A computational model for GPCR-ligand interaction prediction

2.1 Data acquisition

Tuples construction: GPCRs list retrieved based on keyword <Family: “G Protein-Coupled Receptor” AND Reviewed: “Yes”> in UniProt protein database available at www.uniprot.org [20]. Table S1 in the supplementary file shows the list of acquired GPCRs. This list was matched with core GPCRs interactions accessible at the database of interacting protein (DIP) [21]. To do this, all interaction pairs in DIP (Table S2) that at least one of them was GPCR (based on Table S1) were extracted, and the rest of them eliminated. Finally, 224 interactions were extracted (Table S3). These pairs were considered as pairs that there is DIP confirmed evidence for the interaction between them. About 116 receptors and 141 ligands were in the list of selected interactions (Table S4 and S5). From obtained receptors and ligands, 16356 possible GPCR-ligand pairs were constructed (Table S6). ML algorithms need negative tuples (pairs that have no theoretical/practical evidence for interaction between them). Therefore, to make negative examples for the ML algorithms’ training, a random selection method was employed to make a list of negative tuples for the training of the ML algorithms (Table S7).

2.2 Feature extraction and combination

The proteins’ sequences are the basis for feature extraction in many studies [22], [23], [24], [25], [26]. GPCRs and ligands sequences were extracted from UniProt. The sequences are strings made of 20 amino acids. The conserved domains of GPCRs and ligands were the information sources employed in this work to extract features about GPCRs and ligands. NCBI conserved domain database (CDD) [27] was used to extract domains. In this regard, domains with E-VALUE = 0.01 were extracted from CDD. Two types of domains including, specific and non-specific domains in various confidence levels, were extracted. Moreover, two domain model scopes, including super-families and multi-domains, were extracted for each sequence. Definitions below are adopted from the NCBI database webpage [17]:

Specific domain: the top-ranking hit (compared to other hits in overlapping intervals) that meets or exceeds a domain-specific E-value threshold (details and illustration). It represents very high confidence that the query sequence belongs to the same protein family as the sequences used to create the domain model, and therefore a high confidence level for the inferred function of the protein query sequence.

Non-Specific domain: If a specific hit was not found on a query protein sequence, but the protein has an otherwise statistically significant hit (E-value cutoff of 0.01) to any domain model in CDD, the domain model is regarded as a non-specific hit.

Super-family is the domain cluster to which the specific and/or non-specific hits belong. Super-family is a set of conserved domain models that generate overlapping annotation on the same protein sequences and are assumed to represent evolutionarily related domains.

Multi-domains: Multi-domains are domain models that were computationally detected and are likely to contain multiple single domains.

For every sequence, four groups of features, including specific domains, non-specific domains, multi-domains, and super-family, were extracted (see Tables S8–S11). Figure 1 shows extracted features based on conserved domains for guanine nucleotide-binding protein (UniProt id: P10823).

Figure 1:

Extracted domains of guanine nucleotide-binding protein (UniProt id: P10823).

Finally, to determine training capability for all four groups of features, different datasets with all possible compositions of the feature groups were constructed. The algorithms’ performance with different data sets were evaluated based on the area under the curve (AUC).

2.3 Parameter setting

Optimization parameters of the algorithms with different values were examined, and then the best parameters for optimal performance were selected. In this study, for the decision tree (DT) algorithm, K, and D parameters, for the random forest (RF) algorithm I, K, and S parameters, for multi-layered perceptron algorithms number of nodes, hidden layer, learning rate, and momentum parameters, and for support vector machine (SVM) C and gamma parameters were optimized.

2.4 Feature selection

Feature selection (FS) approaches were used to improve the predictive capability of employed models. Relief, information gain, and SVM methods were used for FS. First, all the features ranked based on FS algorithms, and then the feature with the lowest importance (based on FS algorithms) was removed. After each feature elimination, the performance of the prediction models was measured. Elimination of the features and performance measuring was continued until the best set of features was determined.

2.5 Prediction model construction

After determining the best feature group and determining the best parameter setting for prediction models based on their performance over training examples, the best prediction model was selected and constructed. Five common ML algorithms, including SVM, multilayer perceptron (MLP), Naive Bayes (NB), RF, and DT were evaluated based on their performance. All five algorithms with different feature sets were examined and compared to determine the best predictor. Algorithms performance assessed based on the standard assessment criteria such as accuracy and receiver operating characteristic (ROC).

2.6 Measurements and the performance evaluation of models

In this section, details of the performance evaluation criteria were explained. A 10-fold cross-validation method was used to evaluate the performance of the algorithms. Also, 70% of the data were used as a training set, and the remaining 30% were used as a test set. Recruited performance evaluation measures in this study are as follow:

True Positive (TP): interacting receptor-ligand pairs that correctly identified as interacting pairs.

False Positive (FP): non-interacting receptor-ligand pairs that incorrectly identified as interacting pairs.

True Negative (TN): non-interacting receptor-ligand pairs that correctly identified as non-interacting pairs.

False Negative (FN): interacting receptor-ligand pairs that incorrectly identified as non-interacting pairs.

Accuracy (ACC): The percentage of interactive and non-interactive pairs that are predicted correctly.

Sensitivity (SN): The percentage of interactive pairs that are properly predicted.

Specificity: The percentage of non-interactive pairs that are predicted correctly.

The operating characteristic curve area (AU-ROC) was also calculated to measure the performance of the models. AU-ROC is often used to report the performance of binary predictors [28].

2.7 Prediction of potential GPCR-ligand interactions

In this step, the constructed model based on the previous section was used to predict the potential interaction of GPCR-ligand pairs. The model was applied on all GPCR-ligand pairs with unknown interaction label to predict potential GPCR-ligand pairs.

2.8 Software

In this study, Weka (version 3.8.0) was used for data mining purposes, and Gephi (version 0.9.1) was used for graph plotting.

3.1 Data acquisition result

This study utilized the capability of ML approaches to predict GPCR-ligand interaction. Known (experimentally confirmed) GPCR-Ligand interactions used to train the algorithms (see Tables S1–S7 in supplementary file). The overall process of this study is described in Figure 2.

Figure 2:

Overall illustration of the proposed approach for GPCR-ligand interacting pairs.

3.2 Feature extraction and combination

Based on the mentioned feature extraction methodology mentioned earlier, five groups of the features (A total of 3118 features) were extracted. Table 2 shows the extracted feature groups in detail.

The performance of the five ML algorithms with varying groups of features was assessed based on AU-ROC (Table 3).

Among all the examined algorithms, RF algorithm with AU-ROC = 97.1 had the best performance. Moreover, among different feature groups, features based on k-mer method and super-families presented better predictive capabilities (Table 3). All possible combinations of the five feature groups were also examined to utilize various feature groups’ predictive capabilities. Table 4 shows the performance of all ML algorithms with all possible feature groups. More details about the algorithm’s evaluation are available in the Supplementary Table S12.

Table 4 shows that, in some experiments, some algorithm’s performance with combinatory features were improved. However, some algorithm’s performance with combinatory features were deteriorated (i.e., linear SVM and MPL). Table 5 provides information about the algorithms’ performance before and after the combination of the feature groups.

At the end of this step, a set of features with best performance was selected for each algorithm (Table 6).

3.3 Parameter setting

Different parameters and settings for each algorithm were examined to obtain a more accurate model (Table 7). For example, the values for trees, feature, and seed were investigated for the RF algorithm.

3.4 Feature selection

For each algorithm’s best feature set was determined, and then three FS methods (i.e., Info Gain, relief, and SVM) were performed in the last step of the optimization process. At first, features ranked based on FS, and 10 features with the lowest impact were eliminated, and the algorithms’ performance was measured. The following table (Table 9) summarizes the results of performing FS. More details are available in Table S18 in the supplementary file.

3.5 Prediction model construction and identifying GPCRs-ligands pairs with the potential to interact

RF with AU-ROC = 98.1 was selected as the best prediction model. The final model was constructed based on the RF algorithm with selected features from specific domains, non-specific domains, super-family, and multi-domains (870 features). Finally, a constructed model was applied over 16,132 GPCR-ligand pairs, and 13,295 potential interactions were identified.