Network medicine framework for identifying drug repurposing opportunities
Methods and systems for generating drug repurposing predictions for a disease caused by a pathogen, such as a novel pathogen, are provided. A multi-modal system includes a protein-protein interaction network (PPI), a graph neural network (GNN), a diffusion module, a proximity module, and an aggregation module. The GNN is configured to predict new edges between candidate drug nodes and disease nodes in an embedded representation of the PPI to produce a decoded embedding space. The diffusion module is configured to determine a proximity distance for pairs of nodes in the PPI, and the proximity module is configured to determine a proximity distance for pairs of nodes in the PPI, each pair comprising a pathogen-protein node and a drug-protein node. A ranked list of candidate drugs predicted to be effective in treatment of the disease based on candidate drug lists generated by the other modules is generated by the aggregation module.
1 . A multi-modal system for generating drug repurposing predictions for a disease caused by a pathogen, comprising:
a protein-protein interaction network defining pathogen-protein interactions for the pathogen and drug-protein interactions for a plurality of candidate drugs;
a graph neural network comprising an embedded representation of the protein-protein interaction network, the embedded representation including candidate drug nodes and disease nodes, the graph neural network configured to predict new edges between the candidate drug nodes and disease nodes to produce a decoded embedding space, a first list comprising a subset of the plurality of candidate drugs being identifiable from the decoded embedding space; and
a processor, the processor configured to:
determine a diffusion metric for pairs of nodes in the protein-protein interaction network, each pair comprising a pathogen-protein node and a drug-protein node, a second list comprising a subset of the plurality of candidate drugs being identifiable from the determined diffusion metric;
determine a proximity distance for pairs of nodes in the protein-protein interaction network, each pair comprising a pathogen-protein node and a drug-protein node, a third list comprising a subset of the plurality of candidate drugs being identifiable from the determined proximity distances; and
generate a ranked list of candidate drugs predicted to be effective in treatment of the disease based on the first, second, and third lists.
2 . The system of claim 1 , wherein the processor is further configured to generate the ranked list based on a consensus ranking of the first, second, and third lists.
3 . The system of claim 1 , wherein the first list comprises at least two sub-lists, each sub-list generated based on varying decoding parameters applied to the decoded embedding space.
4 . The system of claim 1 , wherein the graph neural network is untrained for the pathogen.
5 . The system of claim 4 , wherein the graph neural network is a graph convolutional neural network trained by a zero-shot learning strategy.
6 . The system of claim 1 , wherein the diffusion metric is a diffusion state distance between nodes, a divergence between vector representations of nodes, or a combination thereof.
7 . The system of claim 6 , wherein the second list comprises at least two sub-lists, each sub-list generated based on varying distance or divergence parameters.
8 . The system of claim 1 , wherein the third list comprises at least two sub-lists, each sub-list generated based on varying drug-inclusion criteria.
9 . The system of claim 1 , wherein the protein-protein interaction network is a human interactome.
10 . A multi-modal method for generating drug repurposing predictions for a disease related to a pathogen, comprising:
with a protein-protein interaction network defining pathogen-protein interactions for the pathogen and drug-protein interactions for a plurality of candidate drugs:
in a graph neural network comprising an embedded representation of the protein-protein interaction network, the embedded representation including candidate drug nodes and disease nodes:
predicting new edges between the candidate drug nodes and disease nodes to produce a decoded embedding space, and
identifying a first list comprising a subset of the plurality of candidate drugs being from the decoded embedding space;
by a processor:
determining a diffusion metric for pairs of nodes in the protein-protein interaction network, each pair comprising a pathogen-protein node and a drug-protein node,
identifying a second list comprising a subset of the plurality of candidate drugs from the determined diffusion metric;
determining a proximity distance for pairs of nodes in the protein-protein interaction network, each pair comprising a pathogen-protein node and a drug-protein node,
identifying a third list comprising a subset of the plurality of candidate drugs from the determined proximity distances; and
generating a ranked list of candidate drugs predicted to be effective in treatment of the disease based on the first, second, and third lists.
11 . The method of claim 10 , wherein generating the ranked list includes consensus ranking of the first, second, and third lists.
12 . The method of claim 10 , wherein identifying the first list comprises identifying at least two sub-lists, each sub-list generated based on varying decoding parameters applied to the decoded embedding space.
13 . The method of claim 10 , wherein the graph neural network is untrained for the pathogen.
14 . The method of claim 13 , wherein the graph neural network is a graph convolutional neural network trained by a zero-shot learning strategy.
15 . The method of claim 10 , wherein the diffusion metric is a diffusion state distance between nodes, a divergence between vector representations of nodes, or a combination thereof.
16 . The method of claim 15 , wherein identifying the second list comprises identifying at least two sub-lists, each sub-list generated based on varying distance or divergence parameters.
17 . The method of claim 10 , wherein identifying the third list comprises identifying at least two sub-lists, each sub-list generated based on varying drug-inclusion criteria.
18 . A method of improving accuracy of drug repurposing predictions for a disease caused by a pathogen, comprising:
with a protein-protein interaction network defining pathogen-protein interactions for the pathogen and drug-protein interactions for a plurality of candidate drugs:
in a graph neural network comprising an embedded representation of the protein-protein interaction network, the embedded representation including candidate drug nodes and disease nodes:
predicting new edges between the candidate drug nodes and disease nodes to produce a decoded embedding space, and
identifying a first list comprising a subset of the plurality of candidate drugs being from the decoded embedding space;
by a processor:
determining a diffusion metric for pairs of nodes in the protein-protein interaction network, each pair comprising a pathogen-protein node and a drug-protein node,
identifying a second list comprising a subset of the plurality of candidate drugs from the determined diffusion metric;
determining a proximity distance for pairs of nodes in the protein-protein interaction network, each pair comprising a pathogen-protein node and a drug-protein node,
identifying a third list comprising a subset of the plurality of candidate drugs from the determined proximity distances; and
generating a consensus ranking of candidate drugs predicted to be effective in treatment of the disease based on the first, second, and third lists, the consensus ranking providing for improved predictive power over each of the first, second, and third lists.