AntiBinder: utilizing bidirectional attention and hybrid encoding for precise antibody–antigen interaction prediction

Abstract

Antibodies play a key role in medical diagnostics and therapeutics. Accurately predicting antibody–antigen binding is essential for developing effective treatments. Traditional protein–protein interaction prediction methods often fall short because they do not account for the unique structural and dynamic properties of antibodies and antigens. In this study, we present AntiBinder, a novel predictive model specifically designed to address these challenges. AntiBinder integrates the unique structural and sequence characteristics of antibodies and antigens into its framework and employs a bidirectional cross-attention mechanism to automatically learn the intrinsic mechanisms of antigen–antibody binding, eliminating the need for manual feature engineering. Our comprehensive experiments, which include predicting interactions between known antigens and new antibodies, predicting the binding of previously unseen antigens, and predicting cross-species antigen–antibody interactions, demonstrate that AntiBinder outperforms existing state-of-the-art methods. Notably, AntiBinder excels in predicting interactions with unseen antigens and maintains a reasonable level of predictive capability in challenging cross-species prediction tasks. AntiBinder’s ability to model complex antigen–antibody interactions highlights its potential applications in biomedical research and therapeutic development, including the design of vaccines and antibody therapies for rapidly emerging infectious diseases.

antibody–antigen interaction prediction, cross-attention, multi-dimension feature confusion

Introduction

Antibodies, or immunoglobulins, are vital proteins produced by the immune system that play a central role in identifying and neutralizing pathogens, such as bacteria, viruses, and toxins. Each antibody is highly specific to its corresponding antigen, binding to it through sites in its variable region. These precise interactions between antibodies and antigens are crucial for effective immune defense. Beyond their natural role in the immune system, antibody–antigen interactions (AAI) are also indispensable in medical diagnostics, therapeutics, and research. For example, the biopharmaceutical industry has developed monoclonal antibodies, like those used to treat COVID-19 and its variants [1, 2], as well as PD-1/L1 inhibitor drugs [3–5] that have significantly advanced cancer treatment.

Accurately predicting antibody–antigen binding is critical for the development of effective treatments. Reliable predictions can improve antibody design by ensuring strong protein–protein interactions (PPI). Computational models that predict these interactions can greatly reduce the time and cost associated with experimental methods, thus accelerating progress in medical and research applications. As a result, the advancement of these predictive models has become a key focus in the field of antibody design.

Given that antigens and antibodies are both proteins, it might seem natural to apply predictive methods for PPI to antigen–antibody binding prediction. A brief overview of some PPI methods is available in Supplementary Note 1. It is important to note that most current methods for predicting PPI are not specifically designed for antigen–antibody interactions. Consequently, these methods often fail to accurately predict the binding of antibody–antigen complexes due to several critical factors.

Firstly, the structural complexity and diversity of antibody binding sites, known as paratopes, pose considerable challenges. Paratopes exhibit significant variability in amino acid (AA) sequences and structures, which generic PPI models often inadequately address. Furthermore, antibodies have flexible and dynamic regions, particularly the complementarity-determining regions (CDRs) [6], which undergo conformational changes upon binding to antigens. This inherent flexibility is crucial for functionality but is difficult for standard binding models to capture.

The diverse nature of epitopes—the specific parts of the antigen recognized by antibodies—adds another layer of complexity. Epitopes can be linear or conformational, and their varied structural forms complicate accurate binding predictions using traditional methods. Generic protein structure prediction tools often fail to accurately model the unique structures of antibodies and antigens, leading to significant inaccuracies. They may overlook essential features, such as the Y-shaped configuration of antibodies and the variability within CDRs [7], which are critical for understanding binding interactions.

Additionally, binding is influenced by intricate energetic contributions from various non-covalent interactions, including hydrogen bonds, van der Waals forces, and electrostatic interactions [8]. Generic models frequently oversimplify these interactions, resulting in inaccuracies.

Although several models have been developed specifically for predicting AAI, they still exhibit limitations. For example, AttABseq [9] and DG-affinity [10] are sequence-based approach; DG-affinity utilizes pre-trained language models to extract effective representations but does not consider the 3D structures of antibodies and antigens, which are critical for understanding their interactions. Additionally, these models do not segment sequences or emphasize crucial regions, such as CDRs, which are central to the binding process. AbAgIntPre [11] extracts features from antibodies and antigens separately, but fails to capture the direct interactions between them. This limitation restricts the model’s ability to fully understand the underlying binding mechanisms. DeepAAI [12] extracts global and local interaction features exclusively from antibody and antigen sequence data, while PESI [13] focuses on paratope and epitope residues but neglects non-binding regions of the antibody sequence, missing contextual information that could enhance understanding of interactions. Furthermore, PESI relies heavily on manually curated datasets, which may limit its scalability and performance in diverse real-world applications, particularly without a robust supporting database.

These limitations underscore the need for more advanced models capable of integrating structural data, applying specialized processing to antibody sequences, and precisely capturing the binding rules between antibodies and antigens to improve prediction accuracy.

In this paper, we propose a novel method for predicting AAI, called AntiBinder. To effectively capture the unique features of antigens and antibodies, we have designed specialized encoders for each, ensuring a comprehensive representation that incorporates both sequence and structural characteristics. For structural feature extraction, we employ IgFold for antibodies and ESM-2 for antigens. IgFold, a protein language model specifically trained on antibody data, offers a precise representation of antibody structural nuances, while ESM-2 predicts protein structures from sequence information, making it ideal for diverse antigens. The use of these two pre-trained models allows our method to obtain structural information even when the dataset contains only AA sequences. This ensures that our method remains applicable and achieves high prediction accuracy even when dataset features are limited. The entire process is fully automated, requiring no prior knowledge or manual feature engineering. To automatically learn the intrinsic mechanisms of antigen–antibody binding, we have developed a bidirectional cross-attention mechanism. This mechanism dynamically models the interactions from both the antibody and antigen perspectives, refining its focus on specific epitopes and paratopes. By capturing these interactions in a context-sensitive manner, our approach eliminates the need for manual feature engineering, which is often necessary to account for intricate energetic contributions and conserved binding sites.

We evaluate the predictive accuracy and generalization capability of the AntiBinder model across several aspects: predicting the binding of known antigens with new antibodies, predicting the binding of unseen antigens and antibodies, and predicting cross-species antigen–antibody interactions. These evaluations were performed using four different datasets to rigorously assess our model’s performance against several existing methods.

The results demonstrate that AntiBinder excels in prediction accuracy and shows superior generalization capability. This indicates that AntiBinder effectively models the interactions between antibodies and antigens and learns the underlying rules of their interactions. Furthermore, AntiBinder can predict the binding for unseen antigens, showcasing its robustness and applicability to unfamiliar scenarios. Significantly, our method’s superior performance in cross-species experiments highlights its potential for broad applications in biomedical research and therapeutic development.

Materials and methods

Overview

AntiBinder is designed to determine whether an antibody can bind to an antigen. Its architecture involves several key steps: extracting features from both the antibody and the antigen, finding associations between them, and ultimately using a classifier to determine their binding potential. The framework is illustrated in Fig. 1, where effective extraction and representation of information from antibodies and antigens, along with understanding the essential mechanisms of their binding interactions, are crucial.

Figure 1

AntiBinder’s architecture. (a) Overview of AntiBinder. AntiBinder takes antibody and antigen sequences as input for interaction prediction. AbSeqEmb extracts and fuses the sequence and structural embeddings of the antibody, while AgSeqEmb performs the same function for the antigen. These embeddings are then combined using the BidAttBlock network, which allows the classifier to predict interactions based on the fused features. (b) AbSeqEmb extracts both sequence and structural embeddings from the antibody heavy chain sequence, incorporating structural information via IgFold. It segments the antibody sequence into regions to emphasize their functional roles in binding. Sequence information is processed through AA and region indexing and is integrated with structural embeddings. The concatenation of these features is indicated by the symbol |$\odot $|⁠. (c) AgSeqEmb generates antigen embeddings by extracting both sequence and structural information without region segmentation. AA indexing captures sequence features, while ESM-2 generates a structural embedding. (d) BidAttBlock integrates antibody and antigen embeddings, using two cross-attention mechanisms to elucidate the binding mechanism. The antibody and antigen embeddings encapsulate the combined sequence and structural information for each. In the antibody-to-antigen attention mechanism, the query is derived from the antibody embedding, while the key-value pairs are sourced from the antigen embedding. Conversely, the antigen-to-antibody attention mechanism utilizes the query from the antigen embedding and the key-value pairs from the antibody embedding. (e) The classifier module predicts the final binding outcome, where Ab_Ag embedding is the output from the antibody to antigen attention in the final BidAttBlock, and the Ag_Ab embedding is the output from the antigen to antibody attention in the same block.

In Fig. 1a, the AbSeqEmb module is used to extract effective representations of the antibody, while the AgSeqEmb module extracts effective embeddings of the antigen. These embeddings are then processed through multiple Bidirectional Attention Blocks (BidAttBlock), which are designed to capture potential rules of AAI. The features extracted by these blocks are then subjected to a weighted fusion process, and subsequently, they pass through fully connected layers. Finally, these features enter the classifier to predict the binding between the antibody and the antigen.

This architecture ensures that both sequence and structural information are effectively integrated, providing a comprehensive understanding of the complex dynamics involved in antibody–antigen binding. The BidAttBlock specifically enhances the model’s ability to learn from the intricate interactions and energetic contributions that characterize these molecular processes. By automating feature extraction and focusing on the key aspects of binding mechanisms, AntiBinder aims to deliver highly accurate predictions, improving both research efficiency and practical applications in immunology and biotechnology.

AbSeqEmb module

The AbSeqEmb module, illustrated in Fig. 1b, is designed to extract effective representations of antibodies. Antibodies are Y-shaped proteins composed of two symmetric sets of chains, each consisting of a heavy chain and a light chain [14]. Each chain includes constant regions and a variable region, which contains three CDRs. The antigen-binding sites are located in the variable regions, primarily in the CDRs, especially CDR-H3. The remaining parts of the variable regions, known as framework regions (FRs), are structurally conserved and provide structural support [15].

Given the significant research highlighting the importance of both sequence and structural characteristics of proteins, we simultaneously extract features from both sequence and structure.

First, we use sequence-to-vector (S2V) embeddings to extract sequence features. Given that proteins consist of 20 distinct AA, we convert sequences into discrete vectors using AA indexing, which serves as the most fundamental sequence representation. This process is illustrated in the AA Indexing section of Fig. 1b.

Recognizing the critical role of CDRs in antigen binding, we partition the antibody heavy chain sequence into FRs and CDRs, assigning distinct value indices to each segment to emphasize their functional roles and biological significance. We adopt the Chothia segmentation scheme [16, 17], which partitions the variable regions into three CDRs and four FRs. This scheme involves aligning the antibody sequence with known structures to determine region positions, verified and adjusted using 3D structural data. The segmented FRs and CDRs regions are then assigned distinct numerical indices, forming discrete vectors termed Region Indexing, as illustrated in Fig. 1b.

These discrete vectors from AA indexing and region indexing are padded or truncated and transformed into vector matrices using an embedding layer in the neural network, resulting in |$X_{\text{Ab}{\_}\text{seq}} \in \mathbb{R}^{L_{\text{Ab}} \times D}$|⁠, where |$L_{\text{Ab}}$| is the maximum sequence length and |$D$| is the embedding dimension. We set the lengths of antibody |$L_{\text{Ab}}$| and antigen sequences |$L_{\text{Ag}}$| to 149 and 1024, respectively, covering over 80% of sequence lengths while reducing the model’s parameter count. These vectors are continuously updated during training to capture complex AA relationships and adapt to different datasets.

Next, we use IgFold [18] to extract unique structural embedding information from antibody sequences. IgFold is a computational method that predicts the 3D structures of antibodies based on their AA sequences, utilizing deep learning algorithms to model variable regions crucial for antigen binding. By inputting the antibody sequence into IgFold, we obtain a structural matrix |$X_{\text{Ab}{\_}\text{stru}}^{\prime} \in \mathbb{R}^{L_{\text{IgFold}} \times 64}$|⁠, where |$L_{\text{IgFold}}$| varies with input sequence length. In practical applications, the matrix |$X_{\text{Ab}{\_}\text{stru}}^{\prime}$| has three dimensions: the first dimension represents the batch size, the second dimension |${L_{\text{IgFold}}}$| corresponds to the AA sequence length, and the third dimension denotes the AA embeddings. To emphasize CDR-H3 and manage data volume, we segment the structural matrix into three parts and concatenated them along the first dimension:

$$ \begin{align*} & p1 = X_{\text{Ab}{\_}\text{stru}}^{\prime}[:,:l_{\text{FR1}\sim\text{FR3}},:], \end{align*} $$

$$ \begin{align*} & p2 = \frac{1}{l_{\text{FR1}\sim\text{CDR3}}} \sum_{i=\text{FR1}\sim\text{CDR3}} X_{\text{Ab}{\_}\text{stru}}^{\prime}[:, i,:], \end{align*} $$

$$ \begin{align*} & p3 = X_{\text{Ab}{\_}\text{stru}}^{\prime}[:, l_{\text{FR1}\sim\text{CDR3}}:l_{\text{FR1}\sim\text{FR4}},:], \end{align*} $$

where |$l_{\text{FR1}\sim \text{FR3}}$| represents the sequence length excluding the CDR3 and FR4 regions, |$l_{\text{FR1}\sim \text{CDR3}}$| excludes FR4, and |$l_{\text{FR1}\sim \text{FR4}}$| is the total sequence length. This process, shown as Slicing in Fig. 1b, resulted in a padded or truncated matrix according to the maximum length |$L_{\text{Ab}}$|⁠, yielding |$X_{\text{Ab}{\_}\text{stru}}^{\prime} \in \mathbb{R}^{L_{\text{Ab}} \times 64}$|⁠. To align with the antibody sequence matrix, we applied a linear transformation:

$$ \begin{align*} & X_{\text{Ab}{\_}\text{stru}} = \sigma(X_{\text{Ab}{\_}\text{stru}}^{\prime} W_{\text{Ab}{\_}\text{stru}} + b_{\text{Ab}{\_}\text{stru}}), \end{align*} $$

where |$W_{\text{Ab}{\_}\text{stru}} \in \mathbb{R}^{64 \times D}$| is the weight matrix, |$b_{\text{Ab}{\_}\text{stru}} \in \mathbb{R}^{D}$| is the bias vector, and |$\sigma $| denotes the activation function ReLU. Finally, we integrate the antibody sequence |$X_{\text{Ab}{\_}\text{seq}}$| and structure matrix |$X_{\text{Ab}{\_}\text{stru}}$| to obtain |$X_{\text{Ab}} \in \mathbb{R}^{L_{\text{Ab}} \times D}$|⁠.

The AbSeqEmb module integrates 3D structural information of antibodies using IgFold, and its region indexing design enhances the model’s ability to capture the evolutionary adaptability of antibodies. This approach enables the model to gain a deeper understanding of the dynamic changes and evolutionary mechanisms of antibodies by leveraging both sequence and structural features.

AgSeqEmb module

As depicted in Fig. 1c, the AgSeqEmb module is designed to extract embeddings for antigens. Unlike antibodies, antigens exhibit significant variation, making it impractical to segment them into distinct regions. Instead, we process the discrete vectors of AA indexing through an embedding layer to obtain |$X_{\text{Ag}{\_}\text{seq}} \in \mathbb{R}^{L_{\text{Ag}} \times D}$|⁠.

The antigen sequence is padded or truncated to a maximum length |$L_{\text{Ag}}$| using a specific padding symbol “pad”. This ensures uniformity in the input sequences, facilitating consistent processing within the neural network.

To capture the distinct structural characteristics of antigens, we employ ESM-2 [19], a model proposed by Meta AI. ESM-2 utilizes large-scale language model training to predict protein 3D structures, effectively generating a structural matrix |$X_{\text{Ag}{\_}\text{stru}}^{\prime} \in \mathbb{R}^{L_{\text{Ag}} \times 1280}$|⁠.

To align the structural matrix with the antigen sequence matrix, we apply a linear transformation to |$X_{\text{Ag}{\_}\text{stru}}^{\prime}$|⁠:

$$ \begin{align*} & X_{\text{Ag}{\_}\text{stru}} = \sigma(X_{\text{Ag}{\_}\text{stru}}^{\prime} W_{\text{Ag}{\_}\text{stru}} + b_{\text{Ag}{\_}\text{stru}}), \end{align*} $$

where |$W_{\text{Ag}{\_}\text{stru}} \in \mathbb{R}^{1280 \times D}$| is the weight matrix, |$b_{\text{Ag}{\_}\text{stru}} \in \mathbb{R}^{D}$| is the bias vector, and |$\sigma $| denotes the activation function ReLU.

By integrating the antibody sequence |$X_{\text{Ag}{\_}\text{seq}}$| and the transformed structural matrix |$X_{\text{Ag}{\_}\text{stru}}$|⁠, we obtain the comprehensive antigen representation |$X_{\text{Ag}} \in \mathbb{R}^{L_{\text{Ag}} \times D}$|⁠. This integrated embedding effectively captures both sequence and structural information, providing a robust representation of the antigen for further processing within the AntiBinder framework.

BidAttBlock module

The BidAttBlock module is central to the AntiBinder architecture, utilizing a bidirectional cross-attention mechanism to enhance the model’s understanding of interaction relationships between antibodies and antigens. This approach allows for a dynamic flow of information, thereby capturing the intricate binding interactions essential for accurate predictions.

The BidAttBlock module, as illustrated in Fig. 1d, uses separate sets of weights for antibodies and antigens to facilitate bidirectional information flow. This is achieved by mapping the embeddings of antibodies and antigens into queries, keys, and values through affine transformations. The attention mechanism then computes the relationships between these elements, enabling the model to learn interaction patterns.

$$ \begin{align*} & Q_{Ab} = X_{Ab}W_{Ab}^{Q}, \quad K_{Ab} = X_{Ab}W_{Ab}^{K}, \quad V_{Ab} = X_{Ab}W_{Ab}^{V}, \end{align*} $$

$$ \begin{align*} & Q_{Ag} = X_{Ag}W_{Ag}^{Q}, \quad K_{Ag} = X_{Ag}W_{Ag}^{K}, \quad V_{Ag} = X_{Ag}W_{Ag}^{V}, \end{align*} $$

where |$ W_{Ab}^{Q} \in \mathbb{R}^{D \times d_{k}}, \; W_{Ab}^{K} \in \mathbb{R}^{D \times d_{k}}, \; W_{Ab}^{V} \in \mathbb{R}^{D \times d_{k}}, \; W_{Ag}^{Q} \in \mathbb{R}^{D \times d_{k}}, \; W_{Ag}^{K} \in \mathbb{R}^{D \times d_{k}} \ \text{and} \ W_{Ag}^{V} \in \mathbb{R}^{D \times d_{k}} $| are learnable parameter matrices, and |$Q$|⁠,|$K$|⁠,|$V$| represent queries, keys, and values, respectively.

Antibody to antigen attention mechanism: The specific binding between an antibody’s variable region and an antigen’s epitope is critical for recognition. Using the antibody as the query, the system learns which antigen features are essential for recognition. In other words, the model learns to pinpoint the specific regions of the antigen that the antibody targets, highlighting potential binding epitopes. This interaction is modeled as follows:

$$ \begin{align*} & \text{Attention}_{Ab \rightarrow Ag}(Q_{Ab}, K_{Ag}, V_{Ag}) = \text{softmax} \left( \frac{Q_{Ab} K_{Ag}^{T}}{\sqrt{d_{k}}} \right) V_{Ag}, \end{align*} $$

where softmax is also an activation function, it converts an unnormalized vector into a probability distribution. This mechanism captures the spatial and chemical interactions necessary for understanding and predicting antibody binding sites and affinities.

Antigen to antibody attention mechanism: Conversely, the antigen–antibody cross-attention mechanism models recognition from the antigen’s perspective. Using the antigen as the query, the system identifies key antibody features involved in binding:

$$ \begin{align*} & \text{Attention}_{Ag \rightarrow Ab}(Q_{Ag}, K_{Ab}, V_{Ab}) = \text{softmax} \left( \frac{Q_{Ag} K_{Ab}^{T}}{\sqrt{d_{k}}} \right) V_{Ab}, \end{align*} $$

This bidirectional approach allows information to flow and integrate between the antibody and antigen, simulating their interaction process.

After passing through |$N$| BidAttBlock modules, the weighted representations of the antibody and antigen are iteratively refined. The updated representations after the |$Nth$| module are as follows:

$$ \begin{align*} & \begin{aligned} &X_{Ab{\_}Ag}^{(N)} = \text{BidAttBlock}^{(N)}\Big(X_{Ab{\_}Ag}^{(N-1)}, X_{Ag{\_}Ab}^{(N-1)}\Big), \\ &X_{Ag{\_}Ab}^{(N)} = \text{BidAttBlock}^{(N)}\Big(X_{Ag{\_}Ab}^{(N-1)}, X_{Ab{\_}Ag}^{(N-1)}\Big), \end{aligned} \end{align*} $$

here |$X_{Ag{\_}Ab}^{(N-1)}$| and |$X_{Ab{\_}Ag}^{(N-1)}$| are the representations after integrating the information from each other, generated by the |$(N-1)th$| BidAttBlock module. This iterative process enhances the model’s capability to accurately predict the binding by thoroughly modeling the AAI. Additionally, such a modular design increases the model’s flexibility, helping it adapt to datasets of different scales.

Classifier module

The classifier module of AntiBinder, depicted in Fig. 1e, is responsible for finalizing the prediction of antibody–antigen binding.

To integrate the information from both antibodies and antigens, the dimensionality of their respective representations is reduced using two linear layers. The outputs of these layers are then concatenated:

$$ \begin{align*} & o=\zeta\Bigl(\operatorname{MLP}\left(\left[\alpha \cdot X_{\mathrm{Ag} {\_}\mathrm{Ab}}^{(N)} ; \; X_{\mathrm{Ab} {\_}\mathrm{Ag}}^{(N)}\right]\right)\Bigr), \end{align*} $$

where |$X_{Ag{\_}Ab}^{(N)}$| and |$X_{Ab{\_}Ag}^{(N)}$| represent the weighted representations of the antigen and antibody after passing through |$N$| BidAttBlock modules, and |$ \alpha $| is a learnable parameter that adjusts the importance of the representation |$X_{\mathrm{Ag} {\_}\mathrm{Ab}}$|⁠. The concatenated vector is fed into a multilayer perceptron (MLP) with two stacked linear layers. The first layer applies a ReLU activation function to introduce non-linearity and capture complex patterns. The second layer transforms the result and uses a sigmoid activation function |$ \zeta $| to produce the final output.

During training, the model is optimized to minimize the binary cross-entropy loss |$ L_{\text{BCE}} $|⁠:

$$ \begin{align*} & L_{BCE} = -\sum_{i=1}^{M} \left( y_{i} \log(o_{i}) + (1 - y_{i}) \log(1 - o_{i}) \right)\!, \end{align*} $$

where |$ M $| is the total number of antibody–antigen samples in the training dataset. |$ y_{i} $| is the ground truth label for the |$ith$| sample, indicating whether the antibody binds to the antigen. |$ o_{i} $| is the output of the sigmoid for the |$i-th$| sample.

This loss function measures the discrepancy between the predicted and the true labels, guiding the model to improve its predictions through iterative updates during training. By minimizing this loss, the classifier module learns to accurately predict the binding of antibody–antigen pairs.

Recommended parameter settings can be found in Supplementary Table 2.

Results

To assess the performance of AntiBinder, we conduct evaluations across three scenarios: predicting the binding of known antigens with new antibodies, predicting the binding of unseen antigens, and predicting cross-species antigen–antibody interactions. We utilize specialized datasets and various experimental setups to comprehensively evaluate the efficacy of our model.

Datasets

Here is the description and detailed information of the datasets used in this study:

COVID-19 dataset [20]: Cov-AbDab is a comprehensive public database containing antibodies and nanobodies targeting coronaviruses, including SARS-CoV-2, SARS-CoV-1, and MERS-CoV. It includes 11 868 studies documenting the neutralizing capabilities of antibodies against specific viral strains. After curation, we focused on 10 720 studies with complete heavy and light chain sequences, resulting in 35 970 antigen–antibody pairs. Specifically, 27 324 pairs related to SARS-CoV-2 antigens were selected for analysis.

HIV dataset [21]: Curated from the Los Alamos HIV database (LANL), this dataset includes sequence data segmented into antibody heavy and light chains. We retained 24 004 sequence pairs after filtering.

BioMap dataset [22]: Originating from a competition hosted by BioMap, this dataset comprises 1706 pairs with binding free energy (Delta G) labels. The dataset includes 473 PDB complex entries and features 638 unique antigens and 1277 unique antibodies from humans, mice, hamsters, chimpanzees, macaques, rabbits, rats, and llamas.

MET dataset [23]: Curated by Makowski et al., the MET dataset focuses on optimizing the affinity and specificity of emibetuzumab, a clinical-stage antibody targeting the MET receptor associated with tumor proliferation and metastasis. We included 4000 sequence pairs from this dataset.

These datasets collectively cover a diverse range of antigen–antibody interactions related to viruses, cancer, and cross-species antibodies, enabling comprehensive evaluation and comparison with existing models. Table 1 provides an overview of the dataset compositions. Figure 2 illustrates the average number of antibodies corresponding to different antigens across the datasets, indicating the dataset sizes and diversity crucial for model learning and evaluation.

Table 1

Detailed information of the four evaluation datasets

Dataset	Interaction pairs	Noninteraction pairs	Total pairs	# Unique antigen
COVID-19	13 471	13 853	27 324	32
HIV	12 650	11 354	24 004	940
BioMap	655	708	1363	594
MET	1516	2484	4000	1
ALL	28 292	28 399	56 691	1567

Figure 2

The average number of antibodies corresponding to different antigens in each dataset.

Baselines

We compared the performance of AntiBinder with 11 benchmark methods, including TAGPPI [24], MARPPI [25], DNN-PPI [26], DeepPPISP [27], PIPR [28], GraphPPIS [29], AbAgIntPre, AttABseq, DeepAAI, DG-affinity, and PESI. A brief introduction to these methods is provided in Supplementary Note 1. The first six methods are designed for predicting PPIs, while the latter five specifically target AAIs. In our experiments, all benchmark methods were implemented according to the guidelines provided in their respective GitHub repositories.

Evaluation metrics

To comprehensively evaluate the performance of AntiBinder with other methods, we employ seven widely recognized metrics for binary classification. These metrics are precision, sensitivity, specificity, F1-score, accuracy, Matthews correlation coefficient (MCC), and area under the ROC curve (AUC). The definitions of these metrics are as follows:

$$ \begin{align*} & \text{Precision} = \frac{TP}{TP + FP}, \quad \text{Sensitivity} = \frac{TP}{TP + FN}, \quad \end{align*} $$

$$ \begin{align*} & \text{F1} = \frac{2 \times TP}{2 \times TP + FP +FN}, \end{align*} $$

$$ \begin{align*} & \text{Specificity} = \frac{TN}{TN + FP}, \end{align*} $$

$$ \begin{align*} & \text{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN}, \end{align*} $$

$$ \begin{align*} & \text{MCC} = \frac{TP \times TN - FP \times FN}{\sqrt{(TP + FP)(TP + FN)(TN + FP)(TN + FN)}}, \end{align*} $$

where TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative, respectively.

AntiBinder outperforms in predicting antibody–antigen interaction

In this section, we conduct 5-fold cross-validation on the four datasets, with the results presented in Fig. 3. Detailed numerical results are provided in Supplementary Table 1.

Figure 3

Average results of four comprehensive metrics obtained from 5-fold cross-validation experiments across the COVID-19, HIV, BioMap, and MET datasets. The average rank of each method for the metrics AUC, F1, MCC, and Accuracy is calculated and displayed in the four figures above. A lower average rank signifies a better performance of the method on the respective metric.

Across all datasets, AntiBinder consistently emerges as the top-performing method, achieving the highest average rank across four key metrics: AUC, F1, MCC, and ACC. The COVID-19, HIV, and BioMap datasets, which are more complex due to their inclusion of multiple antigens, further emphasize AntiBinder’s dominance, with the model ranking first in nearly all metrics. The MET dataset, containing only one antigen with 1516 known binding antibodies and 2484 non-binding antibodies, presents a relatively simpler task. Here, AntiBinder achieves an AUC score of 0.9715, ranking third, with only a slight margin behind the top-performing AttABseq (AUC 0.9775).

In terms of AUC scores, the top five methods are AntiBinder, AttABseq, TAGPPI, PESI, and DeepAAI—four of which are specifically designed for AAI prediction. Overall, AAI models outperform the PPI models, with AntiBinder, AttABseq, and TAGPPI standing out in terms of comprehensive performance, with AntiBinder, AttABseq, and TAGPPI excelling in comprehensive performance. In contrast, GraphPPIS and DG-affinity show weaker performance, consistently ranking lower across most metrics.

In summary, AntiBinder demonstrates superior performance in predicting AAI across a variety of datasets, encompassing multiple species and a broad range of known antibodies. This underscores the robustness and effectiveness of AntiBinder’s specialized design in addressing the complexities of diverse AAI.

AntiBinder excels in prediction for unseen antigens

To evaluate AntiBinder’s generalization ability on unseen antigens, we design experiments using the HIV and COVID-19 datasets.

For the COVID-19 dataset, we conduct additional processing and design two experiments: Cov-exp1 and Cov-exp2. In Cov-exp1, we select various subtypes of COVID-19, including the wild-type (WT), Beta, Alpha, and Delta, as detailed in Table 2. We use WT, Beta, and Alpha as the training set and Delta as the test set to assess the model’s generalization performance. In Cov-exp2, we specifically test the generalization performance for Omicron. Due to the large number of Omicron subtype pairs and their low sequence similarity to other antigens, we include some Omicron subtypes in the training set. Specific details are also shown in Table 2.

Table 2

Details of the Cov-exp1 and Cov-exp2 experiments

Dataset	Exp-name	Type	Subtype	Interaction pairs	Noninteraction pairs	Total pairs	# Unique antigen
COVID-19	Cov-exp1	Training	WT	4236	2579	6815	4
			Beta	283	85	368	4
			Alpha	263	26	289	4
		Test	Delta	256	54	310	2
	Cov-exp2	Training	WT	4236	2579	6815	4
			Beta	283	85	368	4
			Alpha	263	26	289	4
			Delta	256	54	310	2
			Kappa	2	2	4	1
			Gamma	286	242	528	4
			Mu	1	1	2	1
			Omicron	7466	1066	17 532	18
		Test	Omicron	666	789	1455	2

For the HIV dataset, we query the AA sequences of HIV antigens to obtain their corresponding PDB identification and detailed information. Through analysis of this information, we include antigen pairs directly derived from HIV-1, which are frequently used in common research studies, in the training set. These antigens cover key features of HIV-1, such as HIV-1 06TG.HT008 and HIV-1 group M. Antigens related to HIV-1 but with some variation are included in the test set. This dataset is named HIV-exp1 in Table 3.

Table 3

Details of the HIV-exp1 experiment

Dataset	Type	Interaction pairs	Noninteraction pairs	Total pairs	# Unique antigen
HIV-exp1	Training	11 748	10 462	22 210	884
	Test	881	712	1593	41

For all three experiments, we ensure that there was no overlap between the antigens in the training and test sets, maintaining a clear distinction to rigorously test AntiBinder’s generalization capability.

The experimental results for the unseen antigens are shown in Table 4.

Table 4

Comparison of different methods in Cov-exp1, Cov-exp2, and HIV-exp1 experiments

Exp-name	Methods	Spe(%)	Pre(%)	Sen(%)	F1(%)	Acc(%)	MCC(%)	AUC(%)
Cov-exp1	AbAgIntPre	44.44	88.18	87.50	87.84	80.00	31.49	77.00
	AttABseq	20.37	85.12	96.09	90.27	82.90	24.84	64.14
	DeepAAI	72.22	91.93	66.79	77.37	67.74	30.21	74.81
	DG-affinity	92.59	94.59	27.34	42.42	38.70	17.73	69.05
	PESI	37.04	87.16	90.23	88.67	80.96	29.36	69.34
	GraphPPIS	27.78	86.07	94.14	89.92	82.58	28.11	71.74
	DeepPPISP	55.56	89.91	83.59	86.63	78.70	35.16	78.99
	DNN-PPI	46.30	88.01	83.20	85.54	76.77	27.03	75.40
	PIPR	42.59	88.30	91.40	89.82	82.90	36.60	80.30
	MARPPI	61.11	91.35	86.71	88.97	82.25	44.07	75.31
	TAGPPI	37.04	88.11	98.44	92.99	87.74	50.34	81.63
	AntiBinder	46.30	89.67	98.43	93.85	89.35	58.26	82.88
Cov-exp2	AbAgIntPre	7.86	47.73	99.69	64.56	49.89	18.36	96.44
	AttABseq	59.18	65.00	89.78	75.40	73.19	50.60	69.08
	DeepAAI	85.29	81.52	76.87	79.13	81.44	62.53	89.62
	DG-affinity	83.90	79.77	75.22	77.43	79.93	59.47	87.11
	PESI	44.61	55.63	82.28	66.38	61.85	28.65	71.89
	GraphPPIS	8.37	46.64	94.89	62.54	47.97	6.41	57.36
	DeepPPISP	34.60	55.32	95.94	70.18	62.68	37.61	83.24
	DNN-PPI	39.67	56.80	93.99	70.81	64.53	39.12	88.93
	PIPR	34.85	54.83	93.69	69.17	61.78	34.45	81.69
	MARPPI	35.23	52.50	84.83	64.86	57.93	22.78	72.48
	TAGPPI	30.80	50.86	84.83	63.59	55.53	18.33	66.09
	AntiBinder	85.93	85.29	96.69	90.64	90.85	82.39	96.51
HIV-exp1	AbAgIntPre	95.51	91.63	39.72	55.42	64.65	41.02	84.06
	AttABseq	69.24	74.05	70.94	72.46	70.18	40.02	79.17
	DeepAAI	75.98	78.21	69.69	73.70	72.50	45.42	81.29
	DG-affinity	88.48	83.43	46.87	60.02	65.47	37.98	78.64
	PESI	71.77	75.00	68.44	71.57	69.93	39.98	76.93
	GraphPPIS	72.61	75.40	67.87	71.44	69.99	40.26	77.85
	DeepPPISP	78.09	80.50	73.09	76.62	75.32	50.90	84.26
	DNN-PPI	75.84	78.44	71.05	74.56	73.19	46.63	81.73
	PIPR	77.11	79.92	73.66	76.66	75.20	50.49	83.11
	MARPPI	70.65	76.67	77.97	77.32	74.70	48.73	82.21
	TAGPPI	77.53	80.07	72.98	76.36	70.51	50.22	81.94
	AntiBinder	77.53	80.60	75.48	77.96	76.39	52.74	83.89

AntiBinder consistently outperforms other methods across nearly all evaluation metrics on the three datasets, demonstrating its exceptional ability to generalize and accurately predict AAI for newly emerging antigens. Notably, in the Cov-exp2 experiment, AntiBinder’s performance surpasses that of the next best method by over 10%. This experiment focuses on predicting antibodies for the Omicron subtype, a variant with low similarity to other known antigens, which makes it particularly challenging to transfer knowledge from existing AAI data to new antigens. Despite this challenge, AntiBinder achieves outstanding predictive performance, likely due to its ability to capture the fundamental mechanisms of AAI and leverage the advantages of large datasets.

Among the other methods specifically developed for AAI prediction—AbAgIntPre, AttABseq, DeepAAI, DG-affinity, and PESI—none consistently outperform the models designed for PPI prediction. Of these, DeepAAI shows slightly better performance, followed by DG-affinity, while PESI and AttABseq rank in the lower-middle range of the 12 methods. AbAgIntPre, in particular, demonstrates skewed probability outputs, tending toward 0 in the HIV-exp dataset (with a sensitivity below 40%) and toward 1 in the Cov-exp2 experiment (where sensitivity approaches 1, but precision remains around 0.5). Consequently, AbAgIntPre struggles to distinguish positive and negative predictions using a fixed threshold, making it advisable to treat top-ranked predictions as positives instead of relying on absolute probability values.

The PPI prediction methods—DeepPPISP, DNN-PPI, PIPR, MARPPI, GraphPPIS, and TAGPPI—show inconsistent performance across the three datasets. In the Cov-exp2 experiment, where the positive-to-negative ratio is 0.84:1, these methods generally achieve high sensitivity but suffer from low precision and specificity, indicating that their prediction probabilities tend to cluster around 1, creating a bias toward positive predictions. Careful threshold selection is crucial when using these models to avoid inflating false positives.

TAGPPI, which performs well in cross-validation experiments, shows a significant decline in performance when tested on unseen antigens, underscoring its limited generalizability. This contrast further highlights AntiBinder’s advantage in adapting to and accurately predicting AAI, especially for novel or previously unseen antigens.

AntiBinder performs well in cross-species prediction

AntiBinder has demonstrated exceptional performance in making prediction for unseen antigens. To further illustrate its generalizability and practical utility, we test it on cross-species datasets.

We construct a dataset named BioMap+. The test set of BioMap+ comprises homo sapiens antigens from the BioMap dataset. The training set consists of data from the four datasets used in this study, excluding homo sapiens related antigens, ensuring no overlap with the test set. The specific details of the BioMap+ dataset are shown in Table 5.

Table 5

Details of the Bio-exp1 experiment

Dataset	Type	Species	Interaction pairs	Noninteraction pairs	Total pairs	# Unique antigen
BioMap+	Training	Without homo sapiens	22 417	22 664	45 081	1346
	Test	Homo sapiens	224	111	335	178

The experimental results are presented in Table 6. It is evident that all other methods fail on this dataset except for AntiBinder. Although AntiBinder’s performance declines compared to previous experiments in this study, it still exhibits a notable level of predictive capability. This suggests that cross-species prediction is challenging but possible, and that AntiBinder is the only model capable of capturing the fundamental mechanisms of AAI across different species.

Table 6

Comparison of different methods in Bio-exp1 experiment

Exp-name	Methods	Spe(%)	Pre(%)	Sen(%)	F1(%)	Acc(%)	MCC(%)	AUC(%)
Bio-exp1	AbAgIntPre	29.73	41.35	24.55	30.81	26.26	-43.98	28.00
	AttABseq	63.96	76.19	57.14	65.30	59.40	19.86	55.19
	DeepAAI	62.16	76.00	59.37	66.67	60.29	20.29	54.76
	DG-affinity	48.64	68.68	55.80	61.57	53.43	4.20	44.08
	PESI	52.25	69.88	52.25	61.50	54.02	6.70	46.53
	GraphPPIS	50.45	73.42	67.85	70.53	62.08	17.73	57.57
	DeepPPISP	25.23	48.76	35.26	40.93	31.94	−37.21	30.16
	DNN-PPI	47.75	46.29	22.32	30.12	30.74	−30.14	25.82
	PIPR	17.12	64.06	73.21	68.33	54.62	−10.72	34.17
	MARPPI	67.57	48.57	15.17	23.12	32.53	−19.97	40.35
	TAGPPI	48.65	51.28	26.79	35.19	34.02	−24.25	30.59
	AntiBinder	53.15	75.47	71.43	73.39	65.37	24.00	74.84

Analysis of attention weights in AntiBinder

To enhance the interpretability of AntiBinder, we explore the model’s attention mechanism by visualizing the attention weights in Fig. 4. In particular, we focus on the bidirectional cross-attention mechanism that allows AntiBinder to model interactions between antibody and antigen sequences. This attention mechanism is critical for identifying which regions of the antibody and antigen are most relevant to the predicted binding interactions.

Figure 4

Interpretability study of antigen–antibody interaction. (a) Crystal structure of the antibody PGT124 in complex with the HIV-1 JRCSF gp120 core and CD4, highlighting critical interaction sites. The heavy chain (brown) and light chain (blue) of PGT124 interact specifically with the gp120 antigen (purple), while CD4 (green) serves as a spatial reference. Key residues on PGT124’s heavy chain, Y65 and E118, form interactions with residues R349 and Q332 on gp120, suggesting high-affinity binding sites critical to neutralization. (b) Heatmap showing the attention weights generated by AntiBinder’s cross-attention mechanism from AA positions of the HIV-1 JRCSF gp120 antigen to those of the PGT124 antibody. Red denotes regions of higher attention weights, indicating areas where the model concentrates more heavily, while blue represents regions with lower attention. These high-attention areas correspond to key binding sites, highlighting the model’s focus on biologically relevant regions in antigen–antibody interactions.

In Fig. 4a, we show the crystal structure of the neutralizing antibody PGT124 in complex with the HIV-1 JRCSF gp120 core and CD4 receptor. The interaction between antibody residues Y65 and E118 with antigen residues R349 and Q332, which are identified as high-affinity binding sites, is central to this structure. These interactions are important for the neutralizing activity of PGT124 and serve as an example of biologically significant antigen–antibody binding.

To investigate whether the model captures and prioritizes these biologically relevant interactions, we generate an attention weight heatmap using the antigen-to-antibody cross-attention mechanism. In Fig. 4b, the heatmap displays the attention weights assigned to the AA positions of the antigen (x-axis) and antibody (y-axis). Darker colors represent higher attention, which suggests that these AA pairs are more critical in the context of binding.

The model assigns high attention to the Y65-R349 and E118-Q332 pairs, which correspond to the high-affinity sites observed in the structural analysis. This suggests that the model effectively recognizes these biologically important molecular interactions. Importantly, the attention distribution is sparse, with high attention concentrated on only a few critical AA sites. This aligns with biological expectations, where antigen–antibody interactions typically rely on a limited number of key residues rather than the entire sequence.

This sparse attention distribution further supports the idea that antigen–antibody binding is driven by specific interaction hotspots, such as hydrogen bonds, hydrophobic interactions, and electrostatic contacts. While the attention mechanism provides valuable insights into the interaction process, we note that further studies are needed to better understand the correlation between attention weights and specific biological functions.

Ablation study

To investigate the effects of AA encoding methods, the contribution of sequence and structural information, and the impact of the BidAttBlock module on experimental results, we conduct a series of ablation studies.

The study of residues embedding effect: Firstly, we briefly explain our choice of S2V as the AA embedding method. In the COVID-19 dataset, we compare S2V with four other embedding methods: one-hot encoding, |$K$|-mer [30], AC [31], and CKSAAP [32]. For |$K$|-mer, we set |$ K = 2 $|⁠, extracting all subsequences of length 2 from each sequence; for AC, we set the maximum sequence interval to 13, which must be less than the length of the sequence. According to the evaluation results in Fig. 5a, S2V performs best among these embedding methods. Detailed numerical results are provided in Supplementary Table 3.

Figure 5

Results of the three parts of the ablation study. (a) Evaluation of different residue encoding methods. The figure shows the performance of K-mer, one-hot, CKSAAP, AC, and S2V encoding methods in prediction tasks. (b) Impact of sequence and structure information on performance. The figure shows the performance of three types of inputs: using only sequence, only structure, and a combination of sequence and structure in prediction results. (c) Effect of the BidAttBlock module. The figure shows the performance with and without the BidAttBlock module, as well as the effect of replacing this module with two independent Transformers.

The study of the impact of sequence and structure: We use the COVID-19 dataset as an example to illustrate the impact of sequence and unique structural features on prediction performance. As shown in Fig. 5b, the prediction results that utilize both integrated sequence and structural information as input exhibit significant advantages. In comparison, sequence information contributes more to the prediction, but the predicted structures derived from the sequence also enhance the prediction of AAI. This indicates that integrating structural information obtained from sequence-based structural prediction methods can improve the accuracy of predictions. Numerical results in detail can be found in Supplementary Table 4.

The study of the impact of BidAttBlock module: To evaluate the contribution of the BidAttBlock to prediction accuracy, we compare test results from models with and without this module using the COVID-19 dataset, as well as results from models where the module is replaced with two independent Transformers. As shown in Fig. 5c, the BidAttBlock significantly enhances the prediction of interaction relationships. Detailed numerical results are provided in Supplementary Table 5.

Conclusion

In this study, we present AntiBinder, a novel method specifically designed for predicting AAI. AntiBinder captures the unique characteristics of both antibodies and antigens by integrating structural and sequence information, along with region-specific details of the antibody. By employing multiple BidAttBlocks, it effectively models the complex and diverse nature of these interactions, enabling a deeper understanding of the underlying binding mechanisms. Supplementary Table 6 provides a comparative overview of AntiBinder’s features alongside other methods.

We evaluate AntiBinder’s predictive performance through extensive experiments on various datasets, including those featuring known antibodies, unseen antigens, and cross-species data. The results clearly demonstrate that AntiBinder outperforms state-of-the-art methods, achieving significantly higher predictive accuracy across all scenarios.

One of the most notable outcomes is AntiBinder’s superior performance in predicting interactions with unseen antigens. In the Cov-exp2 experiment, which focuses on the Omicron variant of COVID-19, AntiBinder surpasses the next best method by a substantial margin. This highlights the model’s strong generalizability and its potential applications in vaccine development, including for rapidly evolving diseases like COVID-19 and HIV. AntiBinder’s adaptability to emerging infectious diseases, such as new COVID-19 variants, further underscores its real-world relevance.

Additionally, AntiBinder is the only method that demonstrates reasonable predictive accuracy in the cross-species prediction task, a particularly challenging dataset where other methods fail. This ability to generalize across species is highly significant for vaccine research, suggesting that AntiBinder can effectively predict AAI in diverse biological contexts. This capability enhances the translational potential of immunological studies, bridging the gap between preclinical and clinical research.

While there is room for improvement in cross-species predictions, AntiBinder’s ability to capture the fundamental mechanisms of antigen–antibody interactions from large datasets is highly promising. As more data becomes available, we anticipate further improvements in AntiBinder’s predictive power. These advancements will strengthen its practical utility in various real-world applications, including cross-species predictions, ultimately expanding its role in scientific research and development.

Though AntiBinder demonstrates strong performance in predicting AAI, it has certain limitations. Specifically, the model relies on IgFold for antibody structure representation and ESM-2 for antigen embedding, both of which may impose constraints. IgFold, while optimized for antibody sequences, may have difficulty capturing subtle structural variations across different antibody subtypes, potentially affecting paratope representation accuracy. Similarly, while ESM-2 effectively generates protein embeddings, it may not fully reflect antigen diversity, particularly for those with atypical motifs or unique folds. These dependencies could impact AntiBinder’s flexibility and generalizability across highly diverse datasets. Future research could focus on incorporating additional or alternative protein structure models to mitigate these limitations and enhance model robustness. Additionally, developing methods to dynamically fine-tune embeddings from IgFold and ESM-2 in specific contexts could further improve AntiBinder’s adaptability and performance.

Key Points

• AntiBinder is a novel model specifically designed for predicting antibody–antigen binding.

• AntiBinder integrates the unique structural and sequence characteristics of antibodies and antigens into its framework and employs a bidirectional cross-attention mechanism to automatically learn the intrinsic mechanisms of antigen–antibody binding, eliminating the need for manual feature engineering.

• Experiments show that AntiBinder outperforms other methods across multiple datasets and tasks, particularly in generalization tasks involving unseen antigens.

Acknowledgements

We thank all anonymous reviewers for valuable suggestions.

Funding

This work is supported by the National Key R&D Program of China (2021YFC3340703).

Data availability

All datasets and source codes in this study are available at https://github.com/fdu-wangfeilab/AntiBinder.