Towards an AI-based solution of protein structure prediction problem

Yang Zhang

Department of Computer Science at School of Computing; Department of Biochemistry at Yong Loo Lin School of Medicine; Cancer Science Institute of Singapore, National University of Singapore

The past decade has witnessed revolutionary changes in computer-based protein structure prediction, mainly driven by the artificial intelligence (AI) and deep neural-network learning techniques. In this talk, we discuss structure prediction results in recent community-wide CASP experiments, showing that new deep-learning approaches, built on coevolution data from multiple sequence alignments, can result in consistent and successful folding of large proteins with complicated topologies. Of note, AlphaFold2 trained through end-to-end transformer networks could fold nearly all protein domains in CASP14 with 2/3 of them having accuracy comparable to low-resolution experimental solutions. In the most recent CASP experiments, new advancements over AlphaFold2 have been made by integrating end-to-end learning and protein language models with fragment-assembly simulations. These achievements essentially break through the 50-years-old barrier of homology-based modelling and marked a solution, at least at the fold-level, to the single-domain protein structure prediction problem. Nevertheless, constructions of atomic-resolution models for multi-domain and higher-order quaternary protein structures remain challenging. Furthermore, the black-box nature of AI-based approaches remains a barrier to unravelling protein folding dynamics. Given the power of AI and rapid advancement of the field, it is expected that these challenges should be addressed in a foreseeable future by coupling deep learning techniques and metagenome sequencing databases, with the aid of advanced structure assembly simulation algorithms.

Go back