Table 5 Introduction to Deep Learning Methods

Zhao et al.⁷⁸

2020

DrugBank

Improved graph representation learning with multiple data.

The information related to drugs, targets and drug-target pairs are merged into the heterogeneous graph. This method enriches the meaning of node information and optimizes the excessive smoothing phenomenon in traditional graph neural networks.

Monteiro et al.⁷⁹

2020

Private dataset

Based on Convolutional Neural Network (CNN) with inputs as protein sequences and structured compounds SMILES.

This approach is preferable to obtaining a table of data from a traditional descriptor.

Zhao et al.⁵⁰

2020

Davis, KIBA

A semi-supervised approach based on Generative Adversarial Networks (GANs) for predicting binding affinities.

The method consists of two parts, two GANs for feature extraction and a regression network for prediction, which is the first method based on semi-supervised GANs to predict binding affinity.

Jin et al.³⁶

2021

KIBA

Enhanced molecular expression prediction of drug-target interactions by sequence embedding and graph convolutional networks.

For protein sequences, feature embeddings of amino acids are pre-trained using language modeling and fed into a convolutional neural network model for representation learning.

Wang et al.⁸⁰

2021

DUD-E, human, BindingDB

Multitasking neural network consisting of a graph neural network, an attention module, and a multilayer perceptron.

Good multitasking performance, excellent in both binding affinity prediction and interaction categorization.

Huang et al.⁸¹

2021

Biosnap, Davis, BindingDB

The method specifically selects a knowledge-heuristic pattern mining algorithm for the substructures and models the interaction of the data on this basis. An enhanced transformer encoder is also employed in unlabeled data to capture information about its substructures.

Fully exploiting parsing of substructures and potential knowledge of unlabeled data.

Cheng et al.⁸²

2021

human, C.elegans, DUD-E, DrugBank

Based on graph neural networks, the key binding features of sequences are searched for through two engagements of the multiple attention mechanism.

Convolutional neural network is affected by the size of the convolutional window, and can only extract the local correlation information in the window area, but not the long-range contextual relationship. This method can solve the long-range dependency problem of sequences to some extent.

Peng et al.⁴⁴

2021

Luo and Yamanishi

An end-to-end heterogeneous graph learning architecture.

The input to this heterogeneous network contains biological information in multiple modalities, and a low-dimensional feature representation of the data is learned end-to-end in model training.

Wang et al.⁸³

2022

Private dataset

Construction of a knowledge graph of drug-target pairs.

Knowledge graph to get recommendation information.

Li et al.⁴⁵

2022

Luo

Graph neural networks are utilized to synthesize node attributes and topological information.

The extracted low-dimensional features are more representative, taking into account the nodes’ own attributes and the relationships between nodes.

Yu et al.⁴⁷

2022

Luo

A framework for predicting drug-target interactions by information aggregation based on heterogeneous graphical neural networks with incorporation of attention mechanisms.

The method first obtains the molecular fingerprint information of the drug and the pseudo-amino acid composition information of the protein, and then extracts the initial features of the nodes by Bi-LSTM and aggregates the heterogeneous neighbors using the attention mechanism.

Wang et al.⁵²

2022

Luo

A method for predicting drug-target interactions based on the convolution of heterogeneous network graphs integrating network information on drug-target interactions, drug-drug interactions, drug-drug similarity, target-target interactions, and target-target similarity.

Graph convolution operations are performed in the heterogeneous network to obtain node embeddings of drugs and targets. An attention mechanism is also introduced between the graph convolution layers to jointly process the node embeddings from each layer. The model uses fewer network types and has high predictive performance.

Nguyen et al.⁵³

2022

PDBbind, Fragalysis

A method based on graph neural networks that introduces a multi-hop gating mechanism, as well as an atomic enhancement design for data in the face of insufficiently rich coronavirus datasets.

Better atomic representations can be obtained from non-neighbor information to improve the training effect of the model, while the authors mentioned that if a large dataset can be introduced to do migration learning can improve the model performance to better.

Zhao et al.⁸⁴

2022

Kinase, Davis, Metz, KIBA

Deep sequence learning for drug target binding affinity prediction based on attention mechanisms.

Leverage attention mechanisms to focus on important key subsequences in drug and protein sequences to predict affinity.

Zhang et al.⁴

2022

Tang, DrugBank, KEGG and PubChem

Transformer-based deep learning model with convolutional network and graph neural network can effectively extract the local residual information in the sequence, and the interaction data of molecular structure.

Wang et al.⁸⁵

2022

BindingDB

A multi-view strategy with converters is used to achieve dimensionality reduction for matrix dimensions, while an end-to-end model is used to reduce the complexity of the model with feature extraction steps.

Simplifies the characterization learning task for the acquisition of locally important residues.

Huang et al.³⁸

2022

human, C.elegans, BindingDB

A Transformer-based shared attention mechanism model for extracting desired features from raw protein sequences and drug molecule SMILEs, respectively.

Apply pre-trained encoders to protein coding to solve the problem of scarce labeling data. And introduce migration learning as a pre-training guide.

Zhao et al.⁴⁰

2022

DrugBank, Davis, KIBA, BindingDB

In the module used for local feature extraction of sequences, CNN and Transformer are combined to be able to capture and encode feature information over long distances, and then finally unite the global information with the local features as well as inter-class correlation information.

By dividing the drug, target and interaction information into three modules and extracting the key features separately, the local pattern information in the sequence can be better learned.

Wang et al.⁸⁶

2022

human, C.elegans, Davis

Combines graph convolutional neural networks and transformer methods with a multi-head attention mechanism to focus on atomic structure information.

The ability to learn conformationally stable atomic information demonstrates its interpretability at the atomic level.

Li et al.⁸⁷

2022

human, DUD-E, BindingDB

A bipartite interaction messaging module has been used to capture the bidirectional effects of drug-target interactions, and the weight display has been optimized for better interpretability.

The importance of focusing on both sides of the drug and the target.

Ranjan et al.⁸⁸

2022

MOSES

Combining a graph neural network approach with a knowledge graph approach and using an early fusion approach to representational screening of protein tertiary structure and order predicts the binding affinity scores of the generated molecules.

This combined approach effectively reduces the number of generated molecules and retains most of the theoretically well-bound molecules.

Liu et al.³⁷

2022

DUD-E, LIT-PCBA

An intermolecular map conversion method that incorporates a special attention mechanism used to find topological and spatial information between molecules, and uses a three-way transformer to model intermolecular information.

This approach excels at predicting binding positions and has been validated for drug screening of coronaviruses.

Ye et al.⁸⁹

2022

the gold standard dataset, DrugBank

A graph autoencoder and multisubspace deep neural network.

Enhancing its learning capabilities through automatic coding of graphs, subspace layers, and integration layers allows for more features to be extracted from network inputs and for better training of DNN networks.

Bae et al.⁹⁰

2022

Davis, BindingDB

A DTA prediction model that can take into account local substructures. It incorporates an attention mechanism between drugs and proteins that focuses on local information while focusing on global information.

Attention has been given to the interrelationships between protein and drug substructures, and it is thought that it is the interaction between certain specific substructures that is critical for the binding of these drugs and targets.

Joshy et al.⁹¹

2022

KIBA

Based on CNN, the inputs are the FASTA of the protein and the neural fingerprint of the drug, and a multilayer interaction network of the two is constructed to realize the prediction of DTA.

This approach effectively realizes the embedding of drug molecule structure by focusing on chemical properties such as drug molecule structure during drug data processing.

Mehdi et al.⁹²

2022

DUD-E, Human, BindingDB

A graph deep learning model that incorporates an attention mechanism tries to understand this internal semantic-level association and external intermolecular association of drugs and targets from an NLP perspective.

Interpretation of drug-target associativity relationships as sentence-level relationships. Stronger interpretability in the context of applying NLP methods.

Wu et al.⁹³

2022

Drugbank, KEGG

A knowledge graph attention network that decomposes the representation of relationships between nodes with a knowledge graph and adds an attention weighting scheme to filter out important features.

The method cleverly converts the DTIs classification prediction problem into a neighborhood node linkage prediction problem for knowledge graphs.

Li et al.⁵¹

2022

Hetero-A, Hetero-B

A heterogeneous graph attention network with an additional graph attention diffusion layer.

More indirect node information is captured in the intra-layer and inter-layer perspectives, thus expanding the perceptual field and realizing attentional perception enhancement. The method also found that the effect of expanding the perceptual field is different to different degrees in the intra-layer and inter-layer perspectives, and when the perceptual field is expanded to the same degree, the inter-layer perceptual field can be expanded to achieve better model prediction than the intra-layer perceptual field.

Wang et al.⁹⁴

2022

Davis, Biosnap, DrugBank

A parallel decoupling method combining CNN and Transformer.

Local and global features are extracted separately and coupled by a cross-attention mechanism, while the interaction information of drug-target pairs is convolved with a dynamically generated filter to achieve a better representation for local and global features.

Qu et al.⁴⁶

2022

DrugBank, HPRD

A new model for drug-target interaction prediction based on heterogeneous network graph embedding.

Learning nodes (drugs, proteins) and their topological neighborhood representations by extracting higher-order structural information using a GCN-inspired graph autoencoder to form heterogeneous networks.

Ma et al.⁹⁵

2022

PubChem, DrugBank, UniProt

Adding multiple self-attention mechanisms to both drug processing and drug-target co-processing phases.

Multiple attention mechanisms can help improve the model’s handling of specific requirements.

Yin et al.⁴³

2023

DrugBank, TwoSides dataset, DDInter

A generalized deep learning framework for graph-based prediction of drug interactions and drug target interactions.

Learning the structures and sequences of drugs and proteins using Residual Graph Convolutional Networks (RGCNs) and Convolutional Networks (CNNs) to improve the prediction accuracy of Drug-Drug Interaction (DDI) and DTI.

Zhang et al.⁵⁴

2023

DrugBank, UniProt, MalaCar

Two different graph autoencoders are used to form an enhanced deep autoencoder.

Feature inference and label propagation between known DTIs for spaces containing unknown drug targets.

Bian et al.⁹⁶

2023

Davis, KIBA, DrugBank

A multiple cross-attention mechanism based on weight sharing is employed, using only sequence information as input.

The 2D feature maps of proteins and drugs are compressed into 1D feature maps, which are then connected and fed into FCN for classification. It can effectively solve the class imbalance problem and overfitting problem in drug-target datasets.

Wen et al.⁹⁷

2023

Human, C.elegans, GPCR, Davis

Extraction of long-range interdependent features of sequences using a multi-attention mechanism to extract atom-amino acid interaction features.

Focus on the complex interactions between internal atoms and macromolecular compounds.

Kavipriya et al.⁹⁸

2023

KEGG BRITE, BRENDA, SuperTarget, DrugBank

An optimal recurrent neural network-based approach using a semi-supervised approach based on BiLSTM model of RNN to handle inter-drug and inter-target interactions and initiated as weights.

Since the performance of the BiLSTM technique is strongly influenced by the hyperparameters, the Adam optimizer is used to adjust the hyperparameters during the model training process to improve the prediction performance.

Zheng et al.³⁹

2023

DrugBank, BindingDB

A multi-task learning mechanism with two transformer encoder-decoder structures is introduced to enhance the extraction of target features.

The first to apply Mol2Vec, BERT, attentional mechanisms and multitasking mechanisms to a single model.

Zhu et al.⁹⁹

2023

Davis, KIBA

Sequence features and substructural features of individual molecules are extracted by the self-coder to express relationships, while information interaction paths are established between molecules and sequences, forming an associative learning mechanism.

The correlation between substructures is better expressed, which improves the accuracy of the predicted values for the DTA task.

Sofia et al.¹⁰⁰

2023

Davis, KIBA

Based on CNN, inputs are in SMILES and protein residue information.

There is no need to construct a similarity matrix, thus improving computational efficiency compared to some machine learning models.

Chen et al.⁵⁶

2023

DrugBank, HPRD, CTD, SIDER

A graph representation learning method that embeds nearest neighbor relationships between nodes to keep the topology of the relationships between nodes undeformed in the graphical learning process while maintaining the graph structure.

Effectively solves the problem of large number of isolated nodes in traditional graph convolution autoencoder models.