Bonneau lab group meeting will take place on
Monday, June 7, 2021
10:00am ET
Virtual via zoom
Presenter:
Daniel Berenberg, Ph.D. Candidate, New York University
Towards scalable protein structural comparison using graph convolutional networks
Assessing the similarity between protein folds has a variety of applications in the lab including fold detection and discovery as well as function prediction. Existing methods in detecting fold similarity are either hindered by computational complexity or limited by ontological labeling. Standard methods such as TM-score are plagued by the former, requiring a sequence alignment whose runtime is dependent on the lengths of the input proteins, while recent deep learning methods often suffer from the latter - supervised algorithms simply learn to cluster into an existing space of structural labels, excluding the option of fold discovery without heuristics. To address these issues, our proposed solution is to learn intrinsic structural features in a self-supervised way, in particular by learning to extract so-called 'mesoscale' characteristics of the structure - features that express co-location of secondary structural elements common throughout all folds in the dataset. To do so, we re-formulate the optimization problem as a community detection task: assign nodes of the contact map to a set of classes so as to maximize the network modularity. In this talk, I will present some results of the proposed method.
Monday, June 7, 2021
10:00am ET
Virtual via zoom
Presenter:
Daniel Berenberg, Ph.D. Candidate, New York University
Towards scalable protein structural comparison using graph convolutional networks
Assessing the similarity between protein folds has a variety of applications in the lab including fold detection and discovery as well as function prediction. Existing methods in detecting fold similarity are either hindered by computational complexity or limited by ontological labeling. Standard methods such as TM-score are plagued by the former, requiring a sequence alignment whose runtime is dependent on the lengths of the input proteins, while recent deep learning methods often suffer from the latter - supervised algorithms simply learn to cluster into an existing space of structural labels, excluding the option of fold discovery without heuristics. To address these issues, our proposed solution is to learn intrinsic structural features in a self-supervised way, in particular by learning to extract so-called 'mesoscale' characteristics of the structure - features that express co-location of secondary structural elements common throughout all folds in the dataset. To do so, we re-formulate the optimization problem as a community detection task: assign nodes of the contact map to a set of classes so as to maximize the network modularity. In this talk, I will present some results of the proposed method.
