Application of Deep Learning in Protein Structure Prediction and Its Inspirations

The goal of protein structure prediction is to determine, usually based on computer simulations or calculations, the three dimensional structure from a given amino acid sequence. Protein structure prediction is important since the 3D protein structure will further determine its biological functions. Nevertheless, traditional prediction method based on physics can only effectively deal with short proteins with the low accuracy. In the past decade, data-driven methods and methods based on genetic knowledge have become popular. This review covers several important developments about the deep-learning methods on protein prediction in the past ten years. Considering the education background of readers of the journal, we will first present a self-consistent but concise introduction about the prerequisite concepts and methods, related with genetic information and basis of deep learning, to understand the deep-learning methods for the protein structure prediction. The prerequisite concepts and methods include position-specific scoring matrix (PSSM), multiple sequence alignment (MSA), contact map, distogram, protein data bank (PDB), critical assessment of protein structure prediction (CASP), template modelling score (TM score), universal approximation theorem and several important types of neural network closely related with protein structure prediction. Then, we compared the two most popular methods employed in the data-driven protein structure prediction, template-based method and template-free method. As for the deep-learning methods, the AlphaFold method from Deepmind will be specially discussed, which has achieved the prediction accuracy comparable to median or low experimental accuracy that even rendered some people to think that it has resolved the protein structure prediction problem to some extent. Nevertheless, all the above methods are too "overwhelming" and not friendly for beginners. Therefore, this review also introduced a simplest structure prediction problem, HP protein prediction problem, together with the corresponding deeplearning solution, strongly-correlated neural network, to novices in this area. [GRAPHICS]