AI Improves Structural Modeling of Protein-Protein Interactions

2020-01-15

Proteins are often referred to as the working molecules of the human body. An adult body has over 20,000 different types of proteins, each involved in many essential functions of human life. Many important cellular processes also involve the physical interaction of proteins. Therefore, determining the quaternary structure of proteins provides important insights into the molecular mechanisms of complex function.


To better understand the interactions of proteins in the body, researchers at Purdue University have designed a novel method using deep learning to gain a better understanding of how proteins interact in the body. This paves the way for accurate structural models of protein interactions related to various diseases and for the design of better drugs targeting protein interactions. In order to understand the molecular mechanisms of protein complex function, biologists have been using experimental methods such as X-rays and microscopy, but this is a time-consuming and resource-intensive work. Bioinformatics researchers in Purdue University's research laboratory and other institutions have been developing computational methods for simulating protein complexes. A major challenge is that a computational method typically generates thousands of models, and selecting the correct model or ranking the models can be difficult."


We have developed a voxel-based deep neural network (DOVE) method called DOcking bait selection to evaluate protein docking models. To evaluate protein docking models, DOVE uses 3D voxel scanning of protein-protein interfaces and considers atomic interaction types and their energy contributions as input features for the neural network. The deep learning model has been trained and validated on docking models available in the ZDock and DockGround databases. In various combinations of tested functions, almost all functions outperformed existing scoring functions.


DOVE, created by researchers at Purdue University, captures the structure and energy features of protein docking model interfaces with 3D grids and uses a 3D convolutional neural network to determine whether the model is correct or incorrect. Image source: Daiki Muhara/Purdue University


Researchers state, "Our work represents a significant advancement in the field of bioinformatics, potentially the first successful use of deep learning and 3D features to rapidly assess the validity of certain protein models. This information can then be used to create targeted drugs that disrupt interactions between certain proteins."


The researchers have also released their code on GitHub, along with an online demo, allowing anyone to try out the code.

RECENT AI NEWS

RECENT AI TOOLS