On October 13, 2026, a research team led by Director Han Lanqing from Cyagen Biosciences and the Artificial Intelligence Center of the Tsinghua Pearl River Delta Institute published a study in Antibody Therapeutics, introducing the PPBind-3D/1D AI framework for general protein–protein binding affinity prediction. The study developed a dual-model system combining the PPBind-3D structure model and the PPBind-1D sequence model, addressing data heterogeneity, limited high-quality experimental data, and the difficulty of obtaining three-dimensional complex structures in protein affinity prediction. It provides a high-throughput virtual screening tool for antibody and protein drug development.

Image source: Antibody Therapeutics (https://doi.org/10.1093/abt/tbag053)
Dual-model innovation covers both structural and sequence-based scenarios
The therapeutic performance of protein drugs depends greatly on protein–protein binding (PPB) affinity. High-affinity molecules can bind targets more specifically, improve efficacy, and reduce off-target effects. Therefore, rapid and accurate evaluation of protein interaction affinity is a core step in the early screening of protein and antibody drugs. Current affinity measurements mainly rely on wet-lab technologies such as protein microarrays and Octet. Although accurate, these methods are complex and costly, making them difficult to scale for high-throughput screening of large candidate libraries.
Computational prediction has become an important option for high-throughput screening, but practical implementation still faces several bottlenecks. High-quality, experimentally validated affinity samples remain limited; data from SPR, ITC, ELISA, DMS, and other assays are highly heterogeneous, with inconsistent metrics and substantial noise; and many target complexes lack high-resolution crystal structures. Sequence data are readily available, whereas structural data are often missing.
To address these challenges, the team developed the PPBind prediction framework, comprising the PPBind-3D structure model and the PPBind-1D sequence model. Three progressive virtual screening cases were used to evaluate the practical value of the models for antibody and nanobody affinity ranking.

Figure 1. Overview of the PPBind models, including (A) dataset composition, (B) the PPBind-3D architecture, and (C) the PPBind-1D architecture.
1. PPBind-3D (structure-driven model): For three-dimensional protein-complex structures, the model uses a self-and-mutual attention (S&M-Attention) geometric encoder. It introduces monotonic neural network-constrained multi-task learning (MMTL) to accommodate heterogeneous experimental data, including SPR and DMS, and to make use of high-throughput data reported in formats other than ΔG. A strict clustered dataset split based on the iDist algorithm helps avoid data leakage. Under the strict split, the model achieved a Pearson correlation coefficient greater than 0.6, outperforming major baseline tools such as PRODIGY and ProBAN. The model is available through a public web tool.
2. PPBind-1D (sequence-driven model): This model addresses the lack of three-dimensional protein-complex structures in real-world research. It introduces a Feature Alignment training paradigm, using PPBind-3D as the teacher model and knowledge distillation on the unlabeled DIPS-Plus dataset containing 42,108 complexes. With only protein sequences as input, it achieves prediction accuracy close to that of the structure model without requiring complex-structure determination or prediction, making it suitable for large-scale antibody sequence-library screening. Dimensionality reduction shows that aligned PPBind-1D retains a data topology similar to PPBind-3D, whereas features from the non-aligned model are more scattered.
Open resources support collaborative research and development
The study has been deployed on the AbSeek® intelligent antibody computational platform through the ΔG Prediction web tool. The web tool, dataset, and model code are freely available as open resources.
Web tool:
Protein-complex affinity prediction tool: https://abseek.icyagen.com/catalog/%CE%94G%20Prediction
Open-source model code: https://github.com/ChenPy00/PPBind
PPB-Affinity dataset: https://www.nature.com/articles/s41597-024-03997-4
More tools:
AbSeek® intelligent antibody computational platform toolkit: https://abseek.icyagen.com/catalog
RDDC toolkit: https://rddc.tsinghua-gd.org/tool

AbSeek® intelligent antibody computational platform toolkit

Case 1: Antibody–antigen affinity prediction with a true complex structure
Using CTLA-4, PD-1, and PD-L1 antibody-complex data excluded from the training set, PPBind-3D performed affinity prediction from resolved structures and achieved an overall Pearson correlation of 0.52, demonstrating its direct utility for antibody-candidate screening when experimentally resolved structures are available.

Figure 2. Visualization of the feature representations of the three models (A–C), and model predictions using true structures, shown by sample type (D) and PDB ID (E).
Case 2: HEL nanobody affinity ranking from sequence information
Complex structures were predicted with AlphaFold3. When the predicted structure had a low ipTM or interface deviations, PPBind-3D performance was limited. The feature-aligned PPBind-1D model nevertheless maintained stable prediction performance, with a correlation above 0.4.

Figure 3. Correlations between predicted and experimental affinity for (A) PPBind-3D, (B) PPBind-1D-NoAlignment, and (C) PPBind-1D-Alignment; (D) box plots of ipTM and pTM for AlphaFold3-predicted complexes; and (E) two AlphaFold3-predicted structures, with the lowest ipTM on the left and the highest ipTM on the right.
Case 3: Anti-PD-L1 antibody affinity ranking from sequence information
Following the approach in Case 2, PPBind-3D predicted affinity from AlphaFold3-predicted complex structures. Although the AlphaFold3 complexes had relatively high ipTM values, epitope and docking-pose deviations caused PPBind-3D to fail in this setting. PPBind-1D-Alignment performed best, reaching a Pearson correlation of 0.66, far exceeding the PPBind-1D-NoAlignment model. This highlights the robustness of the feature-aligned sequence model when AI-predicted structures are unreliable.

Figure 4. Correlations between predicted and experimental affinity for (A) PPBind-3D, (B) PPBind-1D-NoAlignment, and (C) PPBind-1D-Alignment; (D) box plots of ipTM and pTM for AlphaFold3-predicted complexes; and (E) the two AlphaFold3 structures with the highest ipTM in Case 3. The blue chain is PD-L1 and the green chain is the antibody.
The three cases show that PPBind-3D is preferred when an experimentally resolved structure is available, whereas PPBind-1D is a more reliable high-throughput screening option when AlphaFold-predicted structure quality is insufficient. In practical virtual screening and candidate-ranking tasks, the core advantage of PPBind is its ability to distinguish high- and low-affinity molecules and prioritize candidates in early drug discovery.
The team also described the current limitations of the models. PPBind-3D accuracy depends strongly on complex-structure quality, and predictions should be interpreted cautiously when AlphaFold-predicted structures have low ipTM. Feature alignment improves the overall ranking ability of PPBind-1D, but absolute affinity values for individual samples may still contain bias. Absolute affinity prediction remains a shared challenge in the field because of experimental noise, dataset size, and other factors.
As protein-structure prediction technologies such as AlphaFold continue to improve, the application scope of PPBind-3D without experimentally resolved structures is expected to expand. With the RDE module, PPBind-3D is particularly suitable for rapidly evaluating large numbers of antibody-engineering mutants. Continued expansion of experimentally measured affinity datasets will further unlock the potential of the PPBind model family.
Research team
This study was led by Lance Han, Chief Innovation Officer & Founder of Cyagen Biosciences. An engineer turned entrepreneur in biotechnology, engineering, and artificial intelligence, Lance Han has developed interdisciplinary expertise across these fields. He earned a Bachelor of Engineering from Tsinghua University in 1988, followed by a Master of Engineering from McGill University in Canada. He was also a visiting student at MIT in the United States, where he conducted research on life sciences imaging equipment. This experience sparked his long-term interest in biotechnology and its applications.
In 2006, drawing on his academic background, professional experience in IT and engineering, and an MBA from Queen's University in Canada, Han founded Cyagen Biosciences. Under his leadership as CEO, Cyagen has grown into a provider of innovative in vitro and in vivo solutions for life sciences research and drug development. He also leads a research group at the Innovation Center for Artificial Intelligence of the Institute of Pearl River Delta of Tsinghua in Guangzhou, with work spanning medical imaging technologies and the application of large language models in bioinformatics.
Chen Peiyi and Liu Huaqing were co-first authors. The work was completed in collaboration with the Third Affiliated Hospital of Sun Yat-sen University, the Seventh Affiliated Hospital of Sun Yat-sen University, and Shenzhen Nanshan People's Hospital.
Cyagen Biosciences has developed an integrated wet–dry research and development system combining HUGO-Ab® fully human antibody mice with AbSeek® intelligent antibody computational platform, covering target validation, antibody screening, sequence optimization, and in vitro and in vivo validation. The PPBind-3D/1D general protein-affinity prediction AI framework was extensively validated in an antibody drug development setting, strengthening computational prediction from sequence to affinity and helping reduce early-stage screening costs and development time for antibody and protein drugs.
Cyagen Biosciences continues to promote industry–academia collaboration and welcomes discussions with research institutes and biopharmaceutical companies working on protein engineering and novel antibody drug development.
Publication information
PPBind-3D/1D: Protein-Protein Binding Affinity Prediction for Structural and Sequence Data, published in Antibody Therapeutics by Oxford University Press.


