Cyagen and Tsinghua Research Team Develop PPBind-3D/1D AI Framework for Protein–Protein Binding Affinity Prediction
Guangzhou, China, October 13, 2026 – A research team led by Lanqing Han, Director of the Artificial Intelligence Innovation Center at Tsinghua Pearl River Delta Research Institute and Chairman of Cyagen Biosciences, has published a study in Antibody Therapeutics introducing PPBind-3D/1D (https://doi.org/10.1093/abt/tbag053), a general-purpose artificial intelligence framework for predicting protein–protein binding affinity. Developed in collaboration with researchers from the two institutions and clinical research teams, the study addresses several challenges in computational affinity prediction, including limited experimental data, differences among affinity measurement datasets, and the lack of reliable three-dimensional protein complex structures.
By integrating complementary structure-based and sequence-based models, PPBind-3D/1D provides a computational approach to prioritizing antibody and protein drug candidates. The framework is designed to facilitate high-throughput virtual screening and support early-stage therapeutic discovery.1
A Dual-Model Framework for Affinity Prediction
Protein–protein binding affinity is an important consideration of therapeutic antibody and protein drug performance. Experimental methods such as surface plasmon resonance (SPR), biolayer interferometry (BLI, including Octet systems), remain essential for measuring molecular interactions, but extensive experimental characterization can be resource-intensive when screening large candidate libraries.
To complement these methods, the research team developed two models:
- PPBind-3D: Structure-Based Affinity Prediction. PPBind-3D incorporates a self-and-mutual attention (S&M-Attention) geometric encoder to learn structural features of protein complexes. Its monotonic neural network-constrained multitask learning framework integrates heterogeneous experimental data, including SPR measurements and deep mutational scanning (DMS) datasets. An iDist-based clustering strategy establishes rigorous dataset splits to reduce data leakage. Under strict evaluation, the model achieved a Pearson correlation coefficient above 0.6 and outperformed established baseline tools, including PRODIGY and ProBAN.
- PPBind-1D: Sequence-Based Affinity Prediction. To address scenarios in which experimentally resolved complex structures are unavailable, PPBind-1D uses a feature-alignment training strategy that distills knowledge from PPBind-3D. Training incorporates 42,108 unlabeled protein complexes from the DIPS-Plus dataset. This approach enables sequence-based affinity assessment without requiring experimentally resolved or computationally predicted complex structures as inputs. This approach is particularly suited to large-scale antibody sequence library screening.
Together, the two models provide complementary capabilities for structure-informed prediction and sequence-based candidate prioritization.
Evaluation Across Antibody Discovery Scenarios
The research team evaluated PPBind-3D/1D through three virtual screening case studies involving antibodies and nanobodies, demonstrating how the models can be applied under different structural information conditions.
- Antibody–antigen affinity prediction with experimentally resolved structures: Using antibody complex structures involving CTLA-4, PD-1, and PD-L1 that were excluded from the training dataset, PPBind-3D achieved an overall Pearson correlation coefficient of 0.52, demonstrating its potential for ranking antibody candidates when reliable structural information is available.
- Nanobody affinity ranking with predicted structures: In a hen egg lysozyme (HEL)-targeting nanobody case, PPBind-3D performance was limited when AlphaFold 3-predicted complex structures had low interface predicted TM (ipTM) scores or structural inaccuracies. PPBind-1D with feature alignment maintained more stable ranking performance, achieving a correlation above 0.4.
- Anti-PD-L1 antibody affinity ranking from sequence data: In another sequence-based screening case, PPBind-1D with feature alignment achieved a Pearson correlation coefficient of 0.66, outperforming the non-aligned sequence model. The results highlight the potential of feature alignment to improve ranking robustness when predicted complex structures contain inaccuracies in epitope identification or docking orientation.
These findings suggest a practical model-selection strategy: PPBind-3D is suited to affinity ranking when reliable experimental complex structures are available, while PPBind-1D offers a complementary option for high-throughput screening when structural predictions are uncertain or unavailable. Both models are primarily intended to distinguish and prioritize candidates with different binding affinities rather than provide precise absolute affinity values for individual molecules.
Open-Source Tools for the Research Community
To promote broader adoption and collaborative innovation, the research team has made the PPBind framework available through the AbSeek™ intelligent antibody computing platform. The online prediction tool, model code, and associated dataset resources are publicly accessible.
- Protein Complex Affinity Prediction: AbSeek™ ΔG Prediction
- Open-Source Model Code: PPBindPPB-Affinity on GitHub
- Dataset Resource: Scientific Data
- Additional Computational Tools: AbSeek™ Platform
- RDDC Tool Suite: Tsinghua Pearl River Delta Research Institute
By making these resources openly available, the team aims to facilitate reproducible research, expand access to computational affinity prediction, and encourage collaboration across academic research and therapeutic development.
Advancing AI-Enabled Antibody and Protein Drug Discovery
The PPBind-3D/1D framework strengthens the computational capabilities available for early-stage antibody and protein drug discovery.
PPBind-3D has been integrated into an online affinity prediction tool, allowing researchers to explore structure-based predictions. The research also demonstrates how sequence-based modeling may support screening workflows when reliable complex structures are unavailable.2
Within Cyagen's broader antibody discovery capabilities, computational tools such as those provided through the AbSeek™ platform complement the HUGO-Ab™ fully human antibody mouse platform and downstream experimental screening and validation services.
By connecting computational analysis with experimental research, Cyagen aims to support more informed candidate selection and antibody engineering decisions.
About Cyagen Biosciences
Founded in 2006, Cyagen Biosciences is an AI-enabled biotechnology company advancing biomedical research through proprietary gene-editing technologies, disease models, computational platforms, and integrated preclinical solutions.
Cyagen's antibody discovery capabilities combine the HUGO-Ab™ fully human antibody mouse platform with the AbSeek™ intelligent antibody computing platform, integrating in vivo antibody generation with AI-driven sequence analysis, optimization, screening, and validation.
Through these integrated capabilities, Cyagen aims to help biopharmaceutical innovators prioritize promising candidates, streamline early-stage screening, and advance therapeutic antibody development.
Media Contact
Cyagen: marketing@cyagen.com
[1] The models are intended to support candidate prioritization and affinity ranking, rather than replace experimental affinity measurements or provide precise absolute affinity values for individual molecules.
[2] PPBind-3D performance depends on input structure quality, while PPBind-1D may still show deviations in absolute affinity predictions. With advances in structure prediction and larger experimental datasets, the framework could enable broader applications in affinity screening and antibody optimization, including high-throughput variant evaluation through its RDE module.
