{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# Graph Convolutional Prediction of Protein Interactions in Yeast" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "*(This demo is a part of [Deep Learning for Network Biology](http://snap.stanford.edu/deepnetbio-ismb/) tutorial.)*" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "In this example, we demonstrate the utility of deep learning methods for an important prediction problem on biological graphs. In particular, we consider the problem of predicting [protein-protein interactions](https://en.wikipedia.org/wiki/Protein%E2%80%93protein_interaction) (PPIs). \n", "\n", "Protein-protein interactions (PPIs) are essential to almost every process in a cell. Understanding PPIs is crucial for understanding cell physiology in normal and disease states. Furthermore, knowledge of PPIs can be used:\n", "* for drug development, since drugs can affect PPIs,\n", "* to assign roles (i.e., protein functions) to uncharacterized proteins,\n", "* to characterize the relationships between proteins that form multi-molecular complexes, such as the proteasome.\n", "\n", "We represent the totality of PPIs that happen in a cell, an organism or a specific biological context with a [protein-protein interaction network](https://www.nature.com/subjects/protein-protein-interaction-networks). These networks are mathematical representations of all physical contacts between proteins in the cell.\n", "\n", "The development of large-scale PPI screening techniques, especially [high-throughput affinity purification combined with mass-spectrometry](https://doi.org/10.1016/j.tibtech.2016.02.014) and the [yeast two-hybrid assay](https://en.wikipedia.org/wiki/Two-hybrid_screening), has caused an explosion in the amount of PPI data and the construction of ever more complex and complete interaction networks. For example, the figure below is a graphical representation of three different types of protein-protein interaction networks in [yeast *S. cerevisiae*](https://en.wikipedia.org/wiki/Saccharomyces_cerevisiae). The structure of the binary interaction network is obviously different from the structure of the co-complex interaction network. The network structure of the literature-curated dataset resembles that of the co-complex dataset, even though the literature-curated datasets are reported to contain mostly binary interactions.\n", "\n", "However, current knowledge of protein-protein interaction networks is both [*incomplete* and *noisy*](https://doi.org/10.1016/j.cell.2014.10.050), as PPI screening techniques are limited in how many true interactions they can detect. Furthermore, PPI screening techniques often have high false positive and negative rates. These limitations present a great opportunity for computational methods to predict protein-protein interactions. \n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "