Fig 1.
Adversely interacting drug pairs and non-interacting drug pairs significantly differ with regard to the 11 features selected.
(a) Schematics of calculating indication similarity and side effect similarity features. (b) Indication similarity score of hierarchy level PT, HLGT and SOC between two drugs. (c) Side effect similarity score of hierarchy level HLT and HLGT between two drugs. (d) Schematics of calculating target sequence similarity and genetic interaction features. Genetic interaction scores indicate the deviation from the expected phenotype when two genes are simultaneously knocked out, and were obtained from a global genetic interaction network in yeast by mapping targets of drugs to their yeast homologs. A negative score denotes synergistic interaction while a positive score indicates buffering interaction. (e) Minimum, mean, median and maximum target sequence similarity score between targets of two drugs. (f) Minimum and maximum genetic interaction score between targets of two drugs. Statistical significance was determined by the two-sided permutation test on the sample mean. PT, preferred term; HLT, high level term; HLGT, high level group term; SOC, system organ class. * p < 0.001; ** p < 0.0001; *** p < 0.00001.
Fig 2.
The train-test splitting scheme and model performance on the test set.
(a) The train-test splitting scheme. Drugs are randomly divided into “training drugs” and “test drugs” with ratio of 2:1. Training set only consists of drug pairs constituted by “training drugs” and test set only consists of drug pairs constituted by “test drugs”. Training drugs are further split into “training drugsi” and “validation drugsi” with the same splitting scheme to obtain training seti and validation seti in the training phase. For each iteration of hold-out validation, the classifier is fit with training seti and evaluated with validation seti. Purple squares represent non-interacting drug pairs in training seti. Blue squares represent non-interacting drug pairs in validation seti. Green squares represent non-interacting drug pairs in test set. Red squares represent interacting drug pairs in each set. Grey squares represent unused drug pairs. (b) Approximate receiver operating characteristic (ROC) curves on the training set. (c) Approximate precision-recall curves on the training set. (d) AUROCs and AUPRs on the training set and the test set. (e) Receiver operating characteristic (ROC) curve on the test set. (f) Precision-recall curve on the test set.
Table 1.
Top 20 DDI predictions in the test set.
Table 2.
Top 20 new adverse DDI predictions.
Fig 3.
Genetic interaction provides possible mechanistic insights into DDIs.
(a) Mesalazine inhibits IKBKB, a positive regulator of NF-κB activity, and NF-κB is a transcription factor which induces NOS2 transcription. Dexamethasone can inhibit the transcription of NOS2 and facilitate degradation of NOS2. The combined use of dexamethasone and mesalazine could potentially reduce the amount of NOS2 in cells to a large extent, which may affect neurotransmission, antimicrobial and antitumoral activities. (b) Mexiletine targets NAv1.5, a sodium channel encoded by SCN5A, while arsenic trioxide targets AKT1. The transcription of SCN5A is repressed by the transcriptional repressor FOXO1. AKT1 can activate the transcription of SCN5A by phosphorylating FOXO1. The combined use of mexiletine and arsenic trioxide could inactivate the transcription of SCN5A and at the same time block the existing sodium channel, which may largely reduce sodium influx in cardiac cells.