In the past decades, targeted therapies modulating specific targets were considerably successful. However, recently, the rate of new drug approvals is slowing down despite increasing research budgets for drug discovery. One reason for this is that most human diseases are caused by complex biological processes that are redundant and robust to drug perturbations of a single molecular target. Therefore, the ‘one-drug-one-gene’ approach is unlikely to treat these diseases effectively 
Drug combinations can potentially overcome these limitations: they consist of multiple agents, each of which has generally been used as a single effective drug in clinic. Since the agents in drug combinations can modulate the activity of distinct proteins, drug combinations can help to improve therapeutic efficacy by overcoming the redundancy underlying pathogenic processes. In addition, some drug combinations were found to be more selective compared to single agents 
, thereby reducing toxicity and side effects. Nowadays, drug combinatorial therapy is becoming a promising strategy for multifactorial complex diseases. For example, thiazide diuretics cause hypokalaemia when used to treat hypertension, while this side effect can be prevented by angiotensin-converting enzyme (ACE) inhibitors when they are used concurrently 
. Saracatinib can overcome the resistance of breast cancer to trastuzumab when both drugs are used together, thereby improving the efficacy of trastuzumab 
. Both glyburide and metformin are indicated for type 2 diabetes but work in different ways: glyburide reduces insulin resistance while metformin increases insulin secretion, and therefore the combination of these two drugs can improve therapeutic efficacy due to their complementary mechanisms 
Despite the increasing number of drug combinations in use, many of them were found in the clinic by experience and were not designed as such; the molecular mechanisms underlying these drug combinations are often not clear, which makes it difficult to propose new drug combinations. High-throughput screening was found to be useful to identify possible drug combinations 
; however, it is impractical to screen all possible drug combinations for all possible indications since it leads to an exponential explosion as the number of drugs increases. Therefore, similarly to drug-target predictions 
, a number of computational methods for predicting drug combinations have recently been developed. For example, stochastic search techniques were used to identify optimal combinations within a large parameter space 
in an iterative way, but they only work on small drug sets due to the computational and experimental cost. Mathematical modeling was used to determine synergistic combinations by comparing dose-response profiles of single agents against those of drug combinations 
, but it cannot explain the molecular mechanisms that underlie the drug combinations. Recently, in systems biology, both quantitative 
and qualitative 
models were introduced to investigate drug combinations based on the molecular networks or pathways possibly affected by the drugs. Although network analysis, in principle, can provide insights into the molecular mechanisms of drug actions 
, the incompleteness of molecular networks and the scarceness of the corresponding kinetic parameters limit the application of such approaches to drug combinations considerably.
In general, drugs are combined based on their mechanisms of action, which is characterized by the properties of drugs, such as their targets and pharmacology 
. Taking this into account, we present here a novel concept for the prediction of drug combinations that integrates both molecular and pharmacological features associated with drugs. We treated drug combinations as combinations of their corresponding features, including their target proteins, therapeutic effects, and indication areas. Analysis on the drug combinations approved by the US Food and Drug Administration (FDA) demonstrates that there are some feature patterns enriched in known combinatorial therapies that are both predictive of new drug combinations and provide insights into the mechanisms underlying combinatorial therapy. We consequently predict new drug combinations based on feature patterns enriched in approved drug combinations. Subsequent targeted literature survey revealed that 69% of our predictions were previously reported as effective combinations although they are not approved yet, corroborating the predictive power of our proposed method. In addition, we identify several novel potential drug combinations. For example, we predict a novel combination of promethazine and ibuprofen that could be used as decongestant. Although experimental validation of each individual prediction needs to be provided in the future, we believe that our proposed approach can guide the selection of drug combinations to be tested experimentally.