PersGNN: Applying Topological Data Analysis and Geometric Deep Learning to Structure-Based Protein Function Prediction

Abstract

Understanding protein structure-function relationships is a key challenge in computational biology, with applications across the biotechnology and pharmaceutical industries. While it is known that protein structure directly impacts protein function, many functional prediction tasks use only protein sequence. In this work, we isolate protein structure to make functional annotations for proteins in the Protein Data Bank in order to study the expressiveness of different structure-based prediction schemes. We present PersGNN-an end-to-end trainable deep learning model that combines graph representation learning with topological data analysis to capture a complex set of both local and global structural features. While variations of these techniques have been successfully applied to proteins before, we demonstrate that our hybridized approach, PersGNN, outperforms either method on its own as well as a baseline neural network that learns from the same information. PersGNN achieves a 9.3% boost in area under the precision recall curve (AUPR) compared to the best individual model, as well as high F1 scores across different gene ontology categories, indicating the transferability of this approach.

Encoding protein dynamic information in graph representation for functional residue identification

“…3a). In addition to graphs, we calculate 1D and 2D persistence diagrams of C α atoms in each protein, the coarsened topological features measured by persistent homology 20,28 , and embed the diagrams into persistence images via Gaussian kernels and pixelated integrals 29 (see Methods). Our model takes input from three feature pipelines: (1) spatial information from the contact map where residue pairs are closer than cutoff distance r c such as 8 Å and 12 Å, (2) topological information from the persistence image, which is preprocessed as vectors of 625 dimensions, and (3) dynamic information based on the correlation map such that residue pairs with absolute correlation no less than 0.5 are connected with correlation edges if they are not in contact.…”

“…We also construct two datasets with different cutoff distances r c = 8 Å and r c = 12 Å to observe the effect of the cutoff on the classification performance. The choices of cutoff distance for graph-based protein feature extraction vary from approximately 6 to 15 Å in the literature 17,20,32 . Here we follow the cutoff 8 Å used by PersGNN 20 for a fair comparison.…”

“…The choices of cutoff distance for graph-based protein feature extraction vary from approximately 6 to 15 Å in the literature 17,20,32 . Here we follow the cutoff 8 Å used by PersGNN 20 for a fair comparison. Furthermore, to avoid the influence of imbalanced positive and negative samples, we calculate the balanced accuracy (arithmetic mean of the true positive rate and true negative rate) of the test set as the immediate metric of model performance.…”

“…We regard this phenomenon as the desired outcome because we expect models to capture as many functions as possible. and PersGNN 20 . We evaluate the trained classifiers by leveraging the threshold at the final sigmoid activation.…”

“…Traditional machine learning classifiers, such as support vector machines, random forests, and gradient boosted decision trees, have been extensively used to predict sitespecific or protein-level functions 15,16 . Many of the recent deep learning models focus on the variations in sequence embedding and neural network architecture [17][18][19][20] . One of the most natural choices for learning protein features is the graph neural network (GNN).…”