DPSP: a multimodal deep learning framework for polypharmacy side effects prediction

Abstract Motivation Because unanticipated drug–drug interactions (DDIs) can result in severe bodily harm, identifying the adverse effects of polypharmacy is one of the most important tasks in human health. Over the past few decades, computational methods for predicting the adverse effects of polypharmacy have been developed. Results This article presents DPSP, a framework for predicting polypharmacy side effects based on the construction of novel drug features and the application of a deep neural network to predict DDIs. In the first step, a variety of drug information is evaluated, and a feature extraction method and the Jaccard similarity are used to determine similarities between two drugs. By combining these similarities, a novel feature vector is generated for each drug. In the second step, the method predicts DDIs for specific DDI events using a multimodal framework and drug feature vectors. On three benchmark datasets, the performance of DPSP is measured by comparing its results to those of several well-known methods, such as GNN–DDI, MSTE, MDF–SA–DDI, NNPS, DDIMDL, DNN, DeepDDI, KNN, LR, and RF. DPSP outperforms these classification methods based on a variety of classification metrics. The results indicate that the use of diverse drug information is effective and efficient for identifying DDI adverse effects. Availability and implementation The source code and datasets are available at https://github.com/raziyehmasumshah/DPSP.


DS2 T P (F
. The process of selecting the method for aggregating input features in the neural network in DS1.
Table S43.The process of feature learning based on averaging the outputs on DS1, DS2, and DS3.

Fig. S1 .
Fig. S1.Boxplots of F score, Precision ,and Recall values of 65 events on DS1 for DPSP and GNN-DDI methods.

Fig. S2 .
Fig. S2.Boxplots of F score, Precision ,and Recall values of 100 events on DS2 for DPSP and GNN-DDI methods.

Fig. S3 .
Fig. S3.Boxplots of F score, Precision ,and Recall values of 409 events on DS3 for DPSP and GNN-DDI methods.

Fig. S5 .
Fig. S5.Venn diagram of comaring the impact of results of the method using all features (F) with excluding mono side effect feature (F1) in DS1 by computing false positive (FP).

Fig. S6 .
Fig. S6.Venn diagram of comaring the impact of results of the method using all features (F) with excluding target feature (F2) in DS1 by computing true positive (TP).

Fig. S7 .
Fig. S7.Venn diagram of comaring the impact of results of the method using all features (F) with excluding target feature (F2) in DS1 by computing false positive (FP).

Fig. S8 .
Fig. S8.Venn diagram of comaring the impact of results of the method using all features (F) with excluding enzyme feature (F3) in DS1 by computing true positive (TP).

Fig. S9 .
Fig. S9.Venn diagram of comaring the impact of results of the method using all features (F) with excluding enzyme feature (F3) in DS1 by computing false positive (FP).

Fig. S10 .
Fig. S10.Venn diagram of comaring the impact of results of the method using all features (F) with excluding chemical substructure feature (F4) in DS1 by computing true positive (TP).

Fig. S11 .
Fig. S11.Venn diagram of comaring the impact of results of the method using all features (F) with excluding chemical substructure feature (F4) in DS1 by computing false positive (FP).

Fig. S12 .
Fig. S12.Venn diagram of comaring the impact of results of the method using all features (F) with excluding pathway feature (F5) in DS1 by computing true positive (TP).

Fig. S13 .
Fig. S13.Venn diagram of comaring the impact of results of the method using all features (F) with excluding pathway feature (F5) in DS1 by computing false positive (FP).

Fig. S14 .
Fig. S14.Venn diagram of comaring the impact of results of the method using all features (F') with excluding target feature (F'1) in DS2 by computing true positive (TP).

Fig. S15 .
Fig. S15.Venn diagram of comaring the impact of results of the method using all features (F') with excluding target feature (F'1) in DS2 by computing false positive (FP).

Fig. S16 .
Fig. S16.Venn diagram of comaring the impact of results of the method using all features (F') with excluding enzyme feature (F'2) in DS2 by computing true positive (TP).

Fig. S17 .
Fig. S17.Venn diagram of comaring the impact of results of the method using all features (F') with excluding enzyme feature (F'2) in DS2 by computing false positive (FP).

Fig. S18 .
Fig. S18.Venn diagram of comaring the impact of results of the method using all features (F') with excluding chemical substructure feature (F'3) in DS2 by computing true positive (TP).

Fig. S19 .
Fig. S19.Venn diagram of comaring the impact of results of the method using all features (F') with excluding chemical substructure feature (F'3) in DS2 by computing false positive (FP).

Fig. S20 .
Fig. S20.Venn diagram of comaring the impact of results of the method using all features (F") with excluding mono side effect feature (F"1) in DS3 by computing true positive (TP).

Fig. S21 .
Fig. S21.Venn diagram of comaring the impact of results of the method using all features (F") with excluding mono side effect feature (F"1) in DS3 by computing false positive (FP).

Fig. S22 .
Fig. S22.Venn diagram of comaring the impact of results of the method using all features (F") with excluding target feature (F"2) in DS3 by computing true positive (TP).

Fig. S23 .
Fig. S23.Venn diagram of comaring the impact of results of the method using all features (F") with excluding target feature (F"2) in DS3 by computing false positive (FP).

9344 0.9773 0.9990 0.9309 0.9309 0.9309
Note: The selected model is indicated in bold

Table S5 .
The results of different neural network architectures on DS2.

Table S6 .
The results of different neural network architectures on DS3.
Note: The selected model is indicated in bold

Table S8 .
Results of comparison of the DPSP with some of the machine learning methods on DS3.
Note: Bold numbers show the best performance for each criterion

Table S10 .
Determining the impact of removing individual features by computing the true positives (TP) and false positives (FP) for each feature in DS1.

Table S11 .
This table demonstrates the significance of DS2 data set features by removing one feature and evaluating the results of the DPSP method using the remaining features based on all evaluation criteria.

Table S12 .
Determining the impact of removing individual features by computing the true positives (TP) and false positives (FP) for each feature in DS2.

Table S15 .
This table displays the results of the DPSP method utilizing only the pathway feature from DS1.

Table S16 .
This table displays the results of the DPSP method utilizing only the chemical substructure feature from DS2.

Table S17 .
This table displays the results of the DPSP method utilizing only the mono side effect feature from DS3.

Table S18 .
The results of performing AutoEncoder (AE) dimensionality reduction technique on DS1.

Table S19 .
The results of performing Entropy dimensionality reduction technique on DS1.

Table S20 .
The results of performing PCA dimensionality reduction technique on DS2.

Table S21 .
The results of performing AutoEncoder (AE) dimensionality reduction technique on DS2.

Table S22 .
The results of performing Entropy dimensionality reduction technique on DS2.

Table S23 .
The results of performing PCA dimensionality reduction technique on DS3.

Table S24 .
The results of performing AutoEncoder (AE) dimensionality reduction technique on DS3.

Table S25 .
The results of performing Entropy dimensionality reduction technique on DS3.

Table S26 .
The results of performing PCA dimensionality reduction technique on all methods in scenario1 on DS1.

Table S27 .
The results of performing PCA dimensionality reduction technique on all methods in scenario2 on DS1.

Table S28 .
The results of performing PCA dimensionality reduction technique on all methods in scenario1 on DS2.

Table S29 .
The results of performing PCA dimensionality reduction technique on all methods in scenario2 on DS2.

Table S30 .
The results of performing PCA dimensionality reduction technique on all methods in scenario1 on DS3.

Table S31 .
The results of performing PCA dimensionality reduction technique on all methods in scenario2 on DS3.

Table S32 .
Data availability for each feature type in DS1.

Table S33 .
Data availability for each feature type in DS2.

Table S34 .
Data availability for each feature type in DS3.

Table S35 .
Effect of different feature combinations in DS1 on performance of model We used symbols (M for mono side effects, T for targets, E for enzymes, P for pathways, and S for smiles features) to represent the different types of features (part1).
Note: Bold numbers show the best performance for each criterion

Table S36 .
Effect of different feature combinations in DS1 on performance of model We used symbols (M for mono side effects, T for targets, E for enzymes, P for pathways, and S for smiles features) to represent the different types of features (part2).

Table S37 .
Effect of different feature combinations in DS2 on performance of model We used symbols (T for targets, E for enzymes, and S for smiles features) to represent the different types of features.

Table S38 .
Effect of different feature combinations in DS3 on performance of model We used symbols (M for mono side effects, T for targets features) to represent the different types of features.

Table S39 .
Comparison results between the execution time of the DPSP method and other machine learning methods on all datasets.

Table S44 .
The process of feature learning based on voting the outputs on DS1, DS2, and DS3.

Table S45 .
This table displays the performance of the DPSP method on the five most common drug combinations in DS1 according to all evaluation criteria.

Table S46 .
This table displays the performance of the DPSP method on the five most common drug combinations in DS2 according to all evaluation criteria.

Table S47 .
This table displays the performance of the DPSP method on the five most common drug combinations in DS3 according to all evaluation criteria.

Table S48 .
This table displays, for each of the five most frequent events on DS3, the new interactions predicted by the DPSP method with the highest probabilities.