dyAb: Flow Matching for Flexible Antibody Design with AlphaFold-driven Pre-binding Antigen
Authors: Cheng Tan, Yijie Zhang, Zhangyang Gao, Yufei Huang, Haitao Lin, Lirong Wu, Fandi Wu, Mathieu Blanchette, Stan Z. Li
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that dy Ab significantly outperforms existing models in antibody design involving changing antigen conformations. These results highlight dy Ab s potential to streamline the design process for therapeutic antibodies, promising more efficient development cycles and improved outcomes in clinical applications. We conduct comprehensive experiments that take into account the dynamic nature of antigen conformational changes. This novel experimental setting aims to emulate the flexible conditions under which antibodies must function. While the dataset split is consistent with previous works (Kong, Huang, and Liu 2023b) for a fair comparison, we replace the pre-binding antigen structures with Alpha Fold2 s predictions and predict the ground-truth antigenantibody complexes. We assess the models capabilities in antibody design through three critical tasks: (i) Epitope-binding CDR-H3 generation, assessing the ability to generate the highly variable and functionally crucial CDR-H3 region. (2) Affinity optimization, evaluating the improvements in binding affinity of the designed antibodies to target antigens. (3) Complex structure prediction, predicting the 3D structure of the antibody-antigen complex. Moreover, a detailed ablation study of our proposed dy Ab model is conducted to validate its effectiveness. We summarize the results in Table 1, showing that dy Ab outperforms other methods across various metrics. |
| Researcher Affiliation | Collaboration | 1Zhejiang University 2AI Lab, Westlake University 3School of Computer Science, Mc Gill University 4MILA Qu ebec AI Institute 5Tencent, AI for Life Science Lab |
| Pseudocode | No | The paper describes the methodology using prose, mathematical equations, and figures, but it does not contain a clearly labeled 'Pseudocode' or 'Algorithm' block. |
| Open Source Code | Yes | Code https://github.com/A4Bio/dy Ab |
| Open Datasets | Yes | We train the models on the Structural Antibody Database (SAb Dab) (Dunbar et al. 2014) and evaluate them on the RABD Benchmark (Adolf Bryfogle et al. 2018), which contains 60 antibody-antigen complexes annotated by biological experts. we quantify binding affinity changes using the metric G on the SKEMPI dataset |
| Dataset Splits | Yes | While the dataset split is consistent with previous works (Kong, Huang, and Liu 2023b) for a fair comparison, we replace the pre-binding antigen structures with Alpha Fold2 s predictions and predict the ground-truth antigenantibody complexes. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments (e.g., GPU models, CPU models, or memory specifications). |
| Software Dependencies | No | The paper mentions various methods and models (e.g., Alpha Fold2, Rosetta, Ig Fold), but it does not specify the version numbers for any software libraries, programming languages, or development environments used in their own implementation. |
| Experiment Setup | Yes | Although the default sampling step is set to 10, we also assess performance using smaller steps of 1 and larger steps of 50. Insufficient sampling with smaller steps and suboptimal performance with larger steps indicate that the sampling size significantly influences performance. Nevertheless, the overall performance across different sampling sizes remains comparable to the default setting, suggesting that while sampling size is crucial, the chosen default size of 10 is effective. |