Fig 1.
Tackling the DDI prediction problem as a link prediction problem.
A) A DDI graph is created: nodes represent drugs, and edges represent interactions. B) The DDI graph is represented by an adjacency matrix, rows and columns represent drugs, and a value of one in the matrix indicates an existing interaction; for example, the cell in the first row and the last column represents the interaction between D1 and D5. In a link prediction problem, a score is calculated to every non-existing interaction.
Fig 2.
Link prediction using matrix factorization for DDI prediction.
The dimension of the adjacency matrix is reduced by factorizing it into two lower ranked matrices. By multiplying the matrices a score is calculated for every existing and non-existing interaction. In this case, an interaction between D5 and D3 is very likely to exist. A score is also given to existing links: the interaction between D1 and D4 is stronger than the interaction between D1 and D5.
Fig 3.
Overview of AMF’s architecture.
Drugs are represented as nodes; embedding layers (which act as latent factors) and biases are shared between input nodes. Dropout is used as a regularization mechanism for preventing overfitting.
Fig 4.
Retrospective evaluation scheme.
Parameter tuning is performed using DrugBank release 4.1.0 and 5.0.0. The previous release is used to train the model, and the latter is used to validate the results. The final model is trained using the parameters obtained in the validation stage with the data from release 5.0.0 (which contains the data from release 4.1.0 with some additions and changes) and tested using release 5.1.1.
Fig 5.
A) Receiver operating characteristic curves; B) Per drug average precision @ n; C) Precision @ n.
Table 1.
Area under the ROC and precision-recall curves for the holdout analysis.
Fig 6.
Retrospective analysis results.
A) Receiver operating characteristic curves; B) Per-drug average precision @ n; C) Precision @ n.
Table 2.
Area under the ROC and precision-recall curves for retrospective analysis.
Table 3.
Area under the ROC and precision-recall curves for multi-source data comparison, three-fold cross-validation.
Table 4.
Area under the ROC and precision-recall curves for multi-source data comparison, five-fold cross-validation.
Fig 7.
AMFP’s propagation factor analysis.
A) Retrospective propagation factor analysis. The optimal value selected during validation and used for model training is 0.5. The optimal value for the test set is 0.8. B) Holdout propagation factor analysis. For both validation and training the optimal value is zero—weights are not propagated at all.