Knowledge-based machine learning methods for macromolecular 3D structure prediction

TL;DR

This paper discusses the development of knowledge-based machine learning methods to improve the prediction of 3D structures of macromolecules like proteins and RNA, addressing data scarcity and computational challenges.

Contribution

It introduces a novel knowledge-based machine learning approach that incorporates biological knowledge and constraints to enhance 3D structure prediction accuracy and efficiency.

Findings

Knowledge constraints reduce solution space significantly.

Improved accuracy in structure prediction.

Enhanced computational efficiency.

Abstract

Predicting the 3D structure of a macromolecule, such as a protein or an RNA molecule, is ranked top among the most difficult and attractive problems in bioinformatics and computational biology. Its importance comes from the relationship between the 3D structure and the function of a given protein or RNA. 3D structures also help to find the ligands of the protein, which are usually small molecules, a key step in drug design. Unfortunately, there is no shortcut to accurately obtain the 3D structure of a macromolecule. Many physical measurements of macromolecular 3D structures cannot scale up, due to their large labor costs and the requirements for lab conditions. In recent years, computational methods have made huge progress due to advance in computation speed and machine learning methods. These methods only need the sequence information to predict 3D structures by employing various mathematical models and machine learning methods. The success of computational methods is highly dependent on a large database of the proteins and RNA with known structures. However, the performance of computational methods are always expected to be improved. There are several reasons for this. First, we are facing, and will continue to face sparseness of data.Secondly, the 3D structure space is too large for our computational capability. The two obstacles can be removed by knowledge-based methods, which combine knowledge learned from the known structures and biologists' knowledge of the folding process of protein or RNA. In the dissertation, I will present my results in building a knowledge-based method by using machine learning methods to tackle this problem. My methods include the knowledge constraints on intermediate states, which can highly reduce the solution space of a protein or RNA, in turn increasing the efficiency of the structure folding method and improving its accuracy.