Puxano

Choosing an AI model for antibody–antigen structure prediction

10 November 2026

At PUXANO, we specialize in structure-guided antigen design of challenging (membrane) protein targets for vaccine and antibody development. Our discovery platform combines cutting-edge computational tools with rigorous experimental validation, including protein expression, purification, biophysical analysis, and protein structure determination. A core focus of our development objectives is building better AI models for antibody–antigen prediction. Central to this effort is the integration of experimental data alongside protein sequence information to continuously enhance predictive power (cfr. FF+ innovation project, HARMONY). 
We recently investigated how well current AI-based structure prediction models predict the interaction between antibodies and their targets. To do so, we benchmarked three models — Boltz-2 [1], Protenix-v2 [2] and OpenDDE v1.1.0 [3]— on a dataset of antibody-antigen complexes, with 22,650 predictions in total. We observed clear differences in performance between models, and our data demonstrated the importance of critically evaluating the metrics that are used to rank prediction output. Our findings also highlight why experimental validation remains indispensable in drawing robust biological conclusions. 
 
Building the right benchmark dataset
To setup a proper benchmarking dataset, we filtered SAbDab2 [4]—a collection of all antibody structures in the Protein Data Bank. We applied several experimental-quality filters [5], to ensure that the experimental structures were of sufficient quality to serve as reliable references. We selected 302 antibody and nanobody complexes from 21,914 SAbDab2 records. Targets were released on or after 1 June 2023, the latest training cutoff date of the models being benchmarked, to avoid testing them on structures they may have already encountered during training. Each retained complex contained one antibody or nanobody, and one antigen chain (isolated, when necessary, from a larger antigen assembly). 
We also used Foldseek [6] to check whether any complexes in our dataset were similar to those released before the training cutoff. This additional step aimed to remove examples that, despite being released after the cutoff, closely resembled ones the models may have already encountered during training. Including these could artificially inflate the models’ apparent accuracy, giving a misleading impression of how well they perform on genuinely novel structures [7]. Matches with interface LDDT ≥0.70 and antibody–antigen interface coverage ≥0.50 were removed [8]; structural clustering then selected one representative per cluster, to ensure that the final dataset contained structurally independent examples and avoid biasing the benchmark towards similar complexes. Of course, we cannot guarantee that our benchmark is entirely free from training-data overlaps, as undisclosed training sets may exist.
 
Comparing models under a defined protocol
We generated 25 predictions per antibody-antigen complex for Boltz-2, Protenix-v2, and OpenDDE. Every model completed all targets. To assess the quality of the predictions, we used several metrics: pDockQ, pDockQ2, and ipSAE. DockQ [9][10] measures agreement with experimental structures on a zero-to-one scale. A successful prediction has a DockQ ≥0.23, with scores ≥0.49 indicating medium accuracy and scores ≥0.80 indicating high accuracy. Oracle selection uses the experimental structure to identify the highest-DockQ candidate among our 25 predictions per target. Unlike DockQ, pDockQ, pDockQ2, and ipSAE [11] do not require experimental structures. Instead, they assess the quality of predictions using information generated by the models themselves, allowing us to rank candidates without knowing their actual accuracy. 
OpenDDE achieved the highest success rate, followed closely by Protenix-v2, while Boltz-2 performed worse. This trend was observed both with oracle selection (63.6%, 59.6%, and 27.8%, respectively) and ipSAE selection (53.3%, 48.3%, and 18.2%, respectively) (Figure 1).  For all three models, confidence-based selection sometimes failed to identify a successful prediction, even when one was available among the 25 candidates.
 
How consistently do the models succeed? 
When generating multiple predictions for the same target, OpenDDE was the most likely to succeed across all 25 runs, followed by Protenix-v2, while Boltz-2 failed considerably more often. All 25 candidates succeeded for 119 complexes with OpenDDE, for 83 complexes with Protenix-v2, and for 27 complexes with Boltz-2; every candidate failed for 110, 122, and 218 complexes, respectively (Figure 2). Across all models, 79 targets (26.2%) had no successful prediction among the executed 75 runs. Repeated sampling therefore produced different results across models and targets.
 
How much quality does ipSAE selection leave behind? 
ipSAE generally identified successful predictions when they were available, particularly for OpenDDE and Protenix-v2, but often failed to select the most accurate candidate. Among targets with at least one successful candidate—84 for Boltz-2, 180 for Protenix-v2, and 192 for OpenDDE—ipSAE retrieved a successful prediction in 65.5%, 81.1%, and 83.9% of cases, respectively. These percentages are conditional on a successful candidate existing. 
The DockQ gap is the difference between the best available and the ipSAE-selected DockQ (Figure 3). Median gaps were 0.0895, 0.056, and 0.036, respectively. Half of eligible OpenDDE targets therefore lost no more than 0.036 DockQ through ipSAE selection. Its exact-best recovery was only 4.7%, illustrating that rare recovery of the best candidate can coexist with successful selection.  
This distinction is particularly important for real-world scenarios, where experimental structures are often unavailable. In these cases, we must rely on confidence scores such as ipSAE to select predictions, making it essential to understand how reliably these scores identify accurate structures. 
 
What this means for antibody discovery 
OpenDDE and Protenix-v2 outperformed Boltz-2 in our experiment: The results support a separation of three model properties: generation of at least one successful complex, success stability across repeated samples, and selection of a successful candidate from the generated set. This distinction is important because, in real-world applications where experimental structures are unavailable, a model must not only be capable of generating accurate predictions but also produce them reliably and identify them using its own confidence scores. Further improvement of antigen-antibody prediction will not only require improving AI model architectures and training protocols, but also the development of better rankers (ipSAE, pDockQ).  
 
Finally, experimental structural biology methods, including cryo-EM, remain essential for generating the reference structures needed to develop and rigorously evaluate AI-based structure prediction models, and enable us to transfer from predicted hypotheses into experimentally validated structures. Here at PUXANO we are also investigating the feasibility of integrating additional readily available experimental data to  combine the strengths of both worlds. 
 
If you would be interested in PUXANO running benchmarking tests on your model, or in collaborating with PUXANO for antigen design or with biophysical and structural antibody-antigen validation, please do not hesitate to contact us via info@puxano.com. 
All conclusions apply to the tested SAbDab2-derived dataset, the recorded software versions, model checkpoints and chosen inference parameters. 
 
All work was conducted by Matteo Venturini, a data scientist in PUXANO’s computational team (Australian Office). He holds a Master’s degree in Statistics and Data Science from Hasselt University and is currently involved in the development of a novel AI-based model for antibody–antigen structure prediction.
 
ACKNOWLEDGMENTS
This project has received funding from the European High-Performance Computing Joint Undertaking (JU) under grant agreement No 101163317. The JU receives support from the Digital Europe Programme.  We acknowledge the EuroHPC Joint Undertaking (EuroHPC JU) for providing access to the Karolina supercomputer at IT4Innovations, Czech Republic (project ID: EHPC-REG-2025R01-096), and the JUPITER Booster at the Jülich Supercomputing Centre, Forschungszentrum Jülich, Germany (project ID: EHPC-AIF-2026PG01-1417)  
 
TECHNICAL INFORMATION
Benchmark instructions: our open-source AbAgBench pipeline lets researchers construct antibody–antigen datasets to benchmark their models. To reproduce the dataset underlying this analysis, download the mmCIF structures from the SAbDab website, provide a SAbDab summary CSV, and use the reported benchmark settings. Because the summary is updated, use our recorded sabdab_summary_all.csv. If the mmCIF files themselves have also changed, results may differ slightly from ours even with this summary.  Inference settings: Each complex was predicted using seeds 1998, 2998, 3998, 4998, and 5998, with five samples per seed and 200 diffusion steps, producing 25 candidates per model. Protein MSAs were prepared using ColabFold, and templates were disabled. All other inference parameters used the defaults of Boltz 2.2.1 through LMI4Boltz 1.0.0 [12], Protenix 2.0.0, and OpenDDE 1.1.0. These defaults refer to those specific versions and may differ in later releases. 
 

REFERENCES

  1. Passaro, S., et al. (2025). Boltz-2: Towards accurate and efficient binding affinity prediction. bioRxiv.
  2. Zhang, Y., et al. (2026). Protenix-v2: Broadening the Reach of Structure Prediction and Biomolecular Design. bioRxiv.
  3. Aureka AI OpenDDE project. (2026). Folding, Reasoning, and Scaling with Open-source Drug Discovery Engine. arXiv. Software: OpenDDE, version 1.1.0.
  4. Dunbar, J., et al. (2014). SAbDab: The structural antibody database. Nucleic Acids Research, 42(D1), D1140–D1146.
  5. Zhu, Z., Olson, K. S., & Magliery, T. J. (2024). 50 years of antibody numbering schemes: A statistical and structural evaluation reveals key differences and limitations. Antibodies, 13(4), 99.
  6. Van Kempen, M., et al. (2024). Fast and accurate protein structure search with Foldseek. Nature Biotechnology, 42(2), 243–246.
  7. Bushuiev, A., et al. (2024). Revealing data leakage in protein interaction benchmarks. arXiv.
  8. Kovtun, D., et al. (2024). PINDER: The protein interaction dataset and evaluation resource. bioRxiv.
  9. Basu, S., & Wallner, B. (2016). DockQ: A quality measure for protein-protein docking models. PLOS ONE, 11(8), e0161879.
  10. Mirabello, C., & Wallner, B. (2024). DockQ v2: Improved automatic quality measure for protein multimers, nucleic acids, and small molecules. Bioinformatics, 40(10), btae586.
  11. Dunbrack, R. L. (2025). Rēs ipSAE loquuntur: What’s wrong with AlphaFold’s ipTM score and how to fix it. bioRxiv.
  12. Litfin, T., Caley, J. S., & Michie, K. A. (2025). LMI4Boltz: Optimizing VRAM utilization to predict large macromolecular complexes with consumer grade hardware. bioRxiv.

Read also

13 October 2026

Events

Three Days of Connecting at Festival of Biologics 2026

From 13 to 15 October, Joao Ramos and Wies Lips represented PUXANO at the Festival of Biologics in Basel, Switzerland.

Read more...

4 June 2026

Events

PUXANO at Knowledge for Growth 2026

PUXANO attended Knowledge for Growth 2026 in Antwerp, one of Europe’s key life sciences conferences organised by Biovia. CEO Wouter Van Putte presented PUXANO’s work in antigen design services for vaccine and antibody development companies.
Read more...
Puxano

Grids

Explore collaborations

Antigen design

Membrane protein research

Protein structure determination

About

Resources

Technology

Contact

Sign Up for Our Newsletter

Stay updated on the latest in protein design, AI-driven engineering, and groundbreaking research. Subscribe now to get exclusive insights and updates.