Fig 1.
1) This image was obtained from https://commons.wikimedia.org/wiki/File:Capsule_icon.svg. 2) This image was obtained from https://commons.wikimedia.org/wiki/File:Noun-drugs-1511305-00449F.svg. 3) This image was was obtained from https://f1000research.com/articles/4-178 and is licensed under Creative Commons Attribution License, https://doi.org/10.12688/f1000research.6314.1.
Table 1.
Detailed statistics for each dataset where our newly constructed ArcDFI dataset is an integration of FooDrugs and FDMine.
Table 2.
Number of drug compounds annotated with each type of CYP450 isoenyzme (1A2, 3A4, 2C19, 2C9, 2D6) interaction (substrate, inhibition) label.
Fig 2.
Descriptive illustration of ArcDFI.
a) Model architecture for ArcDFI. The parameters in Compound Substructure Encoder, Compound Graph Encoder and Compound-CYP Interaction Block are shared by both drug and food compounds. b) Detailed illustration of the Compound-CYP interaction block. c) Detailed illustration of the Cross-Modality Fusion Layer. The drug (food) compound-CYP interaction embedding is combined with the food (drug) compound graph embedding using a vector-wise outer product, followed by concatenation of the two embeddings.
Fig 3.
Detailed illustration for Attention Regularization Auxiliary Loss Objective.
Table 3.
Evaluation results for ArcDFI, its ablated version, and baseline models under the cold drug experiment setting. All evaluation scores were averaged over three iterations along with their standard deviation. Best results are bold-faced.
Table 4.
Evaluation results for ArcDFI, its ablated version and baseline models under the cold food experiment setting. All evaluation scores were averaged over three iterations along with their standard deviation. Best results are bold-faced.
Fig 4.
Analysis on ArcDFI’s Compound-CYP Interaction Block for drug-food compound pair Sulfonylurea and Salicylate.
(a), (b): Attention weights and highlighted compound substructures extracted from ArcDFI’s Compound-CYP Interaction Block head related to substrate-type interactions for Sulfonylurea, respectively. (c), (d): Attention weights and highlighted compound substructures extracted from ArcDFI’s Compound-CYP Interaction Block head related to inhibition-type interactions for Sulfonylurea, respectively. (e), (f): Attention weights and highlighted compound substructures extracted from ArcDFI’s Compound-CYP Interaction Block head related to substrate-type interactions for Salicylate, respectively. (g), (h): Attention weights and highlighted compound substructures extracted from ArcDFI’s Compound-CYP Interaction Block head related to inhibition-type interactions for Salicylate, respectively.
Fig 5.
Analysis on ArcDFI’s Compound-CYP Interaction Block for drug-food compound pair Midazolam and Berberine.
(a), (b): Attention weights and highlighted compound substructures extracted from ArcDFI’s Compound-CYP Interaction Block head related to substrate-type interactions for Midazolam, respectively. (c), (d): Attention weights and highlighted compound substructures extracted from ArcDFI’s Compound-CYP Interaction Block head related to inhibition-type interactions for Midazolam, respectively. (e), (f): Attention weights and highlighted compound substructures extracted from ArcDFI’s Compound-CYP Interaction Block head related to substrate-type interactions for Berberine, respectively. (g), (h): Attention weights and highlighted compound substructures extracted from ArcDFI’s Compound-CYP Interaction Block head related to inhibition-type interactions for Berberine, respectively.