Double strength: how can we better predict effective, as-yet untried, drug combinations?

Some things just work better with two – but when it comes to drug combinations, the sheer number of possible pairs means millions go untested. What if a predictive network could help identify the promising couples?

The challenge

Combination therapy has become the standard approach for tackling many diseases, with cancer and infectious diseases such as HIV/AIDS, among others, responding to treatment. It has many advantages over the use of a single drug:

  • the treatment tends to be more effective than monotherapy – the whole is greater than the sum of its individual parts
  • the likelihood of drug resistance decreases, as many drug-resistant mutations are mutually exclusive
  • in some cases lower individual drug dosages can reduce the risk of adverse effects

However, identifying potent combinations, and finding the sweet spot with the right dosage and scheduling, is anything but simple. It’s also expensive; usually each drug has to first be tested alone before the combination can be investigated in a clinical trial – all of which adds up to a time-consuming and costly process. 

Given these conditions, simply trying all possible combinations of all approved drugs is infeasible. The 8,000 or so drugs in the market or in development yield nearly more than 30 million possible pairs – so testing them all, for efficacy against thousands of diseases, is clearly not an option. 

Who, then, chooses which combinations to test – and how do they pick them? At present, combinations tend to be selected based on the pre-existing scientific hypothesis, researchers’ intuition and the availability of proprietary drugs. As well as being a very hit-and-miss approach, it’s also non-transferable to other researchers and diseases.

But what if we could develop a systematic approach using the network of known protein-protein-interactions to identify potentially effective two-drug combinations? 


In their paper, “Network-based prediction of drug combinations”, Cheng et al use a network approach based on known protein-protein interactions of the human body (human interactome) to develop a new approach for identifying effective drug combinations. By taking into account the proximity of drug targets in the human interactome and how they overlap with disease modules, the authors’ model is shown to help identify promising drug pairs for combination therapy. 

Why read this paper?

The paper brings a network perspective to the discovery of drug combination treatments, showing that relying on a protein-protein-interaction network is a promising way to identify potential drug combinations. The paper abstracts from specific disease examples to apply a higher-level conceptualisation that can be used as a valuable starting point in assessing pairs for combination therapy. It suggests a system for drug combination selection that is both scalable and easily interpretable. 

How it works

The interactome dataset – comprising 243,603 protein-protein interactions and 16,677 unique proteins, collected from various sources – formed the basis of the network. The paper’s authors also made use of multiple data sources to collect targets for 1,978 clinically investigated drugs, 681 gold-standard drug combinations and 13,397 clinically reported adverse drug-drug interactions. 

The authors measure how close to each other two drug targets are found in the human interactome; a positive separation measure denotes separated drugs, negative denotes overlapping drugs.

Having defined a measure for drug similarity, the authors then look at how the drugs interact with the disease module (the part of the human interactome that is known to be associated with a particular disease). Six possible configurations emerge: 

  • either the targets of the two drugs overlap (negative separation measure) or they do not (positive separation measure)
  • for each of the above cases there are three variations: that the targets of both drugs overlap with the disease module, that only one drug target overlaps, or that neither overlaps 

Comparing these six configurations with the data of approved drug combinations reveals that combinations in which both drugs overlap with the disease module, but do not overlap with each other – known as complementary exposure – are most effective.

Statistically significant adverse effects were recorded in combinations where the drugs overlap in their targets and with the disease module – known as overlapping exposure.

The authors propose that drug combinations with complementary exposure that have positive, high separation measures are most suitable for trials.

Did the model deliver?

The paper’s model improves on other scores used to evaluate drug combinations, such as chemical similarity, and shows that network approaches using human interactomes are promising.

There are some drawbacks, however. The assumption that complementary exposure is always the desired outcome is very intuitive, and corresponds to what researchers are already doing. Furthermore, the authors assume as ground truth that approved drug combinations work and all other combinations do not – when it may simply be that combinations from the “not working” set have not yet been tested.

Thus, one can claim that the paper just confirms – and tests – what is already known (or intuitive).

Could this work in practice?

We are still some way off from a scenario in which an algorithm can tell us with something close to certainty which drug combination is the most promising. 

Although the model achieves a greater accuracy than the benchmarks cited by the author, its predictive power is not sufficient to consult in the important and costly decision of whether a drug pair should be evaluated in a clinical trial. 

It can, however, serve as a valuable starting point in the early development phases and help to reduce the scope of possible combinations by: 

  • identifying and discounting combinations with adverse effects
  • helping to prioritise those combinations that should progress to quite costly preclinical animal trials
  • providing extra information in cases where preclinical experiments are unlikely to yield much information 

With time, as the human interactome data is cleaned of irrelevant targets and newly discovered targets close existing knowledge gaps, this approach will surely become more valuable – especially in conjunction with other data sources, such as gene expression and physiological and pathological conditions.

The model’s key selling point – a systematic approach to drug combination discovery that doesn’t rely on individuals’ expertise and intuition – may yet prove particularly attractive for newer diseases, for which medication is lacking, but associated proteins have already been identified. 

One such situation is, of course, the Covid-19 pandemic. In fact, the authors used their network-based approach in a consecutive study to propose combinations of existing drugs to treat Covid-19. 

Mona Schirmer was reviewing Cheng et al’s “Network-based prediction of drug combinations” from March 2019, published in Nature.


Paul von Bünau

Managing Director

+49 (0) 173 24 16 000


Potsdamer Straße 68
10785 Berlin