Graph Neural Networks

GINet layer

Graph Interaction Networks layer

This layer is inspired by Sazan Mahbub et al. “EGAT: Edge Aggregated Graph Attention Networks and Transfer Learning Improve Protein-Protein Interaction Site Prediction”, BioRxiv 2020

1. Create edges feature by concatenating node feature

\[e_{ij} = LeakyReLu (a_{ij} * [W * x_i || W * x_j])\]

2. Apply softmax function, in order to learn to consider or ignore some neighboring nodes

\[\alpha_{ij} = softmax(e_{ij})\]

3. Sum over the nodes (no averaging here)

\[z_i = \sum_j (\alpha_{ij} * Wx_j + b_i)\]

Herein, we add the edge feature to the step 1)

\[e_{ij} = LeakyReLu (a_{ij} * [W * x_i || W * x_j || We * edge_{attr} ])\]

class deeprank_gnn.ginet.GINetConvLayer(*args: Any, **kwargs: Any)

reset_parameters()

forward(x, edge_index, edge_attr)

class deeprank_gnn.ginet.GINet(*args: Any, **kwargs: Any)

forward(data)

Fout Net layer

This layer is described by eq. (1) of “Protein Interface Predition using Graph Convolutional Network”, by Alex Fout et al. NIPS 2018

\[z = x_i * Wc + 1 / Ni Sum_j x_j * Wn + b\]

class deeprank_gnn.foutnet.FoutLayer(*args: Any, **kwargs: Any)

This layer is described by eq. (1) of Protein Interface Predition using Graph Convolutional Network by Alex Fout et al. NIPS 2018

Parameters

• in_channels (int) – Size of each input sample.

• out_channels (int) – Size of each output sample.

• bias (bool, optional) – If set to False, the layer will not learn an additive bias. (default: True)

reset_parameters()

forward(x, edge_index)

class deeprank_gnn.foutnet.FoutNet(*args: Any, **kwargs: Any)

forward(data)

sGraphAttention (sGAT) layer

This is a new layer that is similar to the graph attention network but simpler

\[z_i = 1 / Ni Sum_j a_ij * [x_i || x_j] * W + b_i\]

|| is the concatenation operator: [1,2,3] || [4,5,6] = [1,2,3,4,5,6] Ni is the number of neighbor of node i Sum_j runs over the neighbors of node i \(a_ij\) is the edge attribute between node i and j