved between 3.1 Å and 3.8 Å) were retained to preserve antigenic and structural diversity within the benchmark. These entries, which include well-characterized Ab-Ag pairs, such as 1adq (3.15 Å), 2jix (3.20 Å), 2zcl (3.25 Å), 6zer (3.80 Å), and 7kqh (3.50 Å), exhibit clearly interpretable electron-density maps and fully resolved VH/VL domains. The marginal relaxation of the resolution criterion was scientifically justified. These specific complexes provide structurally complete and biologically validated interfaces with minimal backbone ambiguity, ensuring they serve as reliable geometrical inputs for comparative docking. Furthermore, their inclusion was critical for maintaining the dataset’s representativeness across diverse antibody classes and epitope types.

Prior to downstream analyses, protein structures were repaired using PDBFixer (Official Neurosnap server) to ensure structural completeness, including heavy-atom completion, hydrogen addition, and conversion of non-standard residues⁵³. This curated set served as input for all subsequent paratope prediction, docking, and quality-evaluation procedures.

Paratope prediction by ParaDeep

For each antibody, the H- and L-chain amino acid sequences for each antibody were extracted from the cleaned PDB structures. The variable-domain regions (VH and VL) were identified using the IgBlast server (https://www.ncbi.nlm.nih.gov/igblast), which aligns the query sequences against germline V-gene segments to annotate framework and CDR regions⁵⁴. These annotated sequences, typically 110–130 residues in length, were submitted to the ParaDeep model (https://github.com/PiyachatU/ParaDeep), a hybrid deep neural architecture (BiLSTM + CNN and self-attention head)⁴¹. All sequences were formatted as standard single-letter amino-acid codes, ensuring compatibility with the ParaDeep input specification (Table S2).

ParaDeep predicts residue-level paratope probabilities from sequence context using physicochemical, positional, and evolutionary descriptors. Residues with probabilities exceeding the default threshold (0.5) were designated as predicted binding residues or paratope highlights. These highlights were mapped back to the antibody’s 3D structure to define the predicted paratope region, which was used to generate spatial restraints for docking.

ParaDeep was used strictly in its pre-trained form without any retraining, fine-tuning, or modification for this study. Because the ParaDeep training corpus was assembled from publicly available Ab-Ag complexes, some degree of sequence or structural similarity with AACDB entries is unavoidable. The purpose of the present workflow is therefore not to benchmark ParaDeep’s generalization performance, but to evaluate whether any sequence-derived paratope probabilities — regardless of the specific model — can function as practical spatial restraints for guiding rigid-body docking. Consequently, improvements observed in docking accuracy should be interpreted as reflecting the utility of restraint-guided sampling rather than predictive superiority specific to ParaDeep.

Protein-protein Docking by PyDock

PyDockWEB was selected as the docking platform for this study because it provides a robust, well-validated, and reproducible rigid-body docking framework that natively supports externally defined spatial restraints without requiring architectural modification²⁹. Its FFT-based sampling strategy combined with a physics-based scoring function (electrostatics, desolvation, and van der Waals terms) makes it well suited for systematically evaluating the isolated effect of DL-derived paratope restraints in a controlled and interpretable manner. PyDockWEB is also publicly accessible, lightweight, and compatible with automated workflows, enabling consistent large-scale benchmarking across all 50 complexes. Importantly, a general-purpose docking engine was intentionally chosen to avoid antibody-specific heuristics and to provide a neutral testbed for quantifying how sequence-based paratope priors alone influence rigid-body docking performance. While specialized antibody docking platforms exist, the goal here is not to benchmark docking tools themselves, but to assess the generic utility of DL-derived restraint-guided sampling within a standard physics-based docking environment.

Each curated Ab-Ag complex was separated into antibody (chains A and B) and antigen (chain C) components using BIOVIA Discovery Studio Client⁵⁵. Structural integrity, chain continuity, and coordinate completeness were verified prior to docking. Docking simulations were performed using the PyDockWEB server (https://life.bsc.es/pid/pydockweb)²⁹, which combines systematic rigid-body sampling by FTDock with empirical rescoring by PyDock. Two docking protocols were applied to each complex: (i) targeted docking, in which ParaDeep-predicted paratope residues were specified as active residues to impose spatial restraints and bias sampling toward the predicted interface; and (ii) blind docking (control), in which no restraints were provided and unrestricted orientation sampling was performed.

PyDockWEB generates up to 10,000 initial poses using FTDock’s Fast Fourier Transform (FFT) algorithm, and each pose is subsequently rescored using PyDock’s composite energy function incorporating Coulombic electrostatics, desolvation, and van der Waals contributions. For both targeted and blind conditions, the top-ranked model with the lowest total energy was selected for downstream structural analysis. In practice, ParaDeep inference requires approximately 2–5 s per VH/VL sequence on a standard CPU, while each PyDockWEB docking run completes within ~ 5–10 min, yielding a total runtime of ~ 10–15 min per complex. Because both components scale linearly and operate independently, the full workflow is readily parallelizable and can process hundreds of antibody variants within hours using batch job submission.

In addition to the primary AACDB benchmark, we also processed five recently deposited Ab-Ag complexes (PDB IDs: 8FHY, 8RRN, 8XSE, 9EZE, and 9FYT) as a small external validation set to assess whether the ParaDeep-PyDockWEB workflow generalizes to structures not included in AACDB. For each complex, the antibody variable-region sequences were submitted to ParaDeep using the same default settings described above, and the resulting predicted paratope residues were converted into PyDockWEB spatial restraints using the identical protocol employed for the AACDB set. Targeted and blind docking were performed with the same parameters and scoring configuration as the primary benchmark. These additional results, including sequence inputs (Table S1), ParaDeep-predicted paratopes (Table S2), and docking evaluation metrics (Table S3), are reported in the Supplementary Information. To assess robustness beyond the AACDB benchmark, a small external set of recently released Ab-Ag complexes was evaluated under identical docking protocols. For some low-similarity cases (e.g., 8FHY and 9FYT), the fraction of native contacts (Fnat) remained zero, reflecting the limited ability of rigid-body, paratope-only docking to recover exact residue contacts for challenging interfaces.

Docking quality evaluation and statistical analysis

Docking accuracy was evaluated by comparison to the experimental X-ray reference structures using CAPRI-compliant structural metrics as implemented through the DockQ framework by DockGROUND web server (Table S3)^50,56. For each complex, the top-ranked pose from both blind and targeted docking was selected for assessment. Docking quality was evaluated using multiple, partially correlated structural metrics that capture distinct aspects of model accuracy, including: (i) iRMSD to quantify local interfacial accuracy (lower values indicate better agreement), (ii) TM-score to assess global fold similarity, which was converted into a proximity metric (P = |TM – 1|) for analysis, and (iii) DockQ, a composite quality metric integrating iRMSD, ligand RMSD, and fraction of native contacts (Fnat), which enables direct classification into CAPRI quality categories (incorrect, acceptable, medium, and high).

All analyses were performed on paired targeted-blind results across the same 50 complexes. Wilcoxon signed-rank tests were applied to iRMSD and TM-score proximity to assess whether targeted docking produced statistically significant improvements (two-sided, p < 0.05). DockQ values were partitioned into quality classes using optimized thresholds obtained by Matthews correlation coefficient (MCC) maximization, and fractional category distributions were reported to quantify shifts in docking quality. Spearman’s rank correlation (ρ) was used to assess cross-metric consistency between TM-score and DockQ under both docking conditions. Additional nonparametric analyses (Spearman correlation, Mann–Whitney U test, and Cliff’s delta) were used to evaluate the relationship between paratope restraint size and docking performance, and to compare restraint size between improved and worsened cases. All tests were two-sided and nonparametric. Summary statistics (mean, median, interquartile range, and standard deviation) were reported for each metric. No performance-based filtering was applied at any analysis stage.

For methodological clarity, the overall ParaDeep-PyDockWEB workflow is summarized in Fig. 1, which illustrates the sequential integration of ParaDeep paratope prediction with PyDockWEB docking.

Fig. 1.

Overview of the ParaDeep-PyDockWEB docking workflow. This diagram summarizes the conceptual steps of the pipeline — from sequence input and paratope prediction through restraint generation, PyDockWEB docking, and structural evaluation.

Correlation analysis between paratope size and Docking accuracy

Correlation analysis was performed to evaluate whether the number of predicted paratope residues contributed to docking performance. For each of the 50 AACDB complexes, the total count of ParaDeep-predicted residues exceeding the probability threshold (> 0.5) was recorded as the “paratope size.” Docking accuracy was assessed using (i) DockQ of the top targeted model and (ii) TM-score proximity (P = |TM – 1|) to quantify global fold similarity. Spearman correlation coefficients (ρ) and associated p-values were computed to evaluate monotonic relationships between paratope size and each performance metric. To further determine whether restraint size differed between successful and unsuccessful predictions, complexes were stratified into “Improved” (ΔDockQ > 0.01) and “Worsened” (ΔDockQ < − 0.01) groups. Paratope size distributions between these outcome classes were compared using the Mann–Whitney U test and Cliff’s Delta to estimate effect size. Together, these analyses enabled a quantitative assessment of whether restraint set size influences the structural accuracy of ParaDeep-guided docking.

Diagnostic Re-analysis using Structure-Based paratope definitions (AppA)

To evaluate whether docking failures were caused by inaccuracies in ParaDeep-predicted paratopes or by limitations inherent to PyDockWEB’s rigid-body sampling, we performed a diagnostic re-analysis using AppA, a structure-based paratope/epitope annotation tool that identifies residues in direct physical contact within experimentally determined Ab-Ag complexes (https://mspc.bii.a-star.edu.sg/minhn/introduction_appa.html). Importantly, AppA is not used as a sequence-based predictor comparable to ParaDeep; rather, it directly extracts true contact residues, hydrogen-bonding networks, and interfacial water-mediated interactions from native Ab-Ag crystal structures. Consequently, AppA does not operate under realistic prediction conditions and is used here exclusively as a diagnostic reference to provide an approximate upper bound on docking performance when the true interface is known.

For a subset of representative complexes, AppA-identified antibody contact residues were extracted and converted into PyDockWEB-compatible spatial restraint files. These oracle restraints were used to re-dock each complex under otherwise identical conditions to the ParaDeep-guided workflow. Docking outputs were evaluated using TM-score proximity (P = |TM – 1|) and DockQ (Table S4). This diagnostic comparison enabled isolation of performance losses arising from paratope misassignment versus intrinsic limitations of the rigid-body docking algorithm.

Results

Local interface accuracy

To evaluate the impact of ParaDeep-guided paratope residues (targeted docking, Table S2), we compared iRMSD values of top-ranked poses generated by PyDockWEB server²⁹(Table S3). The ΔiRMSD (ΔiRMSD = iRMSD_Blind – iRMSD_Targeted), Fig. 2, distribution showed that targeted docking improved structural accuracy for most complexes (29 improved, 9 unchanged, and 13 worsened). This shift was reflected in central tendency: median-iRMSD decreased sharply from 10.171 Å (blind) to 1.193 Å (targeted), with the mean dropping from 10.511 Å to 6.017 Å (p = 0.0016). As a result, the ΔiRMSD distribution indicates that targeted docking is associated with improved local interface accuracy for most complexes, reflected in reduce structural deviations relative to blind docking.

Fig. 2.

ΔiRMSD per Ab-Ag complex comparing docking with ParaDeep-guided paratope resides vs. unguided blind models. Bars represent the difference in iRMSD (ΔiRMSD = iRMSD_Blind – iRMSD_Targeted) for each complex (N = 50), where positive values (red bars) indicate improved accuracy with targeted docking, and negative values (black bars) indicate worsened accuracy. Overall, ParaDeep-guided docking achieved significantly lower iRMSD values compared with the unguided blind docking condition (p = 0.0016).

Global fold accuracy

Next, we examined the effect of binding-site information on global structural accuracy using TM-scores in term of proximity metric, P = |TM – 1|, where lower values indicate greater similarity (TM-score closer to 1.0), as shown in Fig. 3. Across the dataset, targeted-docking models shifted on average closer to the native state; 22 improved proximity (ΔP > 0.01), 17 unchanged (|ΔP| ≤ 0.01), and 11 became less proximate (ΔP < − 0.01). The median proximity decreased from 0.1760 (blind) to 0.0559 (targeted) with p = 0.0256, reflecting a marked shift toward higher structural similarity. The statistically significant median improvement is consistent with enhanced overall fold similarity and improved interface quality (iRMSD) for many complexes when ParaDeep-predicted paratope residues are incorporated as restraints, although improvements were not observed uniformly across all cases.

Fig. 3.

Proximity-to-1 analysis of TM-scores for targeted- vs. blind-docking models. Each blue dot represents a single protein-protein complex (N = 50). The red dashed diagonal (y = x) indicates no change between the two conditions; points below the diagonal (y < x, ΔP > 0.01) correspond to improved proximity with targeted docking (TM-score closer to 1). The paired Wilcoxon signed-rank test suggested a statistically significant improvement (p = 0.0256).

DockQ distributions

DockQ, a metric integrating multiple structural criteria, further highlighted the quality improvements conferred by targeted docking⁵²(Table S3). In Fig. 4, the median DockQ increased markedly from 0.0523 (blind) to 0.6799 (targeted). This represents a qualitative shift of the central tendency from the “incorrect” range into the “medium-quality” category based on CAPRI thresholds⁵⁰. While blind docking models clustered near zero with almost no acceptable solutions, targeted docking exhibited a broader distribution with an increased proportion of complexes reaching “acceptable” and “medium/high” scores. These findings support the interpretation that ParaDeep-derived restraints are associated with improved overall docking quality on average the benchmark.

Fig. 4.

Distribution of DockQ scores for targeted- vs. blind-docking. Boxplots compare DockQ scores of top-ranked PyDockWEB models obtained with ParaDeep-guided paratope residues (left; DockQ_Targeted) and without binding-site information (right; DockQ_Blind). Individual points are per-complex DockQ scores (N = 50). The horizontal line inside each box is the median; the box spans the interquartile range, whiskers show the range (excluding outliers), and the “⋅” marks the mean. Median values: DockQ_Targeted = 0.6799, DockQ_Blind = 0.0523.

Categorical breakdown of DockQ quality

To precisely quantify improvements, per-complex DockQ scores were categorized into four CAPRI classes (incorrect, acceptable, medium, and high) using dataset-specific cutoffs obtained by an MCC-based analysis. Figure 5 illustrates that targeted docking achieved the optimized cutoffs of C1 = 0.105, C2 = 0.475, and C3 = 0.618 with resulting classifier produced an F1-score of 0.9560, while blind docking required higher C1/C2 (0.173) and C3 (0.670) cutoffs with a slightly lower F1-score (0.9450). Categorically, targeted docking displayed a pronounced reduction in incorrect predictions (< 0.105) and a substantial enrichment of high-quality models (> 0.618), while blind models remained dominated by incorrect predictions with only a modest fraction surpassing the medium threshold.

Fig. 5.

Fractional-count histograms of DockQ quality classes for (A) ParaDeep-targeted docking and (B) unguided blind docking condition PyDockWEB top-ranked models (N = 50). Histograms show the fraction of complexes falling into each DockQ bin; colors indicate categorical quality after thresholding: incorrect (black, DockQ < C1), acceptable (blue, C1 ≤ DockQ < C2), medium (green, C2 ≤ DockQ < C3), and high (red, DockQ ≥ C3). Cutoffs were optimized separately for each dataset using an MCC-based grid search.

Notably, the result shows the quality shift, targeted docking models distributed as 46% high, 6% medium, 4% acceptable, and 44% incorrect. By contrast, blind docking yielded 26% high, 10% medium, and 64% incorrect. The high F1-scores support the reliability of these categorical separations. Thus, ParaDeep-guided docking reduced the proportion of incorrect predictions and increased the fraction of medium- and high-quality models for the majority of complexes.

It can be noted that DockQ scores collectively demonstrate both a continuous distribution shift (broad rightward movement of DockQ) and the categorical enrichment (higher fractions in medium/high), reinforcing that ParaDeep-derived docking systematically elevates prediction quality across complementary analysis levels. The DockQ distribution shows an overall upward shift for targeted docking compared to blind docking, indicating higher median scores across the benchmark. This trend is reinforced by the categorical analysis, where the optimized cutoffs with high F1-scores enable a clear separation of quality classes. These results support the utility of ParaDeep-derived restraints in improving protein-protein docking performance for many complexes.

Cross-metric consistency

Finally, we evaluated the agreement between global (TM-score) and interface-level quality (DockQ) metrics for both targeted and blind docking (Fig. 6). In targeted-docking models (Fig. 6A), the complexes strongly clustered toward the upper-right quadrant, suggesting a strong positive correlation (Spearman’s ρ = 0.8542) between high TM-score (> 0.9) and high DockQ (> 0.8), indicating that higher interface accuracy was generally associated with more native-like global topologies. Acceptable-quality models were sparsely distributed in the midrange, while incorrect predictions occupied the lower-left region with consistently low DockQ and moderate TM values.

Fig. 6.

Two-dimensional plots of TM-score vs. DockQ score for (A) targeted- and (B) blind-docking condition across 50 Ab-Ag complexes. Each point represents a complex, with CAPRI-based classification colored as incorrect (black), acceptable (blue), medium (green), and high (red). The red dashed line indicates the Spearman’s rank correlation trend. Spearman’s ρ = 0.8542 for targeted docking and ρ = 0.7820 for blind docking, resulting a strong positive relationship between TM-score and DockQ across both conditions.

In contrast, blind docking (Fig. 6B) exhibited a more dispersed distribution with a weaker, however, but still significant positive correlation (Spearman’s ρ = 0.7820), indicating less consistent agreement between global and interface metrics. Although several high-quality predictions were retained, the clustering of many complexes in the incorrect region reflects the greater variability of unguided docking and its sensitivity to sampling and scoring limitations.

When integrated with TM-score and DockQ, the strong TM-DockQ correlation observed in the targeted condition is consistent with coordinated improvement across both performance metrics under targeted docking condition and the alignment between global fold and interface accuracy. The coherent pattern across metrics supports the practical utility of the integrated ParaDeep-PyDock framework for many complexes. Compared with blind docking, targeted docking achieved higher median accuracy and improved structural coherence for the majority of complexes, reflected by significant reductions in iRMSD, higher TM-scores, and enrichment DockQ categories in the medium- and high-quality ranges. The strong correlation (ρ = 0.8542 for targeted vs. 0.7820 for blind) suggests that biologically informed restraints can promote coordinated improvements in both interface and global structure for binding interface prediction.

Relationship between paratope size and Docking accuracy

To quantify whether the number of ParaDeep-predicted paratope residues influences docking performance, we examined the relationship between restraint set size and structural accuracy across all 50 complexes (Fig. 7A). Scatter analysis revealed no monotonic association between paratope size and DockQ under targeted docking, as reflected by a near-zero Spearman correlation (ρ = − 0.0188, p = 0.8968). A similar absence of correlation was observed when assessing global fold similarity using TM-score proximity (ρ = 0.0229, p = 0.8744) (data not shown), indicating that neither increasing nor decreasing the number of predicted interface residues systematically enhances or degrades docking accuracy. The LOESS trend line nevertheless suggests a shallow, non-linear optimum occurring around intermediate restraint sizes (≈ 24–27 residues), but this trend is not statistically supported by monotonic correlation analysis. These findings indicate that restraint quantity alone does not drive docking success and that ParaDeep-guided improvements arise instead from the spatial placement of predicted paratope residues rather than from the size of the restraint set.

Fig. 7.

Relationship between paratope size and docking accuracy. (A) Scatter plot showing DockQ of targeted docking vs. ParaDeep-predicted paratope size across 50 AACDB complexes. The LOESS curve (red) illustrates a shallow non-linear trend but Spearman correlation indicates no significant monotonic association (ρ = − 0.0188, p = 0.8968). (B) Comparison of paratope sizes between complexes classified as “Improved” or “Worsened” based on ΔDockQ. Mann–Whitney testing shows no significant difference between groups (U = 136, p = 0.9008) and effect size is negligible (Cliff’s δ = 0.0303).

To determine whether restraint size explains why some complexes improve while others worsen under targeted docking, we compared paratope sizes between the “improved” and “worsened” groups based on ΔDockQ classification (Fig. 7B). The distribution of paratope sizes did not differ significantly between groups (Mann–Whitney U = 136, p = 0.9008), and effect-size analysis showed a negligible difference (Cliff’s δ = 0.0303). A complementary Spearman correlation between paratope size and ΔDockQ across all complexes yielded a similarly weak association (ρ = − 0.1473, p = 0.3074). Collectively, these results support the interpretation that the number of predicted paratope residues is not a distinguishing factor between improved and worsened outcomes. Thus, docking failures are more consistent with paratope mislocalization (incorrect spatial definition) than with paratope size alone, reinforcing that the accuracy and placement of predicted residues — not the restraint count — govern the success of DL-guided docking.

Importantly, this study does not directly model the relationship between residue-level paratope prediction correctness and docking accuracy; establishing such an accuracy-accuracy relationship will require a benchmark with explicit residue-level ground-truth correctness scores and regression/stratified analyses across predictors.

Physicochemical and structural determinants of Docking performance

To investigate whether intrinsic interface properties contributed to the variability observed in docking outcomes, we compared the physicochemical and secondary-structure characteristics of antibody interface residues between the high- and low-DockQ groups (Fig. 8 and Figures S1). High-DockQ complexes exhibited more polar interface compositions (mean Kyte-Doolittle hydrophobicity − 1.33 ± 0.52) compared with low-DockQ complexes (–0.56 ± 0.88; p = 0.0376). This trend suggests that electrostatically enriched surfaces may be more compatible with PyDockWEB’s electrostatics-weighted scoring function, thereby facilitating more accurate ranking of near-native poses.

Fig. 8.

Boxplot of mean Kyte-Doolittle hydrophobicity across DockQ performance groups. Each box represents the distribution of per-complex mean hydrophobicity, with central horizontal lines indicating medians, boxes representing interquartile ranges, and whiskers denoting total data spread. The high-DockQ group exhibited significantly lower mean hydrophobicity (–1.33 ± 0.52) than the low-DockQ group (–0.56 ± 0.88; p = 0.0376).

Secondary-structure patterns further distinguished the two groups. High-DockQ interfaces contained a greater proportion of coil residues (58.6 ± 14.6%) and less α-helical (5.9 ± 8.0%) or β-sheet elements (15.4 ± 6.6%) compared with low-DockQ interfaces (44.3 ± 15.4% coil, 9.7 ± 9.8% helix, 24.7 ± 14.6% sheet). These differences, although modest, indicate that interfaces with moderate flexibility may be more compatible with rigid-body sampling, which is associated with better alignment to the native binding geometry. In contrast, low-DockQ complexes, which tended to exhibit more hydrophobic and structurally rigid interfaces, often generated energetically favorable, however, geometrically incorrect orientations.

Together, these observations show that docking performance is not solely determined by the presence of paratope restraints but is also modulated by intrinsic interface features, particularly polarity and local flexibility, which appear to facilitate more reliable convergence in the ParaDeep-guided docking framework. To further contextualize these structural patterns, we examined representative complexes that exemplify how interface composition and restraint precision jointly determine docking outcomes.

Case studies illustrating structural determinants of Docking performance

To illustrate how physicochemical and structural determinants translate into practical docking outcomes, we examined representative high- and low-performing complexes. In high-DockQ cases such as 5IKC and 1I9R, ParaDeep accurately highlighted paratope residues concentrated within CDR-H3 and adjacent H-chain loops that overlapped with the experimentally observed interfaces. These interfaces were strongly enriched in polar and coil-forming residues, consistent with the broader trends in Fig. 8, and this combination of accurate restraint placement and flexible, hydrophilic surface character were associated with near-native docking orientations in PyDockWEB.

In contrast, low-performing complexes such as 7C88 and 3SKJ exhibited markedly different interface characteristics: the true paratope spanned more rigid β-sheet segments with higher hydrophobicity, and the ParaDeep-predicted residues were partially shifted away from the actual binding patch. This mismatch, combining with the lower electrostatic accessibility of the native interface, caused sampling to be biased toward geometrically plausible but incorrect orientations, resulting in reduced DockQ values despite generally similar restraint sizes. These qualitative examples are consistent with the quantitative trends observed above: accurate spatial alignment of predicted paratopes and the presence of flexible, polar interfaces are associated with more successful targeted docking outcomes, whereas rigid and hydrophobic interfaces amplify the consequences of even modest prediction errors.

While these case examples highlight structural contributors to performance variation, we next asked whether the remaining worsened cases were primarily due to paratope misassignment rather than intrinsic docking limitations.

Diagnostic re-analysis using AppA-derived restraints

To determine whether the few cases of degraded performance under targeted docking were attributable to intrinsic docking limitations or to inaccuracies in the predicted paratope residues, we conducted a diagnostic re-analysis using AppA, a structure-based ML tool that identifies experimentally supported interface residues from crystallographic complexes⁵⁷. For each complex in which ParaDeep-guided docking initially worsened, the predicted restraints were replaced with AppA-derived interface residues and the docking was repeated following the same protocol.

When AppA restraints were used in place of ParaDeep predictions, most of the previously degraded complexes transitioned into the “improved” or “stable” categories (Fig. 9). AppA-guided re-docking reduced the median TM-score proximity from 0.1635 (blind) to 0.0243, with models clustering more tightly around the native-bound diagonal. Although this improvement did not reach statistical significance (p = 0.1401), the trend indicates a clear qualitative enhancement in global docking orientation.

Fig. 9.

Extended validation of misclassified targeted docking cases. Scatter plot comparing the proximity to unity of TM-scores for previously worsened targeted docking models (17 complexes, blue circles) and their re-analyzed counterparts using AppA-derived binding residues (red circles). Each point represents an individual Ab-Ag complex. The red dashed diagonal (y = x) indicates identical proximity between targeted and blind docking. Most red points approach the diagonal, showing measurable “improvement” or transition to the “no-change” category after AppA-guided restraint refinement (restraint moved TM closer to 1), highlighting the importance of accurate paratope identification for improving docking outcomes. The paired Wilcoxon signed-rank test suggested a statistically significant improvement with p = 0.1401.

These findings support the interpretation that paratope correctness is likely necessary but not sufficient for improved rigid-body docking, while also highlighting that rigid-body sampling/scoring constraints limit achievable accuracy even under oracle restraints. When provided with AppA-derived restraints (representing experimentally validated interface residues), most previously degraded cases transitioned to improved or stable categories, indicating that accurate paratope identification is critical for effective docking guidance. However, not all cases achieved perfect docking even with oracle restraints, demonstrating that PyDockWEB’s rigid-body sampling and scoring function impose additional algorithmic constraints on achievable accuracy. Together, these results indicate that the precision of residue-level restraint definitions is an important contributor to docking outcomes within the constraints imposed by the docking algorithm, whereas restraint size alone shows no predictive value.

Discussion

In antibody discovery settings with predefined target epitopes, such as those constrained by receptor blockade or disruption of a specific protein-protein interface, the primary challenge is not epitope identification but rather the prioritization of numerous VH/VL (or scFv/Fab) sequences generated from phage display and next-generation sequencing (NGS). Since NGS yields far more candidates than can be experimentally screened, computational triage is a critical step⁵⁸. The ParaDeep-PyDockWEB strategy is intended to support this by integrating sequence-based paratope probabilities from ParaDeep with PyDockWEB’s FFT-based sampling to guide docking toward predicted CDR contact regions^29,41. This epitope-conditioned screening framework may enable rational down-selection of epitope-consistent candidates for experimental validation⁵⁹. Ab-Ag recognition is governed by the specificity of CDR loops, whose physicochemical diversity and conformational variability make computational docking challenging^1,2. Conventional rigid-body docking approaches like PyDock often fail to predict accurate orientations due to a lack of biological context and inability to accommodate loop flexibility^29,44,45. To assess the influence of deep learning-derived spatial restraints, docking outcomes were evaluated using three complementary structural metrics — iRMSD, TM-score, and DockQ — providing a multi-level assessment of both interface accuracy and global fold similarity. This approach complements epitope-discovery methods that focus on repertoire-level binding site mapping⁵⁹.

ParaDeep-guided targeting improved local interface accuracy for most complexes, reducing the median iRMSD from 10.171 Å (blind) to 1.193 Å (p = 0.0016). Approximately two-thirds of complexes exhibited lower iRMSD, consistent with the idea that focusing sampling onto biologically plausible CDR patches may improve convergence toward more native-like arrangements⁴⁵. In contrast, blind docking must explore a substantially larger search space and therefore generates a broader set of non-native orientations. Global backbone similarity also improved on average under targeted docking, with TM-score proximity decreasing from 0.1760 to 0.0559 (p = 0.0256). This trend is consistent with the possibility that spatially targeted sampling may assist PyDock’s electrostatics-weighted scoring function in identifying orientations that better reflect native fold geometry^44,45. DockQ showed a comparable shift, increasing from a median of 0.0523 (blind) to 0.6799 (targeted), corresponding to a greater fraction of solutions falling within the CAPRI medium-quality range⁵⁰. The categorical breakdown reflected this shift: 46% of targeted models reached the high-quality category, compared with 26% for blind docking. Together, the three metrics indicate that incorporating paratope restraints is associated with improved docking performance for the majority of complexes, though not universally. Consistent with these trends, a small external validation set processed under identical conditions showed mixed but non-uniform outcomes, including both clear improvements to high DockQ/TM regimes (e.g., 8RRN, 9EZE), stable cases with minimal change (e.g., 8XSE, 9FYT), and occasional worsened performance (e.g., 8FHY), indicating that restraint-guided docking enhances accuracy in many but not all unseen complexes.

A cross-metric analysis (Fig. 6) showed that targeted models tended to cluster in regions of simultaneously high TM-score and DockQ, consistent with a strong positive correlation (ρ = 0.8542). This pattern suggests that improvements at the interface were often accompanied by improved global orientations. Although these metrics are not statistically independent, their strong concordance provides internal consistency and supports robust assessment of docking performance across both local and global structural scales. Blind docking displayed a weaker but still significant correlation (ρ = 0.7820), reflecting greater variability in aligning backbone geometry with interface contacts. The residual inconsistencies observed in some targeted models highlight PyDockWEB’s rigid-body limitation: it cannot account for induced fit or loop rearrangements, which can yield native-like contacts but imperfect backbone misalignments⁴⁴. Thus, while targeted restraints generally enhance structural coherence, rigid-body approximations still limit full recovery of native conformations.

The performance of the PyDockWEB pipeline is shaped by methodological constraints despite its efficiency and accessibility for large-scale Ab-Ag docking. The FFT-based rigid-body search and classical energy scoring function assume that the lowest-energy configuration approximates the native state^29,44, yet this approach can fail for interfaces requiring loop adaptation critical phenomena like induced-fit adjustments, conformational adaptation, side-chain rearrangements, or subtle conformational changes. In such cases, near-native poses may be penalized by incomplete packing, while non-native but energetically favorable poses may be over-ranked^44,45. Additionally, PyDock’s scoring emphasis on electrostatics can bias predictions for interfaces where hydrophobic, aromatic, or glycan-mediated contacts dominate^25,60,61. Targeted docking may partially mitigate these issues by restricting sampling, but its effectiveness ultimately depends on the accuracy and parsimony of the restraint set. A large or imprecise set of predicted residues increases computational cost and may misdirect sampling toward non-native regions. Therefore, docking outcomes reflect an interplay between PyDock’s rigid-body constraints and the fidelity of the DL-derived restraints.

Although ParaDeep-guided docking improved structural accuracy for most complexes, its performance remained highly sensitive to the precision with which interface residues were identified. Because ParaDeep applies a fixed probability threshold (> 0.5), antibodies with atypical or extended CDR-H3 loops or non-canonical frameworks can exhibit true contact residues that fall below this cutoff, yielding incomplete or spatially displaced restraint sets that restrict PyDockWEB sampling to suboptimal regions⁴¹. Minor variations in VH/VL boundary annotations or residue indexing can further misalign predicted residues with actual epitope contact regions. As a sequence-only model, ParaDeep provides broad usability for antibodies lacking structural data but does not explicitly capture loop conformations, which can reduce spatial precision for certain interfaces. To evaluate whether the magnitude of the restraint sets contributed to these outcomes, we examined the relationship between paratope size and structural (docking) accuracy across all 50 complexes (Fig. 7). Paratope size showed no meaningful association with either DockQ (ρ = − 0.0188, p = 0.8968) or TM-score proximity (ρ = 0.0229, p = 0.8744), and restraint counts did not differ between improved and worsened cases (Mann-Whitney U = 136, p = 0.9008; Cliff’s δ = 0.0303). Likewise, paratope size displayed no correlation with outcome classification (ρ = − 0.1473, p = 0.3074). Thus, the quantity of predicted residues does not explain performance variation; rather, the results are more consistent with docking success being sensitive to where restraints are placed (spatial alignment) rather than how many residues are restrained. These results indicate that while paratope localization is an important prerequisite for successful restraint-guided docking, it does not fully account for performance variability; docking accuracy remains jointly influenced by the precision of residue placement and by intrinsic limitations of rigid-body sampling and scoring.

To disentangle restraint misclassification from intrinsic docking limitations, the subset of targeted models that initially worsened was re-evaluated using AppA, a structure-based predictor of contact residues. When AppA-derived restraints replaced the ParaDeep predictions, most degraded complexes shifted to the improved or stable category, exhibiting markedly reduced TM-proximity (median 0.0243 vs. 0.1635 for blind; p = 0.1401) and tighter clustering near the native-bound diagonal. This recovery is consistent with residue-level restraint mislocalization being an important contributor to degraded cases under targeted docking. However, because this diagnostic uses structure-derived (oracle) contacts and does not quantify residue-level prediction correctness as a continuous variable, it should not be interpreted as demonstrating a direct accuracy-accuracy relationship.

Several limitations of this study should be acknowledged. Although the benchmark of 50 AACDB complexes spans diverse antigen types and CDR configurations, its modest size and predominance of IgG-like antibodies may limit generalization to nanobodies, noncanonical formats, or highly flexible epitopes. In addition, all docking experiments were performed using bound-state crystal structures, which simplify the problem relative to realistic scenarios involving unbound or predicted Ab-Ag models⁶². ParaDeep was trained on public Ab-Ag structures that partially overlap with AACDB; although it was not retrained or fine-tuned for this study, such overlap may influence paratope prediction accuracy. Consequently, the observed docking improvements should be interpreted as supporting the utility of biologically informed paratope restraints in rigid-body docking rather than as evidence of independent generalization to entirely unseen binding modes, which will require validation on larger, fully independent benchmarks. Furthermore, PyDock relies on rigid-body sampling and electrostatics-weighted scoring and therefore cannot model induced fit or CDR loop rearrangements; docked poses should thus be regarded as candidate starting structures for subsequent flexible refinement rather than fully relaxed binding models. ParaDeep is a sequence-only predictor using a fixed probability threshold, which may yield incomplete or spatially imprecise restraint sets for antibodies with atypical structural features; AppA-guided analyses are consistent with the hypothesis that restraint correctness is more relevant than restraint number, but this relationship was not directly quantified and docking performance therefore appears to reflect an interplay between restraint fidelity and the intrinsic limitations of rigid-body sampling and scoring. Finally, this study did not benchmark against alternative paratope predictors or antibody-specialized docking platforms; accordingly, the reported gains highlight the value of restraint-guided sampling within a standard rigid-body framework rather than superiority over other approaches.

In summary, this study suggests that incorporating sequence-derived paratope information as soft spatial restraints can, on average, improve the performance of rigid-body Ab-Ag docking relative to fully blind sampling. Across the primary AACDB benchmark, ParaDeep-guided docking increased the frequency of near-native solutions while reducing extreme interface deviations, although a subset of complexes showed limited gains or degraded performance, underscoring that improvements are not uniform. Correlation and restraint-size analyses indicate that docking outcomes are more closely associated with the correctness of predicted contact residues than with the number of restrained positions alone. Structural case analyses and AppA-guided re-docking further support the interpretation that inaccuracies in paratope localization may contribute to residual docking errors, alongside the intrinsic limitations of rigid-body sampling. While the current evaluation is restricted to bound structures and a single docking framework, additional external test cases suggest that restraint-guided docking may extend beyond the training-derived benchmark with variable but non-trivial benefit. Collectively, these findings support the view that DL-derived paratope priors can serve as practical and mechanistically interpretable constraints for guiding physics-based antibody docking, while also defining clear boundaries for generalization and future methodological development.

Conclusion

This study presents an integrated ParaDeep-PyDockWEB framework that combines sequence-based paratope prediction with physics-based rigid-body docking to explore improvements in Ab-Ag complex modeling. By incorporating ParaDeep-derived paratope residues as soft spatial restraints within the PyDockWEB sampling pipeline, the approach biases conformational exploration toward regions that are more likely to be biologically relevant. Across the 50 AACDB benchmark complexes, restraint-guided docking achieved, on average, improved agreement with native structures relative to fully blind docking. These gains were reflected across multiple quality metrics, including reduced iRMSD, improved TM-score proximity, and rightward shifts in DockQ distributions, supporting the practical utility of incorporating learned residue-level priors into classical docking workflows.

Beyond overall performance trends, the hybrid strategy also enabled exploration of structural features associated with docking outcomes. Comparative analysis of high- and low-performing models revealed that higher DockQ accuracy tended to be associated with moderately hydrophilic, coil-enriched interfaces, suggesting that balanced polarity and limited local adaptability may be favorable under rigid-body docking conditions. In contrast, interfaces dominated by rigid secondary structure or strong hydrophobic character were more frequently associated with mis-ranked or geometrically inaccurate solutions. These observations should be interpreted as associations derived from structural analysis rather than as direct evidence of modelled flexibility.

Performance analysis further helped contextualize sources of residual inaccuracy. PyDockWEB’s rigid-body sampling and electrostatic-desolvation-based scoring provide computational efficiency but inherently limit representation of induced fit, side-chain rearrangement, and CDR loop plasticity. At the same time, docking outcomes were sensitive to restraint fidelity: incomplete or spatially misassigned paratope predictions were often associated with degraded local accuracy. The recovery of many initially degraded cases following AppA-guided diagnostic re-docking supports the interpretation that inaccuracies in paratope localization contribute to residual errors, alongside intrinsic limitations of rigid-body docking.

Collectively, these results support the view that deep learning-derived paratope priors can serve as practical and interpretable spatial constraints for guiding physics-based antibody docking. Rather than proposing a new docking paradigm, the ParaDeep-PyDockWEB framework illustrates a general strategy for augmenting existing rigid-body docking engines with biologically informed residue-level information. This hybrid approach may be useful for early-stage antibody modeling workflows and provides a foundation for future extensions incorporating flexible refinement, epitope prediction, and emerging generative docking methodologies.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Kanchanok Kodchakorn (KK) contributed to conceptualization, methodology, investigation, data curation, visualization, and preparation of the original draft. Piyachat Udomwong (PU) was responsible for algorithm development and implementation (ParaDeep). Thanathat Pamonsupornwichit (TPa) and Thanyaluck Phitak (TPh) provided help in data collection, and discussion on results. Chatchai Tayapiwatana (CT), the corresponding author of this paper, supervised the overall project and advised on all parts of this paper. All authors contributed to the manuscript, reviewed the final draft, and approved the submitted version.

Funding

This research was supported by CMU Proactive Researcher, Chiang Mai University (Grant No. 768/2567), awarded to Kanchanok Kodchakorn. The financial supports were also provided by Chiang Mai University (2569A10412029, TGCMU2567P007, and R68IN00231) and National Research Council of Thailand (N34E670096) to Chatchai Tayapiwatana.

Data availability

The original contributions presented in this study are included in the article and Supplementary Material. The datasets analyzed during this study are available at the Antigen-Antibody Complex Database (AACDB);https://i.uestc.edu.cn/AACDB/⁴⁸. The ParaDeep source code is available at the project’s GitHub repository;https://github.com/PiyachatU/ParaDeep.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval

Not applicable.

Consent for publication

Not applicable.

Consent to participate

Not applicable.