Table 2 Hyperparameters related to model topology for models with molecular features as input screened in grid search for each data set

From: Graph neural networks learn emergent tissue properties from spatial molecular profiles

dataset

depth feature embedding

width feature embedding

depth node embedding

width node embedding

depth graph embedding

width graph embedding

IMC - breast cancer (Jackson)

{1, 2, 3}

{4, 8, 16, 32, 64}

{1, 2, 3}

{4, 8, 16, 32, 64}

{1, 2, 3}

{16, 64}

IMC - breast cancer (METABRIC)

{1, 2, 3}

{4, 8, 16, 32, 64}

{1, 2, 3}

{4, 8, 16, 32, 64}

{1, 2, 3}

{16, 64}

CODEX - colorectal cancer

{1, 2, 3}

{4, 8, 16, 32, 64}

{1, 2, 3}

{4, 8, 16, 32, 64}

{1, 2, 3}

{16, 64}

  1. node embedding: The node embedding describes transformations of node-wise feature vectors and is used in MI and GCN/GIN models. All layers have the same width. graph embedding: The graph embedding describes the layer stack that transforms the graph representation to a graph label prediction. The input representation derives from an aggregation over nodes for MI and GCN/GIN models and is the input feature vector for MLP models. All layers have the same width.
Back to article page