Surf2Spot: a surface-informed geometry-aware model for predicting partner-independent binder and nanobody design hotspots Authors: Anwen Zhao, Yu Li, Zixuan Wang, Tianhao Wu, Ran Fu, Yuxuan Lou, Yanfen Xu, Xuliang Han, Wenying Wang, Jun Yan, Xiangfeng Wang Published: 2026 ABSTRACT AbstractProtein–protein interactions (PPIs) and nanobody–antigen interactions (NAIs) play essential roles in cellular function, yet potential hotspot prediction remains challenging. We present Surf2Spot, a deep learning framework that integrates sequence embeddings, structural features, and protein surface properties to predicts surface regions with a high propensity to contain PPI hotspot residues. Unlike complex-prediction approaches that require predefined binding partners, Surf2Spot identifies interaction-prone regions directly from target sequences or predicted monomeric structures, enabling partner-independent hotspot prediction. By jointly modeling structural and physicochemical determinants, Surf2Spot achieves strong hotspot prediction performance on curated PPI and NAI datasets, outperforming existing methods in terms of F1-scores and the area under the precision-recall curve (AUPRC). Case studies on NbPDS1 and VdPDA1 demonstrate that Surf2Spot can identify putative hotspot residues within functional domains that are enriched in experimentally validated binder designs. For the tested targets, Surf2Spot-guided designs (with RFdiffusion and BindCraft) yielded a four-fold increase in successful design throughput and enhanced binding affinities compared to baseline strategies in a target-dependent manner. These results establish Surf2Spot as a powerful tool for hotspot discovery and rational protein engineering.Keywords: surface-centric graph, dynamic graph CNN, hotspot prediction, binder design, nanobody design Introduction Protein–protein interaction (PPI) is a fundamental process underlying cellular activities such as signal transduction, transport, and metabolism [1]. Accurate identification of the structural surfaces and residue-level hotspots that mediate PPIs is critical for understanding disease and for engineering molecular tools in therapeutic development and biotechnology. These interaction hotspots often contribute disproportionately to the total binding free energy of a PPI, despite comprising only a small subset of surface residues, and their accurate identification is essential for structure-based approaches to modulating protein interfaces. Recent advances in generative modeling have revolutionized protein design, particularly through the development of diffusion-based frameworks that generate binding scaffolds followed by inverse folding [2]. Notable examples include RFdiffusion for plausible backbone generation coupled with ProteinMPNN for de novo sequence design [3, 4], BindCraft for integrated sequence-structure design workflows [5], and IgGM for antibody and nanobody modeling [6]. All of these methods rely on an initial definition of “hotspot” residues within the target surface to initiate the diffusion process [7–9]. These hotspots act as anchors for guiding binding site generation, and their accurate identification directly impacts the success of downstream binder design. While co-crystal structures enable direct hotspot extraction from known interfaces [10, 11], and AlphaFold2-Multimer or AlphaFold3 can predict complexes from partner sequences [12] when the binding partner is known, these approaches require specifying both partners and predict a particular complex. This contrasts with the task of identifying interaction-prone regions without a predefined partner, for which monomeric-structure-based methods remain valuable—especially for de novo binder design where the partner is unknown. With the emergence of proteolysis-targeting chimera (PROTAC) strategies [13–16], and their biological equivalents (bioPROTACs [17, 18]), there is increasing interest in designing modular binders capable of targeting endogenous proteins for degradation. Nanobodies, in particular, offer several advantages due to their small size, stability, and capacity to recognize flat or cryptic epitopes [19, 20]. However, hotspot prediction remains a bottleneck for applying such strategies in nonmodel organisms, especially given the scarcity of specific structural and interaction data. Several computational tools have been developed for hotspot prediction. Structure-based methods like HotPoint [13], KFC2 [21], and the PredHS series [22, 23] use energy scores or structural descriptors to identify energetically important residues, but they require experimentally determined complexes and are not applicable without a known binding partner. Other approaches work directly from single proteins. PPI-hotspotID [24], DeeProSite [25], GraphPPIs [26], and ScanNet [27] predict interaction interfaces or binding-related surface residues from structural or evolutionary features; interface localization, however, does not always equate to energetic hotspots defined by mutational experiments, and additional work is needed to bridge that gap. Sequence-based methods such as SPOTONE [28] integrate evolutionary information for hotspot-related prediction, yet their effectiveness depends on sufficient homologous sequences or reliable structural models. On the antibody side, epitope predictors including SEMA [29], DiscoTope3 [30], BepiPred3 [31], and RoBep [32] are designed to identify antigen surface residues involved in immune recognition, though they are not built for general NAI prediction. Taken together, existing methods have advanced the field, but many still depend on partner information and evolutionary data or focus on interface localization rather than partner-agnostic hotspot identification. A framework that predicts potential hotspots directly from a single sequence and its predicted structure—without requiring any knowledge of the binding partner—would address a clear unmet need. To address this need, we developed Surf2Spot, a partner-independent computational framework that predicts potential PPI hotspot regions directly from a protein sequence or monomeric structure. Built on a dynamic graph convolutional neural network (DGCNN) model [33], Surf2Spot is trained on both PPI and nanobody–antigen interaction (NAI) data to model the shared and task-specific determinants of interaction hotspots, without requiring the sequence of a predefined binding partner. While NAI can be considered a specialized form of PPI, the distinct spatial and energetic characteristics of nanobody–antigen interfaces motivate task-specific adaptation of the hotspot prediction model. In contrast to complex prediction approaches such as AlphaFold2-Multimer and AlphaFold3, which model a specific target–partner pair, Surf2Spot focuses on identifying interaction-prone regions across the target surface before a binding partner is specified [12]. Surf2Spot further uses a point cloud–based representation of protein surfaces, integrating geometric and physicochemical features while explicitly excluding buried residues unlikely to participate in binding [33]. This design enables rapid, partner-independent potential hotspot identification from predicted monomeric structures and provides a direct input for downstream binder and nanobody design. Results Surf2Spot: a surface-aware graph neural network for hotspot prediction of monomeric proteins We developed Surf2Spot, a deep learning framework that predicts PPI and NAI hotspots using only monomeric sequence and structure. Unlike prior methods that rely on co-crystal complexes, Surf2Spot leverages a surface-aware graph representation built from predicted 3D structures and point cloud features (Fig. 1A). Each input protein sequence is first processed by two pretrained models: ProtTrans [34], which generates informative sequence embeddings, and ESMFold [35], which predicts the 3D structure. From the predicted structure, we extract residue-level properties, including secondary structure, torsion angles, and relative solvent accessibility, using the Define Secondary Structure of Proteins (DSSP) [36]. These are combined with one-hot amino acid encoding and a spatial distance featurizer that generates adjacency matrices based on residue proximity [37] (Fig. 1B). To manage challenges associated with long sequences, such as data imbalance and high spatial complexity [38], we employed two distinct but complementary approaches: Chainsaw segments proteins exceeding 400 AAs into subdomains to manage structural complexity [39], while GPSite scans complete structures for high-probability interaction residues [40] (Fig. 1A). Overview of the Surf2Spot pipeline. (A) The input protein sequence is processed by the pretrained language model ProtTrans and the folding model ESMFold to generate sequence embeddings and predicted structures, respectively. For oversized structures, we use domain segmentation or PPI probability enrichment for segmentation. (B) One-hot encoding is constructed from this sequence, while DSSP-derived secondary structure properties and the residue adjacency matrix are extracted from the predicted structure. We apply the MaSIF-based method to sample and extract surface features (shape index, Poisson–Boltzmann continuum electrostatics, hydrogen-bond potential, and hydropathy) and to generate a triangular mesh of surface point clouds. ProtTrans embeddings and adjacency matrix features are projected into lower-dimensional spaces. These features are aligned via Euclidean distance-based mapping and integrated with the surface point cloud to form a graph structure. (C) This graph is then input into a three-layer DGCNN and, for nanobody–antigen interaction (NAI) hotspots, a GAT network. Protein–protein interaction (PPI) hotspots omit the GAT step (dotted path). Residual connections are applied throughout. A final MLP predicts hotspot likelihood. Applications of Surf2Spot include hotspot identification for both PPI and NAI. (D) Surf2Spot facilitates key downstream applications in protein design by accelerating the design–build–test cycle, thereby streamlining workflows across biology, biotechnology, and synthetic design. The protein surface is represented as a graph using the MaSIF pipeline [41]: surface points are sampled and encoded with physicochemical properties (e.g. shape index, hydrogen-bond potential, hydropathy, Poisson–Boltzmann continuum electrostatics) [42], then connected by a triangular mesh [43]. Residue-level features are mapped to these surface nodes via Euclidean distance, producing a surface graph with rich node and edge attributes, including residue distances, sequence embeddings extracted by ProtTrans, structural features constructed by DSSP, and surface features sampled with MaSIF (Fig. 1B). These graphs are processed using a dynamic graph convolutional neural network (DGCNN) [33]. For NAI prediction, an additional graph attention network (GAT) captures long-range spatial context (Fig. 1C). Surf2Spot processes heterogeneous features via dedicated multilayer perceptron (MLP) encoders to generate initial graph embeddings [44], which are passed through a three-layer DGCNN module with edge recomputation at each step. For NAI prediction, an additional three-layer GAT module operates directly on the original point cloud topology to capture long-range dependencies [45]. Residual connections are applied throughout, culminating in a fully connected layer that outputs residue-wise hotspot probabilities. This architecture culminates in a fully connected layer that transforms the final graph representation into hotspot probability predictions (Fig. 1C). These predicted hotspots can directly inform the design of binders and nanobodies, providing actionable insights for targeting specific epitopes (Fig. 1D). Surf2Spot distinguishes itself from conventional approaches through three fundamental innovations: (i) it bypasses multiple sequence alignments (MSAs) and native complex structures by using ProtTrans and ESMFold, enabling rapid and scalable prediction [46]; (ii) it predicts interaction-prone regions from unbound or predicted monomeric structures without requiring a predefined binding partner, in contrast to complex-prediction approaches that model a specific target–partner pair; and (iii) its surface-based graph construction inherently excludes buried amino acids, improving hotspot balancing and structural robustness. Together, these features make Surf2Spot particularly well suited for practical applications in binder and nanobody engineering, where rapid and accurate epitope identification is essential. Task-specific dataset design and model validation for hotspot prediction To train Surf2Spot for both PPI and NAI hotspot prediction, we curated two distinct datasets reflecting their biological differences (Supplementary Figs S1 and S2). PPI hotspots, traditionally defined as residues whose alanine substitution reduces binding energy by >2 kcal/mol [47–53], were sourced from 507 nonredundant complexes (filtered from 1334 complexes from PDBbind2020 and GPSite [54]). NAI hotspots, by contrast, reflect epitope functionality and are enriched in charged, hydrophilic residues (e.g. Arg, Asp, Glu), requiring distinct modeling from the hydrophobic-dominated PPI hotspots (e.g. Leu, Tyr, Phe) (Fig. 2A) [55]. A total of 248 high-quality nanobody–antigen complexes were selected from 1248 entries from the SabDab database, using identical structural filters [56]. Due to the distinct structural and energetic characteristics of NAI interfaces, we applied a less stringent FoldX threshold (0 kcal/mol) for NAI hotspot annotation; applying the PPI threshold (2 kcal/mol) would have eliminated more than half of the complexes entirely. Sensitivity analyses using alternative thresholds confirmed the robustness of this choice, with model performance remaining stable across a reasonable range (Supplementary Figs S3 and S4). Dataset construction and training of Surf2Spot. (A) Amino acid frequency distributions in PPI and NAI hotspots. Frequencies of the 20 amino acids are plotted in ascending order of PPI frequency. (B, C) Distribution of hotspot residues by surface and sequence in the PPI and NAI datasets, respectively. Contour plots show the density of surface-to-sequence hotspot ratios normalized per protein. The dashed line (y = x) represents equal distribution. (D, E) Comparison of residue RMSD between ESMFold-predicted and experimental structures for the (D) PPI and (E) NAI datasets. The binding interface is defined by a Euclidean distance cutoff of 5 Å for PPI and 8 Å for NAI from the partner nanobody. Significance codes: ***P < .001, **P < .01, *P < .05. We present a presegmentation surface point cloud approach that aligns protein sequence and structure data exclusively on the molecular surface, discarding buried residues irrelevant to PPIs. This surface-centric organization inherently mitigates label imbalance and feeds into a graph neural network for accurate PPI prediction, offering a biologically principled alternative to conventional full-structure models (Fig. 2B and C) [55]. Furthermore, residue conformations differed markedly between monomeric and complexed states, with the most pronounced discrepancies occurring at the interface (Fig. 2D and E). To avoid overfitting to complex-bound conformations, all samples were supplemented with ESMFold-predicted monomer structures, and each complex–monomer pair was co-assigned to training or test sets for unbiased evaluation. This surface-centric approach focuses on binding interface features, enhancing prediction accuracy and efficiency. We applied five-fold cross-validation to assess model performance for predicting PPI and NAI hotspots. Each fold trained a DGCNN model with the Adam optimizer (learning rate = 1e-4, weight decay = 1e-5) and a learning rate scheduler. Metrics including accuracy, precision, recall, F1, Matthews correlation coefficient (MCC), and the area under the receiver operating characteristic curve (ROC-AUC) were monitored throughout training, with F1-based thresholds selected for final evaluation. The optimal threshold was determined within each training fold during cross-validation and fixed before evaluation on the corresponding held-out test fold. Results from all folds were aggregated to evaluate generalization. Surf2Spot achieved high predictive performance on both tasks: F1 scores of 0.322 and 0.320, and MCC values of 0.256 and 0.228, respectively (Tables 1 and 2). These results demonstrate robust generalizability across diverse binding modalities and highlight the value of surface point cloud modeling, which excludes buried residues and enhances hotspot signal quality. Table 1PPI test evaluation.ModelBasis of predictionACCPrecisionRecallF1MCCAUROCAUPRCPCRE Surf2Spot Surface0.9130.2250.5670.3220.3200.8660.3350.188 DeepPPISite Structure0.8980.2040.6260.3080.3170.8870.2390.198 ScanNet Structure0.9140.2100.4910.2940.2830.8400.2440.182 Surf2Spot-whole surface Surface0.9480.2770.2640.2700.2440.8490.2010.087 GraphPPIs Structure0.9120.1900.4350.2640.2470.8440.1800.205 Surf2Spot-GAT Surface0.9570.3370.2050.2550.2420.8030.1980.064 Surf2Spot-GNN Surface0.9530.2990.2230.2550.2340.7990.1810.074 Surf2Spot-GCN Surface0.9570.3460.2010.2540.2430.7970.1860.062 Surf2Spot-ProtTrans only Structure0.9580.3540.1860.2440.2370.8210.2190.074 PPIHotspotID Structure0.6300.0770.8360.1410.1750.7740.1690.441 SpotONE Sequence0.9420.1630.1420.1510.1220.6170.0780.232 Atom Graph Structure0.9390.1370.1270.1320.1010.7380.0940.190 Residue Graph Structure0.8030.0720.3690.1200.0900.6450.0670.384 Surf2Spot-DSSP only Surface0.9640.4620.0150.0280.0770.8300.2030.197 Surf2Spot-surface only Surface0.9640.0000.0000.0000.0000.7590.1150.192 PPI test evaluation. Table 2NAI test evaluation. Evaluation.ModelBasis of predictionACCPrecisionRecallF1MCCAUROCAUPRCPCRE Surf2Spot Surface0.9220.2170.5500.2560.2280.7340.2410.222 BepiPred3 Structure0.9750.0000.0000.0000.0000.6830.0520.111 RoBep Structure0.8670.0720.3670.1210.1160.6220.0790.228 ScanNet Structure0.9740.0000.0000.000−0.0100.6480.0440.071 DiscoTope3 Sequence0.9540.0540.0500.0520.0280.7180.0580.180 NAI test evaluation. Evaluation. Ablation analysis and benchmarking reveal the advantages of Surf2Spot’s surface-aware graph framework We first performed ablation experiments to assess the contribution of Surf2Spot’s surface-centric representation, multimodal input features, and dynamic graph architecture (Fig. 3A). To evaluate the surface-centric graph representation, we compared Surf2Spot with alternative residue-level, atom-level, and whole-molecular-surface graphs. The residue-level graph performed poorly (AUROC 0.645; the area under the precision-recall curve (AUPRC) 0.067), and while the atom-level graph improved both metrics (0.738 and 0.094), it still lagged far behind the full model. Expanding to the whole molecular surface substantially boosted performance to 0.849 AUROC and 0.201 AUPRC, yet still fell short of the complete Surf2Spot representation (0.866 AUROC, 0.335 AUPRC). This gap suggests that surface restriction alone is insufficient; integrating sequence-derived and structural features provides additional predictive information beyond what surface geometry alone can offer. Surf2Spot outperforms state-of-the-art hotspot predictors. (A) AUROC curves from ablation experiments comparing alternative graph representations and graph neural network architectures, including atom-level, residue-level, whole-surface, GCN, GAT, generic GNN, features alone, and the full Surf2Spot model. (B) Precision–recall curves for the ablation on PPI models and established hotspot prediction methods. (C) Ablation and baseline comparison on the test set. Global IOS measures overall interface overlap; Cluster IOS evaluates overlap on spatially clustered predicted sites. (D) F1 score comparison of model variants with and without PPI pre-training and GAT layers. The combination of PPI pretraining and GAT yields the best performance. (E) Precision–recall curves for the established NAI hotspot prediction methods. (F) Structure of the large subunit of ribulose-1,5-bisphosphate carboxylase/oxygenase (RuBisCO) from Halothiobacillus neapolitanus during proteinaceous shell assembly, composed of ribulose bisphosphate carboxylase small subunit and large subunit. Shown below are the contact structures of a set of complexes within the assembly. (G–L) Spatial correspondence between predicted hotspots from DeepProSite, ScanNet, GraphPPIs, PPI-hotspotID, SpotONE, and Surf2Spot and ground-truth hotspot clusters on the protein surface. Predicted hotspot residues are depicted as red triangles, while ground-truth hotspots are shown as color-coded spheres, representing different spatial clusters. (M) Global IOS and Cluster IOS of DeepProSite, ScanNet, GraphPPIs, PPI-hotspotID, SpotONE, and Surf2Spot on the representative protein. We next examined whether the dynamic graph architecture contributes to performance. Replacing DGCNN with conventional graph convolutional network (GCN), GAT, or a generic graph neural network (GNN) resulted in substantially lower AUROC (0.797–0.803 versus 0.866) and AUPRC (0.181–0.198 versus 0.335), confirming the importance of the DGCNN design (Fig. 3A and B). We further evaluated individual feature groups by restricting Surf2Spot to surface, DSSP, or ProtTrans features alone. None of these single-feature models matched the full model’s performance (AUROCs of 0.759, 0.830, and 0.821, respectively), indicating that surface geometry, structural environment, and sequence-derived representations are complementary rather than redundant. We next benchmarked Surf2Spot against existing partner-independent hotspot predictors, including the sequence-based SPOTONE [28] and structure-based PPI-hotspotID [24]. To broaden the comparison, we additionally included other PPI site prediction approaches—DeepProSite (sequence + structure) [25], GraphPPIs (structure-based graph convolution) [26], and ScanNet (geometric deep learning)—which represent diverse methodological strategies [27] (Fig. 3B). All models were evaluated on the same PPI test set (Test_52) under comparable input conditions. On this benchmark, Surf2Spot achieved an AUPRC of 0.335, outperforming DeepProSite (0.239), ScanNet (0.244), GraphPPIs (0.180), PPI-hotspotID (0.169), and SPOTONE (0.078) (Fig. 3B). These results indicate that Surf2Spot maintains a substantially better precision–recall trade-off across a broad range of competing methods. We also considered AlphaFold3 as complementary references rather than direct baselines [12], since they require specified partner sequences and predict a particular complex (Supplementary Table S1), whereas Surf2Spot identifies interaction-prone regions without a predefined binding partner. Because hotspot prediction is inherently spatial, residue-level classification metrics alone may not fully reflect whether predicted hotspots form the correct binding region. We therefore introduced a cluster-aware Interface Overlap Score (Cluster IOS) to evaluate spatial overlap between predicted and experimentally defined hotspot clusters, alongside a Global IOS for reference. Surf2Spot consistently outperformed all other methods on both metrics, achieving a Cluster IOS of 0.133 (versus 0.124 for runner-up DeeProSite) and a Global IOS of 0.278 (versus 0.260), corresponding to relative improvements of 7.3% and 6.9% (Fig. 3C). The higher Cluster IOS suggests that Surf2Spot’s advantage is not limited to residue-level classification but extends to the spatial localization of hotspot regions. To investigate whether PPI knowledge can improve NAI hotspot prediction and to assess the contribution of the additional GAT module, we performed a two-factor ablation study on the NAI dataset (Fig. 3D). Without PPI pretraining, adding GAT slightly decreased the F1 score from 0.258 to 0.243. Similarly, without GAT, PPI pretraining alone resulted in an F1 score of 0.227. In contrast, combining PPI pretraining with the GAT module achieved the highest F1 score of 0.303. These results suggest that PPI pretraining alone or GAT alone is insufficient to improve NAI prediction, whereas their combination provides complementary information that benefits hotspot prediction on the NAI dataset. For NAI hotspot prediction, Surf2Spot achieved an AUROC of 0.734, outperforming DiscoTope3 (0.718), BepiPred3 (0.683), ScanNet (0.648), and RoBep (0.622) (Fig. 3E). Although these methods were primarily developed for antibody–antigen interface or epitope prediction rather than nanobody-specific hotspot identification, the comparison demonstrates that Surf2Spot provides a competitive approach for identifying NAI-associated hotspot regions. To evaluate the impact of structural confidence on prediction performance, test proteins were stratified according to the predicted local distance difference test (pLDDT) values of their input structures. Surf2Spot performance was evaluated separately on high-, medium-, and low-confidence groups. For PPI prediction, hotspot residues were distributed predominantly in high-confidence regions (64.4%), followed by medium-confidence (31.6%) and low-confidence regions (4.0%). The corresponding AUPRC values were 0.453, 0.373, and 0.295, respectively (Supplementary Fig. S5). For NAI prediction, 71.6% of hotspot residues were located in high-confidence regions, while 9.4% and 19.0% were located in medium- and low-confidence regions. The corresponding AUPRC values were 0.222, 0.272, and 0.180, respectively (Supplementary Fig. S6). To examine whether Surf2Spot can identify interaction-prone regions in a protein with multiple known interfaces, we performed a case study using the large subunit of ribulose-1,5-bisphosphate carboxylase/oxygenase (RuBisCO) from Halothiobacillus neapolitanus (PDB ID: 1SVD, chain A). This protein participates in two distinct interaction interfaces: one with the ribulose bisphosphate carboxylase small chain and another with itself during proteinaceous shell assembly (Fig. 3F). Compared with the predictions from SPOTONE, PPI-hotspotID, DeeProSite, GraphPPIS, and ScanNet, Surf2Spot produced more spatially compact predictions around the experimentally characterized interaction regions (Fig. 3G–L). The residue-wise IOS analysis further supported the spatial enrichment of Surf2Spot predictions around the known interaction interfaces (Fig. 3M). Surf2Spot enhances binder design accuracy and efficiency Hotspot residues on a target protein’s surface are critical for guiding structure-based binder design tools such as RFdiffusion [3], ProteinMPNN [4], and BindCraft [5]. We hypothesized that Surf2Spot’s predicted hotspots, derived solely from monomeric sequence and structure, could enhance both success rate and computational efficiency, particularly for targets lacking co-crystal structures. To evaluate this, we conducted two case studies: one involving a plant biosynthetic enzyme (NbPDS1) [57] and the other a fungal virulence factor (VdPDA1) [58]. NbPDS1 is an enzyme involved in the carotenoid biosynthesis pathway of Nicotiana benthamiana [57] for which no known complex structure has been reported. BLAST alignment identified 5MOG [59], a homologous pentamer sharing 85% sequence identity, as the closest structural match to NbPDS1 (Fig. 4A). NbPDS1 was divided into three domains (D1–D3), and hotspot residues were independently predicted for each domain (Fig. 4B, Supplementary Fig. S7). Notably, Surf2Spot accurately identified hotspots within a helical region of D3 that mediates oligomerization (Fig. 4C), despite 5MOG being absent from the training set and having <20% similarity to any training sample. This suggests that Surf2Spot effectively captures generalizable physicochemical features of protein interaction interfaces. Based on the predicted hotspots (residues 170, 173, and 174), we selected the D3 segment and performed binder design using both BindCraft and a combined RFdiffusion-ProteinMPNN workflow (Fig. 4E and F, Supplementary Fig. S8). In the BindCraft pipeline, full-length NbPDS1 without hotspot constraints yielded no successful binders after 7.5 days. Restricting design to D3 produced five successful binders in 30.7 h, and adding Surf2Spot hotspots further reduced design time to 15.6 h, with a 41% decrease in required iterations (Fig. 4D–G). This demonstrates that hotspot guidance with Surf2Spot improves both efficiency and success rate. Surf2Spot improves the accuracy and efficiency of binder design. (A) Architectural features of NbPDS1. Sequence conservation with 5MOG and secondary structure elements are shown along the primary sequence. (B) Domain-wise prediction of PPI hotspots in NbPDS1. The protein was segmented into three domains—D1, D2, D3—with predicted binding hotspots shown for each domain. (C) Surface-aligned hotspot predictions for 5MOG and the D3 domain of NbPDS1, with predicted hotspots highlighted as sticks. (D) Computational time required to obtain five successful designs under three strategies: Surf2Spot-guided D3, D3 without hotspot constraints, and full-length NbPDS1. “+” indicates timeout. (E, F) D3-targeted binder designs for NbPDS1 guided by Surf2Spot-predicted hotspots using RFdiffusion and BindCraft. (G) Number of attempts required to obtain five successful designs under three strategies: Surf2Spot-guided D3, D3 without hotspot constraints, and full-length NbPDS1. “+” indicates timeout. (H, I) Success rate and throughput of seven design strategies with or without geometric and AF2-based filtering. (J) Distribution of interface pAE values among NbPDS1 binders generated under different design strategies. (K) Structure of VdPDA1 and Surf2Spot-predicted PPI hotspots. The protein backbone is shown in blue-white, with predicted hotspot residues highlighted as red sticks. (I) Distribution of interface pAE values in binders targeting full-length VdPDA1 with and without Surf2Spot hotspot guidance. We next benchmarked Surf2Spot-guided designs using RFdiffusion (structure generation) and ProteinMPNN (sequence design), generating 300 candidates per domain on a single NVIDIA A100 GPU. The results revealed strikingly different domain-dependent outcomes. D1 contained two surface hotspots and, although sparse, they enabled 59 viable designs, 49 of which passed all filters. In D2, Surf2Spot localized hotspots predominantly to a disordered region, but all initial and guided designs failed filtering due to structural instability. In D3, the most favorable domain, Surf2Spot produced 231 filtered candidates, including 125 of 125 hotspot-guided designs, with a 65% reduction in generation time (4 versus 14 min) (Fig. 4H and I, Table 3). While guidance slightly reduced design diversity, it maintained near-perfect success and yielded the best interface predicted aligned error (iPAE) scores, with most clustering in the 0–10 range (Fig. 4J). The D3 strategy achieved 0.104 successful designs/min—four times the rate of full-length designs—establishing domain-focused, hotspot-guided design as the preferred paradigm for large proteins (Supplementary Table S2). Together, these findings from the NbPDS1 case study highlight how Surf2Spot-guided hotspot prediction can streamline domain-specific binder design by identifying structurally favorable regions that enhance both success rate and computational efficiency. Table 3Domain-wise residue prediction with or without PPI hotspots for comparing design efficiency on NbPDS1.Design strategyAmino acid rangeHotspots Whole Protein 1–586/ D3 136–238/ D3 + Surf2Spot 136–238170, 173, 174 D2 1–32/75–158/238–273/334–441/ D2 + Surf2Spot 1–32/75–158/238–273/334–4415, 6, 7, 8, 9, 378 D1 11–115/250–333/445–508/ D1 + Surf2Spot 11–115/250–333/445–50816, 32 Domain-wise residue prediction with or without PPI hotspots for comparing design efficiency on NbPDS1. By contrast, binder design targeting VdPDA1 [58]—a chitin deacetylase secreted by Verticillium dahliae—proved highly resistant to binder design. Only 3 of 300 (1%) unguided candidates passed filtering, and none of the 300 hotspot-guided designs (residues 47, 83, 85, 89, 90, 92, 96, 99) were viable (Fig. 4K and L, Table 4). These results suggest that structural or electrostatic features of VdPDA1 may preclude conventional binder formation. In such cases, alternative scaffolds like nanobodies, which tolerate flexible targets and offer enhanced stability, may represent a more effective design route. Table 4Domain-wise residue prediction with or without PPI hotspots for comparing design efficiency on VdPDA1.Design strategyAmino acid rangeHotspots Whole Protein 1–395/ Whole Protein 1–39547, 85, 89, 90, 92, 96, 99 Domain-wise residue prediction with or without PPI hotspots for comparing design efficiency on VdPDA1. Surf2Spot enhances nanobody design efficiency and interface quality We further assessed the utility of Surf2Spot in guiding nanobody design using the IgGM pipeline, targeting the same two proteins, VdPDA1 and NbPDS1. IgGM generated 300 candidate nanobodies within 11.3 h on a single NVIDIA A100 GPU, both with and without incorporating the predicted hotspots. Design success was evaluated using two Rosetta-based performance metrics [60, 61]: (1) dG_cross/dSASA × 100, which is inversely correlated with binding strength [62]; and (2) hbonds_int, which represents the number of interface hydrogen bonds and is positively correlated with binding affinity. For VdPDA1—a relatively small protein not subdivided into domains—Surf2Spot predicted three surface hotspots (residues 262, 265, and 268) as likely NAI epitopes (Fig. 5A). With Surf2Spot-guided hotspots, 32.7% of VdPDA1 nanobody designs passed both thresholds, a 3.37-fold increase over the 9.7% success rate without hotspot guidance (Fig. 5B). For NbPDS1, Surf2Spot predicted 18 hotspot residues across four NAI epitopes in the D2 and D3 domains (Fig. 5C). Guided design yielded higher success rates for three of the four epitopes, with improvements of up to 1.75-fold (Fig. 5D). Surf2Spot enhances the accuracy and efficiency of nanobody design. (A) Structure of VdPDA1 showing NAI hotspots at the surface and residue levels (sticks). (B) Design success rates for VdPDA1 under three filtering criteria: (1) dG_separated/dSASA × 100, (2) hbonds_int, and (3) both combined. (C) Domain-level NAI hotspot predictions for NbPDS1, with clustered regions highlighted as sticks. (D) Design success rates for NbPDS1 under three filtering criteria: (1) dG_separated/dSASA × 100, (2) hbonds_int, and (3) both combined. (E and F) Distribution of interfacial residues in (E) VdPDA1 binders and (F) VdPDA1 nanobodies across different design strategies. The dashed line indicates the peak density for designs without hotspot guidance. (G and H) Binding poses of nanobody designs generated with and without Surf2Spot hotspot guidance for VdPDA1. Binder interface hotspots are shown as sticks. To further assess binding interface quality, we computed the number of antigen-contacting residues using a ΔSASA >1.0 Å2 threshold. NbPDS1 designs similarly showed 1.2- to 1.e interfacial residues when guided by Surf2Spot predictions (Fig. 5E, Supplementary Table S3). For VdPDA1, Surf2Spot guidance led to a sharp increase in interfacial residue count: guided designs peaked at 25 contacts—nearly five times more than the peak of four contacts in unguided designs (Fig. 5F–H, Supplementary Table S4). These results demonstrate that by providing more informative interface localization, Surf2Spot improves not only binder design but also nanobody design, enhancing both success rates and predicted binding interface quality. Validation of designed protein binding to cognate targets Based on the 3D structural features of NbPDS1 and VdPDA1, we designed and screened a set of specific binders and nanobodies. For NbPDS1, five binders (NbPDS1-B1 to NbPDS1-B5) were generated. For VdPDA1, five binders (VdPDA1-B1 to VdPDA1-B5) and five nanobodies (VdPDA1-N1 to VdPDA1-N5) were produced. The interactions between each target protein and its corresponding candidates were evaluated using a luciferase complementation assay (LUC). The LUC screening results demonstrated that, in the NbPDS1 group, the binders NbPDS1-B2 and NbPDS1-B3 specifically recognized NbPDS1 and yielded strong luminescence signals (Fig. 6A). In the VdPDA1 group, the binder VdPDA1-B5 and the nanobody VdPDA1-N1 exhibited significant restoration of luminescence upon co-expression with VdPDA1, indicating specific interaction with the target protein (Fig. 6B and C). To validate these findings, pull-down assays were performed by immobilizing NbPDS1 or VdPDA1, incubating with the respective candidates, and detecting the retained proteins via western blotting. The results showed that VdPDA1-B5 and VdPDA1-N1 were efficiently captured by VdPDA1 (Fig. 6D and E). Collectively, these results identify VdPDA1-B5 and VdPDA1-N1 as specific VdPDA1 binding molecules. Designability of surf2spot-selected hotspots: evidence from NbPDS1/VdPDA1 binding with engineered proteins. (A) LUC signals from the co-expression of NbPDS1 with each binder (NbPDS1-B1 to NbPDS1-B5), respectively. Significant luminescence enhancement was observed for NbPDS1 with NbPDS1-B2 and NbPDS1-B3, indicating direct interactions. (B) LUC signals obtained from co-expression of VdPDA1 with each binder (VdPDA1-B1 to VdPDA1-B5), respectively. Significant luminescence enhancement was observed for the VdPDA1-B5, indicating a direct interaction. (C) LUC signals for VdPDA1 co-expressed with each nanobody (VdPDA1-N1 to VdPDA1-N5), respectively. VdPDA1-N1 showed a marked increase in luciferase activity, suggesting specific binding to VdPDA. (D) Pull-down assay confirming the interaction between VdPDA1 and VdPDA1-B5. EG: Experimental group containing both the flag-tagged target protein and the GST-tagged candidate binder. CG: Control group containing only the GST-tagged candidate binder without the target protein, used to assess background signals and exclude nonspecific binding. A clear GST signal was observed in the EG but not in the CG, indicating a specific positive interaction between VdPDA1 and VdPDA1-B5. (E) Pull-down assay confirming the interaction between VdPDA1 and the VdPDA1-N1. A clear GST signal was observed in the experimental group (EG), whereas the signal in the control group (CG) was weaker than that in EG, indicating a positive interaction between VdPDA1 and VdPDA1-N1. Discussion Protein binders, engineered molecules that selectively attach to a specific protein target, are foundational tools in structural biology, therapeutics, and synthetic biology. By recognizing precise surface features on a target protein, binders can modulate function, block interactions, recruit effectors, or serve as diagnostic or imaging tools. Their success, however, depends on accurately identifying surface regions that are not only accessible but also energetically favorable for binding. These critical regions, known as interaction hotspots, often contribute disproportionately to binding free energy despite comprising only a small subset of surface residues. Identifying such hotspots remains a significant challenge. Traditional approaches rely on co-crystallized protein complexes, which are expensive and time-consuming to obtain, and typically represent only one of many possible binding modes [63]. Moreover, experimental methods like alanine scanning mutagenesis are labor-intensive and low-throughput. Recent advances in protein complex prediction, particularly AlphaFold2-Multimer and AlphaFold3, provide powerful approaches for predicting interfaces when the interacting partners are specified. Surf2Spot addresses a complementary problem by identifying interaction-prone regions on a target surface without requiring prior knowledge of the binding partner. This distinction is particularly relevant to de novo binder design, in which the target is known but the eventual binder sequence and binding mode are not predetermined. In addition, Surf2Spot avoids partner-specific complex prediction and MSA generation, enabling rapid screening of candidate interaction regions across proteins. Thus, Surf2Spot should be viewed not as a replacement for complex prediction but as a lightweight and partner-independent hotspot identification layer that can be integrated upstream of generative binder-design pipelines. Surf2Spot adopts DGCNN as its backbone architecture. Unlike conventional GCNs that operate on a fixed graph topology, DGCNN dynamically recomputes local neighborhoods as feature representations evolve across layers. Dynamic neighborhood construction allows the model to adaptively aggregate locally relevant surface patterns rather than relying on a fixed connectivity structure. For NAI prediction, incorporating GAT layers on top of PPI pretraining consistently improves performance (Fig. 3D). One possible explanation is that NAI hotspots tend to be more spatially localized, making adaptive attention over neighboring surface features potentially beneficial for concentrating predictions around relevant epitope regions (Supplementary Table S4). However, this interpretation remains a mechanistic hypothesis and would require additional analysis of attention weights or spatial prediction distributions to establish directly. We investigated whether target-domain truncation alone could provide sufficient localization information for binder design. However, domain-only designs showed a strong bias toward the artificial truncation surface, with generated binders repeatedly concentrating on the exposed truncation regions rather than biologically accessible interaction sites. We hypothesize that the newly exposed truncation surfaces contain buried hydrophobic residues that resemble favorable binder interfaces, resulting in a high frequency of structurally plausible but biologically irrelevant designs. Therefore, explicit hotspot specification becomes necessary under these circumstances. By leveraging predicted structures and surface-based graph representations, Surf2Spot predicts potential PPI hotspot regions on protein surfaces in a partner-agnostic manner, without requiring the sequence or structure of a predefined interaction partner. The predicted hotspots therefore represent surface regions with a higher propensity to participate in PPIs, rather than partner-specific energetic hotspots associated with a single known complex. This architecture dramatically reduces both time and cost compared to traditional methods such as alanine scanning or crystallographic structure determination, which typically require months of experimental work. Surf2Spot can generate widely distributed hotspots in multiple domains in minutes, making it well suited for high-throughput screening and rapid iteration in protein engineering workflows. By combining surface-aware modeling with pretrained sequence and structure embeddings, Surf2Spot unlocks scalable, partner-agnostic hotspot discovery and enhances hotspot-guided binder design efficiency, even for structurally uncharacterized or “hard-to-drug” targets [64]. Surf2Spot’s design addresses key technical challenges that have historically limited hotspot prediction tools. One of the central difficulties is the severe label imbalance in available datasets: true interaction hotspots are sparse, and experimentally validated complexes capture only a subset of potential interactions for any given protein. To mitigate this, Surf2Spot adopts a multi-pronged strategy. First, it employs surface-focused representations that exclude buried residues unlikely to participate in PPIs. Second, it decomposes proteins into functional substructures using GPSite and Chainsaw, narrowing the prediction space to regions more likely to be involved in binding (Supplementary Fig. S7). Finally, it uses a focal loss function during training to emphasize hard-to-classify examples, improving model performance under class imbalance [65]. Together, these innovations allow Surf2Spot to outperform state-of-the-art sequence-based and structure-based methods across multiple evaluation metrics, including F1 score and AUPRC. Besides, Cluster IOS was introduced for evaluating each hotspot cluster independently (Fig. 3C). Beyond predictive accuracy, the practical utility of Surf2Spot is reflected in its capacity to guide and accelerate protein design campaigns. Within binder engineering pipelines, confinement of the design space to functional domains reduced computational expenditure by more than half, attributable to the substantial reduction in search space. Yet, domain restriction in the absence of hotspot constraints was found to generate a considerable proportion of nonfunctional or misdirected designs, thereby limiting its standalone utility. The subsequent incorporation of Surf2Spot-derived hotspot predictions markedly enhanced the success rates of design campaigns, with the NbPDS1 case study exhibiting an approximately four-fold increase in viable candidates. This enhancement, however, proved to be target-dependent: for the VdPDA1 target, hotspot-guided RFdiffusion failed to produce any functional binders, underscoring intrinsic limitations of the strategy under certain target contexts. In contrast, Surf2Spot-guided IgGM-based nanobody designs for VdPDA1 yielded superior calculated interface quality and an increased number of predicted contact residues, indicating workflow-specific outcomes. Collectively, these observations suggest that the combination of domain restriction and hotspot prediction constitutes a mutually reinforcing design framework, although its efficiency and success remain contingent upon both target properties and pipeline selection. Despite its performance, Surf2Spot has several limitations. First, the accuracy of hotspot prediction depends on the quality of the input structural model. Our pLDDT-stratified analysis indicates reduced performance in low-confidence regions, suggesting that structural uncertainty remains a limiting factor (Supplementary Figs S5 and S6). Second, PPI and NAI interfaces exhibit different energetic and structural characteristics (Supplementary Figs S1 and S2); therefore, separate models are currently required. Developing a unified model capable of capturing diverse interaction patterns will require larger and more heterogeneous training datasets. Third, Surf2Spot predicts potential hotspot locations on protein surfaces but does not directly model partner-specific binding processes or binding energetics. Future improvements integrating dynamic structural information and experimentally validated interaction data may further enhance prediction accuracy. An important limitation of the current labeling strategy is that hotspot annotations are derived from experimentally characterized protein complexes. A residue labeled as nonhotspot because it does not contribute substantially to one observed interaction may still participate in another biologically valid interaction with a different partner. Thus, the negative labels should not be interpreted as evidence that these residues are universally incapable of mediating PPI. Accordingly, Surf2Spot should be interpreted as predicting regions with a higher propensity to contain PPI hotspot residues rather than exhaustively identifying all possible interaction sites on a protein surface. Proteins with multiple interaction interfaces therefore represent an important challenge for evaluation and future model development. Besides, a limitation of this study is the absence of a large, independently curated NAI benchmark. Existing tools address related but distinct tasks (e.g. antibody–antigen interface or B-cell epitope prediction), rendering direct comparisons imperfect. Future work with larger independent NAI datasets will be valuable for further validating the generalizability of Surf2Spot. Although LUC complementation assays support the existence of interactions between designed candidates and target proteins, they do not directly quantify binding affinity or thermodynamic parameters. Similarly, AlphaFold-derived confidence metrics such as pLDDT and pAE reflect structural confidence rather than experimental binding strength. Therefore, while Surf2Spot-guided designs show improved interaction compatibility, quantitative affinity measurements such as SPR, BLI, or ITC will be required to determine whether hotspot-guided designs achieve enhanced binding affinity. Future adoption of positive-unlabeled learning could address a key limitation of Surf2Spot’s current formulation: treating unannotated surface residues as negatives, despite their unknown functional status. Within the DGCNN architecture, unlabeled samples could be incorporated via tailored graph losses without disrupting spatial representation. This aligns naturally with Cluster IOS, where cluster-level uncertainty can be estimated. Larger experimental datasets and PU learning together promise more biologically realistic predictions. In conclusion, Surf2Spot offers a scalable and interpretable approach to hotspot prediction that enables more efficient and effective protein binder design. By operating directly on predicted structures and bypassing the need for experimental complex data, it expands the range of tractable targets and accelerates the design–build–test cycle across diverse applications in biology, biotechnology, and synthetic design. Material and methods Surface feature representation In our framework, a protein surface represented as a point cloud is modeled as an undirected graph , where V denotes the surface points (nodes) and E denotes the contacts between these points derived from the triangular mesh (edges). Thus, the surface graph G can be represented by a node feature matrix X and an adjacency matrix. To train our model, we extracted three groups of protein correlation features: sequence properties (ProtTrans embedding and AA type), structural properties (DSSP), and surface properties (point cloud sampling), which were concatenated after processing by MLPs to form the final node feature matrix. Sequence properties Sequence embeddings were computed using the pretrained protein language model ProtTrans, which was trained on 393 billion amino acids from UniRef and BFD [66, 67]. Each amino acid was encoded as a 1024D vector using ProtTrans with default parameters. In parallel, amino acid types were encoded using one-hot encoding. Structural properties Three types of structural properties were computed using the DSSP program: (i) a 9D one-hot encoded secondary structure profile, where the first eight dimensions represent eight categories of secondary structure states and the last dimension represents unknown secondary structure; (ii) peptide backbone torsion angles PHI and PSI; and (iii) solvent-accessible surface area (ASA), which was normalized to relative solvent accessibility (RSA) using the maximal possible ASA of the corresponding amino acid type. Surface properties One geometric and three energetic features were extracted and attached to the nodes of the pocket surface. The geometric feature included the shape index; the energetic features included Poisson–Boltzmann continuum electrostatics, hydrogen bond potential, and hydropathy. Shape index The shape index describes local surface topology in terms of principal curvature and was proposed by Koenderink and van Doorn as a single-valued angular measure. The definition is as follows: where and are the principal curvatures of the surface at a specific point, which can be calculated from Gaussian curvature and mean curvature as follows: Poisson–Boltzmann continuum electrostatics The Poisson–Boltzmann equation describes the continuous charge distribution of biomolecules in solution as follows: where ϕ is the system potential; ε is the dielectric constant in solution; and are the bulk phase concentration and charge of the i-th atom, respectively; and , where is the Boltzmann constant. The calculation was performed by PDB2PQR and APBS following the MaSIF method. Hydrogen-bond potential A hydrogen bond is essentially a dipole–dipole interaction, usually represented as X-H …Y, where X is covalently (or ionically) bonded to hydrogen and has high electronegativity, while Y is generally an atom containing lone pairs of electrons. Hydrogen bonding interactions are widely used in the construction of force fields. Similar to previous work, hydrogen-bond potential was assigned to surface nodes. Hydropathy Hydrophobic interaction is a noncovalent interaction present between nonpolar groups and is prevalent in biological systems. Similar to previous work, hydrophobic interactions were assigned to surface nodes based on the residues they belonged to. Compilation of training datasets The benchmark dataset for evaluating PPI-binding hotspot prediction was derived from the supplementary materials of the PDBBIND2020 and GPSite datasets [40, 68]. The benchmark dataset for predicting NAI-binding hotspots was derived from the VHH antigen–antibody conjugate structures in SabDab [56]. A total of 1200 protein dimer complexes corresponding to target proteins were collected from PDBBIND2020, and 446 were supplemented from the GPSite dataset. SabDab included 1248 antigen VHH antibody complex structures corresponding to antigens. We removed redundant proteins with sequence length differences > 20% and shared sequence identities > 80%, and removed proteins shorter than 30 residues. To reduce redundancy between training and evaluation datasets, sequence similarity filtering was performed during dataset construction. The dataset sizes retained under different sequence identity thresholds are summarized (Supplementary Table S5). Because experimentally characterized protein complexes deposited in structural databases are often derived from related protein families and biological systems, applying extremely stringent sequence identity thresholds substantially reduces the available dataset size and limits model training. Therefore, an 80% sequence identity cutoff was initially used to balance dataset diversity and sufficient training samples. To further assess potential data leakage, pairwise protein sequence similarities between training and test proteins were calculated and are summarized (Supplementary Table S6). Sequence similarity We adopted a k-mer profile-based approach for sequence comparison. For each protein sequence extracted from PDB structures, we computed the frequency distribution of all possible 3-mer (tripeptide) subsequences, yielding an 8000D vector representing the normalized occurrence of each k-mer (203 possible amino acid combinations). Sequence similarity between two proteins was then quantified as the cosine similarity between their respective k-mer profile vectors. This method offers computational efficiency for large-scale comparisons and is robust to sequence length variation. As a reference, we also computed global alignment-based sequence identity using the Needleman–Wunsch algorithm (match: +2, mismatch: −1; gap open: −0.5; gap extension: −0.1) implemented in Biopython, with identity defined as the number of matches divided by the length of the longer sequence. Structural similarity Structural similarity was evaluated using GDT-TS (Global Distance Test—Total Score). Cα coordinates were extracted from PDB files, and structures were optimally superimposed using the Kabsch algorithm. For each aligned pair, we calculated the proportion of Cα atoms with root-mean-square distances (RMSDs) below three thresholds: 1, 2, and 4 Å, yielding GDT-1, GDT-2, and GDT-4 values, respectively. For both the PPI and NAI datasets, we performed three pairwise comparisons: within the training set, within the test set, and between training and test sets. For each comparison, we reported the mean sequence similarity and mean GDT-TS values, as well as the proportion of protein pairs exceeding a 50% sequence similarity threshold. The final training, validation, and testing datasets were generated based on this redundancy-controlled dataset. Hotspot definition Hotspot residues were defined according to their FoldX-calculated energetic contributions [69]. For the PPI dataset, residues with a FoldX interaction energy contribution above 2 kcal/mol were considered hotspots, whereas a less stringent threshold of 0 kcal/mol was used for the NAI dataset. Although NAI can be considered a specialized form of PPI, the two datasets exhibit distinct interface characteristics. NAI interfaces are enriched in CDR-loop-mediated contacts, whereas conventional PPI interfaces are generally broader and can involve residues across α-helical, β-sheet, and loop regions. Consequently, the distributions of residue-level FoldX energetic contributions differ between the two datasets (Supplementary Fig. S1, Supplementary Table S7). Applying the PPI threshold directly to the NAI dataset resulted in a substantial number of complexes containing few or no hotspot labels because of the smaller interface size and different energetic distribution of NAI interfaces. Therefore, the 0 kcal/mol threshold was adopted to retain energetically favorable NAI interface residues while avoiding excessive loss of positive labels. Other residues were treated as nonhotspot residues for model training. Importantly, this negative label indicates the absence of a FoldX-defined hotspot in the experimentally characterized interface and does not imply that these residues are incapable of mediating other PPIs. Complex structures with fewer than three hotspot residues were removed to ensure the quality of the dataset. To prevent the model from over-fitting the amino acid conformation of the binding state on the contact surface of the complex, ESMFold was used to predict the monomer conformation of the target protein in the complex and supplement it into the dataset. The final PPI dataset contains 1020 proteins, and the NAI dataset contains 498 proteins. Each pair of complex structure and its corresponding monomer prediction was assigned to either the training set or the test set together, ensuring consistency and fairness in the evaluation process. Compilation of model evaluation datasets (Test_52) We adopted three publicly available and widely used benchmark datasets from previous studies: Dset_72 [70], Dset_164, and Dset_186 [71]—named according to the number of proteins they contain. We selected protein structures from these established benchmark datasets that were absent from our training data. Initial screening identified 11, 24, and 50 candidate proteins from Dset_72, Dset_186, and Dset_164, respectively. We processed the dataset by removing duplicate sequences using the same method applied to the training dataset and required each entry to contain at least six hotspot residues. These candidates were further refined using ESMFold to retain only high-confidence monomeric structures with pLDDT scores >80, yielding a final fused test set (Test_52) comprising 10 proteins from Dset_72, 18 from Dset_186, and 24 from Dset_164. This rigorous selection process ensured that Test_52 represented a diverse, nonredundant collection of structurally reliable targets for robust model validation. To minimize potential information leakage, samples from the training data that overlapped with the Test_52 proteins were removed from the Surf2Spot training set. This conservative filtering was applied before model training to ensure that the reported Surf2Spot performance was not obtained from direct exposure to test proteins during training. Because this filtering reduces the number of available training samples and may result in a smaller independent test set than some existing benchmark settings, we additionally quantified protein- and sequence-level similarity between the training and test sets (Supplementary Table S6) to assess the degree of potential distributional overlap. For NAI, we curated an additional 21 complexes as an independent test set for model evaluation. Because existing prediction tools are primarily designed for conventional antibody–antigen or B-cell epitope prediction rather than nanobody–antigen hotspot prediction, we did not consider their outputs to constitute directly comparable external hotspot-prediction benchmarks. We therefore evaluated Surf2Spot on the held-out NAI test set and explicitly acknowledge the absence of an independent external NAI benchmark as a limitation. Assessment of train–test sequence similarity To evaluate potential sequence-level overlap between the training and test sets, we calculated pairwise sequence identity between each Test_52 protein and the training proteins. To ensure the robustness of model evaluation and prevent data leakage, a sequence similarity filtering procedure was performed to remove proteins from the training and test sets that exhibit high sequence similarity to the opposite set. The detailed protocol is as follows: Pairwise sequence alignment For each protein in the source set (e.g. the training set), a global pairwise sequence alignment was performed against every protein in the target set (e.g. the test set) using the Needleman–Wunsch algorithm. Alignment parameters The sequence alignments were conducted with the following parameter settings: Substitution Matrix: BLOSUM62, Gap Open Penalty: −10, Gap Extension Penalty: −0.5. Sequence identity calculation For each pair of aligned sequences, the sequence identity was calculated as: Similarity scoring For each protein in the source set, its similarity score was defined as the maximum sequence identity observed across all alignments with the target set: Filtering criterion A protein was retained in the source set if its similarity score was below the predefined threshold and filtered out if the score was equal to or exceeded the threshold. This filtering procedure was applied bidirectionally to both the training and test sets to ensure no highly similar sequences remained across sets. Threshold selection Based on common practice for data leakage prevention, the filtering threshold was set to threshold specifies an interval of 20% sequence identity, with any protein in the opposite set excluded. For each test protein, the maximum sequence identity to any training protein was recorded, and the resulting distribution was summarized across the test set (Supplementary Table S8). Quantifying the spatial distribution of hotspots To evaluate the spatial organization and predictive accuracy of hotspot residues, we introduced several quantitative metrics that assess how well the predicted hotspots capture key structural and distributional characteristics. To evaluate the spatial organization and predictive accuracy of hotspot residues, we introduced several quantitative metrics that assess how well the predicted hotspots capture key structural and distributional characteristics. These include the Spatial Consensus Index (SCI), which measures the consistency of hotspot locations across predictions; the clustering score (Sclustering), which quantifies spatial compactness and uniformity; and the similarity score (Ssimilarity), which measures the spatial overlap between predicted hotspots and reference hotspots (Supplementary Fig. S9). SCI reflects the degree of consistency in hotspot spatial distribution and quantifies the comprehensive quality of hotspot distribution and is defined as: The clustering score is defined as the product of the and , which requires hotspots to be both compact and evenly distributed: measures the uniformity of hotspot distribution, with values closer to 1 indicating a more uniform distribution (i.e. all points are at similar distances from the center). In contrast, lower values indicate a less uniform distribution, where some points are very close to the center while others are much farther away: : CA atom coordinate vector of the i-th hotspot residue. : geometric center of the hotspot residue. : coordinate vector of the j-th atom in the protein. n : number of hotspot residues. m : number of protein atoms. Distance calculation: Statistical calculation measures the spatial compactness of hotspots within a protein. A value closer to 1 indicates that the hotspots are more tightly clustered, whereas a smaller value suggests a more dispersed spatial distribution: : maximum diameter of the protein The similarity score quantifies the proximity between the geometric centers of two hotspot residue sets in 3D space. A value close to 1 indicates that the two hotspot centers are nearly coincident, reflecting high prediction accuracy, whereas a value close to 0 indicates that the centers are far apart, implying low prediction accuracy. Standardization ensures comparability of the scores among proteins of different sizes: However, because SCI incorporates geometric assumptions regarding hotspot distribution, it was not used as the primary performance metric. To quantitatively evaluate the spatial overlap between predicted and true hotspot residues in 3D space, we developed a cluster-aware Interface Overlap Score (IOS) that accounts for the spatial distribution of hotspot residues across multiple binding clusters instead (Supplementary Fig. S10). Surface hotspot filtering. Surface hotspots were first filtered based on their relative solvent accessibility (RSA) computed from DSSP. Only predicted hotspots with RSA > 0.2 were retained as surface residues for subsequent analysis. Spatial clustering of true hotspots. To avoid overestimation of prediction quality when hotspot residues are distributed across multiple spatial clusters, we performed hierarchical clustering with complete linkage on the true hotspot residues. The distance threshold was set to 10 Å, based on the typical spatial extent of binding interfaces. Global IOS. For reference, we also computed a global IOS without clustering, where all predicted and true hotspot residues were pooled together for a single IOS calculation. The cluster-aware IOS provides a more conservative and realistic assessment than the global IOS, as it penalizes predictions that correctly identify individual hotspots but fail to capture complete spatial clusters. Cluster IOS. For each true hotspot cluster, predicted hotspot residues within a radius of 20 Å (2× clustering distance) from the cluster center were identified, and the standard IOS was computed independently for each cluster. Specifically, a 3D voxel grid with a voxel size of 1.0 Å was constructed, and spheres of radius 5.0 Å were drawn around each Cα atom. The IOS was then calculated as the intersection-over-union (IoU) of the voxelized spheres: where and denote the voxelized volumes of predicted and true hotspot residues, respectively. Weighted aggregation. The per-cluster IOS values were averaged with weighting by the number of true hotspot residues in each cluster: where Cluster\ IOS is the IOS of cluster and is the number of true hotspot residues in cluster c. Construction and training of DGCNN and GAT networks Model of DGCNN We computed a directed graph representing the local point cloud structure, where and are the vertices and edges, respectively. In the simplest case, we constructed G as the k-nearest neighbor (k-NN) graph of X in . The graph includes a self-loop, meaning each node also points to itself. We defined edge features as , where : is a nonlinear function with a set of learnable parameters ϴ. Then, a channel-wise symmetric aggregation operation □ (e.g. sum or max) is applied to re-characterize the central point, aggregating additional neighborhood information. Overall, given an F-dimensional point cloud with n points, EdgeConv produces an F′-dimensional point cloud with the same number of points: It has been shown that recomputing the graph using nearest neighbors in the feature space produced by each layer is beneficial, which is a crucial distinction between Dynamic Graph CNNs (DGCNNs) and graph CNNs that operate on a fixed input graph. Construction of models The above-mentioned attributed graph, incorporating features from ProtTrans, DSSP, and surface properties, is input into several MLP layers. The processed surface is represented by , where and are the vertices and edges, respectively. We constructed G as the k-nearest neighbor (k-NN, k = 20) graph of X in . The graph includes a self-loop, meaning each node also connects to itself. Edge features are defined as , where : is a nonlinear function with a set of learnable parameters ϴ. The new representations are aggregated and input into a three-layer DGCNN. Each DGCNN layer includes message passing, edge update, and global node update modules to learn residue representations by considering multi-scale interactions at the node, edge, and global context levels. DGCNN applies a channel-wise symmetric aggregation operation to re-characterize each central point and recomputes the graph using nearest neighbors in the feature space produced by each layer. Overall, given an F-dimensional point cloud with n points, EdgeConv produces an F′-dimensional point cloud with the same number of points. In this way, the three-layer DGCNN progressively updates the structural edge representations of the protein surface map and aggregates node features. The node features output by DGCNN are combined with the surface representations via a residual connection before being input into the graph neural network module. For NAI hotspot prediction, the representation after this residual connection is fed into a three-layer Graph Attention Network (GAT), and its output is then combined with the DGCNN output through a second residual connection. The output of the last graph layer is passed through an MLP to predict the protein interaction probabilities for all amino acid residues: where is the output of the Lth graph neural network layer; W is the weight matrix; b is the bias term, and is the prediction of the surface node. To improve logit separation under severe class imbalance, we applied an empirically selected scaling factor α of 5 to the output logits for the binding probability value of each surface point before sigmoid activation, using the same scaling during both training and inference. Training process We performed five-fold cross-validation on the training data, where in each iteration, the model was trained on four folds and evaluated on the remaining fold. This process was repeated five times, and the average validation performance was used to optimize the network’s hyperparameters. In the test phase, all five trained models from cross-validation were used to make predictions, which were then averaged to produce the final Surf2Spot output. The Adam optimizer (learning rate = 1e-4) was used for model optimization on the focal loss: The training process (batch_size = 1) lasted at most 120 epochs, and we performed early stopping based on validation performance, which took ~18 h on an NVIDIA A100 GPU (Supplementary Table S9). Model performance evaluation We used accuracy (ACC), precision, recall, F1-score (F1), Matthews correlation coefficient (MCC), area under the receiver operating characteristic curve (AUROC), and area under the precision–recall curve (AUPRC) to measure predictive performance: PCRE (Predicted Count Relative Error) is defined as the absolute value of the difference between the number of predicted positive samples and the number of true-positive (TP) samples, divided by the number of TP samples. This metric reflects the relative error in the count of positive samples predicted by the model. TPs and true negatives (TNs) denote the number of interacting and noninteracting sites identified correctly, and false positives (FPs) and false negatives (FNs) denote the number of incorrectly predicted interacting and noninteracting sites, respectively. AUROC and AUPRC are independent of thresholds, thus revealing the overall performance of a model. Other metrics were calculated using a threshold to convert predicted interaction probabilities to binary predictions, which was determined by maximizing the F1-score for each model. Our model defines a hotspot amino acid based on the principle of spatial proximity: any residue is identified as a hot spot if its 10 Å spatial sphere contains surface points predicted as hot spots. This cutoff was selected as a compromise between interface localization specificity and hotspot label coverage under the severe class imbalance of the dataset. The robustness of this threshold was evaluated using alternative cutoffs ranging from 0 to 12 Å (Supplementary Fig. S3). This rule effectively mitigates identification biases stemming from point cloud mapping variations, aligning predictions more closely with the authentic spatial patterns of biophysical interactions. Protein-level bootstrap confidence interval analysis To perform statistical inference on model performance metrics while accounting for the hierarchical structure of the data (residues nested within proteins), we employed a protein-level bootstrap resampling procedure to estimate 95% confidence intervals (CIs) for threshold-independent metrics (Supplementary Tables S10 and S11). For each protein *i* in the test set, we computed the AUROC and AUPRC using the predicted probabilities and ground-truth labels. Proteins containing only a single class in the ground-truth labels (all hotspots or all nonhotspots) were excluded from the analysis, as AUC is undefined for single-class problems. We then performed bootstrap resampling with replacement at the protein level. In each bootstrap iteration *b* (where *b* = 1, …, 1000), we resampled N proteins with replacement from the test set (where N equals the number of proteins with valid per-protein metrics) and computed the mean AUROC and mean AUPRC across the resampled proteins. This procedure was repeated 1000 times to obtain bootstrap distributions for both metrics. Confidence intervals were estimated using the percentile method with α = 0.05: the lower bound was defined as the 2.5th percentile and the upper bound as the 97.5th percentile of the bootstrap distribution, yielding 95% CIs for AUROC and AUPRC. This protein-level bootstrap approach was chosen over residue-level resampling to avoid violating the independence assumption, as residues within the same protein are spatially and evolutionarily correlated. By resampling entire proteins, the hierarchical structure of the data is properly preserved, enabling valid statistical inference for model comparisons across independent proteins. All bootstrap analyses were performed with B = 1000 iterations and a random seed of 42 for reproducibility. Ablation experiments To systematically evaluate the contribution of Surf2Spot’s surface-centric representation, input features, and graph neural network architecture, we performed a series of ablation experiments. For graph representation ablation, we compared the full Surf2Spot surface representation with atom-level graphs, residue-level graphs, and whole-surface graphs. For feature ablation, we separately evaluated models using surface features, DSSP-derived structural features, or ProtTrans sequence embeddings alone. To assess the contribution of the graph neural network architecture, we replaced the DGCNN module with conventional GCN, GAT, or other graph neural network architectures while keeping the remaining components unchanged. All ablation models were evaluated under the same dataset partitioning and evaluation protocol as the full Surf2Spot model. Comparative methods To provide a broader assessment of Surf2Spot, we additionally evaluated DeeProSite, GraphPPIs, and ScanNet alongside SPOTONE and PPI-hotspotID. These methods were selected to represent complementary approaches for protein interaction-site or hotspot prediction. Each method was evaluated using AF2 predicted structure, and predictions were mapped to the same residue-level test set for quantitative comparison. Cross-task transfer and fine-tuning We found that a model trained exclusively on PPI data generalized poorly to NAI prediction (F1 ≈ 0.02, AUROC ≈ 0.56) (Supplementary Table S12), likely due to the distinct interface properties of NAIs, including smaller contact areas and different energetic determinants. We therefore trained a separate NAI-specific model. To assess whether knowledge learned from PPI hotspot prediction can generalize to NAIs, we evaluated the PPI-trained Surf2Spot model directly on the NAI dataset without additional training. We then compared this cross-task transfer setting with models trained exclusively on the NAI dataset and models initialized with PPI-trained parameters, followed by NAI-specific fine-tuning. All models were evaluated using the same NAI test set and evaluation metrics. Surface partitioning Before hotspot prediction, the accessible protein surface can be partitioned into spatially localized surface regions using either GPSite or ChainSaw. Users can select either method, with GPSite used as the default option. This preprocessing step serves two purposes: (i) to restrict the prediction space to localized surface regions that are more likely to participate in PPIs and (ii) to alleviate the severe label imbalance caused by the large number of nonhotspot surface residues in the training data. Importantly, the GPSite prediction scores themselves are not provided as input features to Surf2Spot (Supplementary Fig. S8). Instead, they are used only to define the surface regions on which the Surf2Spot model performs hotspot prediction. Training. During training, we partitioned the protein surface into localized candidate regions using either GPSite or ChainSaw to mitigate the dominance of nonhotspot residues in the training samples. Surf2Spot was subsequently trained on these localized surface graphs. Inference. The same surface-partitioning procedure is applied during inference to generate localized surface regions prior to hotspot prediction. Notably, GPSite was used as an upstream surface-partitioning step during inference, but its prediction scores were not provided as input features to Surf2Spot. For users who require a holistic prediction covering the entire protein surface, we recommend using ChainSaw for domain-based segmentation instead of GPSite, as it yields structurally meaningful partitions that better preserve the spatial continuity across multi-domain proteins. Structure source for Surf2Spot prediction Surf2Spot was trained and evaluated using ESMFold-predicted monomer structures as input. Although AlphaFoldDB provides high-confidence predicted structures and can be used when available, directly replacing ESMFold structures with AlphaFoldDB models introduces a structural distribution shift relative to the training data. In our control analysis, using AlphaFoldDB structures without retraining resulted in reduced prediction performance (Supplementary Table S13). Therefore, we maintained ESMFold as the default structure source for Surf2Spot inference to ensure consistency between training and application settings. For downstream binder design experiments, AlphaFoldDB structures were used because high-quality target structures are beneficial for subsequent complex modeling and binder optimization. This choice does not affect the Surf2Spot prediction procedure, which uses ESMFold-generated structures as defined in the prediction pipeline. We further evaluated the effect of replacing ESMFold inputs with AlphaFoldDB structures without retraining. The performance decrease observed under this setting suggests that maintaining consistency between training and inference structure sources is important for Surf2Spot prediction. Target preparation and hotspot-guided design For binder design, target structures were prepared before diffusion-based binder generation. Surf2Spot-predicted hotspot residues were provided as spatial constraints to guide binder generation toward biologically accessible interaction regions. In addition, target domain truncation was applied when necessary to remove irrelevant structural regions and focus the design process on the intended binding domain. Importantly, domain truncation and hotspot guidance influence different aspects of the design process: domain truncation reduces the structural search space, whereas hotspot guidance specifies the preferred interaction region on the accessible target surface. Binder design pipeline and filtration criteria RFdiffusion design comparison To compare the design efficiency of RFdiffusion with or without Surf2Spot, we designed binders to two targets: NbPDS1 (AlphaFoldDB ID: AF-B1NYI4-F1-v4) and VdPDA1 (AlphaFoldDB ID: AF-G2XC45-F1-v4). In the case of Surf2Spot guidance, we targeted binders to specific sites on the target protein using predicted hotspot residues. The selected hotspots and truncated substructures are listed below (with chain and residue indices from PDB); all other parameters were kept at their default values. We set the length of the generated binders to 100 amino acids and generated 300 binder structures per run, with all other parameters left as defaults. For sequence design and refinement, we applied the previously described ProteinMPNN-FastRelax protocol [72], which starts with an initial ProteinMPNN round and then iteratively alternates between FastRelax and ProteinMPNN to progressively optimize sequence-structure agreement. One binder sequence was generated for each structure produced by RFdiffusion. To filter the designs, we ran AlphaFold2 with initial guess and target templating. A design was classified as successful if it met two criteria: (i) the predicted aligned error (pAE) of interaction between the binder and target <10, and (ii) the pLDDT > 70. All AlphaFold2 parameters were left at their default settings. BindCraft design comparison To compare the design efficiency of BindCraft with or without the guidance of Surf2Spot, we designed the same target proteins as in the efficiency comparison experiment of RFdiffusion. Under the guidance of Surf2Spot, we used the predicted hotspot residues to direct the conjugate to a specific site on the target protein. The selected hotspots and truncated substructures are as follows (chain and residue indices from PDB): For the BindCraft design, the length of the adhesive was fixed at 100 amino acids, and the expected number of final filtered designs was set to 5. All other design and filtering parameters were kept at the default settings. Default filtering parameters that were used include: a) AF2 confidence pLDDT score of the predicted complex (>0.8) a) AF2 confidence pLDDT score of the predicted complex (>0.8) b) AF2 interface predicted confidence score (i_pTM) (>0.5) b) AF2 interface predicted confidence score (i_pTM) (>0.5) c) AF2 interface predicted alignment error (i_pAE) (>0.35) c) AF2 interface predicted alignment error (i_pAE) (>0.35) d) Rosetta interface shape complementarity (>0.55) d) Rosetta interface shape complementarity (>0.55) e) Number of unsaturated hydrogen bonds at the interface (<3) e) Number of unsaturated hydrogen bonds at the interface (<3) f) Hydrophobicity of binder surface (<35%) f) Hydrophobicity of binder surface (<35%) g) RMSD of binder predicted in bound and unbound form (<3.5 Å) g) RMSD of binder predicted in bound and unbound form (<3.5 Å) Nanobody design pipeline and filtration criteria To evaluate the impact of Surf2Spot on the design efficiency of IgGM, we generated nanobodies targeting two distinct antigens: NbPDS1 (AlphaFoldDB ID: AF-B1NYI4-F1-v4) and VdPDA1 (AlphaFoldDB ID: AF-G2XC45-F1-v4). In the Surf2Spot-assisted design group, nanobodies were strategically engineered to bind predicted hotspot residues within specific antigenic epitopes. The selected hotspots and their corresponding truncated structural domains were annotated based on chain identifiers and residue indices from reference PDB structures. The selected hotspots and truncated substructures are listed below (chain and residue indices are based on the PDB entry); all other parameters were kept at their default settings. Nanobody binding free energy and interaction analysis The binding energy and interface properties of nanobody–antigen complexes were analyzed using Rosetta (v3.13) through the following steps: (i) Energy minimization: The minimize.static.linuxgccrelease module was executed with parameters -min_type lbfgs_armijo_nonmonotone (L-BFGS optimization), -score:weights ref2015_cst (ref2015_cst scoring function), -min_tolerance 0.001 (convergence threshold), and backbone constraints (-relax:bb_move false). (ii) Interface analysis: The InterfaceAnalyzer.static.linuxgccrelease tool quantified interface metrics, specifying the interaction chains (-interface A_H), enabling repacking of separated states (-pack_separated), and outputting energy/structural features. Nanobody design filtering criteria Key parameters were extracted and filtered based on: (i) Binding efficiency: Complexes with dG_cross/dSASA × 100 < 0 were retained, indicating favorable binding free energy (dG_cross) per unit solvent-accessible surface area (dSASA). (ii) Hydrogen bonds: On the basis of ensuring dG_cross/dSASA × 100 < 0, interfaces with hbonds_int > 0 were preferred to ensure specific polar interactions. Results were sorted by dG_cross/dSASA × 100 (ascending) and exported as a CSV file, including comprehensive metrics (total score, dG_cross/dSASA × 100, hbonds_int, total_score, etc.). Filtering criteria for binder design Since different binder-generation methods produce different structural confidence metrics and quality scores, method-specific filtering criteria were applied according to the original workflows. However, for each individual target protein, the same filtering strategy was consistently applied among candidate designs generated by the same method to ensure fair comparison. For RFdiffusion-based designs, the interaction predicted alignment error (pae_interaction) threshold was adjusted according to the target-specific score distribution. Although a value around 10 is commonly used, previous studies have indicated that this threshold can be optimized based on preliminary design runs for different targets [3]. Luciferase complementation assay Positive control setup The co-transfection of the SV40 Large T antigen (TAg) expression plasmid and the p53 expression plasmid served as the positive control for the luciferase (LUC) reporter system. p53, a key human tumor suppressor protein, specifically binds to p53-responsive elements in promoters and strongly activates downstream transcription; it is a commonly used transcriptional activation positive control in luciferase reporter assays. The SV40 Large T antigen is a core transforming protein encoded by Simian vacuolating virus 40 (SV40) that directly binds to and inhibits p53, blocking its DNA-binding and transcriptional activation functions. Co-transfection of TAg and p53 serves as a positive control for p53 functional inhibition, verifying the validity of the experimental system [1]. Plasmid construction The coding sequences of the target and design were cloned into the upstream region of the nLuc multi-cloning site (MCS) of the p1300-nLuc vector and the downstream region of the cLuc MCS of the p1300-cLuc vector, respectively. The nLuc fusion was constructed without a stop codon, while the cLuc fusion retained the stop codon. This resulted in the experimental plasmids p1300-nLuc-target and p1300-cLuc-design. Agrobacterium transformation and tobacco infiltration These constructs, empty vectors, and positive controls were transformed into Agrobacterium tumefaciens strain GV3101 and cultured on LB plates containing 50 μg/ml kanamycin. When the bacterial suspension reached an OD600 of 0.8, cells were harvested and resuspended in buffer containing 10 mM MgCl₂, 10 mM MES (pH 5.6), and 100 μM acetosyringone. Paired bacterial suspensions were mixed in equal volumes and co-infiltrated into leaves of N. benthamiana using a needleless syringe. Different areas of the same leaf were assigned to different treatment groups. At 36–48 h postinfiltration, 1 mM D-luciferin potassium salt working solution was injected into the corresponding areas, and signals were immediately captured using a chemiluminescence imaging system. Luminescence intensity reflected the strength of PPIs. The coding sequences of target and design were individually cloned into the p1300-nLuc vector upstream of the nLuc MCS and the p1300-cLuc vector downstream of the cLuc MCS, respectively, ensuring in-frame fusion without a stop codon at the nLuc terminus and retention of the stop codon at the cLuc terminus to generate the experimental plasmids p1300-nLuc-target and p1300-cLuc-design. These constructs, along with empty vectors and positive controls, were transformed into Agrobacterium tumefaciensGV3101 and cultured on LB agar plates containing 50 μg/ml kanamycin. When the optical density at 600 nm (OD600) reached 0.8, bacterial cells were harvested and resuspended in an infiltration buffer containing 10 mM MgCl₂, 10 mM MES (pH 5.6), and 100 μM acetosyringone. Equal volumes of paired bacterial suspensions were mixed and co-infiltrated into N. benthamiana leaves using a needleless syringe, with distinct regions on the same leaf allocated to different treatment groups to avoid signal overlap. At 36–48 h postinfiltration, 1 mM D-luciferin potassium salt working solution was injected into the corresponding leaf areas, and luminescence signals were immediately acquired using a chemiluminescence imaging system, where signal intensity reflected the strength of PPI. Four LUC combinations were included in each leaf assay. The upper-left position contained nLuc-target + cLuc-design, representing the interaction between the target protein and the designed protein. The upper-right position contained nLuc-T + cLuc-p53 as a positive control. The lower-left position contained nLuc-target + cLuc, and the lower-right position contained nLuc + cLuc-design, serving as negative controls (Supplementary Fig. S11). Flag(his)/GST pull-down assay for evaluating physical interaction between target protein and designed protein​ First, 20 μl of Flag(His)-conjugated magnetic beads were washed and equilibrated with Tris-buffered saline (TBS) buffer. Approximately 50 μg of the target protein tagged with Flag (or His) was then added to the beads. The mixture was incubated at 25°C with constant shaking at 1200 rpm for 1 h to allow capture of the target protein by the beads. After incubation, the beads were extensively washed with TBS buffer to remove unbound proteins. Subsequently, 100 μl of the GST-tagged designed protein to be validated was introduced into the system, and incubation was continued under the same conditions for 1.5 h. Following this incubation, the beads were washed repeatedly with TBS, and the tube was changed to minimize nonspecific binding. Next, 30 μl of 1 mg/ml 3 × Flag peptide was added for competitive elution (for bait proteins with a His tag, direct sodium dodecyl sulfate (SDS) elution can be used instead). The eluted products were mixed with 5× SDS loading buffer and heated at 98°C for 10 min. Finally, Western blot analysis was performed using anti-Flag (or anti-His) and anti-GST specific antibodies. By comparing the co-precipitation of the target protein in the experimental group (containing both the target protein and the designed protein) with that in the control group (containing the empty control and the designed protein), the physical interaction between the designed protein and the target protein was comprehensively evaluated. Key Points Surface-Centric Graph Construction. Integrates sequence, structure, and physicochemical data directly onto the protein’s 3D surface (using the MaSIF pipeline). This inherently excludes buried residues, addressing label imbalance and improving potential hotspot signal quality by focusing only on biologically relevant interaction zones. Surface-Centric Graph Construction. Integrates sequence, structure, and physicochemical data directly onto the protein’s 3D surface (using the MaSIF pipeline). This inherently excludes buried residues, addressing label imbalance and improving potential hotspot signal quality by focusing only on biologically relevant interaction zones. Dynamic Graph CNN (DGCNN) for Local Patterns. Applies a three-layer DGCNN that recomputes edges at each layer, allowing the model to dynamically learn relationships based on evolving feature representations rather than static distances. Dynamic Graph CNN (DGCNN) for Local Patterns. Applies a three-layer DGCNN that recomputes edges at each layer, allowing the model to dynamically learn relationships based on evolving feature representations rather than static distances. Improved hotspot prediction performance. Surf2Spot consistently outperforms current state-of-the-art methods on independent benchmarks, achieving superior F1-score and AUPRC. Improved hotspot prediction performance. Surf2Spot consistently outperforms current state-of-the-art methods on independent benchmarks, achieving superior F1-score and AUPRC. Acceleration of generative binder design. Domain restriction halved computational cost, while Surf2Spot hotspot guidance improved binder success rates in a target-specific manner. These findings support hotspot prediction as a promising design enhancer, contingent on target compatibility. Acceleration of generative binder design. Domain restriction halved computational cost, while Surf2Spot hotspot guidance improved binder success rates in a target-specific manner. These findings support hotspot prediction as a promising design enhancer, contingent on target compatibility. Experimental validation. Designed binders and nanobodies targeting NbPDS1 and VdPDA1 were validated for target binding via luciferase complementation and pull-down assays, confirming that Surf2Spot-guided strategies enhance the success rate of functional binder generation. Experimental validation. Designed binders and nanobodies targeting NbPDS1 and VdPDA1 were validated for target binding via luciferase complementation and pull-down assays, confirming that Surf2Spot-guided strategies enhance the success rate of functional binder generation. Supplementary Material Contributor Information Anwen Zhao, State Key Laboratory of Maize Bio-breeding, National Maize Improvement Center, Frontiers Science Center for Molecular Design Breeding, College of Agronomy and Biotechnology, China Agricultural University, Beijing 100094, China. Yu Li, State Key Laboratory of Maize Bio-breeding, National Maize Improvement Center, Frontiers Science Center for Molecular Design Breeding, College of Agronomy and Biotechnology, China Agricultural University, Beijing 100094, China. Zixuan Wang, Molbreeding Biotechnology Co., Ltd, Shijiazhuang, Hebei Province 051430, China. Tianhao Wu, State Key Laboratory of Maize Bio-breeding, National Maize Improvement Center, Frontiers Science Center for Molecular Design Breeding, College of Agronomy and Biotechnology, China Agricultural University, Beijing 100094, China. Ran Fu, State Key Laboratory of Maize Bio-breeding, National Maize Improvement Center, Frontiers Science Center for Molecular Design Breeding, College of Agronomy and Biotechnology, China Agricultural University, Beijing 100094, China. Yuxuan Lou, State Key Laboratory of Maize Bio-breeding, National Maize Improvement Center, Frontiers Science Center for Molecular Design Breeding, College of Agronomy and Biotechnology, China Agricultural University, Beijing 100094, China. Yanfen Xu, Molbreeding Biotechnology Co., Ltd, Shijiazhuang, Hebei Province 051430, China. Xuliang Han, Molbreeding Biotechnology Co., Ltd, Shijiazhuang, Hebei Province 051430, China. Wenying Wang, State Key Laboratory of Maize Bio-breeding, National Maize Improvement Center, Frontiers Science Center for Molecular Design Breeding, College of Agronomy and Biotechnology, China Agricultural University, Beijing 100094, China. Jun Yan, State Key Laboratory of Maize Bio-breeding, National Maize Improvement Center, Frontiers Science Center for Molecular Design Breeding, College of Agronomy and Biotechnology, China Agricultural University, Beijing 100094, China. Xiangfeng Wang, State Key Laboratory of Maize Bio-breeding, National Maize Improvement Center, Frontiers Science Center for Molecular Design Breeding, College of Agronomy and Biotechnology, China Agricultural University, Beijing 100094, China. Author contributions Xiangfeng Wang (Conceptualization, Investigation, Supervision, Writing—review & editing), Anwen Zhao (Conceptualization, Biological information methodology, Investigation, Visualization, Writing—original draft, Writing—review & editing), Yu Li (Molecular biological methodology, Visualization), Zixuan Wang (Molecular biological methodology, Visualization), Tianhao Wu (Supervision), Ran Fu (Supervision), Yuxuan Lou (Supervision), Yanfen Xu (Supervision) Xuliang Han (Supervision), Wenying Wang (Supervision), and Jun Yan (Supervision, Writing—review & editing) Funding This research was supported by multiple funding sources. We gratefully acknowledge the financial support from: Beijing Rural Revitalization Agricultural Science and Technology Project [NY2601490000], Major Science and Technology Special Project of Liaoning Province [2025JH1/11700012], Pinduoduo-China Agricultural University Research Fund [PC2024A02002], and Chinese Universities Scientific Fund [2026TC100/2026TC101]. Data availability The predicted structures of NbPDS1 and VdPDA1 can be accessed at the AlphaFold Protein Structure Database under accession codes B1NYI4 and G2XC45, respectively. Protein structures from the PDB used for training data construction are available at https://huggingface.co/datasets/anwzhao/Surf2Spot_raw_data. Code availability The Surf2Spot source code is freely available for academic use at https://github.com/AnwZhao/Surf2Spot. References Associated Data Supplementary Materials Data Availability Statement The predicted structures of NbPDS1 and VdPDA1 can be accessed at the AlphaFold Protein Structure Database under accession codes B1NYI4 and G2XC45, respectively. Protein structures from the PDB used for training data construction are available at https://huggingface.co/datasets/anwzhao/Surf2Spot_raw_data.