Skip to main content
Advertisement

< Back to Article

Fig 1.

Violin plots displaying metric scores of all models.

More »

Fig 1 Expand

Table 1.

Comparison results in % of the proposed DGNN-DDI and baselines on the dataset.

More »

Table 1 Expand

Fig 2.

The significant difference between DGNN-DDI and other models in terms of predicted scores.

More »

Fig 2 Expand

Fig 3.

The ROC curves and P-R curves of all models, where T = 3 or L = 3 for all models.

More »

Fig 3 Expand

Fig 4.

The effects of steps T and layers L.

More »

Fig 4 Expand

Fig 5.

Parametric analysis.

(A) Effects of batch size. (B) Effect of hidden dimension. (C) Effects of learning rate. (D) The significance analysis of batch size in terms of predicted scores. (E) The significance analysis of hidden dimension in terms of predicted scores. (F) The significance analysis of learning rate in terms of predicted scores.

More »

Fig 5 Expand

Table 2.

Investigating the contributions of substructure-attention mechanism and co-attention layer.

More »

Table 2 Expand

Fig 6.

Analysis of the substructure attention mechanism (SA) and co-attention layer (CA).

(A)-(C) The metric scores of DGNN-DDI and without SA and/or CA. (D)-(F) The training and testing losses for DGNN-DDI and without SA and/or CA.

More »

Fig 6 Expand

Fig 7.

Heat maps of the atom similarity matrix for drug sildenafil.

More »

Fig 7 Expand

Fig 8.

Heat maps of the atom similarity matrix for drug phenindione.

More »

Fig 8 Expand

Fig 9.

Performance for each DDI type.

More »

Fig 9 Expand

Fig 10.

The key substructures contributing to the SARS-CoV-2 drug combinations.

The center of the most important substructure and its receptive field are shown as red circle and green colors respectively.

More »

Fig 10 Expand

Fig 11.

Drug combination to treat SARS-CoV-2.

More »

Fig 11 Expand

Fig 12.

The overview of proposed DGNN-DDI for DDI prediction.

(A) The workflow of DGNN-DDI. (B) The SA-DMPNN updates the node-level features with T steps where T is 4 in this example.

More »

Fig 12 Expand

Fig 13.

Molecule representation and graph embedding.

(A) Preprocessed the smiles into graph. (B) Graph message passing phase. (C) Graph readout phase.

More »

Fig 13 Expand

Table 3.

Atom and bond features.

More »

Table 3 Expand

Fig 14.

Both structure information and DDIs are important for GNN.

(A) The sight of GNNs in the second layer is shown in blue as we take the carbon with orange as the center. In this example, a GNN with two layers fails to identify the ring structure of zearalenone. (B) The GNN should preserve local structure information (orange ellipse) (C) The interaction type of ‘blood calcium increased’ between drug pair ‘Carnitine’ and ‘Budesonide’ is caused by their partial significant substructures (elliptical parts).

More »

Fig 14 Expand

Fig 15.

The overall computational steps for graph-level representation of dx and dy.

More »

Fig 15 Expand