Upcoming Events

Academic Report:Prediction of Constrained Protein and Complex Structures Based on Deep Learning

Time: September 12, 2023(Tue.) 10:00 AM

Location: Conference Room, 18th Floor, Building M 

Presenter: Liu Sirui, Researcher from Changping Laboratory

Proteins, as an important class of biological macromolecules, their monomer and complex structures are significant for understanding their structure-activity relationships and drug design, among other downstream applications. Although deep learning structure prediction methods such as AlphaFold and AlphaFold-Multimer are efficient, they may not always be consistent with experimental observations; on the other hand, experimental techniques such as NMR face challenges in identifying and constructing structures that are time-consuming and difficult to learn. The symbiosis of experimental techniques and AI methods may be a way to address these issues. For single protein prediction and analysis, we developed the RASP model, which directly utilizes experimental or empirical constraint information to assist structure prediction. With the assistance of constraint information, RASP can better predict the structures of multi-domain proteins and proteins lacking MSA. Its score can also reflect the quality of constraint well. Based on RASP, we developed the NMR NOE peak assignment workflow FAAST, which shortens the NOE spectrum peak assignment time to minutes to hours, and provides high-quality structural ensembles and peak assignment results in return, thus using AI methods to help and accelerate experimental analysis. In complex systems, we developed ColabDock to improve the accuracy of interface prediction. The method adopts a "generation-prediction" architecture and achieves structure prediction assisted by various constraint signals through re-engineering the ColabDesign framework and training ranking methods. ColabDock achieves interface prediction results that surpass HADDOCK and ClusPro on simulated residue and interface constraints, and performs well on applying NMR CSP and covalent labeling real experimental data. It can also improve the accuracy of antigen-antibody interface prediction with simulated antigen-antibody DMS constraint assistance. ColabDock provides a unified framework for integrating various sparse constraint signals, and the predicted framework itself can be applied to more constraint forms.