Antifungal synergistic drug combinations

First, we investigated hundreds of studies on drug combinations and selected 69 compounds involved in antifungal drug combination experiments (see S1 Table). Then we searched literatures with the keywords ‘synergy’, ‘synergic’, ‘synergistic’, ‘synergism’, ‘interaction’ and ‘combination’ in the PubMed, Google Scholar and Web of Knowledge and collected 75 experimentally confirmed synergistic antifungal drug combinations (dataset 1, see S2 Table). Therefore, all the antifungal compounds involved in antifungal drug combination experiments are considered in this study. We do not require they must have known synergistic partners. We also classified these compounds into principal drugs and adjuvant drugs according to the following rules. If one compound in the synergistic combination shows activity in the antifungal assay, but the other does not, as reported before, then the former compound is considered as the principal drug, and the latter is considered as the adjuvant drug. If both compounds in the synergistic combination show activity in the antifungal assay, or neither one shows activity, as reported before, then these two compounds are considered as both principal and adjuvant drug. If one compound does not have antifungal synergistic effect with any other compound, then this compound is classified according to its antifungal activity.

To further confirm the predictive ability of NLLSS, we further obtained two other antifungal drug combination datasets from the dataset 1: synergistic antifungal drug combinations against Candida albicans (dataset 2, the drugs are the same as the ones in dataset 1 and the combinations are the ones in dataset 1 against Candida albicans) and synergistic azole drug combinations (dataset 3, the drugs are the ones in dataset 1 which are azole drug and the combinations are the azole drug combinations in dataset 1). Related information on these two synergistic drug combination datasets can be obtained from the supplementary materials (See S3–S6 Tables: drugs and synergistic drug combinations in dataset 2 and 3, respectively). Statistics of three drug combination datasets were listed in Table 1, including the number of drugs, principal drugs, adjuvant drugs, known synergistic combinations (A), and drug pairs without known synergistic relationship (B) and the ratio A/B.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Statistics of three drug combination datasets were listed, including the number of drugs, principal drugs, adjuvant drugs, known synergistic combinations (A), and drug pairs without known synergistic relationship (B) and the ratio A/B.

https://doi.org/10.1371/journal.pcbi.1004975.t001

In this study, NLLSS was developed based on the framework of Laplacian Regularized Least Square (LapRLS), which required drugs in the same combination must be divided into principal drug and adjuvant drug. However, according to the classification rules mentioned before, many drugs have been considered as both principal and adjuvant drugs. This fact ensure that we still can obtain plenty of potential synergistic drug combinations composed of two principal drugs or two adjuvant drugs.

Drug chemical structure similarity

Chemical structure similarities between compounds were calculated by SIMCOMP [36] based on chemical structure information from the DRUG and COMPOUND Sections in the KEGG LIGAND database [37]. The similarity calculated from SIMCOMP is a global score based on the ratio between the size of the common substructures and the size of the union structures [36]. Applying this operation to all compound pairs, chemical structure similarity scores between principal (adjuvant) drugs can be obtained

Drug-target interactions

Target proteins of all the drugs in three drug combination datasets were obtained from the Drug Bank database [38] and related literatures. The drug-protein interactions in three datasets were shown in S7–S9 Tables, respectively.

Drugs for biological experiments

Ketoconazole, Fluconazole, Voriconazole, Posaconazole, Itraconazole, Terbinafine, Flucytosine, Radicicol, Disulfiram, Lovastatin, Geldanamycin, Caspofungin and Micafungin were purchased from a local pharmaceutical company. Amphotericin B, Beauvericin and FK506 were purchased from Sigma.

Principles of NLLSS

NLLSS assumes that principal drugs which obtain synergistic effect with similar adjuvant drugs are often similar and vice versa, which is referred to as the similar nature of drug combinations (see the examples in Fig 1 and S1 Fig). Based on this assumption, the following conclusion can be deduced: principal drugs which obtain synergistic effect with the same adjuvant drug are often similar and vice versa. In this paper, similarity between two drugs is established on the basis of drug chemical structure, drug-target interactions, and known synergistic drug combinations. The similar nature of drug combination was further formulated into two classifiers based on the framework of LapRLS in the principal and adjuvant drug space, respectively. The classifiers in the principal drug space and adjuvant drug space both considered all possible drug pairs. They both used the information of known synergistic drug combinations and unlabeled drug combinations. The difference between these two classifiers was that they adopted the different drug similarity. Classifier in the principal drug space only used principal drug similarity. Correspondingly, classifier in the adjuvant drug space only used the information of adjuvant drug similarity. Finally, two classifiers were combined into a single classifier to give a final predictive result. Based on the model, a score to assess how likely two drugs will obtain synergistic effect can be obtained. Drug combination pairs with high scores can be expected to have a high probability of obtaining synergistic effect when combined, thus having priority in subsequent biological experiments and, in turn, reducing the cost of identifying potential synergistic drug combinations. The flow chart of NLLSS is shown in Fig 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Schematic illustrating the similar nature of synergistic drug combinations.

Blue nodes represent principal drugs, and yellow nodes represent adjuvant drugs. The arc from the principal drug to the adjuvant drug means that two drugs have synergistic effect when combined in the antifungal assays. The edge between two principal (adjuvant) drugs represents the similarity between two drugs. Thickness of edges linking drugs indicates degree of similarity between them. This figure shows adjuvant drugs which obtain synergistic effect with similar principal drugs are often similar. Similar principal drugs, Cyclosporine A and FK 506 gain synergistic effect with seven adjuvant drugs, including Fenpropimorph, Fluconazole, Itraconazole, Ketoconazole, Tunicamycin, Voriconazole, and Terbinafine. It can be observed that those seven adjuvant drugs are similar. Also the adjuvant drugs which obtain synergistic effect with Ketoconazole are similar and form a module. On the contrary, most of the adjuvant drugs which obtain synergistic effect with dissimilar principal drugs (Cyclosporine A and Ketoconazole, FK506 and Ketoconazole) are dissimilar.

https://doi.org/10.1371/journal.pcbi.1004975.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. The basic idea of NLLSS is shown.

First, the principal and adjuvant drug similarity are calculated based on drug chemical structure, drug-target interactions, and known synergistic drug combinations. Next, we construct the synergistic drug combination classifiers in the principal and adjuvant spaces, respectively. Finally, two classifiers are combined into the final classifier to select potential synergistic drug combinations for experimental validation.

https://doi.org/10.1371/journal.pcbi.1004975.g002

Drug similarity calculation

NLLSS first calculates the similarity between drugs. In this model, the similarity between two drugs depends on three factors: drug chemical structure similarity, drug target similarity, and drug synergistic similarity.

Drug chemical structure similarity can be obtained by SIMCOMP as noted before. We defined principal (adjuvant) drug chemical structure similarity matrix as . Next we want to extract the information from drug-target interactions for the measurement of drug similarity. The underlying assumption made here was that two drugs are similar if they share more common target proteins (see Fig 3). Based on this assumption, the principal (adjuvant) drug target similarity matrix was defined. The entity of the matrix was the number of target proteins shared by two drugs. Third, we extracted the information from known drug synergistic combinations, assuming that if two principal (adjuvant) drugs obtain synergistic effect with more common adjuvant (principal) drugs, they will have greater similarity (see S2 Fig). The principal (adjuvant) drug synergistic similarity matrix was defined as . The entity of the matrix was the number of common adjuvant (principal) drugs which have synergistic effect with two principal (adjuvant) drugs.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. In order to calculate the similarity between two drugs, we extract the information from drug-target interactions.

The underlying assumption made here is that two drugs that share more common target proteins are more similar.

https://doi.org/10.1371/journal.pcbi.1004975.g003

Drug-target similarity matrix and drug synergistic similarity matrix must be normalized. For , we defined a diagonal matrix such that was the sum of row i of . We set which yielded a symmetric matrix where . A similar operation was applied to another three matrices. Now, the principal drug similarity matrix S^P can be obtained by linear combination as follows: where combinatorial coefficient means the weight of various similarity measures for the final integrated principal drug similarity. Similarly, the adjuvant drug similarity matrix S^A can be obtained by the following form.

Here, we have adopted the method of weighted averaging for the drug similarity integration, which means all the drug similar measures have equal weight (i.e. α^P = β^P = γ^P = 1/3, α^A = β^A = γ^A = 1/3) for the final drug similarity matrix.

Construction of the classifier

For the employment of the LapRLS, Laplacian operation must be applied to the similarity matrix. The diagonal matrices D^P and D^A were defined such that D^P(i,i) and D^A(i,i) were the sum of row i of S^P and S^AS^A, respectively. The normalized Laplacian matrices were defined as follows:

Let matrix Y represents prior synergistic drug combination information. If principal drug i and adjuvant drug j were known to produce synergistic effect, then Y(i,j) = 1; otherwise Y(i,j) = 0. The aim was to obtain a continuous classification function, which reflected the probability that two drugs could obtain synergistic effect when combined. Intuitively, it is anticipated that when similar principal (adjuvant) drugs are combined with the same adjuvant (principal) drug, these combinations can obtain similar synergistic probability scores. Also this classification function should comply with prior synergistic information. LapRLS defines a cost function and wants to minimize this cost function in order to obtain an optimal classification function. The classification function was composed of optimal functions in the principal drug space and adjuvant drug space.

We first address how to obtain optimal classification function in the principal drug space. Cost function was defined as follows:

Where ‖.‖_F is Frobenius norm and η_P is the trade-off parameter in the principal drug space. Then, we can get the optimal classification function [39,40] as follows:

We also can get the optimal classification function in adjuvant drug space in a similar manner: where η_A is the trade-off parameter in the adjuvant drug space. We set these two trade-off parameters as 0.3 in this study according to previous literatures [40–43]. Hence, the classification function can be obtained by combining the prediction results in both principal and adjuvant drug space, as follows:

We converted the probability of candidate drug combinations to the Rank Probability (RP). The probabilities of combinations were ranked in ascending order, and each candidate combination obtained Rank (R). Rank Probability (RP) of a drug combination was calculated by Rank (R) divided by the total number of candidate drug combinations (N). In this case, the most probable synergistic drug combination will get the RP of 1.

Synergistic antifungal bioassay

Candida albicans SC5314 was used as a test strain for antifungal and synergistic antifungal bioassay. All procedures were described previously [16]. The experiments were carried out in flat bottom, 96-well microtiter plates (Greiner), using a broth microdilution protocol modified from Clinical and Laboratory Standards Institute M-27A methods [44]. Overnight cultures were selected to prepare the strain suspension with RPMI 1640 medium (Gibco) at the concentration of 1×10⁴ cells/mL counted by hemocytometer. To the test wells in 96-well plates, 2 μL of the samples were added, followed by an additional 80 μL of the strain suspension. The test plates were incubated at 35°C. The antifungal MICs were determined by measuring and comparing the optical densities of the positive control and test wells at different time points. For the synergistic antifungal assay, checkerboard assay was used, and beauvericin combined with ketoconazole served as positive control [16]. The MICs were determined by measuring and comparing the optical densities of the positive control and test wells at different time points.

Cross validation

We evaluated the predictive performance of NLLSS using leave-one-out cross validation (LOOCV). To do so, each known synergistic drug combination was treated as a test dataset in turn, while the remaining known synergistic drug combinations were used as the training dataset. First, we calculated the enrichment score to measure the performance of NLLSS. When LOOCV is implemented, if there are n candidate drug combinations without known synergistic evidences, the enrichment score is calculated by dividing n/2 by the rank of the left-out drug combination among candidate drug combinations. For example, if NLLSS gives the left-out known synergistic drug combination the highest ranking (ranked 1st in the candidate drug combinations), there would be an enrichment score of n/2. Furthermore, if the left-out known synergistic drug combination is ranked by random, it would have the rank of n/2 and therefore have an enrichment score of 1. Therefore, enrichment score could represent the difference between prediction accuracy obtained by NLLSS and random. Here, the average of enrichment scores for all the left-out known combinations is calculated for the final evaluation. Next, receiver-operating characteristic (ROC) curve was used as another evaluative measure. The ROC curve plots the true-positive rate (TPR) versus the false-positive rate (FPR). The area under the ROC curve (AUC) was calculated to reflect predictive accuracy. Here, the ROC curves of NLLSS based on the combination of two classifiers and only based on a single classifier in three drug combination datasets were compared (Fig 4, S3 Fig, S4 Fig). The AUCs for the combination of two classifiers in three dataset were 0.9054, 0.8963, and 0.8819, respectively, which shows reliable ability to predict potential synergistic drug combinations. The AUCs for the classifier in the principal and adjuvant drug space were significantly inferior to the combined classifier, which shows the reasonableness of combining the classifiers in the principal and adjuvant drug space. Also, the comparisons of combined and single classifiers in terms of fold enrichment score in the three drug combination datasets were shown in Fig 5, still illustrating the prefect performance of NLLSS.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. LOOCV was implemented in the antifungal synergistic drug combinations dataset 1.

AUC was calculated to evaluate the performance of the method. Here known synergistic drug combinations were used as a test dataset. The ROC curves of NLLSS based on the combination of two classifiers and based only on a single classifier were compared. The results confirmed the performance advantage of combining the classifiers in the principal and adjuvant drug space into a single classifier.

https://doi.org/10.1371/journal.pcbi.1004975.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. LOOCV was implemented in three antifungal synergistic drug combinations datasets.

Fold enrichment score was calculated to evaluate the performance of the method. Here known synergistic drug combinations were used as a test dataset. The fold enrichment score of NLLSS based on the combination of two classifiers and based only on a single classifier were compared. The results confirmed the performance advantage of combining the classifiers in the principal and adjuvant drug space into a single classifier.

https://doi.org/10.1371/journal.pcbi.1004975.g005

Performance comparison between NLLSS with current state-of-the-art computational models can’t be reasonably implemented. Different synergistic drug combination benchmark datasets and drug data sources for drug similarity calculation have been used in the different studies. For example, we only paid attention to antifungal synergistic drug combinations and integrated drug chemical structure, drug-target interactions, and known synergistic drug combinations to calculate drug similarity. In order to compare different computational models based on the same benchmark dataset, we must obtain different data sources of all the drugs in the benchmark dataset, such as drug-target interactions, drug side-effect information, and drug chemical structure. It is difficult to obtain all the datasets. Furthermore, some studies didn’t use known synergistic drug combinations to predict potential ones. It is unreasonable to directly compare them with NLLSS. The AUCs of NLLSS was 0.9054, which has been better than AUCs reported in the previous studies.

Furthermore, in order to confirm NLLSS is robust to the training sample selection, we implemented 10-fold, 5-fold, and 3-fold cross validation (CV) in all the three datasets, respectively (See Table 2). Here, all the known synergistic combinations were randomly divided into 10-fold, 5-fold, and 3-fold, which means 90%, 80%, and 66.67% of the known synergistic combinations were regarded as the training samples for model learning and the other 10%, 20%, and 33.3% were used as test samples for performance validation, respectively. Considering the potential influence caused by sample division, we implemented 100 different random divisions and calculated the mean and the standard deviation of all the obtained AUCs. As the results listed in Table 2, NLLSS has a reliable and robust performance in all the validation schemas.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2.  Performance of NLLSS under the validation framework of LOOCV, 10-fold CV, 5-fold CV, and 3-fold CV, which have demonstrated the NLLSS has a reliable and robust performance.

https://doi.org/10.1371/journal.pcbi.1004975.t002

Experimental validation of the antifungal drug combinations in fungal pathogen Candida albicans

To experimentally validate the predicted combinations which have potential synergistic antifungal activities, we tested all combinations in vitro on the leading human pathogen Candida albicans. Here, we implemented experiments for the top 10 potential drug combinations in all the three datasets (See Table 3). The synergis tic activities were judged by fractional inhibitory concentration index (FICI) values, which were calculated by comparing MICs in combinations with MICs of the each drug used alone (See S13 Table) at different time points. Biological experimental results indicated that 6, 5, and 6 out of the top 10 potential combinations in three datasets did indeed obtain antifungal synergistic effect (See Table 3). After the removal of duplicate combinations, we found 7 synergistic combinations (Fig 6) and proved that 6 groups were nonsynergistic combinations (S5 Fig) in total. Considering the validated 7 synergistic combinations are totally new combinations, which have not been reported in any publicly published literatures, this prediction accuracy could be considered high. To identify 7 synergistic combinations, we only need implement experiments for 13 candidate combinations, which have greatly reduced the cost and time of pure experimental research. Most of previous computational studies for synergistic drug combination prediction didn’t implement any experimental validation. Only Jansen et al (2009) implemented biological experiments for predicted synergistic combination. However, training samples and criterion of selecting potential combinations for experimental validation were totally different between their studies and NLLSS. Therefore, it is also difficult to compare these two models based on the accuracy of independent prediction.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Synergistic antifungal combinations: The x and y axis indicated the concentrations of the combination drugs used in the synergistic screening.

The dots were the active concentrations for inhibiting the growth of C. albicans in the combinations. All experiments were performed on 96-well plates and incubated at 35 oC for 48 h. The data from three independent experiments were measured at different time points (16 h, 24 h and 48 h).

https://doi.org/10.1371/journal.pcbi.1004975.g006

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3.  Biological experimental results for potential top 10 antifungal drug combinations in three drug combination datasets, and the targets for the drugs are shown here.

In total, 13 drug combinations have been predicted to be the top 10 potential combinations. Seven combinations have been confirmed to be synergistic by antifungal biological experiments.

https://doi.org/10.1371/journal.pcbi.1004975.t003

Synergistic drug combinations

Among the predictions, 7 pairs showed synergistic effects. For the Group 1 in Fig 6, FK506 is a novel immunosuppressant isolated from Streptomyces [45] and has been demonstrated to bind to FKBP12 to inhibit calcineurin [46], which is the key pathway for the cells responding to different stresses [47–50]. Radicicol and geldeanamycin from Groups 5 and 13, respectively, can bind to Hsp90 and alter its function. Also, Hsp90 can act as the molecular chaperone to calcineurin [51,52]. Synergistic activity between FK506 and ketoconazole occurred at 16 h (Fig 6 and S14 Table), but only at 48 h for radicicol with fluconazole and geldeanamycin with fluconazole (Fig 6, S15 and S16 Tables).

Synergism between beauvericin and ketoconazole from Group 7 was identified based on our high-throughput synergistic screening platform [16]. Beauvericin can inhibit drug efflux pumps to reduce the accumulation of azoles and present synergistic activity [53], and the synergistic activity was observed at 48 h (Fig 6 and S17 Table).

In Group 8, caspofungin is a lipopeptide antifungal drug that inhibits the enzyme β(1,3)-D-Glucan synthase, thus interfering with the integrity of the fungal cell wall [54]. Our results proved that it could synergize with voriconazole and that synergistic activity started at 24 h (Fig 6 and S18 Table).

In Group 10, posaconazole is a triazole antifungal drug that inhibits lanosterol 14α-demethylase (ERG11) to block ergosterol biosynthesis [55,56]. Terbinifine also inhibits ergosterol biosynthesis by inhibiting squa lene epoxidase (ERG1) [57]. These two drugs, which have the same pathway, showed synergistic activity that occurred at 24 h (Fig 6 and S19 Table).

Lovastatin from Group 12 is a member of the drug class of statins, used for lowering cholesterol by inhibiting the 3-hydroxy-3methylglutaryl-coenzyme A reductase (HMG-CoA reductase), an enzyme that catalyzes the conversion of HMG-CoA to mevalonate [58]. Lovastatin could synergize with itraconazole in this study, and the synergistic activity started at 16 h (Fig 6 and S20 Table).

Non-synergistic drug combinations

We found that 6 pairs of the predictions showed non-synergistic synergistic effects. Amphotericin B in Group 2 is a kind of polyene antifungal drug by binding to ergosterol to destroy the cell membrane [59,60]. Some researchers observed that a combination of azoles (such as posaconazole and itraconazole) can decrease the fungal infection burden [61–63]. However, others have observed no difference using a combination of ketoconazole or fluconazole and amphotericin B [64–66]. Ayse et al. reported that some isolates even showed antagonistic activities between amphotericin B and fluconazole [67]. Our results proved the absence of synergistic activity, even after 48 h (S5 Fig and S21 Table).

In Group 3, no synergistic activity was observed between terbinafine and ketoconnazole on C. albicans, which is in agreement with a previous report [68]. However, it does show synergism with itraconazole [16,69,70], fluconazole, voriconazole [71] and posaconazole (S5 Fig and S22 Table).

Flucytosine from Group 4 is a fluorinated pyrimidine analogue, and it inhibits fungal RNA and DNA synthesis [72]. The combination of ketoconazole and flucytosine can increase the antifungal effect, but no synergistic activity was observed [73]. Our result demonstrated that this combination only had “additive effect” (0.5<FICI<1), but no synergistic activity (S5 Fig and S23 Table).

Disulfiram in Group 6 is an antifungal and can inhibit the drug efflux pump from C. albicans [74]. Ann et al. found antagonistic activities in clinical isolates when combining disulfiram with fluconazole [75]. Our results show that the combination of fluconazole and disulfiram had no synergistic activity during two days of incubation (S5 Fig and S24 Table).

Micafungin, the second available agent in the echinocandins class, is a potent inhibitor of 1,3-β-D-glucan synthase. It is used against fungal pathogens, such as Candida spp. and Aspergillus spp [76]. Our result showed that the combination of voriconazole and caspofungin had synergistic activity (Group 8), but micafungin did not show synergy with ketoconazole (Group 11, S5 Fig and S25 Table) during two days of incubation, indicating that the interactions between echinocandins and triazoles may be related to the chemical structures of individual agents.

For the same reason, in this study, we confirmed that caspofungin has synergistic activity with voricanazole (Group 8). However, the combination of ketoconazole and caspofungin had no synergistic activity (Group 9, S5 Fig and S26 Table), just like the nonsynergistic combination of ketoconazole and micafungin (Group 11, S5 Fig and S25 Table), suggesting that the synergistic activity may be related to chemical structures.