DTIP-WINDGRU a novel drug-target interaction prediction with wind-enhanced gated recurrent unit | BMC Bioinformatics

The proposed model

The DTIP-WINDGRU model uses pre-processing procedures to get the data ready for prediction. Additionally, class labeling is involved, which entails giving known interactions labels (such as positive or negative) for unlabelled instances in the data. D-D and T-T interactions are taken into consideration during the actual prediction process using the GRU. On the account that the protein sequences are lengthy, a basic CNN is unable to adequately capture the context dependencies within them. Consequently, we enhance it by interpreting the protein sequences as time series and use the GRU to extract their features. In order to ensure that the model performs better in protein feature extraction, we employ GRU to capture the long-term dependencies in order to handle the feature extraction problem for lengthy amino acid sequences. The findings of the four models have been improved to varying degrees. A GRU, a kind of recurrent neural network (RNN), is at the centre of the model. The model’s weights are initialized using these interactions, facilitating the model’s ability to learn and produce accurate predictions. The model uses the WDO method in the concluding step. The GRU model’s optimal hyperparameters are chosen using the model behavior-influencing settings, which is essential for obtaining reliable findings. A new method for DTI prediction combines a GRU model with D-D and T-T interactions. It is potentially beneficial to optimize hyperparameters using the WDO technique. The model’s versatility in tackling DTI prediction difficulties is indicated by its ability to handle labeled and unlabeled data. Overall, the proposed DTIP-WINDGRU model provides a unique viewpoint on DTI prediction and includes innovative components to enhance precision and effectiveness in this significant field of study. The process model for the DTIP-WINDGRU approach is shown in Fig. 1.

The step-by-step explanation of the model’s workflow includes:

Step 1: Data Pre-processing:

Class Labeling: We assign labels to known interactions (positive or negative) for previously unlabeled instances in the dataset.

Interaction Consideration: The D-D and T-T interactions are taken into account to enrich the data.

Step 2: Feature Extraction:

Protein Sequence Processing: The protein sequences are interpreted as time series data.

GRU: A GRU is employed to extract features, capturing long-term dependencies within the sequences.

Step 3: Model Initialization:

Weight Initialization: The model’s weights are initialized using the interaction data, facilitating effective learning.

Step 4: Prediction Process:

GRU-Based Prediction: The GRU model processes the input data to predict drug-target interactions.

Step 5: Hyperparameter Optimization:

WDO Method: The WDO technique is used to optimize the GRU model’s hyperparameters, ensuring reliable and accurate predictions.

Dataset description

The performance of the proposed method is validated by utilizing the gold standard dataset. The databases DrugBank, KEGG BRITE, SuperTarget, BRENDA, and others are used to find drug-target interactions but due to bias concerns caused by the greater number of negative samples than positive data, DTI pairings only include positive interactions since negative interactions have not been empirically proven. The negative interactions must be reprocessed to return to their original state since they lack experimental support, which may result in the emergence of new positive interactions.

The gold standard dataset was originally compiled by Yamanishi et al. (2008), integrating information from multiple databases including DrugBank, BRENDA, SuperTarget, and KEGG BRITE. It provides benchmark datasets for four classes of target proteins: Enzymes, Ion Channels, G-Protein Coupled Receptors (GPCRs), and Nuclear Receptors. Each dataset consists of known positive drug-target interactions, with drugs and targets represented using chemical structure and protein sequence similarity matrices respectively. The dataset is publicly available at: http://web.kuicr.kyoto-u.ac.jp/supp/yoshi/drugtarget/

Pre-processing and class labeling

In the proposed work, we have used four DTIP data sets such as Ion Channel, Enzyme, Nuclear Receptor, and GPCR. The drug chemical data is attained in DRUG AND COMPOUND Section in the KEGG LIGAND. The chemical formation resemblance amongst the components is defined by SIMCOMP which provides a scoring value based on the size of substructure with graph alignment. The target protein sequence similarity is evaluated using the Smith-Waterman technique. It is used to achieve local sequence alignment and to find similarities between two sequences (target protein amino acid sequences). The Smith-Waterman algorithm concentrates on locating the best local alignments, useful for locating homologous or functional domains in proteins by understanding the connections and probable functional similarity between various proteins. The sequence similarity amongst the target is defined by the normalized Smith–Waterman method, based on the data of amino acid order of target protein derived from KEGG GENE data base. A pair of proteins \(A_{i}\) and \(A_{j}\), the sequence similarity amongst them can be defined by the following equation:

where \(SW\left( {A_{i} , A_{j} } \right)\) denotes the score of Smith–Waterman method. Next, the proposed approach undergoes class labelling where the target interaction exists in the dataset are utilized for labelling the unknown instance.

The class labeling method categorizes instances in a dataset, particularly those connected to target interactions. The process of class labeling entails categorizing or labeling the data instances using target interactions. The target interactions and their weights are used to label the unlabeled instances in the DTI dataset. The unknown instances, or data points without labels indicating whether the interaction exists or not, are subjected to the class labelling procedure. This technique labels ambiguous occurrences by using the positive interactions from known target interaction information. The unknown instances are classified as positive interactions if they have similarities with known target interactions; otherwise, they are classified as negative interactions. In essence, this class labeling phase uses weights from RNN, where the algorithm learns from labeled examples to generate predictions or classifications on fresh, unlabeled data, to help categorize data instances in the dataset.

Design of GRU based predictive model

In a RNN, the NN utilizes recurrent connection in all the units. The activations of neurons are fed into weights, and each unit incorporates a time delay along with a hidden memory value. This enables the network to learn the temporal dynamics of successive information by retaining historical activations [22]. Assuming a temporal input sequence \(a^{l} = \left( {a_{1}^{l} ,{ } \ldots ,a_{T}^{l} } \right)\) of length \(T\) (\(a_{t,i}^{l}\) the activation of unit \(i\) in hidden layer \(l\) at time \(t\)), an RNN maps to a series of hidden value \(h^{l} = \left( {h_{1}^{l} ,{ } \ldots ,h_{T}^{l} } \right)\) and outputs a sequence of activation \(a^{{\left( {l + 1} \right)}} = \left( {a_{1}^{{\left( {l + 1} \right)}} ,{ } \ldots ,a_{T}^{{\left( {l + 1} \right)}} } \right)\) by iterating through the following equation:

where \(\sigma\) refers the nonlinear activation function, \(b_{h}^{l}\) signifies the hidden bias vector and \(W\) stands for the weight matrices, \(W_{xh}^{l}\) represents the input‐hidden weight matrix and \(W_{hh}^{l}\) implies the hidden‐hidden weight matrix. It represents the weights connecting the hidden layer at time t − 1 to the hidden layer at time t in the RNN equation [9]. Here, \(W_{hh}^{l}\) captures the influence of the previous hidden state on the current hidden state. It is a matrix of weights governing the recurrence in the hidden layer. It was determined in the following:

where \(W_{ha}^{l}\) denotes the hidden activation weight matrix and \(b_{a}^{l}\) signifies the activation bias vector. It is worth highlighting the weight matrix \(W^{l}\) determined for the MLP corresponds to \(W_{xh}^{l}\) matrix in Eq. (2). Figure 2 depicts the structure of GRU.

These kinds of networks have Turing abilities [23] and, therefore, they are suitable for learning sequences but the memory model makes learning difficult to handle real‐time sequence processing. The LSTM extended RNN with memory cell, rather than recurrent unit, for storing and resultant data, easing the learning of temporal connections on long time scale. The GRU has a basic form of LSTM that is also a type of RNN. Unlike LSTM, GRU combines forget and input gates. In consideration of a network with h hidden layers, a small‐batch input at a given time step of \(t\) is denoted as \(X_{t} \in {\mathbb{R}}^{{n{*}d}}\), and the hidden state at the preceding time step t1 is \(H_{t – 1} \in {\mathbb{R}}^{{n{*}h}}\). The resulting hidden layer \(h\) of single GRU at existing time step \(t\):

Here, \(\sigma\) represents the sigmoid activation function, defined as \(\sigma \left( x \right) = 1/1 + e^{ – x}\). The weights connecting the hidden layer and update gate are represented by \(W_{hz}\). The weights \({\text{connecting input layer and reset gate are denoted by }}W_{xr} , while W_{hz} {\text{and }}W_{xr}\) represents the hidden layer and reset gate, input layer and update gate. Here, \(b_{r}\) and \(b_{z}\) indicate the bias of reset and update gates, while \(H_{t}\) points toward the candidate hiding state of existing time step denoted by \(t.{ }\) The symbol \(\odot\) represent the matrix multiplication of two components, and \(Tanh\) characterizes the hyperbolic tangent activation function:

Hyperparameter optimization

The optimal adjustment of the hyperparameters involved in the GRU model, includes WDO algorithm. The WDO algorithm is derived from the concept of atmosphere driven by forces such as wind, exhibit optimization-like characteristics. It is believed that by integrating these ideas into an optimization algorithm, the WDO method provides special benefits in terms of effectiveness and scalability [24]. The potential utility of the WDO algorithm lies in its capacity to navigate complex solution spaces and identify optimal configurations [31]. The algorithm has characteristics advantageous for maximizing the selection of therapeutic targets inspired by atmospheric dynamics. It is used to improve prediction accuracy with specific biological targets, contributing to the drug discovery process by precisely identifying potential therapies. The basis of WDO approach is Newton’s second law of motion that is utilized for providing precise results for analyzing atmospheric motion in the Lagrangian [24]:

where \(\vec{\alpha }\) represent the acceleration, \(\rho\) indicates the air density for infinitesimal air parcel, and \(\vec{F}\) denotes each force act on the air parcel. The air pressure establishes the relation with air parcel temperature and density, it can be shown as follows

Now \(P\) indicates the pressure, \(R\) denotes the universal gas constant, and \(T\) represents the temperature. The reason of air movement is because of the integration of several forces, largely including gravitational force \(\left( {\vec{F}_{G} } \right)\), pressure gradient force \(\left( {\vec{F}_{PG} } \right)\), Coriolis force \(\left( {\vec{F}_{C} } \right)\), and friction force \(\left( {\vec{F}_{F} } \right)\). it is described in the following

where \(\delta V\) denotes finite volume of the air, \(\vec{g}\) indicates the gravitational acceleration, \(\nabla P\) characterizes the pressure gradient, \(\Omega\) denotes rotation of the Earth, \(\vec{u}\) epitomizes the velocity vector of the wind, and \(\alpha\) signifies the friction coefficient. The abovementioned forces are added to (10).

where the acceleration \(\vec{\alpha }\) in (1) is expressed as \(\vec{\alpha } = \vartriangle \vec{u}/\vartriangle t\); for simplicity set \(\vartriangle t = 1\); for infinitesimal air parcel, set \(\delta V = 1\), that simplify (4) to

Based on (11), the density \(\rho\) is expressed interms of the pressure;

where \(P_{cur}\) indicates the pressure of present position. It is considered in the WDO approach [32] that position and velocity of the air parcel change at all the iterations. Therefore, \(\vartriangle \vec{u}\) is expressed as \(\Delta \vec{u} = \vec{u}_{new} – \vec{u}_{cur}\), in which \(\vec{u}\) indicates the velocity in following iteration and \(\vec{u}\) indicates the velocity at the present iteration. \(\vec{g}\) and \(\nabla P\) denotes vectors, they are broken down in direction and magnitude as \(\vec{g} = \left| g \right|\left( {0 – x_{cur} } \right),\) \(- \nabla P = \left| {P_{opt} – P_{cur} } \right|\left( {x_{opt} – x_{cur} } \right),\) \(P_{opt}\) denotes the optimal pressure point, \(x_{opt}\) indicates the optimal position, and \(x_{cur}\) represent the existing position; update (14) with the new equation, (14) as follows

At last, there are three further substitutions required. The actual pressure value is substituted with rank amongst each air parcel according to the pressure value, the resultant formula of updating the velocity is defined in the following [19]:

whereas \(i\) denotes the ranking amongst each air parcel and \(\vec{x}_{new}\) indicates the novel position for the following iteration.

The WDO system develops a fitness function (FF) for reaching enhanced classifier results. It defines a positive integer for demonstrating the best efficiency of the candidate solution. In case of, minimizing of the classification error rate was regarded as FF is offered in Eq. (18).

In the earlier sections, a prediction model based on the GRU with wind-driven optimization techniques [25] to carry out a thorough hyperparameter optimization was designed to predict DTIs. The complex process of parameter tuning plays a vital link between the design of our proposed approach and the validation stage that follows. As we proceed with the validation phase, our aim is to provide a smooth transition between these stages described in Fig. 1 and clarify how the improved parameters and the model’s approach supports the stability and effectiveness of the prediction model.