A deep learning-based method for drug-target interaction prediction based on long short-term memory neural network

Abstract

Background The key to modern drug discovery is to find, identify and prepare drug molecular targets. However, due to the influence of throughput, precision and cost, traditional experimental methods are difficult to be widely used to infer these potential Drug-Target Interactions (DTIs). Therefore, it is urgent to develop effective computational methods to validate the interaction between drugs and target. Methods We developed a deep learning-based model for DTIs prediction. The proteins evolutionary features are extracted via Position Specific Scoring Matrix (PSSM) and Legendre Moment (LM) and associated with drugs molecular substructure fingerprints to form feature vectors of drug-target pairs. Then we utilized the Sparse Principal Component Analysis (SPCA) to compress the features of drugs and proteins into a uniform vector space. Lastly, the deep long short-term memory (DeepLSTM) was constructed for carrying out prediction. Results A significant improvement in DTIs prediction performance can be observed on experimental results, with AUC of 0.9951, 0.9705, 0.9951, 0.9206, respectively, on four classes important drug-target datasets. Further experiments preliminary proves that the proposed characterization scheme has great advantage on feature expression and recognition. We also have shown that the proposed method can work well with small dataset. Conclusion The results demonstration that the proposed approach has a great advantage over state-of-the-art drug-target predictor. To the best of our knowledge, this study first tests the potential of deep learning method with memory and Turing completeness in DTIs prediction.

DeepH-DTA: Deep Learning for Predicting Drug-Target Interactions: A Case Study of COVID-19 Drug Repurposing

“…Until recently, the task of modeling DTI has been addressed as a binary classification problem ignoring a vitally significant section of characteristics regarding protein-ligand interactions, specifically the binding affinity scores which represent interactivity strength between DT pairs. Such scores are regularly quantified with measures such as halfmaximal inhibitory concentration (IC50) which relies on the attentiveness of the ligand and target, dissociation constant ( ), and inhibition constant ( ) [6]. Lower values of IC50,…”

“…and values are typically used to compute the negative logarithm of the dissociation or inhibition constants and denoted as or [7], [29]. In DTI binary classification studies, dataset construction is a significant stage, since the selection of the non-binding instances directly influences the performance of the model [6,8,10]. Recently, four datasets have been widely used in several DTI studies in which pairs of DTs with unknown binding evidence are considered as non-binding instances.…”

“…However, adding additional shallow features does not lead to increase performance because of the probable intensification of features. Therefore, to guarantee the effective recognition of compound-target interactions, the typical procedure is to extract a vast quantity of such shallow features ignoring whether they are finally exploited for identification of interactions, and then feature selection techniques-such as Principle Component Analysis (PCA)-are adapted to form the critical DT features into a uniform vector space [6]. Such traditional learning-based DTI schemes, however, are unable to perform well for modeling complex interactions.…”

“…Recurrent neural network (RNN) architectures are proposed for sequential data, where the current data element state is calculated depending on the preceding one or the upcoming one. Mainly, long short-term memory (LSTM) models are talented to capture and remember longer sequences compared to simple RNN or Gated Recurrent Unit (GRUs) which make it the best choice to learn sequential dependencies within molecule sequences or amino acid sequences [6].…”

“…Each of the filters convolve along their receptive field, which restrains the calculated convolutional output = [ 1 , ⋯ , ] from making use of correlation information outside of this region, where ∈ ℝ × passed as input to squeeze and excite (SE) module to be combined using global average pooling (GAP) to produce a channel descriptor of the entire context of input channels. Hence, the spatial squeeze of the f-th channel is calculated using the spatial squeeze function sq as expressed in equation (6),…”