Introduction

Protein-protein interactions (PPIs) are fundamental to understanding vital cellular processes, such as enzyme regulation and signaling. The ability to predict which proteins form complexes with each other is crucial for biomedical research, as PPIs are directly involved in many diseases and are strategic targets for drug development.

Antibodies, or immunoglobulins, are proteins produced by the immune system to identify and neutralize pathogens such as viruses and bacteria. Each antibody consists of two light chains and two heavy chains, forming two main functional regions: Fab (fragment antigen-binding) and FC (fragment crystallizable). The Fab region contains the antigen-binding site, responsible for the specific recognition of antigenic molecules. When an antibody binds to an antigen ant, it can directly neutralize it or mark the pathogen for elimination by other components of the immune system [1,2]. In the case of a virus like SARS-CoV-2, responsible for the COVID-19 pandemic, the virus uses its spike protein (S) to bind to receptors on host cells, triggering infection. The spike protein is composed of two main subunits, S1 and S2, with the receptor-binding domain (RBD) located within the S1 subunit. Antibodies can bind to the RBD of the spike protein, thus preventing the virus from attaching to the ACE2 receptor on human cells, blocking viral entry and thereby neutralizing the infection [3].

The area or interface between the two proteins (an example is shown in Fig 1 for a spike RBD-antibody pair) is the contact point where their molecular surfaces meet and interact, forming a complex. These interaction areas consist of amino acid residues that can establish various types of bonds (such as hydrogen bonds, hydrophobic interactions, ionic bonds, or Van der Waals forces), therefore determining the strength and stability of the complex [4]. Understanding the chemical structure of interfaces involves examining the types of interactions and the chemical properties of the involved protein surfaces, such as the distribution of electric charges, polarity, and the presence of reactive residues. In parallel, the geometric aspect concerns the three-dimensional shape of the interface, its complementarity and its accessibility. Indeed, the two surfaces must fit together optimally for the interaction to be effective; if the geometry is not favorable, i.e., the two surfaces are not shape-complementary, the interaction will be weak or non-existent [5].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Structural context for RBD–antibody binding.

Left: SARS-CoV-2 spike protein (PDB 7OAN), shown as a cartoon, and schematic of an antibody structure. Right: the spike receptor-binding domain (RBD) bound by a representative Fab (PDB 6W41); heavy (H) and light (L) chains are labeled. The cartoon views emphasize the antigen-binding interface on the RBD and the Fab paratope formed by the CDR loops, providing structural context for our 2D inputs (grayscale surface images) and the per-pixel descriptor maps used throughout the study. Created with BioRender.com.

https://doi.org/10.1371/journal.pone.0335270.g001

Over the years, several approaches, both experimental and computational, have been developed to study and characterize these types of interactions. In the early 1980s, Kuntz et al. [6] pioneered a geometry-based approach to analyze biomolecular interactions, using geometric alignment of rigid-body representations for proteins and ligands (approximated by spheres) to identify compatible binding sites and model protein-ligand interactions based on spatial and molecular compatibility. Jones and Thornton [7] used a combined scoring system focused on interface residues to predict protein-protein interaction sites. Later, Lee et al. [8] refined this by classifying the binding sites according to the geomorphic features of the Earth (cave, crater, canyon, plain, and valley).

Computational approaches have gained ground thanks to the increase in computing power and the availability of large datasets. Molecular docking, sequence analysis, and protein structure modeling have become key tools for predicting and analyzing interactions. Furthermore, the application of machine learning methods has opened new frontiers in PPI prediction, allowing for more accurate and sophisticated models. The emergence of geometric deep learning, which, unlike traditional deep learning (focused on structured data like grids or sequences), can handle complex non-Euclidean data structures, including graphs and surfaces, was a significant development. However, many of these methods primarily rely on either sequence data or structural abstractions, often neglecting detailed geometric representations of binding interfaces. A notable mention in PPI prediction is the application of Siamese neural networks. Introduced in the 1990s for signature verification [9], Siamese architectures have become integral to tasks like face recognition, image retrieval, and object tracking, particularly after their successful application in one-shot learning by Koch et al. [10]. In PPI prediction, Siamese networks have been adapted to model protein sequence relationships and structural information to capture subtle but essential differences between interacting protein pairs.

Several models have been proposed using protein sequences as inputs, or features engineered thereof. Chen et al. implemented a Siamese network model using pairs of protein sequences to capture mutation effects on binding affinity. This model outperformed traditional sequence-based models such as support vector machines (SVM), k-nearest neighbors (kNN), and random forest by providing binary PPI predictions along with multiclass interaction type classification and binding affinity estimation [11]. Yang et al. further improved Siamese networks for PPI prediction by integrating a transfer learning module, enhancing performance in human-virus PPI predictions, including SARS-CoV-2, by combining pre-acquired evolutionary sequence features with a Siamese CNN and a multilayer perceptron (MLP) [12]. Mahapatra et al. developed a hybrid classifier combining a Siamese network with a light gradient boosting machine (LGBM) algorithm. Their model leverages descriptors like pseudo-amino acid composition (PseAAC) and conjoint triad (CT) to transform variable-length protein sequences into fixed-length numerical arrays. This approach enabled accurate PPI network reproduction, as well as high-performance predictions across various metrics [13]. Wu et al. introduced a Siamese Pyramidal network inspired by feature pyramid networks (FPNs) [14], incorporating an attention module to account for protein folding and self-binding, optimizing prediction accuracy by analyzing primary sequences and spatial conformations [15]. Chen et al.’s 2022 model leveraged natural language processing-inspired encoding to convert amino acids into vectors and predict binary PPI classifications [16]. This model applied a dissimilarity threshold to identify negative samples, ensuring non-homologous protein comparisons by removing sequences with more than 40% similarity, as introduced by Eid et al. [17]. Madan et al. expanded on these methods by integrating Elnaggar et al.’s ProtBERT [18] deep sequence embedding model into a Siamese network architecture. Their model provides binary classification and multiclass predictions for interaction types, with binding affinity scores and explainability features using integrated gradients to track the influence of individual amino acids on the model predictions [19].

A notable limitation of these models is the lack of local spatial and structural information; as mentioned above, the geometrical compatibility between protein pairs is an important driving force for binding.

Geometric deep learning approaches for PPI prediction have emerged that take into account the structural information of proteins. Gainza et al. developed the MaSIF framework, applying geometric deep learning to molecular surfaces and hypothesizing that proteins with similar interaction functions may share ’fingerprint’ patterns on their surfaces, regardless of their evolutionary background. MaSIF addresses three tasks: prediction of protein-ligand binding sites, identification of PPI sites, and scanning of protein surfaces to predict PPI complexes [20,21]. Later, the authors introduced a framework for real-time protein surface learning with efficient geometric convolutional layers, followed by a de novo protein interaction design model that generates molecular surface descriptors to guide interaction predictions [22]. Dai and Bailey-Kellogg introduced PInet, a neural network model based on point clouds encoding structural information of protein pairs. This model captures both geometric and physicochemical complementarity between protein surfaces, providing information on which regions mediate interactions [23]. Das and Chakrabarti also contributed with a machine learning classifier, built on support vector machines, aimed at differentiating native from non-native complexes [24]. DockNet, a Siamese graph neural network model, addresses limitations in traditional docking methods, which often struggle with protein flexibility. DockNet predicts contact residues at the protein interfaces by using graph representations where nodes represent amino acid residues, and edges capture chemical or spatial relationships, providing residue-level predictions that incorporate flexibility [25]. Recent developments include the PeSTo framework, a parameter-free approach to geometric deep learning for the prediction of binding interfaces, which avoids the reliance on predefined thresholds or heuristics by leveraging geometric insights to improve prediction accuracy [26]. The EGGNet framework extends geometric learning to various molecule types interacting with protein targets, accommodating small molecules, synthetic peptides, and natural proteins, thereby expanding the applicability of the model to diverse molecular interactions [27].

In this work, we propose an AI-based approach using images within a Siamese neural network, specifically designed to predict SARS-CoV-2 spike-targeting antibodies. Understanding these interactions is crucial for COVID-19 therapeutic and vaccine development, and our model is tailored for this purpose. To achieve this, we curated a dataset containing both spike protein binders and non-binders, ensuring that the model distinguishes between interacting and non-interacting pairs. Unlike existing approaches that focus solely on sequence or structural abstraction, our method uniquely combines precise spatial features of binding interfaces. Our model exploits a representation of protein interfaces in the form of depth maps and descriptors that capture the crucial geometric features of the protein surface. The network, a deep learning architecture specifically designed to learn similarity relations between input pairs, is optimized to discriminate between proteins that form a complex and those that do not. This dual-layered approach improves predictive power and reduces the risk of bias. The core of the network training is the use of contrastive loss, an optimization technique specific to Siamese networks, which encourages the network to learn spatial representations where protein pairs that form complexes are close to each other in the feature space, while those that do not are far apart. Our high test accuracy supports that the model effectively captures meaningful patterns relevant to antibody-spike interactions rather than memorizing known structures. While broader applications may be explored in future research, this study deliberately focuses on COVID-19 spike-targeting antibodies, with the chosen dataset being specifically designed for this scope. The proposed framework represents a valuable step forward in leveraging deep learning for targeted biomedical applications, offering potential insights into SARS-CoV-2 therapeutic strategies.

Materials and methods

Geometrical descriptors

Differential geometry descriptors have been mapped on the proteins’ depth maps to retrieve their shape information. Here, we use a set of geometric descriptors that have been successfully applied in previous research regarding 3D human face analysis [40] and protein shape complementarity [41,42]. Specifically, eight descriptors were selected to characterize the geometry and topology of the protein surfaces (described in [40]): mean curvature (denoted as H), principal curvatures (k₁ and k₂), Gaussian curvature (K), shape index (S), curvedness (C), the sine of the third coefficient of the second fundamental form (sing) and a descriptor that highlights symmetry properties inspired by the second coefficient of the first fundamental form called F_den2. These descriptors depend on the surface derivatives, enabling them to effectively illustrate alterations in surface curvature, surface depressions and peaks, and symmetry aspects of the surface itself.

The six primary descriptors arise from the first and second fundamental forms:

(1)(2)

where E,F,G,e,f,g are the coefficients defined by:

(3)(4)(5)(6)(7)(8)

where h denotes a differentiable function representing the three-dimensional surface, and h_x,y,xx,yy,xy are the surface derivatives.

Curvatures are used to measure how a surface bends. The principal curvatures k₁ and k₂ are the roots of the following quadratic equation:

(9)

Thus, k₁ and k₂ are given by:

(10)(11)

where K and H represent the Gaussian and mean curvatures, respectively, defined as follows:

(12)(13)

The shape index S, which characterizes the surface shape, and the curvedness index C, providing a measure of the curvature, were defined by Koenderink and van Doorn [43] as:

(14)(15)

The descriptor F_den2 is a derived descriptor that was formulated as [40]:

(16)

The descriptors have been evaluated point by point on the depth maps. In some cases (k₁, k₂, g, H, S), their mean- or median-filtered versions have been employed in place of the original descriptors for smoothing purposes, as described in [40].

Siamese neural network and training objective

Siamese neural networks learn to compare two inputs by mapping them into a shared feature space with weight-tied subnetworks and then evaluating a similarity measure [44]. In our setting, each input image x (RBD or Fab) may contain an image channel and, optionally, descriptor channels. We therefore employ two parallel modality towers per branch: an image tower and a descriptor tower. The image tower consists of three convolutional blocks:

• 
• 
• 

A small fully connected head follows: , with the output L₂-normalized to yield (d = 128). The descriptor tower has the same layout except for the first convolution , producing z^desc. LeakyReLU is used throughout to avoid zero-gradient saturation and improve early optimization. When only one modality is present, we use its embedding directly. When both are present, we apply concat fusion: passes through a small fusion multi-layer perceptron (MLP, ) and the result is L₂-normalized to obtain the fused embedding z. A paired forward pass encodes into with shared weights across branches. Similarity is computed via cosine similarity:

(17)

with cosine distance . Training uses a cosine contrastive loss, defined as:

(18)

where indicates binders (1) versus non-binders (0), and m is a margin (set to 1.0). The objective pulls embeddings of binding pairs together () while pushing non-binding pairs at least m apart. A schematic representation of the architecture is shown in Fig 2.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Overview of the Siamese two-tower architecture used.

For each item (Fab and RBD), the raw image (top/third rows) and the z-scored descriptor stack (second/fourth rows) are encoded by CNNs. The image and descriptor embeddings are fused (, implemented as a small MLP) to produce a single representation. The two fused embeddings are compared with cosine similarity and converted to a distance metric d = 1−cosine; training minimizes a contrastive loss with margin m = 1.0, and the operating threshold thr is chosen on the validation ROC. Channel counts are annotated beneath blocks; aqua tones denote Fab branches and purple tones denote RBD branches.

https://doi.org/10.1371/journal.pone.0335270.g002

Training was performed with the Adam optimizer (learning rate 1e⁻⁴) and Xavier initialization. Unless otherwise noted, inputs are depth maps; two- or more-channel variants concatenate one or more geometric descriptors with the depth map. For analysis and calibration, we also expose per-modality embeddings to compute image-only and descriptor-only distances; at validation, when both modalities are available, we apply late fusion by linearly blending these distances with a weight chosen to maximize ROC–AUC. PyTorch was used for implementation and training [45–47]. A class-balanced sampler ensured equal representation of binders and non-binders in each batch. For evaluation, cosine distances were converted into binary classifications using thresholds selected on the validation set, as described below.

Dataset splitting and evaluation

Homology-aware data splits.

As a pairwise learning framework trained on images was employed, our approach does not rely on sequence information; instead, it leverages geometric and spatial features extracted from depth maps and geometric descriptors of antigen-antibody interfaces. As a result, sequence homology does not directly influence the learning process, since the model does not receive sequence data as input. Nonetheless, we adopted a homology-aware split to reduce potential information leakage arising from highly similar antibodies across folds. We implemented this at the CDR-H3 level, as H3 is the most diverse and paratope-defining region. Heavy-chain CDRs were extracted from FASTA sequences using ANARCI [48] with IMGT numbering, with a robust fallback to index-based slicing when region labels were unavailable. For homology grouping, we retained CDR-H3 sequences and wrote per-antibody H3 FASTAs. We clustered H3 sequences with CD-HIT [49] at 90% identity and 80% coverage, yielding H3 clusters that served as the group units for splitting. This enforces that antibodies sharing ≥90% H3 identity are assigned to the same split. To reduce label imbalance before splitting, we undersampled the global majority label (non-binding antibodies) at the H3-group level while preserving H3 diversity. Concretely, we selected one representative H3 per majority group, meta-clustered these representatives with CD-HIT at 95% identity, then greedily retained at most one (and, if needed, additional) group per meta-cluster until the target majority proportion of 50% was reached. All minority-label groups were kept. We performed a stratified split of H3 clusters into 70%/20%/10% train/validation/test sets, optimizing (i) total weight per split, (ii) per-split label proportions within a tolerance (default ) relative to the global label distribution, and (iii) a minimum number of groups in validation and test (defaults 30/30). The procedure iterates many random orders of groups and selects the assignment minimizing a size-and-label deviation objective, subject to the constraints. After the split, we quantified potential leakage by computing, for each validation/test H3, the maximum identity to any training H3 using a simple global aligner (match 1, mismatch 0, gap 0). The median of these nearest-neighbor identities was ∼60% for both validation and test, indicating limited homology to the training examples. Grouping by CDR-H3 identity captures the principal source of paratope similarity, thereby limiting near-duplicate antibodies from straddling splits. The stratified, group-aware optimization preserves overall label balance while enforcing this homology constraint.

Positive/negative pairs.

Positive pairs were defined as co-crystallized RBD–Fab complexes (binders, label 1), whereas negative pairs were obtained by pairing RBD structures with non-cognate Fabs (non-binders, label 0). For each pair, we extracted two interface representations (RBD interface and Fab interface), rendered as depth maps and, where applicable, concatenated with one or more geometric descriptors.

Evaluation.

Model selection and calibration were performed on the validation set. Cosine distances were converted to binary predictions using Youden’s statistic, defined as:

(19)

where TPR is the true positive rate (sensitivity) and FPR is the false positive rate (1 - specificity). The threshold maximizing on the validation set was selected, then fixed for the test set to ensure unbiased evaluation. We report mean ± standard deviation of accuracy, precision, recall, F1, and ROC–AUC across five independent seeds.

An overview of the adopted pipeline is provided in Fig 3.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. End-to-end pipeline for training and evaluation.

Schematic overview of the modular framework implemented in this work. Each block corresponds to a distinct stage of the workflow, with the internal labels reflecting the actual classes and functions implemented in the code. From left to right: (i) CLI and Data Preparation: command-line arguments specify dataset root and index files (index_train/val/test.csv), with optional interface augmentations. (ii) Preprocessing: geometric descriptors are z-scored, either eagerly or on-the-fly, to standardize input channels. (iii) Dataset Construction: inputs are wrapped into dataset classes that enable flexible handling of preloaded and lazy modes. (iv) Dataloaders: balanced train/validation/test loaders are created to enforce class balance across epochs. (v) Model: a Siamese neural network encodes Fab and RBD interfaces into a shared embedding space, combining image and descriptor representations via fusion. Initialization and training rely on weights_init and ContrastiveLoss. (vi) Training: optimization is performed with Adam, with warmup, ReduceLROnPlateau scheduling, and early stopping on validation AUC. (vii) Evaluation: cosine distances are converted to binary predictions using Youden’s statistic, with final metrics (accuracy, precision, recall, F1, ROC–AUC) reported on the test set.

https://doi.org/10.1371/journal.pone.0335270.g003

Results and discussion

Ablations results

Loss formulation.

We compared the standard Euclidean contrastive loss with the cosine-based variant. As shown in Table 1, the Euclidean formulation led to substantially lower accuracy and AUC, with higher variance across seeds. In contrast, the cosine-based loss consistently achieved more stable convergence and improved discrimination between binding and non-binding interfaces. This indicates that cosine distance is a better match for the normalized embedding space learned by our Siamese architecture.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Ablation results comparing Euclidean and cosine contrastive loss functions. Metrics are reported on the test set as mean ± standard deviation across five seeds. Best values per metric are highlighted in bold.

https://doi.org/10.1371/journal.pone.0335270.t001

Fusion strategy.

We next evaluated the effect of feature fusion when combining depth maps with geometric descriptors. The tests were conducted using all descriptors as input, without data augmentation. The results of the ablation are reported in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Ablation results comparing MLP fusion with simple concatenation. Metrics are reported on the test set as mean ± standard deviation across five seeds. Best values per metric are highlighted in bold.

https://doi.org/10.1371/journal.pone.0335270.t002

Without the fusion MLP (simple concatenation), the model underperformed, with test accuracy ∼0.76 and AUC ∼0.88 across seeds. Conversely, including the fusion MLP substantially improved performance, raising accuracy to ∼0.91 and AUC to ∼0.95. Therefore, the fusion model was retained in the final model.

Data augmentation.

Additional flips and rotations reduced validation AUC in multi-seed runs. Therefore, augmentations were disabled in the final model to prevent degradation of generalization. The results of the ablation comparing the image-only baseline with and without data augmentation, over five independent seeds, are shown in Table 3.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Ablation results comparing the performance on the test set with and without including data augmentations during training. The metrics are reported as mean ± standard deviation across five independent seeds. The best value for each metric is highlighted in boldface.

https://doi.org/10.1371/journal.pone.0335270.t003

Surface preprocessing.

We compared models trained on normal versus smoothed interface representations, both in the image-only setting without augmentation. Results are summarized in Table 4. Both settings performed comparably, with no consistent advantage from smoothing.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Ablation results comparing normal versus smoothed interface representations (image-only, no augmentation). Metrics are reported as mean ± standard deviation across five independent seeds. The best value for each metric is highlighted in boldface.

https://doi.org/10.1371/journal.pone.0335270.t004

Overall, smoothing did not yield a consistent improvement over normal surfaces, so the final model retained the normal interface representation.

Input modality and descriptor contribution.

We next investigated whether geometric surface descriptors provided complementary information to the shape-based depth map baseline. The effect of the addition of each descriptor was evaluated individually over five independent seeds (Table 5). Metrics are reported on the test set (mean ± std across five seeds). Validation AUC is also included, as it served as the selection criterion in the greedy procedure.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Ablation results comparing input modalities on the test set. Each model was trained without data augmentation using normal (non-smoothed) surfaces. Results are reported as mean ± standard deviation over five independent seeds. Best values per metric are highlighted in bold.

https://doi.org/10.1371/journal.pone.0335270.t005

While the image-only model achieved strong baseline performance (AUC ), several descriptors further improved the predictive accuracy. In particular, curvedness index (C), mean curvature (H) principal curvatures (k_1/2), and sing consistently enhanced the AUC, whereas other descriptors such as F_den2 and S had less pronounced or inconsistent effects. Notably, combining all descriptors jointly did not yield a substantial gain over individual descriptors and, in some cases, slightly reduced performance, suggesting that redundant or noisy channels may offset the benefits of informative ones.

To further isolate the most informative subset of descriptors, we performed a greedy forward feature selection procedure, starting from the image-only baseline and sequentially adding the descriptor that maximized validation AUC at each step. The first addition, principal curvature (k₂), improved mean validation AUC from 0.968 to 0.979. At the default selection threshold (), the procedure terminated after this step, identifying the image+k₂ configuration as the optimal one. When the threshold was relaxed to , principal curvature k₁ was added at step 2, yielding a marginal further increase to 0.980. Given the minimal improvement beyond k₂, we retained image+k₂ as the final descriptor configuration for subsequent experiments.

Baseline comparison

Table 6 reports the results of our image-only CNN with sequence-based baselines.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Comparison with sequence-only baselines. Metrics are calculated on the test set, as mean ± standard deviation over five seeds. Best values per metric are highlighted in boldface.

https://doi.org/10.1371/journal.pone.0335270.t006

The best sequence-based Siamese baseline (all CDRs with ESM-3 embeddings) achieved AUC (accuracy ). Whole-chain and H3-only variants performed worse, with AUC values of and , respectively. In contrast, our structure-based Siamese CNN consistently achieved an AUC of at least . These results highlight the added value of surface geometry: explicit 3D surface representations provide substantially stronger discrimination of binding versus non-binding interfaces.

Final model performance

During training, we monitored loss and accuracy metrics for the final selected configuration (image + k₂). Fig 4 shows the trends of training and validation loss (solid lines) and accuracy (dashed lines), using the seed that achieved the highest validation ROC AUC. Loss curves display the expected behavior, with training loss lower than validation loss. As accuracy is threshold-dependent, while loss reflects the continuous margin between positive and negative pairs, the adaptive ROC-based threshold used during validation favored higher classification accuracy, despite a slightly larger validation loss, resulting in consistently high validation accuracy.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Training and validation curves for the final selected model (image + k₂ setting, best validation seed).

Solid lines show training and validation losses (left axis), while dashed lines show accuracies (right axis). As expected, training loss remains lower than validation loss. Validation accuracy is slightly higher than training accuracy across epochs; this arises because accuracy is computed after applying an adaptive ROC-based threshold, whereas loss reflects continuous margins between positive and negative pairs. Training was extended to 100 epochs with early stopping (patience = 20), ensuring convergence and preventing overfitting. The early stopping epoch (36) is indicated with a grey dashed vertical line.

https://doi.org/10.1371/journal.pone.0335270.g004

To further assess how the model distinguishes between binding and non-binding pairs, we analyzed the distribution of distances between embeddings on the test set and compared them with the corresponding Receiver Operating Characteristic (ROC) curve. This analysis, shown in Fig 5, complements the training and validation trends by showing the degree of separation achieved in the learned representation space and the resulting classification performance across thresholds.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Test-set separation of binding and non-binding pairs.

(a) Distribution of cosine distances between embeddings of binding (purple) and non-binding (teal) pairs in the test set. Binder pairs show lower distances, consistent with geometric complementarity. The vertical dashed line indicates the decision threshold optimized on the validation set. (b) Receiver Operating Characteristic (ROC) curve for the same test set, with area under the curve (AUC) reported. The curve demonstrates strong discrimination between classes, with an AUC over 0.95.

https://doi.org/10.1371/journal.pone.0335270.g005

Taken together, these results demonstrate that the proposed Siamese CNN consistently outperforms sequence-based baselines and benefits from the selected architectural and input design choices. The ablation experiments confirmed that cosine loss and the fusion module were beneficial for robust performance, while smoothing and heavy augmentation were not. Among descriptors, k₂ emerged as the most informative complement to depth maps, providing a simple yet effective final configuration. The final model achieved clear separation of embedding distances and good ROC performance on the held-out test set. These findings underscore the effectiveness of geometry-aware representations.

Conclusion

This work addressed the challenge of classifying RBD–antibody interfaces as binding or non-binding pairs using geometric and image-based deep learning. We proposed a Siamese convolutional neural network trained on depth-map renderings of protein interfaces, optionally combined with geometric surface descriptors. The model was optimized with a cosine contrastive loss and evaluated on homology-aware splits to prevent information leakage across folds. Our ablation experiments clarified the design choices underlying the final model. Data augmentation with flips and rotations reduced generalization, and surface smoothing was not beneficial. In contrast, the fusion module significantly improved cross-modal integration, and cosine contrastive loss aligned well with normalized embeddings. Among individual descriptors, the mean principal curvature k₂ emerged as the most effective, representing the strongest improvement over the baseline. Curvature information is biologically intuitive in this context, as k₂ encodes local convexity and concavity patterns that are relevant for complementarity. Other descriptors provided moderate improvements but fell short of k₂. For example, the Shape Index S captures qualitative surface geometry (peaks, ridges, saddles, valleys) that is known to influence binding affinity, yet it performed slightly worse than k₂ in our benchmarks. Similarly, K and F_den2 emphasize global surface characteristics or isolated critical points, which may be less discriminative for antibody binding sites. Principal curvature k₁ was also evaluated and, while informative, added only marginal benefit beyond k₂ under a greedy selection procedure. These findings underscore the importance of choosing descriptors aligned with the biological context of protein recognition. Local curvature features complement the global depth-map representation and provide a meaningful signal. This suggests that future work could extend the framework by incorporating descriptors of physicochemical properties (e.g., electrostatics or hydrophobicity) alongside geometric cues, thereby improving generalizability and robustness. Compared to a sequence-only baseline using ESM-3 embeddings, our structure-based approach achieved substantially higher performance (ROC–AUC ∼0.95 vs ∼0.76), underscoring the added value of 3D geometric information. Our results highlight the potential of geometric deep learning for practical antibody discovery. Accurate discrimination between binding and non-binding antigen–antibody interfaces can accelerate therapeutic antibody development by prioritizing promising Fab candidates, and can inform vaccine design by identifying stable immunogenic epitopes. These findings serve as a proof-of-concept for the translational potential of integrating artificial intelligence with structural bioinformatics, paving the way for broader applications in biomedical and drug discovery research. While broader applications are possible, this study’s primary focus was to improve predictive performance within COVID-19-related antibody interactions. The good test performance confirms that the model effectively generalizes within this focused dataset, aligning with the intentional design choice to target this specific biological challenge. Additionally, the method relies on experimentally solved antigen-antibody complexes, where binding interfaces are well-defined. While this ensures precise characterization of binding site geometry, it does not explicitly account for protein flexibility or binding-induced conformational changes. Future work could explore integrating docking simulations or flexible protein modeling to extend this method to unbound structures. Incorporating these strategies could improve the model’s applicability to cases where experimental structures are unavailable or incomplete. Finally, increasing data complexity by incorporating additional types of information across multiple channels and using higher-resolution images may refine the model’s ability to distinguish between complex and non-complex pairs.

Acknowledgments

This research was enabled in part by support provided by the Prairies DRI and the Digital Research Alliance of Canada (alliancecan.ca).