[CS] Songhao Jiang MS Presentation/July 7, 2021

[nitayack at cs.uchicago.edu]

This is an announcement of Songhao Jiang's MS Presentation.
===============================================
Date: Wednesday, July 07, 2021

Time: 9:00 am CST

Location: https://uchicago.zoom.us/j/95670990073?pwd=UWFoaHpBYTVNZURPWjJsL0VmbTE4UT09

The password is:
496014

M.S. Candidate: Songhao Jiang

M.S. Paper Title: Protein Point Cloud

Abstract: Point cloud is one of the most significant data format for 3D representation. However, the unstructuredness of point clouds poses a challenge for deep learning. This is because deep neural networks, specifically convolutional models, require a structured grid to best leverage spatial information. As a result, point cloud methods are not yet competitive for protein problems even though they seem a natural fit. Most approaches in this area today sidestep this challenge by transforming point clouds to structured 3D voxel grids, which benefits from convolution operations at the cost of resolution loss, conversion artifacts, and increased computational cost. In this paper, we design a novel model architecture that directly consumes proteins atomic coordinates as point clouds. Our network, consisting of nearest-neighbor convolutions and self-attention, is designed to represent both surface and interior protein structure and connects local and global features in a more effective way. We evaluate this model on three representative prediction problems in protein secondary structure, solvent accessible surface area and binding site. We demonstrate outperformance and better stability in all three benchmarks over the state-of-the-art point cloud and sequence models.


Advisor: Rick Stevens

Committee Members: Ian Foster, Rick Stevens, and Fangfang Xia