Peking : On October 27, Professor Xu Jinbo from Toyota Technological Institute at Chicago (TTIC) gave an academic lecture on Artificial Intelligence and Protein Folding as invited by the Peking University (PKU) Global Fellowship program.
Xu briefly introduced the background knowledge of the use of artificial intelligence to predict protein structure at the beginning of the lecture. He said that there are great challenges faced by traditional prediction methods, because the protein contains thousands of atoms with high degrees of freedom and the energy picture is not smooth, making it difficult to optimize.
Xu also reviewed the development history of protein structure prediction. He noted that the prediction approach before 2016 was inefficient and consumed a large amount of computing resources. To promote the efficiency, the scientists changed their research thinking: starting from the protein amino acid sequence that needs to be predicted, searching related databases to obtain multiple-sequence alignments (MSAs), and then obtaining the relationship matrix of amino acid residue pairs (such as contact matrix and distance matrix), and finally predict the structure.
In 2016, Xu’s team developed the RaptorX-Contact method based on ResNet. This method, for the first time, proved the feasibility of deep learning method on protein structure prediction. It can also be used to predict the structure of membrane protein and protein interaction.
From 2017 to 2019, Xu’s team successfully achieved a leap from contact matrix prediction to distance matrix prediction, making protein structure prediction more accurate.
Regarding the future development trend of artificial intelligence predicting protein structure, Xu believes that it will mainly focus on the better use of sequence and structural information and new deep learning network architectures. He also believes that the newly released AlphaFold2 is superior to other algorithms at the residue level. But for proteins with high molecular weight and multiple domains, it is still challenging to accurately predict the spatial position relationship between their domains.
After the lecture, Xu and the on-site teachers and students had a lively exchange and discussion on issues related to artificial intelligence prediction of protein structure.