Synergy Maps: exploring compound combinations using network-based visualization

Abstract

Background: The phenomenon of super-additivity of biological response to compounds applied jointly, termed synergy, has the potential to provide many therapeutic benefits. Therefore, high throughput screening of compound combinations has recently received a great deal of attention. Large compound libraries and the feasibility of all-pairs screening can easily generate large, information-rich datasets. Previously, these datasets have been visualized using either a heat-map or a network approach-however these visualizations only partially represent the information encoded in the dataset.

Results: A new visualization technique for pairwise combination screening data, termed "Synergy Maps", is presented. In a Synergy Map, information about the synergistic interactions of compounds is integrated with information about their properties (chemical structure, physicochemical properties, bioactivity profiles) to produce a single visualization. As a result the relationships between compound and combination properties may be investigated simultaneously, and thus may afford insight into the synergy observed in the screen. An interactive web app implementation, available at http://richlewis42.github.io/synergy-maps, has been developed for public use, which may find use in navigating and filtering larger scale combination datasets. This tool is applied to a recent all-pairs dataset of anti-malarials, tested against Plasmodium falciparum, and a preliminary analysis is given as an example, illustrating the disproportionate synergism of histone deacetylase inhibitors previously described in literature, as well as suggesting new hypotheses for future investigation.

Conclusions: Synergy Maps improve the state of the art in compound combination visualization, by simultaneously representing individual compound properties and their interactions. The web-based tool allows straightforward exploration of combination data, and easier identification of correlations between compound properties and interactions.

Keywords: Compound combinations; Dimensionality reduction; Mixtures; Network; Synergy; Visualization.

Fig. 1

Heatmap representation of the NCATS malaria dataset. The heatmap, created using the Python visualization package matplotlib [47], is constructed as a matrix; rows and columns map to individual columns of the dataset, and the intersecting elements to their combination. The heatmap used the pGamma metric described in Table 3. Compounds were clustered according to their predicted targets, using predictions from an inhouse target prediction tool, such that compounds with a similar bioactivity profile, such as Artesunate and Artemether, cluster together.

Fig. 2

Network Representation of the NCATS malaria dataset. This network visualization was created using Cytoscape [48]. Nodes represent compounds, whilst edges represent combinations, with thickness indicating degree of non-additivity, and red and blue indicating antagonism and synergy respectively. The layout was generated using Cytoscape’s “organic” layout routine.

Fig. 3

Synergy Maps work flow. The work flow employed by the Synergy Maps application. The raw compound and combination data is transformed in steps to yield processed data in JSON, which is then used by the JavaScript App to create the visualization. Specifically, the chosen descriptors (Table 1) are generated from the supplied chemical graphs, and then reduced to two dimensions by the selected dimensionality reduction techniques (Table 2). The combination data is assigned synergy values. The processed data is packaged into a JSON file.

Fig. 4

Improbable combination surface. The surface yields a suspiciously strongly antagonistic (−0.7) value of pGamma. The surface implies that the growth of P. falciparum is rescued by a low concentration of Artemeter, a known antimalarial. In fact, it seems much more likely that the zero concentration row has simply been contaminated, causing an incorrect value of pGamma.

Fig. 5

Synergy Maps. Sample static networks. Nodes represent compounds, with radius indicating relative pIC50. Edges represent combinations, with thickness indicating degree of non-additivity, and red and blue indicating antagonism and synergy respectively. It appears that whilst PCA is a passable dimensionality reduction algorithm for physicochemical and structural space (despite concentrating points in the centre), it does not differentiate the compounds well in biological space. MDS does a little better, yet ultimately still concentrates points towards the centre, preventing compounds from being easily being differentiated. In the authors’ opinion, t-SNE performs well in all spaces; clear clusters can be seen, identifying groups of compounds similar in that space, yet points are still spread across space helpfully so as not to clutter the visualization.

Fig. 6

Synergy Map, represented under t-SNE reduced biological space. Biological space, with various clusters annotated according to hypothesized mode of action or drug function. It appears that the HDAC inhibitor cluster (Including Quisinostat, Trichostatin A and Panobinostat) tends to be disproportionately synergistic compared to other clusters, whilst the PI3K/mTOR inhibitors exhibit disproportional antagonism.

Fig. 7

Screenshot of the interactive web visualization. A screen shot of the interactive web visualization. The representation, reduction type, synergy metric and activity metric may be set by drop down menus in the top bar. Compounds may be searched for using the search box. A slider, shown in the top left, may be used to select threshold levels above and below which combinations should be shown. Individual compounds and combinations may be selected to bring up a tooltip, as shown for Dihydroartemisinin in this example. The tooltip will display any extra property information supplied, such as the primary mode of action in this example. Additional metadata specific to whether a compound or a combination has been selected is also given, such as the activities for a compound, and the synergies for a combination.

Fig. 8

Validation through randomization and scrambling. In order to assess the meaningfulness of potential hypotheses drawn from a synergy map, it is advantageous to compare with random data, such as those in the figure, to protect against spurious correlations being interpreted for more than they are. The random compound positions (leftmost maps) are generated using random feature vectors, which were then reduced using t-SNE in an identical fashion to the Bayes Affinity vectors (rightmost maps). The random positioning of compounds appears not to produce the clusters observed when real data is used. Randomly shuffling the combinations values (topmost maps) reveals more realistic maps—there are several clusters which appear to share many synergies, for which hypotheses may have been proposed, illustrating the danger of overinterpretation of synergy maps.