Ethics statement

The in vitro experiments in this study utilized protoscoleces of Echinococcus granulosus. These samples were obtained from the livers of naturally infected sheep. No animals were subjected to additional specialized breeding or in vivo procedures specifically for this research. Therefore, this study did not require specific approval for animal ethics or biosafety handling.

Overview of the MetaSynMT

MetaSynMT is developed for drug combination synergy prediction and consists of three core components: drug feature construction, parasitic disease feature construction, and synergy effect and side effect prediction module. The overall architecture of the MetaSynMT is illustrated in Fig 1. MetaSynMT first constructs a parasitic drug-target graph that integrates diverse types of interactions. To capture the semantic relationships within the drug–target network, multi-hop propagation paths are generated to model meta-paths, and a meta-path-guided aggregation mechanism is subsequently applied to hierarchically extract semantic drug features. Then leverages a gene–disease association matrix and a disease similarity matrix to generate parasitic diseases features. Finally, drug features and parasitic diseases features are used to predict drug combination synergy and potential side effects.

Download:

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

Fig 1. Flowchart of the MetaSynMT framework.

A. Drug feature construction: Drug representations are learned by aggregating information in the parasitic drug-target graph guided by meta-paths. B. Parasitic disease feature construction: The parasite disease–gene association matrix and the disease similarity matrix are independently processed by Multilayer Perceptron (MLP) to generate two feature matrices, which are subsequently concatenated. C. Prediction module: The drug–drug features and drug–drug–disease triplet features are input into the task-specific (side effect and synergy) prediction modules to produce corresponding prediction scores.

https://doi.org/10.1371/journal.pntd.0013991.g001

Data acquisition

Currently, no publicly available dataset is specifically designed for predicting drug combinations against parasitic diseases. To fill this gap, we constructed a benchmark dataset that integrates synergy information for drug combinations, drug–drug interaction data, and drug- and disease-associated targets collected from the MalaCards and Harmonizome databases [30,31]. Detailed statistics are summarized in Table 1.

Download:

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

Table 1. Summary of data types and numbers.

https://doi.org/10.1371/journal.pntd.0013991.t001

Parasitic Disease Benchmark Dataset. This dataset contains drug combination information relevant to parasitic diseases, annotated with both synergy and side effects. We selected 25 parasitic diseases with rich annotations and high clinical relevance from the MalaCards database—such as malaria, ascariasis, schistosomiasis, and echinococcosis. Drug combination data for these diseases were collected through a comprehensive PubMed literature search [32]. In total, 336 synergistic drug combinations were extracted from more than 190 research articles and labeled as positive samples, while 337 antagonistic combinations were labeled as negative samples. Across all samples, 393 unique drug pairs were identified, of which 117 are reported to induce adverse interactions and were thus annotated as side-effect positive.

Drug-Related Target Data. The dataset involves 232 unique drugs. For these drugs, a total of 3781 direct and indirect targets were retrieved from the Harmonizome database [31].

Parasitic Disease–Related Gene Data. Disease-associated genes were obtained from the GeneCards database [33], which provides comprehensive annotations for both predicted and validated human gene–disease associations. A union of genes associated with various parasitic diseases was collected, resulting in 1741 unique genes.

Parasitic drug-target graph construction

MetaSynMT constructs a parasitic drug-target graph using drug and target information. From the perspective of synergy, each drug can interact with multiple targets, and therapeutic effects are generally achieved through binding to specific biological targets. From the perspective of safety, overlapping target modules may lead to shared exposures, potentially causing toxic side effects.

In this section, we integrate drug–drug, drug–target, and target–target interaction data to construct a parasitic drug-target graph, denoted as , where V is the set of nodes and E is the set of edges. The node set V consists of two disjoint subsets: the drug set and the target set , where N₁ and N₂ represent the total numbers of drugs and targets, respectively.

The specific data sources used to construct the parasitic drug-target graph G are summarized in Table 2. An example of the constructed parasitic drug-target graph is illustrated in Fig 2(a), which includes three types of undirected edges:

• Drug–drug interaction (DDI),
• Drug–target interaction (DTI),
• Target–target interaction (TTI).

Download:

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

Fig 2. Schematic illustration of drug representation learning.

(a) An example of parasitic drug-target graph. (b) types of meta-paths. (c) Representative instance sequences corresponding to meta-paths M₂ and M₃. (d) An example of and based on (c). (e) The process of hierarchical information aggregation from neighboring nodes at each layer to obtain the node embedding based on the specific meta-path M_k. (f) Aggregate information from all different types of meta-paths to obtain the final feature representation of drug node d_i.

https://doi.org/10.1371/journal.pntd.0013991.g002

Download:

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

Table 2. Summary of data types and numbers in the parasitic drug-target graph.

https://doi.org/10.1371/journal.pntd.0013991.t002

Meta-path construction for parasitic drug-target graph

Since both tasks require learning drug representations, drugs are designated as the endpoints (i.e., start and end nodes) of meta-paths to capture their attribute features. By exploiting the topological relationships between drugs and targets, the model can learn high-order associations that strengthen semantic representation and reasoning. Specifically, we design -type meta-paths, where both ends are drug nodes and the intermediate nodes consist of an arbitrary number of targets, thereby approximating potential interaction routes and facilitating information propagation.

We begin with direct drug-drug interactions, represented by the meta-path M₁: Drug-Drug (DD). To further capture indirect associations, we construct the two-hop meta-path M₂: Drug-Target-Drug (DTD), which reflects relationships mediated by a single target. However, such two-hop structures only model direct drug-target links and fail to incorporate target-target interactions. To address this limitation, we incorporate TTI to construct higher-order meta-paths that encode more complex semantic dependencies.

To generalize this design, we introduce a parameter K (), denoting the number of intermediate target nodes in a -type meta-path. Accordingly, types of meta-paths are generated, as illustrated in Fig 2(b). A meta-path instance is defined as a specific node sequence that conforms to a meta-path pattern, where the start and end drugs are connected through edges following that pattern. Each meta-path type contains multiple such instances, with examples for M₂ and M₃ shown in Fig 2(c). Based on the meta-paths , we then construct the corresponding subgraphs , as shown in Fig 2(d).

Information aggregation of parasitic drug-target graph construction

After constructing the meta-paths, we adopt a meta-path-guided information aggregation mechanism to generate feature representations for drug nodes. This process consists of two stages: (1) drug embeddings are extracted for each type of meta-path, and (2) these embeddings are then aggregated across all meta-path types to obtain the final representations of the drug nodes.

Aggregation of information within each type of meta-path.

In the parasitic drug-target graph G, drug and target nodes are initialized as one-hot vectors unique to their node types, allowing the model to distinguish between them during feature extraction. We then project these one-hot vectors into a higher-dimensional space to capture more complex patterns and relationships. For a node i of type t, the transformation is:

(1)

where W^t is the type-specific transformation matrix, b is the bias vector, and is the activation function.

For the subgraph induced by meta-path M_k, we denote the drug set as and the target set as . As show in Fig 2(e), at the l-th layer, the embedding of drug node d_i is updated by aggregating drug- and target-type neighbors:

(2)

Since different neighbors contribute unequally, we apply an attention mechanism. The attention score of node i at layer l is:

(3)

where a_k is the learnable attention vector and b_a is the bias term.

The final embedding of drug node d_i at the last layer L is:

(4)

After L layers, the embeddings of all drugs are denoted as , and similarly for targets .

Aggregation of information between different types of meta-paths.

After aggregating information from each individual meta-path type, the drug representations obtained from all meta-paths are integrated to derive the final drug features. Since meta-paths capture diverse semantic information of varying importance, we employ an attention mechanism to adaptively assign their weights.

Fig 2(f) illustrates the process of aggregating representations across meta-paths. For a given subgraph constructed from meta-path M_k, we first compute a summary vector by averaging the transformed features of all nodes in the subgraph:

(5)

where n is the number of nodes in , is a learnable weight matrix, and b_n is the bias term. The importance score for meta-path M_k is computed using a trainable attention vector q:

(6)

Next, the softmax function is used to normalize the importance weights to obtain the weight coefficients:

(7)

where K+1 is the number of meta-path types. After obtaining the normalized importance weights, we aggregate the drug representations across all meta-path types. Finally, we can obtain the final feature representation of drug node d_i based on all types of meta-paths:

(8)

Parasitic disease feature construction

The input features for the 25 parasitic diseases are composed of two parts. The first is the parasite–associated gene matrix R₁, a binary matrix where rows correspond to diseases and columns to protein-coding genes associated with them. Entries are set to 1 if a disease is associated with a gene.

The second is the parasite disease similarity matrix R₂, a matrix derived from R₁ using the Jaccard similarity coefficient [34]. Both R₁ and R₂ are processed by a two-layer MLP with ReLU activations. Each input is projected into a 64-dimensional feature space.

The final representation of each parasitic disease is obtained by concatenating the outputs of the two MLPs:

(9)

where H_e denotes the feature vector for disease e.

Predict module

For the drug pair , we have constructed their respective meta-path-based features. The model is trained to predict both side effects and synergy effects. The side effect and synergy prediction modules both consist of multi-layer perceptron with ReLU activation function. For the side effect prediction task, the features of two drugs are concatenated to form drug-drug feature , which is fed into the side effect prediction module to predict the side effect score p_side.

For the synergy prediction task, the features of two drugs and feature of parasitic disease are concatenated to form drug-drug-parasitic disease triplet feature. Then the triple feature is fed into the synergy prediction module to produces synergy probability score p_syn. The detailed process is as follows:

The binary cross-entropy (BCE) loss is employed as the objective function. Let n be the total number of training samples, and let y_i and p_i denote the true label and the predicted probability for the i-th sample, respectively. The BCE loss is defined as:

(10)

The total loss of the model is computed as a weighted sum of the losses for the two tasks:

(11)

where and denote the BCE losses for the synergy prediction and side effect prediction tasks, respectively. The scalar is a learnable parameter that balances the two objectives.

Experiment settings

We employ the hold-out method to split the parasitic disease benchmark dataset into training, validation, and test sets with a ratio of 6:2:2. To ensure robustness, the dataset is randomly partitioned ten times, and the average performance across these runs is reported.

To prevent information leakage, drug–drug–disease triplets in the test set are excluded from the training set. Since drug combinations are order-independent (i.e., drug_row-drug_col and drug_col-drug_row are equivalent), each training sample is duplicated in both orders. For validation and test sets, predictions from both orders are averaged to mitigate any bias introduced by drug ordering.

MetaSynMT is implemented in PyTorch based on a meta-path-guided information aggregation framework. The embedding dimension of node features is set to 64, and the MLP hidden layers contain [2048, 1024, 512] units. The learning rate is 0.0005, and training runs for up to 50 epochs with early stopping if validation metrics do not improve for 10 consecutive epochs. The model is optimized using the Adam optimizer.

Model performance is evaluated using multiple metrics: area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPR), accuracy (ACC), precision, recall, and F1 score. AUC measures the model’s ability to distinguish positive and negative samples, while AUPR reflects the trade-off between precision and recall. ACC indicates the proportion of correctly classified samples. Precision denotes the fraction of true positives among all predicted positives. The F1 score, as the harmonic mean of precision and recall, captures both accuracy and completeness. Higher values for these metrics indicate better predictive performance.

Baseline methods

To evaluate the performance of MetaSynMT, we compare it against several representative baseline methods. Among them, the recently published MLSyPred relies primarily on traditional machine learning algorithms, such as Random Forest, Support Vector Machines, and Gradient Boosting, to generate predictions.

In addition, we include six state-of-the-art deep learning-based methods for drug combination prediction. GraphSynergy [21], Graph Convolutional Network (GCN) [35], Graph Attention Network(GAT) [36], LGSyn [37], and MRHGNN [23] are are graph-based models that utilize heterogeneous network structures, while GAECDS [19], MFSynDCP [38], and DeepDDS [18] rely solely on the molecular structures of drugs for feature extraction. The details of these baselines are as follows:

GCN is a neural network model specifically designed to operate on non-Euclidean graph-structured data using convolutional operations.

GAT introduces an attention mechanism to adaptively weigh the importance of neighboring nodes, thereby enabling more informative feature aggregation for each node.

GraphSynergy constructs a heterogeneous graph architecture and applies spatial graph convolution along with attention mechanisms to encode high-order structural information from drug–cell line and protein modules.

GAECDS integrates graph autoencoders and convolutional neural networks to model the drug synergy prediction task.

MFSynDCP designs a graph aggregation module with an adaptive attention mechanism that dynamically focuses on key information within drug molecular graphs. This allows the model to comprehensively capture crucial interaction patterns between drug pairs.

DeepDDS is a deep learning framework that combines graph neural networks and attention mechanisms. Baseline models are implemented using the hyperparameters reported in their respective original publications.

LGSyn is a novel framework for drug synergy prediction that integrates local molecular features with global biological interaction networks. It employs three fusion strategies to effectively combine these heterogeneous features before feeding them into a deep neural network for prediction.

MRHGNN is a novel dual-channel multimodal relational hypergraph neural network designed for synergistic drug combination prediction. It leverages attention mechanisms to integrate multimodal drug features and employs a unified framework combining primary and self-supervised learning tasks to enhance prediction robustness.

To ensure robustness and reproducibility, each experiment is repeated ten times with random splits, and average performance across these runs is reported.

Comparative performance experiment

We compare MetaSynMT with selected baseline models in terms of synergistic prediction performance on the parasitic disease benchmark dataset, using the synergy prediction score for binary classification analysis. The results, presented in Table 3, show that MetaSynMT consistently outperforms all competing methods across multiple evaluation metrics, including ACC, AUC, AUPR, precision, F1-score, and recall, demonstrating its strong capability in predicting synergistic drug combinations. For example, MetaSynMT achieved a higher AUC than the second method, MFSynDCP (mean difference = 0.041, 95% CI [0.025, 0.057], paired t-test, p < 0.05).

Download:

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

Table 3. Performance comparison of MetaSynMT and baseline methods on the parasitic disease benchmark dataset.

https://doi.org/10.1371/journal.pntd.0013991.t003

Among the baseline methods, the traditional machine learning model MLSyPred exhibits relatively poor predictive performance. In contrast, deep learning-based models perform better, likely due to their superior ability to capture complex, non-linear patterns in high-dimensional data. However, models such as GAECDS, DeepDDS, and MFSynDCP, which rely solely on drug chemical structure, show limited performance due to their inability to integrate relevant target-related information. Although GraphSynergy incorporates protein–protein interaction networks to model topological relationships between drugs, its performance remains inferior to that of MetaSynMT, suggesting that the inclusion of the auxiliary side effect prediction task helps improve overall prediction accuracy. Although GraphSynergy, LGSyn, and MRHGNN incorporate protein–protein interaction networks to model topological relationships between drugs, its performance remains inferior to that of MetaSynMT, suggesting that the inclusion of the auxiliary side effect prediction task helps improve overall prediction accuracy.

Additionally, attention-based models—such as GAT, DeepDDS, GAECDS, and MFSynDCP—demonstrate better performance than the standard GCN, underscoring the importance of attention mechanisms in capturing critical structural and neighborhood-level information.

To verify the robustness and generalization performance of our model, we used a widely applied cancer synergy dataset, Oncology-Screen (On-Screen) dataset [21] for the synergy prediction task. The On-Screen dataset encompasses 4176 drug combination samples involving 21 drugs and 29 cancer cell lines. It includes 2257 positive samples, 1919 negative samples, and 405 drug-related targets. Synergy scores in the On-Screen dataset are calculated using the ZIP metric. Gene expression data for cancer cell lines are obtained from Cancer Cell Line Encyclopedia [39], an independent project designed to characterize genomic profiles, mRNA expression, and anticancer drug responses across cancer cell lines.

The comparison results of MetaSynMT and baseline methods on On-Screen dataset are shown in Table 4. As can be seen from the table, the binary classification evaluation metric of MetaSynMT ranks among the top in all baseline methods. This demonstrates that MetaSynMT exhibits stable predictive performance across different types of datasets, strongly confirming that the MetaSynMT framework possesses excellent generalization capabilities.

Download:

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

Table 4. Performance comparison of MetaSynMT and baseline methods on the on-screen dataset.

https://doi.org/10.1371/journal.pntd.0013991.t004

Comparative experiment of four initial feature extraction methods

To determine the optimal strategy for initial feature extraction, we additionally employed three initial drug/target feature extraction methods, i.e., (1) the Word2Vec [40] chemical language model trained on SMILES representations and the Word2Vec model trained on protein sequences; (2) the Rdkit [41] cheminformatics toolkit and the BERT [42] deep protein sequence representation model; (3) the specially pre-trained ChemBERTa-2 [43] chemical large language model, and the ESM-2 [44] evolutionary scale protein language model. In addition, the SMILES string of the drug is obtained from DrugBank [45], and the protein sequence of the target is obtained from UniProt [46]. The evaluation results on the parasitic disease benchmark dataset are presented in Table 5. The comparison results of four methods revealed that one-hot encoding method achieved the best performance across all three core prediction metrics. Therefore, we selected one-hot encoding approach for its advantages of having few parameters, high computational efficiency, and the ability to quickly provide foundational features for downstream processing. This result also indicates that the initial representation is not the decisive factor for performance, the core strength of our model lies in the subsequent meta-path-guided feature extraction and aggregated process, which learns more discriminative representations from complex biological associations. Pre-trained large models like ChemBERTa-2 and ESM-2 still possess significant advantages, as the rich prior knowledge they encapsulate may play a prominent role in more complex deep learning models.

Download:

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

Table 5. Performance comparison of four different initial feature extraction methods on the parasite benchmark dataset.

https://doi.org/10.1371/journal.pntd.0013991.t005

Ablation experiment and meta-path effectiveness validation experiment

To evaluate the contribution of key components in MetaSynMT, we conducted ablation experiments on the parasitic disease benchmark dataset. The ablated variants include:

1. (1) MetaSynMT-side, which removes the side effect prediction module and performs only the synergy prediction task;
2. (2) MetaSynMT-atten, which removes the attention mechanism from the meta-path-based information aggregation module. In this variant, embeddings of neighbors at different hierarchical levels are aggregated using simple averaging, and drug embeddings across different meta-path types are directly summed without applying attention weights.

As shown in Table 6, MetaSynMT-side consistently underperforms the full MetaSynMT model across all evaluation metrics, indicating that the auxiliary task of side effect prediction enhances the accuracy of synergy prediction. Similarly, MetaSynMT-atten also exhibits inferior performance compared to MetaSynMT, demonstrating that the attention mechanism improves learning over graph structures by enabling the model to focus on more informative node features.

Download:

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

Table 6. Performance comparison of MetaSynMT and its variants for drug combination prediction.

https://doi.org/10.1371/journal.pntd.0013991.t006

To assess the effectiveness of the meta-path-based information aggregation module in MetaSynMT, we replace it with several widely used heterogeneous graph representation learning methods, including heterogeneous GCN, GAT, and DeepWalk [47]. Accordingly, we construct three model variants: MetaSynMT-GCN, MetaSynMT-GAT, and MetaSynMT-DeepWalk. Each variant replaces the original aggregation component with a different method.

As shown in Table 7, MetaSynMT consistently outperforms all variants across the parasitic disease benchmark dataset. The inferior performance of these alternative models may stem from several limitations. Specifically, heterogeneous GCN and GAT often struggle to capture long-range dependencies, as information propagation weakens with increasing hop distance. In contrast, DeepWalk generates node embeddings through random walks, but tends to introduce high levels of information redundancy, which can negatively impact prediction accuracy.

Download:

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

Table 7. Performance comparison of MetaSynMT and its variants with different meta-path aggregation strategies.

https://doi.org/10.1371/journal.pntd.0013991.t007

These results indicate that the meta-path-based aggregation mechanism is more effective for modeling parasitic drug-target graph structures in the context of synergistic drug combination prediction.

Parameter sensitivity analysis

We conduct a sensitivity analysis to evaluate the impact of three key hyperparameters in MetaSynMT: (1) the maximum number of intermediate targets K in the designed meta-paths, (2) the node embedding dimension d, and (3) the learning rate lr. The corresponding results are presented in Fig 3.

• Effect of intermediate target number K. As shown in Fig 3(A), the AUC performance initially increases and then decreases as the number of intermediate targets K increases from 1 to 5, with the best performance observed at K = 3. This trend suggests that incorporating moderate levels of target-hopping enhances the expressiveness of meta-path-based feature representations. However, overly long meta-paths (i.e., larger K) introduce excessive noise, which hinders the effectiveness of information aggregation.
• Effect of embedding dimension d. Fig 3(B) demonstrates a similar pattern for the node embedding dimension d. As d increases from 16 to 128, model performance improves and peaks at d = 64, after which it declines. A small embedding size (e.g., d = 16) may lead to underfitting due to insufficient model capacity, while an overly large dimension (e.g., d = 128) may cause the model to overfit the training data by capturing noise rather than meaningful patterns.
• Effect of learning rate lr. In Fig 3(C), we evaluate four different learning rates: 0.01, 0.001, 0.0005, and 0.0001. The model achieves the best performance when lr = 0.0005, indicating that this setting provides a good balance between convergence stability and optimization efficiency.

Download:

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

Fig 3. (A) Performance metrics (AUC, AUPR, and ACC) under varying values of parameter K.

(B) Performance metrics (AUC, AUPR, and ACC) under varying embedding dimensions d. (C) Performance metrics (AUC, AUPR, and ACC) under different learning rates lr.

https://doi.org/10.1371/journal.pntd.0013991.g003

Key meta-paths analysis

In the parameter sensitivity analysis, we found that the model achieves optimal performance when the maximum number of intermediate targets in the -type meta-paths is set to 3. Accordingly, we design four meta-paths: Drug-Drug (DD), Drug-Target-Drug (DTD), Drug-Target-Target-Drug (DTTD), and Drug-Target-Target-Target-Drug (DTTTD).

To further investigate the contribution of these meta-paths to drug node embedding aggregation, we randomly selected 120 drug combination samples from the training set. The attention scores assigned to each meta-path for these drug pairs are visualized in a heatmap (Fig 4). The analysis reveals that the DTD and DD meta-paths receive the highest attention scores, indicating that the model regards drug-drug and drug-target-drug interactions as the most critical factors during the learning process.

Download:

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

Fig 4. Heat map of attention scores for four types of meta-paths.

https://doi.org/10.1371/journal.pntd.0013991.g004

Further, the necessity of the two key meta-paths is analyzed by biological instances analysis. Taking the drug combination of albendazole and tetrandrine as an example, the meta-path distribution analysis shows contribution scores of 0.31 (DTD), 0.21 (DTTD), 0.22 (DTTTD), and 0.26 (DD), respectively. The main interactions are shown in Fig 5. From a biological perspective, CYP3A4 is the key metabolic enzyme responsible for activating albendazole [48], while tetrandrine significantly inhibits CYP3A4 activity, thereby affecting the plasma concentrations of other drugs metabolized by this enzyme [49]. When used in combination, tetrandrine inhibits CYP3A4, slowing down the further metabolism of the active metabolite of albendazole, which may lead to increased systemic exposure and prolonged retention of the active metabolite. The “Drug-Target-Drug" meta-path may provide the intrinsic mechanisms of the synergistic anti-echinococcal effects of albendazole and tetrandrine. The usage of the “Drug-Target-Drug" meta-path may effectively capture the essential information between two drugs (albendazole and tetrandrine) and the target (CYP3A4).

Download:

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

Fig 5. Meta-path visualization of the albendazole and tetrandrine combination.

https://doi.org/10.1371/journal.pntd.0013991.g005

Case study on predicting synergistic and safe drug combinations

This section presents a case study to demonstrate the capability of MetaSynMT in predicting synergistic and safe drug combinations for parasitic diseases. In the parasitic disease benchmark dataset, the known drug combinations involve 232 drugs and 25 parasitic diseases, yielding a theoretical total of drug–drug–disease triplets. After removing the collected positive and negative samples, the remaining triplets are used to construct the test set. For each sample in the test set, the model generates a synergy score and a side effect score. A higher synergy score indicates stronger predicted synergistic efficacy, while a lower side effect score suggests a safer combination. We first rank the samples based on synergy scores and then sort them by side effect scores, ultimately selecting eight drug combinations with both strong synergy and low toxicity, as shown in Table 8.

Download:

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

Table 8. The eight candidate synergistic and safe drug combinations targeting parasitic diseases.

https://doi.org/10.1371/journal.pntd.0013991.t008

The combination of miltefosine and curcumin has been reported for the treatment of mucocutaneous leishmaniasis [50]. This drug pair exhibits synergistic activity against both promastigote and amastigote forms in vitro, and experimental results further showed increased production of reactive oxygen and nitrogen species, as well as enhanced lymphocyte proliferation. The combination of rifapentine and moxifloxacin has been investigated for the treatment of onchocerciasis. An in vivo screening model was established to evaluate the anti-Wolbachia efficacy of different antibiotic combinations [51]. The results showed that rifapentine (15 mg/kg) combined with moxifloxacin (2×200 mg/kg) produced the most pronounced treatment-shortening effect, regardless of administration route. Although no direct studies currently support its efficacy against ascariasis, both onchocerciasis and ascariasis are caused by nematodes with similar morphology, life cycles, and parasitic behaviors [52], suggesting potential therapeutic relevance. Clinical trials have demonstrated that the combination of moxidectin and albendazole is effective for treating elephantiasis, with no serious adverse events reported [53]. Oral terbinafine combined with biweekly cryotherapy (meglumine antimoniate) has been evaluated in clinical trials for the treatment of cutaneous and mucosal leishmaniasis [54]. Follow-up of the patients during these trials showed full recovery without any reported adverse events. Previous studies have indicated that the combination of antioxidants and anthelmintic drugs, such as artesunate, exhibits synergistic anthelmintic effects against Schistosoma mansoni [55]. As allicin is also a natural compound with significant antioxidant activity, the combination of allicin and artesunate may likewise possess potential anthelmintic efficacy.

These examples confirm the effectiveness of MetaSynMT in predicting synergistic and safe drug combinations. Although the model-generated high-scoring combinations show promising therapeutic potential in theory, their safety and synergy still require further validation through clinical trials.

In vitro experimental analysis of the combination of allicin and sodium stibogluconate for echinococcosis

After applying MetaSynMT to predict the synergy and side effect scores of drug combinations targeting parasitic diseases, we further selected several combinations that demonstrated both strong synergistic effects and low toxicity for the treatment of echinococcosis. The six drug combinations are shown in the Table 9 below. These six combinations collectively involve eight drugs. Cystic echinococcosis is a zoonotic neglected tropical parasitic disease caused by Echinococcus granulosus, while alveolar echinococcosis is caused by Echinococcus multilocularis. Since only the protoscoleces of Echinococcus granulosus are available, in our work we validated the efficacy of the detected drug combinations against cystic echinococcosis. The protoscoleces were harvested from the livers of sheep infected with echinococcus granulosus and subsequently cultured, process is detailed in S1 Text S1.1. Under aseptic conditions, protoscoleces were isolated and cultured in an incubator at 37^°C with 5% CO₂. Protoscoleces with a viability exceeding 90% were transferred to a 96-well plate, with approximately 200 larvae per well. Similarly, the experimental groups were set up with three replicate wells per group, and the culture medium and solutions were replaced every two days. Mortality of the protoscoleces was assessed under an optical microscope using 1% eosin staining, and the average of the three replicate results was taken as the final statistical data.

Download:

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

Table 9. The synergistic and safe drug combinations targeting echinococcosis.

https://doi.org/10.1371/journal.pntd.0013991.t009

We first conducted single-drug maximum concentration efficacy experiments against echinococcosis to select the combination where both individual drugs demonstrated the highest efficacy. The survival curve graph of protoscoleces on single drug is shown in Fig 6. Survival rate records of protoscoleces over 7 days showed varying degrees of anthelmintic efficacy for azithromycin, melarsoprol, moxidectin, moxifloxacin, N-acetylcysteine, and levofloxacin at their maximum single-drug concentrations. Detailed experimental procedures are provided in S1 Text S1.2. Furthermore, among these drugs, allicin, sodium stibogluconate, azithromycin and melarsoprol showed strong efficacy. Allicin was identified as a broad-spectrum antimicrobial agent with no significant hepatorenal toxicity in vivo [56]. Therefore, we focused our experimental validation on the combination of allicin and sodium stibogluconate, highlighting the potential of repurposing traditional drugs for new therapeutic indications.

Download:

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

Fig 6. Survival curves of protoscoleces over 7 days with individual drugs at maximum concentrations.

https://doi.org/10.1371/journal.pntd.0013991.g006

The LD₅₀ values of each drug were determined through gradient concentration assays, where LD₅₀ refers to the concentration required to induce 50% mortality in protoscoleces. These values serve as a critical reference for evaluating the inhibitory effects of each drug. Based on the LD₅₀ values predicted by the SPSS software, the gradient concentrations used were set at [0 μM, 7.0 μM, 12.7 μM, 22.6 μM, 36.3 μM] for allicin and [0 μM, 425.0 μM, 850.0 μM, 893.0 μM, 936.2 μM] for sodium stibogluconate. A dose–response matrix of inhibition rate for the allicin and sodium stibogluconate combination was constructed (Fig 7). The value of LD₅₀ and the design of dose-response matrix are detailed in S1 Text S1.3. This matrix serves as an important tool for evaluating combination efficacy and provides a visual representation of protoscolex survival under varying drug concentrations.

Download:

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

Fig 7. The dose-response matrix of the allicin and sodium stibogluconate combination.

https://doi.org/10.1371/journal.pntd.0013991.g007

From the dose-response matrix, it was determined that the combination of allicin (850 μM) with sodium stibogluconate (36.3 μM) achieved an inhibition rate of 100%. We statistically analyzed the effects on protoscoleces viability in the control group, two monotherapy groups and combination group. As shown in Fig 8, the experiments demonstrated that compared with monotherapy, the combination group (allicin and sodium stibogluconate) significantly increased the inhibition rate of protoscoleces (P < 0.001). Specifically, compared with the second most potent monotherapy, i.e., allicin (inhibition rate = 55.0% ± 3.0%; 95% CI: 49.6%-60.4%; n=3), the combination regimen demonstrated complete inhibition against protoscoleces with an inhibition rate of 100.0% (n=3). These results highlight the synergistic effect of allicin and sodium stibogluconate combination therapy against echinococcosis.

Download:

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

Fig 8. Compared with the negative control, DMSO solvent (1%), albendazole (15 μM), allicin (36.3 μM) or sodium stibogluconat (850 μM) alone, the combination of allicin (36.3 μM) and sodium stibogluconat (850 μM) has a significantly stronger inhibitory effect on echinococcosis protoscoleces.

https://doi.org/10.1371/journal.pntd.0013991.g008

We then employed SynergyFinder [57] to calculate the synergy scores at various concentration combinations. As shown in Fig 9, the heatmap and 3D synergy plots generated using the ZIP, Bliss, HSA, and Loewe models indicate that the ZIP, Bliss, and HSA scores all exceed 10. These results demonstrate that the allicin and sodium stibogluconate combination exhibits significant synergistic effects within a specific concentration range. Notably, the strongest synergy was observed at a concentration of 850 μM allicin and 36.3 μM sodium stibogluconate, as indicated in the heatmap.

Download:

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

Fig 9. Heatmap and 3D surface plot of four synergy scores (ZIP, Bliss, HSA, and Loewe) for allicin and sodium stibogluconate combination.

Synergistic: synergy score > 10; Additive: –10 < synergy score < 10; Antagonistic: synergy score <–10. A t-test was used to calculate the P-values of the four synergy scores (p < 0.09), suggesting a trend toward significance.

https://doi.org/10.1371/journal.pntd.0013991.g009

Overall, these findings confirm the therapeutic potential of the allicin and sodium stibogluconate combination for treating echinococcosis and further validate the practical applicability of MetaSynMT in predicting effective and safe drug combinations. However, further investigations are required to assess the long-term synergy and safety of this combination therapy.