Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Rationalized All-Atom Protein Design with Unified Multi-Modal Bayesian Flow
Authors: Hanlin Wu, Yuxuan Song, Zhe Zhang, Zhilong Zhang, Hao Zhou, Wei-Ying Ma, Jingjing Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To validate, our method delivers consistently exceptional performance in both peptide and antibody design tasks. Our code, checkpoint, and designed PDBs can be found in https://github.com/Gen SI-THUAIR/Pro Bayes. ... We verify the effectiveness of Pro Bayes on two protein design tasks, including peptide design and antibody design. ... We report the results on Pep Bench and Pep BDB in Table 1 and Table 2 for seq-structure codesign and binding conformation generation task, respectively. ... We validate our design considerations through the following ablation experiments using the antibody design task. Results can be found in Tab. 4. |
| Researcher Affiliation | Academia | 1 Institute of AI Industry Research (AIR), Tsinghua University 2 Dept. of Comp. Sci. & Tech., Tsinghua University EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 Training ... Algorithm 2 Sampling ... Algorithm 3 Invariant point attention with quaternion implementation (IPAq) |
| Open Source Code | Yes | Our code, checkpoint, and designed PDBs can be found in https://github.com/Gen SI-THUAIR/Pro Bayes. ... We provide our code in https://github.com/Gen SI-THUAIR/Pro Bayes. |
| Open Datasets | Yes | The datasets used for evaluation include Pep Bench and Pep BDB [Wen et al., 2019]. Pep Bench is constructed from the Protein Data Bank [Berman et al., 2000], containing 6105 complexes and the LNR dataset Tsaban et al. [2022] is utilized as the test set with 93 complexes. Pep BDB is a protein-peptide complex dataset... We utilize the SAb Dab database [Dunbar et al., 2014] of antibody antigen complexes as our training dataset and evaluate model performance using the RAb D benchmark [Adolf-Bryfogle et al., 2018]. |
| Dataset Splits | Yes | Pep BDB is a protein-peptide complex dataset containing 8434, 370, and 190 items for training, validation, and test, respectively. ... the LNR dataset Tsaban et al. [2022] is utilized as the test set with 93 complexes. |
| Hardware Specification | Yes | All experiments in this paper are conducted on a node with 8 NVIDIA A100 80GB. Each training task requires 4 GPUs. |
| Software Dependencies | No | The paper does not explicitly state specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | For all experiments, the network is configured with a node embedding size of 128 and an edge embedding size of 64. ... The IPAq attention mechanism comprises 8 heads, with 8 query-key points and 12 value points for geometric attention. ... The entire encoder architecture consists of 6 stacked IPAq blocks... Finally, the network consists of 7.02 million parameters. For optimization, we set the learning rate to 5e-4 and use a batch size of 40 per distributed node with Adam optimizer. |