In this submission, we employ a hybrid approach integrating three models: OriginFlow, CUTEDGE, and ProDESIGN-LE.

OriginFlow[1] is a de novo protein design model based on flow-matching diffusion. It generates protein backbones with high designability, diversity, and functional relevance. The model uses a hybrid architecture that combines global invariant-point attention with local graph neural network (GNN) refinement, enhancing both long-sequence generation and structural detail. OriginFlow achieves a better balance among stability, novelty, and controllability compared to prior methods. A key distinguishing feature is its specialized capability in binder design. When tested on targets such as PD-L1, SARS-CoV-2 RBD, and VEGF, the authors reported approximately 90% wet-lab success in expression, solubility, and measurable affinity, with most binders exhibiting binding affinities in the micromolar KD range.

CUTEDGE (Chain-free Underinformed Targeted Electron Density Generative Explorer) is also a diffusion model for binder protein design. It takes electron density maps of potential binding sites as input and outputs electron densities of candidate binder proteins. This approach enables the construction of reliable local binding configurations with diverse amino acid compositions.

ProDESIGN-LE[2] is a deep learning–based method for protein sequence design, aiming to generate amino acid sequences that fold into a desired backbone structure. It operates on the principle that a protein will adopt a target structure if each residue is compatible with its local environment—defined by neighboring residues and their geometric relationships. ProDESIGN-LE uses a transformer model trained to predict the conditional probability of an amino acid given this local context. The algorithm iteratively selects positions, predicts optimal residues via the transformer, and updates the sequence until convergence, thereby maximizing the likelihood that all residues fit their structural environments.

We integrate these algorithms to target RBX1 as a binding protein. First, OriginFlow generates candidate binding backbones. CUTEDGE then designs local residue configurations with sufficient binding affinity. Finally, ProDESIGN-LE performs sequence redesign to ensure overall sequence consistency. The sampled designs range from 70 to 150 residues in length, and we have selected a total of 95 sequences for submission.

[1] Junyu Yan, Zibo Cui, Wenqing Yan, Yuhang Chen, Mengchen Pu, Shuai Li, and Sheng Ye. Robust and reliable de novo protein design: A flow-matching-based protein generative model achieves remarkably high success rates. bioRxiv, pages 2025–04, 2025.

[2] Bin Huang, Tingwen Fan, Kaiyue Wang, Haicang Zhang, Chungong Yu, Shuyu Nie, Yangshuo Qi, Wei-Mou Zheng, Jian Han, Zheng Fan, et al. Accurate and efficient protein sequence design through learning concise local environment of residues. Bioinformatics, 39(3):btad122, 2023.