Figure 1.
Examples of interaction profile fingerprints (IPFs) calculated for the drugs oxybutynin and dicyclomine.
The similarity of both fingerprints is measured through the TC coefficient. The drugs corresponding to the non-intersecting interactions for the pair are assigned the TC score and form part of the prediction of the model. The effect associated by the interaction is the same as the original interaction source that generated the prediction.
Figure 2.
The model generates interactions through the multiplication of the matrix M1 (Established DDI matrix) by the matrix M2 (Interaction profile similarity matrix.
Note that each cell shows the TC between drugs A, B and C but interactions with more drugs are considered to calculate the TC value). The values in the diagonal of the matrices are set 0 since drug interactions with themselves are not taken into account. In the final matrix M3 only the maximum value in the multiplication-array in each cell is preserved and a symmetry-based transformation is carried out retaining the highest TC value. In the example, the initial interactions A–B and A–C (red color) have a TC score of 0.9 in the matrix M3. The system generated a new predicted interaction between B and C with a TC score of 0.8 (green color).
Table 1.
Model performance in the four independent test sets A, B, C and D along with random results.
Figure 3.
ROC curves in the hold-out validation process: a) training set with the 85% of the DrugBank interactions; b) test set with the 15% of the extracted DrugBank interactions; c) training set with the 70% of the DrugBank interactions; d) test set with the 30% of the extracted DrugBank interactions.
Figure 4.
ROC curves for test set D: a) ROC curve generated by the IPF model for test set D.
Interactions for the top 50 drugs (41 generic names) confirmed in drugs.com/drugdex were considered as true positives within all the possible interactions in a matrix of 41×928 drugs. Interactions already in the initial DrugBank DDI database (matrix M1) were not included in the analysis; b) ROC showed by a model applied to test D using MACCS fingerprints; c) ROC curve calculated by the IPF model for test set D but excluding CYP interactions; d) ROC showed by the MACCS fingerprints model applied to the test D without CYP interactions.
Figure 5.
Enrichment factor (a) and precision (b) achieved by the model regarding random results for top drugs sold in 2010 (test set D).
The test set of drugs are sorted according to the enrichment factor.
Table 2.
Some examples of correct interactions predicted for the 50 most frequently sold drugs in 2010 in which the model generated interactions through the comparison of drugs belonging to different pharmacological classes.
Figure 6.
Comparison between the TC for all the pairs of drugs in a matrix of 928×928 using MACCS and IPF fingerprints.
The correlation coefficient (r) calculated through linear regression is 0.167 and p<.0001.