Fig 1.
The network-interaction hypothesis.
The effects of two drugs (Drug1, Drug2) reach their imminent targets first (arrows) and the effects will then propagate to their network neighborhoods (subnetworks) indicated in red and green, respectively. Targets in the overlap are affected by both drugs, and we suppose that drugs affecting a number of common targets will influence the effects of each other. The overlap is quantified as the proportion of jointly affected targets within all affected targets (in set theory terms: intercept divided by union).
Fig 2.
Ranking performance of the TOS score on known drug interactions and therapeutic combinations.
The ranking performance was measured via ROC analysis as described in Data and Methods. The standard deviation of AUC values (not shown) are between 0.0001 and 0.006 for the different datasets. Note that the tendencies of drug combination groups are the same between cancer-related and not cancer-related drugs. Also, combinations of drugs with identical targets or with similar chemical structures give high TOS scores. These combinations were left out from the statistics of the other groups so they do not influence the AUC values of the other groups.
Table 1.
Datasets.
Fig 3.
Performance of combined predictors on different training sets.
The short titles TOS, TOS+ATC, TOS+GO or TOS+GO+ATC refer to the combination used. The curves represent the AUC value distribution (as a probability density function) obtained via a kernel density estimation (KDE) approach. The data were obtained by a 5 fold cross-validation procedure described in Methods (section 4.5). Note that the distributions are quite similar to the TOS values (top left) which indicates that TOS effectively captures the drug combination phenomenon.
Fig 4.
Scatter plot of TOS scores and Overall Response.
The predicted scores are on the x axes, the clinical outcome, Overall Response (for the definition of outcome measures see the RECIST [35]) are on the y axes. Each data point corresponds to a multicomponent combination. The generalized TOS score of multicomponent combinations was calculated as described in Data and Methods.
Table 2.
Spearman correlation between the clinical outcome measures and the generalized TOS scores of multicomponent combinations.
Table 3.
TOS scores of binary and multicomponent combinations.
Fig 5.
Flow chart of the training procedure.
The input is a list of candidate combinations (i.e. combinations selected for clinical trials) and the set of known combinations (i.e. previously approved cancer combinations). The first step is to compute the Target Overlap Score (TOS) and the drug interaction measures (GO, ATC) for all possible drug combinations. The database consists of the random generated drugs and of the components of the candidate and the known combinations. After the selection of the training sample (both the positive—known cancer combinations—and the negative one—random combinations) a logistic regression was trained using the previously computed TOS and similarity values. In the next step the trained model is used for ranking a set of candidate combinations. The output is the ranked list of the drug combinations.