AbAgKer: a unified semi-supervised framework for antigen-antibody binding affinity and kinetics prediction Authors: Gang Luo, Junkai Wang, Sizhe Zhang, Zhilin Zhu, Zhangli Lu, Min Li, Pier Luigi Martelli Published: 2026 ABSTRACT AbstractMotivationThe rapid advancement of generative artificial intelligence has enabled the high-throughput design of therapeutic antibody candidates. However, the precise evaluation of these candidates remains a significant challenge due to the scarcity of high-quality activity data and the structural flexibility of antibody complementarity-determining regions (CDRs).ResultsTo address these challenges, we propose AbAgKer, an antibody screening model leveraging pre-trained representations and biological prior guidance for antigen-antibody affinity and kinetics prediction. Specifically, we design a biological prior-guided feature fusion framework that integrates pseudo-structural epitope knowledge and CDR-specific attention mechanisms via a mixture-of-experts architecture to effectively capture complex binding landscapes. To mitigate data scarcity, we employ a semi-supervised learning strategy for data self-distillation, which significantly enhances affinity prediction performance. Additionally, we demonstrate that the interaction representations learned by AbAgKer can be effectively transferred to the data-scarce task of predicting dissociation rates via few-shot learning. Extensive experiments demonstrate that AbAgKer outperforms baseline models and exhibits strong generalization capabilities in antibody screening and drug residence time analysis.Availability and implementationThe source code and dataset are available at https://github.com/CSUBioGroup/AbAgKer and https://doi.org/10.5281/zenodo.19691211. 1 Introduction The development of therapeutic antibodies has revolutionized modern medicine, offering precise interventions for diseases ranging from cancer to viral infections (Pantaleo et al. 2022, Paul et al. 2024). With the advent of deep learning, antibody discovery is shifting from labor-intensive experimental screening to antibody de novo design. Recent years have witnessed an explosion of antibody generative models, ranging from sequence-based language models (Shuai et al. 2021, He et al. 2024, Wasdin et al. 2025) to structure-based diffusion models (Luo et al. 2022, Wang et al. 2024, Team et al. 2025), capable of generating millions of antibody candidates. The pre-trained antibody generative large language model (PALM-H3) (He et al. 2024) focuses specifically on sequence design for the heavy chain complementarity-determining region 3 involved in antigen-antibody interactions, successfully generating antibodies targeting SARS-CoV-2. Furthermore, DiffAb (Luo et al. 2022) emerged as the first deep learning model to generate antibodies targeting specific antigen structures. RFantibody (Bennett et al. 2026) fine-tunes RFdiffusion (Watson et al. 2023) on antibody complex structure data to achieve precise sampling of complementarity-determining regions (CDRs) loops targeting specified epitopes, with binding states confirmed via cryo-electron microscopy. However, the lack of accurate, high-throughput computational screening tools to evaluate these generated candidates remains a significant challenge (Cheng et al. 2024). The gold standard for in silico verification is binding affinity prediction, which predicts the equilibrium dissociation constant (Kd) to reflect the strength of the interaction. With the advent of Protein Language Models (PLM), recent affinity prediction studies have attempted to model affinity based on domain-specific pre-trained representations, such as AntiBERTy (Ruffolo et al. 2021) and AntiBERTa2 (Barton et al. 2024). For instance, DG-Affinity (Yuan et al. 2023) employs ConvNeXt to integrate representations from the general protein model TAPE (Rao et al. 2019) and the antibody-specific model AbLang (Olsen et al. 2022). Similarly, prior study (He et al. 2024) proposed a high-precision predictor by fusing its pre-trained antibody features with antigen ESM-2 (Evolutionary Scale Modeling 2) representations to evaluate binding strength against SARS-CoV-2. Meanwhile, MVSF-AB (Li et al. 2025) utilizes residue AAindex features and ProteinBERT (Brandes et al. 2022) antibody representations, and achieves excellent correlation performance across multiple datasets. However, the scarcity of high-quality data continues to limit the development of this field. Moreover, antigen-antibody affinity prediction is insufficient to screen for excellent antibodies in the real world. The dissociation rate (koff) provides critical kinetic information (Vauquelin 2016) in antigen-antibody binding, as shown in Fig. 1c, revealing the stability of the complex and the duration of biological action. Even for interactions with identical Kd values, varying koff can lead to significant differences in biological behaviors, such as drug efficacy and signal persistence (Piccoli et al. 2020). However, traditional experimental methods, such as Surface Plasmon Resonance (SPR) or Bio-Layer Interferometry (BLI), are prohibitively expensive for large-scale screening. (a) The overall framework of AbAgKer. (b) The pipeline of the MEI Module. (c) Antigen-antibody binding affinity is typically quantified by the equilibrium dissociation constant (Kd), which is derived from the kinetic association (kon) and dissociation (koff) rate constants. (d) Based on the SKEMPI v2.0 and SAbDab databases, we constructed three datasets for this study: the affinity dataset AbAgA, the interaction dataset AbAgI, and the antibody dissociation rate dataset AbAgKoff. Although recent studies have explored methods to predict antigen-antibody affinity (Li et al. 2025) or binding free energy (Hummer et al. 2025) to assist in high-throughput screening, end-to-end deep learning methods face several critical challenges. First, the availability of high-quality antigen-antibody complexes paired with experimental affinity or kinetic data is extremely limited compared to the vast space of single sequences. This scarcity makes it difficult to construct high-performing and generalizable expert prediction models using only raw data. Second, beyond static affinity (Kd), binding kinetics koff are determinants of drug efficacy and safety (e.g. residence time) (Lee et al. 2019). Predicting koff is notoriously difficult due to the even greater scarcity of data and the complex temporal dynamics of molecular unbinding. Existing antibody screening tools often neglect these biological behaviors associated with efficacy persistence, limiting the practicality for real-world virtual screening. To address the challenges, we propose AbAgKer, a novel model for antigen-antibody interaction prediction. We incorporate pre-trained representations and design multiple sub-modules guided by biological prior knowledge to reduce interaction modeling difficulty. To tackle the scarcity of activity data, we employ semi-supervised learning, significantly expanding the training dataset and achieving stable affinity prediction. Experimental evaluations indicate that our model achieves the best overall performance among the evaluated baselines. Finally, to address the lack of computational screening methods for kinetic metrics in antibody engineering, we propose leveraging the learned binding interaction representations to predict the dissociation rate constant koff. In summary, we provide an effective screening method for predicting the strength of antigen-antibody interactions and their functional persistence. 2 Methodology 2.1 Overview The architecture of AbAgKer is illustrated in Fig. 1a. First, we extract residue-level representations for both the antigen and antibody by embedding their paired sequences using domain-specific PLMs. Next, the antigen residue-level representations are fed into the Multi-scale feature adaptive Epitope Identification (MEI) module to delineate binding epitope regions within complex antigen proteins. Subsequently, we design the CDR-Guided Global-Local Attention (CAttention) mechanism to perform antibody feature re-fusion and extract CDR loop representations from antibody chains. Finally, the pooled representations of the antigen and antibody are concatenated and input into a Mixture-of-Experts (MoE) architecture (Rajbhandari et al. 2022). The MoE architecture facilitates fine-grained representation learning for diverse antigen-antibody binding scenarios. The results from multiple experts are then input into an MLP regression head to predict interaction affinity and kinetic information. 2.2 Pre-trained protein language model representation To capture the rich semantic information hidden in protein sequences, we employ state-of-the-art PLMs as the feature backbone. For the Antibody Encoder, we use PALM-H3 (He et al. 2024) specifically optimized for antibodies. Specifically, we employ its Heavy-chain RoFormer and Light-chain RoFormer to encode heavy and light-chain sequences, respectively. Unlike general protein models, PALM-H3 is pre-trained on massive antibody repertoires, effectively capturing recombination patterns and loop hypervariability. For the antigen encoder, we employ ESM-2 (Lin et al. 2023), a transformer protein language model trained on millions of diverse protein sequences. We use the residue-level representation from the pre-trained protein language models. To handle variable-length sequences of antibodies and antigens, we use the truncation and padding strategies from the protein tokenizer to maintain fixed-length inputs for the neural network. The heavy- and light-chain residue embeddings from PALM-H3 are concatenated along the sequence dimension to form the antibody representation XAb∈RLAb×dAb, where LAb and dAb are set to 512 and 768. The antigen encoder provides evolutionary-rich residue representations XAg∈RLAg×dAg for characterizing antigen epitope features, where LAg and dAg are set to 896 and 640, respectively. Then, these residue-level representations are projected into a unified hidden dimension d of 512 via a linear layer before being passed to the downstream modeling modules. 2.3 Multi-scale feature adaptive epitope identification module Antigen epitopes are often composed of discontinuous residues that form specific 3D structural motifs, and the standard global pooling operations tend to obscure this fine-grained structural information. To address this, we propose the MEI module, which dynamically fuses the antigen’s PLM embedding with pre-calculated sequence pseudo-structure features to identify the latent binding residue location of the antigen. As shown in Fig. 1b, we use the encoder to obtain residue-level embeddings for the antigen, projecting the features to XAg∈RLAg×d, where d is the token representation dimension of 512. Second, the sequence-based physicochemical property feature shows a great capability for epitope prediction (Wang et al. 2026). Therefore, we incorporate biological prior knowledge to pre-analyze potential binding information at the sequence level for locating possibly binding token embeddings of conformational or linear epitopes. We employ the bioinformatics toolkit B2bTools (Kagami et al. 2021) to calculate six physicochemical and structural features from the antigen sequence, including the backbone disordered properties (Orlando et al. 2022), sidechain properties (Cilia et al. 2013), and early folding propensity (Raimondi et al. 2017). Additionally, we design a mask vector within the maximum sequence length to mark the antigen’s valid length. These are stacked to construct the sequence pseudo-structure feature matrix Ssef. Subsequently, we employ a fused gated weighting mechanism via a dual-path architecture that combines local residue context and biological sequence bias. Firstly, we apply a convolution operation with a kernel size of 9 in the token dimension of the antigen representation to capture local residue weight wconv from antigen fragment context: Then, we design a learnable projection to extract pseudo-structural information for the per-residue property weight wsef of antigen from the pre-calculated matrix Ssef at the residue level: where Ws∈R7×1 is a learnable parameter matrix, bs is the bias term. The final epitope token weights wepitope are then adaptively fused via a learnable gate, defined as: where α represents the learnable weight, and Tanh is the activation function. Finally, we perform Top-K pooling based on importance weight wepitope to select the most significant K tokens, yielding Agpool∈RK×d, where K is 64. 2.4 CDR-guided global-local attention mechanism Antigen-antibody binding is predominantly determined by the complementarity-determining regions (Pantazes and Maranas 2010). To leverage this, as shown in Fig. 2, we propose a novel CAttention mechanism that reinforces the model’s focus on antibody CDR token representations. Unlike standard attention mechanisms (Vaswani et al. 2017), we explicitly define a mask tensor M from antibody CDR information to constrain attention in specific heads. We set M to negative infinity when the position is not the CDR residue, and to 0 otherwise. We construct a multi-layer encoder architecture based on CAttention to learn CDR-guided re-encoding of the antibody representation. The CAttention is constrained to enhance focus on antibody CDR regions via CDR-local heads, effectively forcing the model to prioritize paratope features. The CAttention head outputs are calculated as: The CAttention encoder module and CAttention mechanism, where the Xcdr is the CDR local attention heads, and XGlobal is the global attention heads. The CAttention encoder is composed of multiple layers of CAttention, and its output is the re-encoded antibody representation. where Qi,Ki,Vi are derived from XAb via different learnable weights, dk is the dimension of each attention head, and Bi is the bias term for the i-th attention head, if i