Steven (Szu-Han) Chen
Dept. Of Electrical Engineering & Computer Science
Steven Chen is a 4th year student in cognitive and computer science at York University. Having formal training in machines learning, Steven is spending the summer apply deep learning on protein images taken from cryo-electron microscopy with Dr. Marcus Brubaker. Specifically, Steven will be building deep convolutional neural networks to apply image super resolution on protein structures. By the end of the summer, Steven is hoping to have furthered the research to gain a better picture of computer vision and deep learning. The project is important because being able to transform low resolution protein structures to high resolution ones can provide greater understanding of structural biology and improve drug discovery.
Volumetric Super-Resolution for Learning Detailed Protein Structure Prediction
The goal of this research project is to predict high-resolution detail of a protein structure from low-resolution one. In computer vision when working with images, this technique is commonly known as super-resolution. For instance, when zooming in on an image, normally the image becomes blurry or pixelated but the goal with super-resolution is to create new details that look natural but may not exist in the original image. The basis of my research is to transfer this concept to the domain of structural biology, particularly exploring the use of super-resolution methods on atomic resolution molecular structures. In the case of molecular structures, the regularity of atomic structure suggests that super-resolution in this domain should be plausible as the existence of these structural motifs may be predictable at lower resolutions. One approach to this problem is through the use of deep learning, specifically the application of convolutional neural networks. My research so far has been focused on generating training data from the protein images stored in the online protein databank where I am using them to creating high resolution structures and corresponding low-resolution structures through blurring and down sampling. Currently I am enhancing previous state-of-the-art architectures of deep convolutional networks for image super-resolution and I will be exploring variations on the network architecture and problem formulation to see their influence on the final performance. If the results are promising, this technique can be applied to help biologists study atomic resolution molecular structures and we can integrate this into a fully functional solution.