PRO‐LDM: A Conditional Latent Diffusion Model for Protein Sequence Design and Functional Optimization
Sitao Zhanga, Zixuan Jiangb, Rundong Huangc, Wenting Huanga b, Siyuan Penga, Shaoxun Moa, Letao Zhua, Peiheng Lid, Ziyi Zhangb, Emily Pane, Xi Chenf, Yunfei Longf, Liang Qi
- aState Key Laboratory of Microbial Metabolism School of Life Sciences and Biotechnology Shanghai Jiao Tong University Shanghai 200240 China
- bFunction Hub The Hong Kong University of Science and Technology (Guangzhou) Guangdong 511453 China
- cDepartment of Mathematics School of Computation Information and Technology Technical University of Munich 80333 Munich Germany
- dElectronic and Information Engineering School of Electronic Information Engineering Beihang University Beijing 100191 China
- eThe Lawrenceville School Lawrenceville NJ 08648 USA
- fXtalPi Inc. New Area Free Trade No. 1 Life Science and Technology Industrial Park Phase 1 No. 2‐1, Pudong District Shanghai 200120 China
- gResearch Center for Intelligent Computing Platforms Zhejiang Lab Hangzhou Zhejiang 311121 China
- ✉Corresponding author
Papers open inside SciRah from the original publisher or repository; SciRah never hosts them.
Abstract
The diffusion model has grasped enormous attention in the computer vision field and emerged as a promising algorithm in protein design for precise structure and sequence generation. Here PRO-LDM is introduced: a modular multi-tasking framework combining design fidelity and computational efficiency, by integrating the diffusion model in latent space. The model learns biological representations at local and global levels, to design natural-like species with enhanced diversity, or optimize protein properties and functions. Its modular nature also enables the integration with alternative pre-trained encoders for enhanced generalization capability. Outlier design can be implemented by adjusting the classifier-free guidance that enables PRO-LDM to sample vastly different regions in the latent space. The approach is demonstrated in generating a novel green-fluorescence-protein variant with notably enhanced fluorescence in multiple working scenarios along with increased solubility and stability. The model provides a versatile tool to effectively extract physicochemical and evolutionary information in sequences for designing new proteins with optimized performances.