[1] ProDualNet: Dual-Target Protein Sequence Design Method Based on Protein Language Model and Structure Model, Briefings in Bioinformatics, 2025.
After several months of validation and trial-and-error, we have observed that the mean architecture does not appear to be the optimal choice. The cross_attention architecture seems to perform better in certain scenarios, such as in some agonist design cases. Unfortunately, we no longer have the resources to conduct further in vitro experimental validation. Although we suspect few may read this, it must be said: peptide drug design is truly challenging.
We have recently updated a preprint, primarily focused on peptide drug design, with the aim of exploring how to better leverage existing tools to meet the more urgent design demands in industrial development. Due to experimental cost constraints, we were unable to establish a comprehensive experimental platform encompassing protein, cellular, and animal models for rigorous quantitative analysis.
Overall, for current structure-based design and screening tools, adopting a multi-path integrated design strategy can help improve success rates across different experimental stages. For example, in the design of PTHR agonists, the best-performing design path in the zebrafish efficacy model failed completely in rat PK studies, while another design path yielded positive hits [2]. We conducted this research in collaboration with Intelligent Medicine Original (Shanghai) Co., Ltd., Shanghai, China.
[2] Ensemble Strategy for Robust AI-Driven Peptide Design，2025.
Important update: We have successfully designed a dual agonist using this model and are currently progressing with efficacy experiments. Please wait for our preprint—August 1, 2025.
Install Python>=3.0, PyTorch, Numpy.
-
The main folder includes the execution code and test cases for ProdualNet. You can use it to design dual-target protein sequences, such as GLP-1/GCGR dual agonists, or proteins designed to bind with different receptors, causing conformational changes, using the weights produalnet_02.pt.
-
The mutation_task folder is for a zero-shot protein function prediction task, including thermal stability and DDG.
Similar to recent work on conformational bias in mutations.
- The baseline folder on this project contains a modified multi-state design model based on ProteinMPNN, supporting multiple target protein sequence design and multiple protein complex conformations sequence design.
- The current model only supports the design of natural amino acids.
- You may not use the material for commercial purposes.
This project is based on ProteinMPNN/Pifold/esm/BERT-pytorch, under their License.
Source: https://github.com/dauparas/ProteinMPNN, https://github.com/A4Bio/PiFold, https://github.com/facebookresearch/esm, https://github.com/codertimo/BERT-pytorch/tree/master