iScience ll

OPEN ACCESS

Review
Drug-drug interactions prediction based on deep
learning and knowledge graph: A review
Huimin Luo,1,2 Weijie Yin,1 Jianlin Wang,1,3 Ge Zhang,1,2 Wenjuan Liang,1 Junwei Luo,4 and Chaokun Yan1,3,*

SUMMARY
Drug-drug interactions (DDIs) can produce unpredictable pharmacological effects and lead to adverse
events that have the potential to cause irreversible damage to the organism. Traditional methods to
detect DDIs through biological or pharmacological analysis are time-consuming and expensive, therefore,
there is an urgent need to develop computational methods to effectively predict drug-drug interactions.
Currently, deep learning and knowledge graph techniques which can effectively extract features of en-
tities have been widely utilized to develop DDI prediction methods. In this research, we aim to systemat-
ically review DDI prediction researches applying deep learning and graph knowledge. The available
biomedical data and public databases related to drugs are firstly summarized in this review. Then, we
discuss the existing drug-drug interactions prediction methods which have utilized deep learning and
knowledge graph techniques and group them into three main classes: deep learning-based methods,
knowledge graph-based methods, and methods that combine deep learning with knowledge graph.
We comprehensively analyze the commonly used drug related data and various DDI prediction methods,
and compare these prediction methods on benchmark datasets. Finally, we briefly discuss the challenges
related to drug-drug interactions prediction, including asymmetric DDIs prediction and high-order DDI
prediction.

INTRODUCTION
The use of drug combinations is common and necessary to treat patients with complex diseases.1,2 However, when drugs are concomitantly
administered to a patient, the effects of the drugs may be enhanced or weakened, which may also cause side effects, these kinds of interac-
tions are called drug-drug interactions (DDIs).3 For example, the serum concentration of dofetilide decreases when it is taken with dabrafenib
together, whereas its serum concentration increases when taken with dalfopristin.4 Better knowledge of the incidence of DDIs and the drugs
most frequently involved, can be helpful in a more accurate assessment of their overall clinical importance.5 Although a large number of DDIs
are found during the clinical trials, there are still many DDIs on the market.6 Unknown DDIs can lead to unsafe treatments and even medication
errors in those patients who are receiving polypharmacy.7 According to the US Centers for Disease Control and Prevention, more than 10% of
people take five or more drugs at the same time, and even worse, 20% of older adults take at least 10 drugs,8 this phenomenon makes DDIs
more likely to occur. Moreover, DDIs were estimated to be responsible for 4.8% of hospitalization in the elderly, an 8.4-fold increase compared
to the general population.9 According to relevant statistics, DDIs cause a large number of deaths every year, and cause DDI-related costs of
177 billion US dollars.10 Therefore, it is very important to discover more potential DDIs.
Predicting potential DDI helps reduce unanticipated drug interactions and drug development costs and optimizes the drug design pro-
cess.11 However, traditional experimental methods for DDIs identification and prediction, such as testing cytochrome P450 or transporter-
related interaction, face challenges such as high cost and long duration.4 In some cases, researchers may suffer from limitations.12 Therefore,
machine learning13 and deep learning14 based computational methods are proposed to solve these problems in traditional DDIs identifica-
tion. The workflow of DDIs prediction using computational methods is shown in Figure 1. First, researchers need to collect available data from
publicly available biomedical data sources such as databases or relevant literatures, including DDIs, targets, genes, proteins, etc. These data
can provide valuable information for identifying potential DDIs. In the second step, advanced models utilizing machine learning and deep
learning techniques are developed to identify DDIs. Then, in order to evaluate and validate the predictive performance of these proposed
methods, it is necessary to compare them with state-of-the-art methods, perform prediction tasks on different gold standard datasets, and
analyze them. Finally, these predicted interactions are validated in vitro and in vivo. Most methods generally classified DDI prediction tasks
into binary classification task, multi-class task and multi-label task. The binary-class task is to predict whether there is an interaction between

1School of Computer and Information Engineering, Henan University, Kaifeng, China

2Henan Key Laboratory of Big Data Analysis and Processing, Henan University, Kaifeng, China
3Academy for Advanced Interdisciplinary Studies, Zhengzhou, China
4College of Computer Science and Technology, Henan Polytechnic University, Jiaozuo, China

*Correspondence: [email protected]
https://doi.org/10.1016/j.isci.2024.109148

iScience 27, 109148, March 15, 2024 ª 2024 The Authors. 1

This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
 ll iScience
OPEN ACCESS Review

Figure 1. The workflow of drug-drug interactions prediction

two drugs, multi-class task is to predict the type of DDI between drug pairs, and multi-label task is to predict the interaction types when there
are two or more interactions between drug pairs.
Machine learning is used to teach machines how to handle data more effectively and relies on various algorithms to address data-related
problems.15 Machine learning methods, including support vector machine (SVM), random forest, decision tree, naive Bayes, and K-nearest
neighbors, have been widely applied in various fields such as computer vision, prediction, semantic analysis, natural language processing,
and information retrieval.16 Traditional machine learning-based methods have made significant contributions to the field of DDIs prediction.
Cheng et al.17 calculated four types of similarities of drugs as the features of drug pairs, including drug phenotypic similarity, therapeutic sim-
ilarity, chemical structure similarity and genomic similarity, and finally applied naive Bayes, decision tree, K-nearest neighbors, logistic regres-
sion and support vector machine to predict DDIs. Li et al.18 have developed a probability ensemble approach (PEA) to build DDIs prediction
model using drug molecules and pharmacological characteristics. Mei et al.19 developed a logistic regression model with L2 regularization,
which used a simple f drug target profile representation to depict drugs and drug pairs to predict DDIs. Song et al.20 integrated the 2D mo-
lecular structure similarity, 3D pharmacophoric similarity, interaction profile fingerprint (IPF) similarity, target similarity and adverse drug effect
(ADE) similarity to obtain the feature representation of drugs, and utilized SVM to predict candidate DDIs.
Deep learning techniques have been widely developed and applied due to their excellent performance on large-scale and high-dimen-
sional datasets.21 Deep learning allows computational models that are composed of multiple processing layers to learn representations of
data with multiple levels of abstraction,22 which can effectively find complex structures in high-dimensional data, and is more flexible than
traditional methods based on Bayesian, random walk, support vector machine, and so on. Therefore, deep learning has been well applied
in image recognition,23 computer vision,24 NLP,25 and speech recognition.26 Furthermore, deep learning has also gained extensive usage
in the field of drug discovery including drug molecular activity prediction,27 molecular property prediction,28 target identification29 and
DTI (drug-target interaction) prediction.30
In addition, there has been increasing interest in extending deep learning methods to graph data,31 the use of graph has also brought new
breakthroughs in deep learning. Knowledge graph (KG) has been widely used in various business and scientific fields.32 Knowledge graph is a
multi-relationship graph containing multiple types of entities and edges, in which nodes correspond to entities, and edges correspond to
relations between the two connected entities.33 KG is mainly composed of triples, generally represented as G = ðE;R;T Þ, where E denotes
the set of entities, R denotes the relations, T denotes the set of edges, the edge ðs; r; oÞ denotes the existence of a relation r between entities
s to o, where r˛R, s is the head entity and o is the tail entity. The rich information provided by KG, containing structured and unstructured
knowledge, can be input into deep learning models to find hidden connections between entities.
The use of drug related biomedical data is critical for developing high-performance DDI prediction models. Although many public data-
bases have provided various drug-related information, there is still lacking of comprehensive summary of drug related entities and their in-
teractions in current popular data sources, and the required data cannot be accurately and quickly obtained when developing computational
prediction methods. In addition, many effective DDI prediction methods have been proposed, but there are still many issues that need to be
considered. Certain models are unable to effectively utilize DDIs data, and the imbalance between positive and negative samples also needs
to be resolved. Finally, the datasets and experiments used to analyze and evaluate the proposed prediction methods are generally not unified
in many studies.

2 iScience 27, 109148, March 15, 2024

iScience ll
Review OPEN ACCESS

The advances in DDI prediction have been reviewed in detail from different aspects in recent years.8,34–38 For example, Zhang et al.35 re-
viewed deep learning-based methods for extracting DDIs from biomedical literatures. Han et al.8 reviewed DDI prediction models based on
machine learning and organized into: traditional similarity, traditional classification, network diffusion, matrix factorization, ensemble-based
approach and based on literature approach. This review can provide useful guidance for interested researchers to further promote bioinfor-
matics algorithms to predict DDI. Compared with previous reviews, our study provides a more comprehensive and integrative analysis of
deep learning-based and knowledge graph-based prediction methods and related biomedical data used in DDI prediction. Firstly, we sum-
marize various sources of biomedical data used in DDI prediction methods. Secondly, the popular DDI prediction methods based on deep
learning and knowledge graph are analyzed. Then, these computational methods are compared on the same datasets briefly. Finally, we
discuss the limitations and challenges for developing DDI prediction methods.
In the recent review by Lin et al.,37 chemical structure based, network based, NLP based and hybrid methods for DDI prediction have been
summarized, and provided an updated and accessible guide to the broad researchers and development community with different domain
knowledge. In our review, we focus on analyzing and classifying existing computational methods for potential DDI prediction using deep
learning and knowledge graph, respectively. Although the methods covered in our survey have certain overlaps with Lin et al.,37 we comple-
ment it with computational methods based on KG, which have recently demonstrated attractive prediction precision enhancement in DDI
prediction. We further discuss the advantages and disadvantages of various prediction models. Furthermore, we have organized some
benchmark datasets from the literatures that commonly used in DDI prediction tasks. The common validation strategies and evaluation met-
rics used in DDI prediction studies are also included to guide researchers to efficiently evaluate and verify the predictive ability of their devel-
oped methods in future studies.

DRUG-RELATED DATA
The explosive growth of large-scale genomic and phenotypic data, as well as data of small molecular compounds with granted regulatory
approval,39 has enabled new developments for DDIs prediction. In addition to DDIs information, drug features are often utilized in DDI
prediction, including chemical substructures, targets, enzymes, pathways, genes, transporters, side effects, indications, etc. The effective
use of these drug features enables the model to learn comprehensively and improve performance. This study briefly summarizes the common
public available databases and datasets involved in current researches on DDI prediction.

Drug-related public databases

Predicting DDIs often requires the use of multiple characteristics of drugs as well as known DDIs. The most commonly used databases
include DrugBank,40 Drug Repurposing Knowledge Graph(DRKG),41 Kyoto Encyclopedia of Genes and Genomes(KEGG),42 Bio2RDF,43
TWOSIDES,44 SIDER,45 PubChem46 and DrugCentral.47 A detailed description of these public databases has been presented in Table 1.
These public databases can be classified based on the content and function into drug omics data, drug adverse effects and drug knowledge
graph databases.
Drug omics databases mainly include DrugBank,40 KEGG,42 PubChem46 and DrugCentral.47

(1) DrugBank40 is an open-access DB, which collects data from various sources, including journal articles, electronic databases, and text-
books.48 DrugBank (version 5.1.10) contains 16,565 drug entries including 2,761 approved small molecule drugs, 1,610 approved bi-
ologics, 135 nutraceuticals and over 6,723 experimental (discovery-phase) drugs. Moreover, 5,302 non-redundant protein (i.e., drug
target, enzyme, transporter and carrier) sequences are linked to these drug entries. The data have been validated by DrugBank cura-
torial staff, multiple automated data consistency checks have also been performed to ensure a uniformly high level of data integrity.49 It
can provide rich and high-quality data that enables significant advances in the bioinformatics field. Moreover, DrugBank team has
constantly optimized the interface of the database and enriched the search functions to make it more convenient for researchers to
access and use.
(2) KEGG42 is a daily updated, free database resource to help understand the high-level functions and utilities of the biological system,
which contains fifteen manually curated databases and a computationally generated database,42 and the PATHWAY database is the
most important component. KEGG has powerful graphical functions to provide a comprehensive understanding of the interaction net-
works of genes, proteins and compounds, based on graphical representation of biological objects and graphical computation tech-
nologies.50
(3) PubChem46 is key open chemical information resource at the US National Institutes of Health (NIH), which has collected information of
drugs from hundreds of data sources, including university labs, patent documents, government agencies, pharmaceutical companies,
chemical vendors, publishers and a number of chemical biology resources. PubChem contains small molecules, chemical structures,
identifiers, chemical and physical properties, and many others. In addition, it also offers rich query and analysis functions that allows
chemical compounds to be searched by name and structural formula.
(4) DrugCentral47 is a well-rounded drug information resource that integrates a wide range of drugs, chemical substructures and indica-
tions information. The majority of the data are collected and aggregated from online public resources, combined with manual curation
of literature and drug label information.51 DrugCentral is updated every 2–3 years, and its team regularly monitors new drug approval
from FDA, EMA, and PMDA to provide accurate and high-quality data for related researches.

DRKG41 and Bio2RDF43 are two comprehensive drug knowledge graph databases.

iScience 27, 109148, March 15, 2024 3

4

Table 1. The available databases

OPEN ACCESS
Available data
iScience 27, 109148, March 15, 2024

Category Database Entities Drug properties Drug-related interactions URL API

ll
40
Drug omics data DrugBank Drug, Target, Enzyme, Type, Chemical structure, Drug-drug interaction, https://go.drugbank.com/ https://docs.drugbank.
Transporter, Protein, Disease, Category, Approval status, Drug-food interaction, Drug- com/v1/
Gene, Carrier, Chemical identifiers, metabolite interaction,
Metabolite, Pathway, Indication, Function, Action Drug-protein interaction,
Compound, ATC, Adverse Drug-transcript
response interaction, Drug-target
interaction
KEGG42 Pathway, Gene, Genome, Structure, Drug class, Drug-drug interaction, Drug- https://www.kegg.jp/ https://www.kegg.jp/kegg/
Protein, Compound, Glycan, Chemical reaction, Chemical gene interaction, Drug- rest/keggapi.html
Enzyme, Variant, Disease, structure similarity disease interaction
Drug, ATC, Target,
Metabolism
PubChem46 Taxonomy, Compound, Structure, Indication Chemical-chemical https://pubchem.ncbi.nlm. https://pubchem.ncbi.nlm.
Protein, Gene, Pathway, Cell interaction, Chemical-gene nih.gov/ nih.gov/rest/pug
line, Substance, Side effect, interaction, Chemical-disease
Bioactivity, Target interaction
DrugCentral47 Drug, Target, Disease, Substructure, Adverse event, Drug-drug interaction, Drug- https://drugcentral.org/ https://drugcentral.org/
Protein Similarity, Active ingredient, disease interaction, Drug- OpenAPI
Indication, Drug mode of target interaction
action, Pharmacologic action,
Pharmacokinetic properties,
Bioactivity
Drug knowledge graph DRKG41 Anatomy, Biological process, SMILES Compound-Compound https://github.com/gnn4dr/ –
Cellular component, interaction, Compound-side DRKG
Compound, Disease, Gene, effect interaction,
Molecular function, Pathway, Compound-ATC
Pharmacologic class, Side classification interaction,
effect, Symptom, ATC, Tax Compound-pharmacologic
class interaction, Compound-
disease interaction,
Compound-gene interaction
Bio2RDF43 Drug, Gene, Protein, Chemical structure, Drug-adverse event http://bio2rdf.org/ https://github.com/bio2rdf/
Compound, Disease, Cell Pharmacological property interaction bio2rdf-api
Drug adverse effects TWOSIDES44 Drug, Side effect Indication Drug-side effect interaction https://www.tatonettilab. –

iScience
org/resources/tatonetti-

Review
stm.html
45
SIDER Drug, Side effect, ATC, Chemical structure, Drug-side effect interaction http://sideeffects.embl.de –
Target Indication
Note: ’-’ indicates that the database does not provide API.
iScience ll
Review OPEN ACCESS

(1) DRKG41 is a comprehensive biological knowledge graph, which contains genes, compounds, diseases and side effects, ATC, biolog-
ical process, etc. It integrates data from six existing biological databases and publications to ensure integrity and usefulness of data,
which includes 97,238 entities belonging to 13 entity-types, and 5,874,261 triplets belonging to 107 edge-types.
(2) Bio2RDF43 is an open knowledge graph that uses Semantic Web technologies to build and provide the largest network of Linked Data
for the Life Sciences, which contains 11 billion triples across 35 biomedical databases. Bio2RDF can create RDFs compatible Linked
Data from a diverse set of heterogeneously formatted sources to support complex biomedical researches.

Both TWOSIDES44 and SIDER45 are drug adverse effects databases.

(1) TWOSIDES44 is a multi-drug side effect resource that contains significant associations between drugs and adverse events, this data-
base contains 868,221 significant associations between 59,220 pairs of drugs and 1,301 adverse events. But these associations are
limited to only those that cannot be clearly attributed to either drug alone.44
(2) SIDER45 (version 4.1) contains 1430 drugs, 5,868 side effects (SEs) and 139,756 drug-SE pairs, the information about marketed drugs
and their recorded side effects is extracted from public documents and package inserts. The available information also includes fre-
quency of side effect and side effect classifications, etc., can provide high-quality and valuable data for related studies. Moreover, the
high quality of the extracted entities has also been ensured by manually annotating names, adding synonymous names and using an
additional Natural Language Processing step.45

Standard datasets collected from the literatures

With the development of DDIs prediction research, the existing studies have provided many benchmark datasets to facilitate the evaluation of
prediction methods. Tables 2 and 3 show some benchmark datasets collected from the existing literatures, which are applied on three
different tasks.
In this review, we evaluate the performance of the models using different datasets in Tables 2 and 3. For the binary classification prediction
task, six benchmark datasets including Dm_l4,80 Db_2,53 Db_3,54 Db_4,55 Db_556 and Db_657 are chosen for model performance evaluation.
The multi-class prediction task is performed on three benchmark datasets including Dm_c1,79 Dm_c252 and Dm_c3.81 For the multi-label pre-
diction task, three benchmark datasets including Dm_l1,52 Dm_l285 and Dm_l381 are selected for model evaluation.

DDIs PREDICTION METHODS

We group the existing DDI prediction methods into three classes: methods based on deep learning, methods based on knowledge graph,
and methods that combine deep learning with knowledge graph. Table 4 lists the methods corresponding to each class. The graphical sum-
mary of the overall methods is shown in Figure 2.

DL and KG techniques
Deep learning uses large amounts of unsupervised data to automatically extract complex representations.111 It has been demonstrated that
that deep learning is effective in discovering complex structures in high-dimensional data and is therefore applicable to many areas of sci-
ence, business, and government.22 In this section, we review the various deep learning and KG techniques.

DNN
DNN can learn more complex and abstract high-level features than shallow neural networks.112 Given an input sample fixed at the input layer,
the other units of the network compute their values based on the activity of the units that they are connected to in the next layer.113 Further-
more, whether it is a linear or nonlinear relationship, DNN has the ability to discover suitable parameters to convert inputs into corresponding
outputs. In the DDI prediction field, DNNs are widely used to develop prediction frameworks. Figure 3 provides a specific example of DDI
prediction method using DNN. DNN can be processed through two main phases: Forward Propagation (FP) and Backward Propagation (BP).

(1) FP: The input data are propagated from the input layer to the output layer. Each layer utilizes the outputs of the previous layer as inputs,
and the predicted outputs for given inputs are obtained by fully connected between layers.
(2) BP: Backward propagation is the process of calculating the gradients of the loss function with respect to the weights and biases of the
neural network. It allows the network to learn and adjust its parameters during the training process. BP also involves Weight Gradient
Calculation (WG) and Weight Update (WU). WG refers to the computation of the gradients of the loss function with respect to the
weights of the neural network and WU focuses on adjusting the weights of the neural network based on the calculated weight gradient.

GNN
Graphical representation is a useful tool to represent potential relationships among entities in the field of science and engineering, such as
computer vision, pattern recognition, data mining.114–118 Graph Neural Networks (GNNs) are deep learning-based methods that operate on
graphs domain, due to its convincing performance, GNNs have been widely used in recent years.119 Figure 4 reviews popular GNN models
include GCN, GAT. For DDI prediction, drugs are regarded as nodes in the graph, these nodes are connected to form a network, an edge

iScience 27, 109148, March 15, 2024 5

 ll iScience
OPEN ACCESS Review

Table 2. Benchmark datasets collected from existing literatures for binary classification prediction task

Dataset name Drugs DDIs Drug-related information Data resource

52
Db_1 – 1,178,210 SMILES DRKG
Db_253 548 48,584 Substructure: 881 TWOSIDES, SIDER, OFFSIDES, PubChem,
Target: 780 DrugBank, KEGG
Enzyme: 129
Transporter: 78
Pathway: 253
Indication: 4,897
Side effect: 4,897
Off side effect: 9,496
Db_354 1,537 34,282 SMILES DrugBank
Db_455 2,578 612,388 – DrugBank
56
Db_5 – 48,548 SMILES DrugBank,53
57
Db_6 1,752 504,468 Morgan fingerprint DrugBank
Db_758 1,562 180,576 Chemical structures, ATC, DBP (899 drug DrugBank,59,60
Db_858 1,934 230,887 targets and 222 non-target proteins)
61
Db_9 2,367 209,494 Target: 2,411 DrugBank, KEGG, PubChem
Enzyme: 285
Pathway: 314
Substructure: 699
Db_1055 1,925 56,983 – KEGG
Db_1162 613 80,702 Enzyme: 454 DrugBank, CTD,63 KEGG, SIDER,
Pathway: 533 LINCS64,65,66,67
Side effect: 4,859
Substructure: 811
Target: 2,670
Node2vec: 613
PRL: 978
Db_1268 841 82,620 Target: 1,333 DrugBank, KEGG, PubChem
Enzyme: 214
Pathway: 307
Substructure: 619
69
Db_13 1,850 443,046 SMILES Drugbank
Db_1469 1,322 83,040 BIOSNAP70
Db_1571 – 2,898,937 – DrugBank, KEGG, TWOSIDES, MEDLINE,72
OFFSIDES, PharmGKB53,73–76
77
Db_16 10,533 1,195,972 SMILES DrugBank, OGB-biokg78
77
Db_17 1,925 56,983 DrugBank, KEGG
Note: ‘Drugs’ represents the number of drugs in the dataset, ‘DDIs’ represents the numbers of drug-drug interactions, ‘Drug-related information’ represents
other drug-related features in the dataset, ‘Data resource’ represents the source of the data, ‘-’ represents no clear explanation in the original literature.

denotes an interaction between drugs. GNN can effectively capture the relationships among nodes and edges in DDI graphs, thereby
enhancing the accuracy of DDI prediction through comprehensive training and modeling.

GCN. Graph convolutional neural network (GCN) is one of the most commonly used methods in GNN. GCN can effectively leverage the
graph structure and aggregate node information from neighboring nodes in a convolutional manner. It has significant expressive power to
learn graph representations and achieve superior performance in various tasks and applications.120 Figure 4B) shows a simple example of
GCN. After constructing a DDI graph, the node representations are updated by aggregating the features of neighbor nodes, and the
commonly used aggregation methods include Sum and Average. Then, the obtained node representations are input into the activation
function for nonlinear transformation, which can help the model learn more complex feature representations. The new node representa-
tions finally obtained are utilized as the inputs of the next layer, the local and global information of nodes are gradually fused through
several iterations.

6 iScience 27, 109148, March 15, 2024

iScience ll
Review OPEN ACCESS

Table 3. Benchmark datasets collected from existing literatures for DDI events prediction tasks

Task Dataset name Drugs DDIs Types Drug-related information Data resource
79
Multi-class Dm_c1 572 74,528 65 Chemical substructure, Target, Pathway, DrugBank, KEGG
Enzyme
Dm_c252 – 172,426 81 SMILES Ryu et al.80
81
Dm_c3 1,697 190,728 86 Chemical structure Ryu et al.80
82
Dm_c4 1,704 191,400 86 SMILES DrugBank
Dm_c583 1,258 323,539 100 Substructure, Target, Enzyme DrugBank
84
Dm_c6 1,935 589,827 2 Chemical structure DrugBank
Dm_c781 1,013 114,204 71 Chemical structure Ryu et al.80, Lin et al.83
52
Multi-label Dm_l1 – 99,002 200 SMILES TWOSIDES
Dm_l285 10,533 1,195,972 39 – OGB-biokg78
Dm_l381 751 53,888 200 Chemical structure TWOSIDES,83
80
Dm_l4 1,861 192,284 86 SMILES DrugBank
Dm_l585 3,797 1,236,361 2 – DrugBank
Dm_l685 1,925 56,983 2 – KEGG
Dm_l786 1,918 30,979 100 – TWOSIDES
87
Dm_l8 1,597 188,258 106 SMILES, Gene: 22,032 DrugBank, BioGrid88–90
GO terms: 29,692
Dm_l981 1,314 103,938 200 Chemical structure TWOSIDES
Dm_l1091 1,317 198,697 86 Protein, Substructure DrugBank, PubChem
Fingerprint
Note: ‘Task’ represents prediction task, ‘Drugs’ represents the number of drugs in the dataset, ‘DDIs’ represents the numbers of drug-drug interactions, and
‘Types’ indicates the number of DDI types, ‘Drug-related information’ represents other types of data, ‘Data resource’ represents the source of the data, ‘-’ rep-
resents no clear explanation in the original literature. There are a total of 2,322 drugs in 52.

GAT. Graph Attention Network (GAT) introduces an attention mechanism into the process of computing weights between nodes. By stack-
ing layers in which nodes are able to attend over their neighborhoods’ features, GAT enables (implicitly) specifying different weights to
different nodes in a neighborhood, without requiring any kind of computationally intensive matrix operation (such as inversion) or depending
on knowing the graph structure upfront.121 Figure 4C) shows a single graph attentional layer, which inputs a set of node features and applies a
weight matrix to every node, then uses a shared self-attention for each node to compute attentional coefficients. The softmax function is uti-
lized to normalize the neighbor nodes and to compute linear combination of node features.

AE
AE122 is a nonlinear unsupervised neural network model, which consists of encoder and decoder. The encoder maps the input to the repre-
sentation space, while the decoder reconstructs the original input from the representation.123 Due to its advantages in data dimensionality
reduction and feature extraction,124 autoencoder has been applied to DDI prediction with favorable results. By setting appropriate dimen-
sions for the encoder, the most important features can be learned. Figure 5 demonstrates the reconstruction of the input data using an AE
model. There are two phases in the training process of the autoencoder: the encoder phase and the decoder phase.

(1) Encoder phase: The Encoder consists of a series of hidden layers that receive the input data and gradually compress the data. The pa-
rameters of encoder are learned by minimizing the reconstruction error, which ensures to capture the main features of the input data.
(2) Decoder phase: In this phase, the representations of hidden layers are mapped back to the original input space through the decoder.
The parameters of decoder are also learned by minimizing the reconstruction error to recover the original data as possible.

KG
Knowledge graph (KG) is a multi-relational semantic network, which exhibits robust expressive capabilities and modeling flexibility. Figure 6
illustrates an example of a node embedding and its updating process on knowledge graph, which obtains the node’s multi-hop neighbor-
hood features by propagating and aggregating information, and utilizes these features to update the node’s embedding vector.
However, when using KGs to calculate semantic relations between entities, it is often necessary to design special graph algorithms to
achieve this, but such graph algorithms have high computational complexity and poor scalability.125 Therefore, some knowledge graph
embedding methods have been proposed. KGE models ingest graph data in triplets form where they learn global graph low-rank latent fea-
tures, which preserve the graph’s coherent structure, the common KGE models include TransE, DistMult, HolE, etc.126

iScience 27, 109148, March 15, 2024 7

8

Table 4. Methods related to deep learning and knowledge graph

OPEN ACCESS
iScience 27, 109148, March 15, 2024

Technology Method Descriptions Advantages Limitations Validate Links

DeepDDI80

ll
DNN The model predicts potential DDIs It breaks the limitation of not being DNN used in the model needs to be N https://bitbucket.org/
using only the drug name and able to obtain detailed information upgraded based on the training with kaistsystemsbiology/deepddi
structure, and can also be utilized to about drugs. more data on drug pair interactions.
predict drug-food interactions.
DDIMDL79 Four sub-models are constructed by DDIMDL has taken advantage of Ignoring the problem of unbalanced N https://github.com/YifanDengWHU/
using each drug feature and a joint deep learning and diverse drug- datasets, and the fewer number of DDIMDL
DNN framework is used to combine related features to predict DDI some event interactions may lead to
the sub-models to learn cross- events. underfitting of the model.
modality representations of drug-
drug pairs.
DANN-DDI68 An attention neural network is DANN-DDI can combine multiple The imbalanced data and the noise Y https://github.com/naodandandan/
designed to learn the drug information to predict novel bring challenges, and for some DANN-DDI
representations of drug-drug pairs, drug-drug interactions and DDI- events without detailed descriptions
which considers the different associated events. and proved interactions, the model
contributions of different features need the interpretability.
and their dimensions.
BioChemDDI92 A computational method that Graph collapse is innovatively The more reasonable way of N http://120.77.11.78/BioChemDDI/
integrates multi-level information by introduced to extract network selecting negative samples needs to
applying the self-attention structure, and biochemical be considered to reduce the noise
mechanism to efficiently integrates information is utilized during the pre- from the imbalance in the original
biochemical and network features. training process. dataset.
GNN BI-GNN93 The model treats the data as a bi- The transductive setting of drug The introduction of extra features N –
level graph, which the highest level repurposing is considered and can may improve the performance of this
represents the interactions between also be extended to other biological model.
biological entities, and each entity link prediction tasks with different
itself is further expanded to its interaction biological entities.
intrinsic graph representation.
SSI-DDI82 The task of DDI prediction between The model can learn substructures When the order of drugs of DDIs are N https://github.com/kanz76/SSI-DDI
two drugs is decomposed into directly from drug molecular graphs, changed, the performance is
identifying pairwise interactions and can improve the ability of both affected even during the training
between their respective expert and non-expert users to phase; there is some noisy
substructures, and directly operating interpret the results of the prediction. information leaked in during the
on the raw molecular graph substructure extraction phase, which
representations of the drugs. might have affected performance in
inductive setting.

iScience
(Continued on next page)

Review
 Review
iScience
Table 4. Continued
Technology Method Descriptions Advantages Limitations Validate Links
DSN-DDI94 The model iteratively learns modules DSN-DDI shows the usefulness for The model’s ability to generalize to Y https://github.com/microsoft/Drug-
using local and global real-world DDI applications and can new drugs in the inductive learning Interaction-Research/tree/DSN-DDI-
representations, while learning drug also serve as a generalized setting needs to be further improved. for-DDI-Prediction
substructures from intra-view and framework in the drug discovery
inter-view. field.
Chen et al.95 Only the compressed structural It achieves higher performance on The graph convolution operator can N –
information extracted from both small-scale and large-scale only operate on flat 2D molecular
molecular graphs is utilized to datasets, and is more robust to the graphs, which may lose some vital
predict the DDI. extremely low pairwise similarity information conveyed by 3D
information. structure.
GCNMK62 Two DDI graph kernels are used for Benefiting from the two graph This model can’t identify DDIs Y –
the graph convolutional layers, kernels, GCNMK model can be used among isolated drugs.
namely, increased DDI graph to predict DDIs effectively.
consisting of ‘increased’-related
DDIs and a decreased DDI graph
consisting of ‘decreased’-related
DDIs.
RS-GCN86 It is a new relation-dependent RS-GCN specifically provides an There is still room for further N –
sampling model. The core of this advantage in scalability. It is also improvement in predicting on graph
approach is to assign a learnable shown that learning edge type with complex relationships and less
probability to each relation type and probabilities is indeed beneficial. dense.
update it.
MIRACLE96 A novel unsupervised contrastive It is superior on small-scale, medium- The performance of the model N https://github.com/isjakewong/
learning component is proposed to scale, and large-scale datasets. becomes gradually worse as the size MIRACLE
balance and integrate multi-view of the dataset gets larger and larger.
information. It can capture inter-view
molecule structure and intra-view
iScience 27, 109148, March 15, 2024

interactions between molecules

simultaneously.
MFFGNN97 A new feature extraction module is MFFGNN can effectively fuse the The model cannot be extended to N https://github.com/
proposed to capture the global topological information in molecular multi-type DDI prediction tasks. kaola111/mff
features for the molecular graph and graphs, the interaction information
the local features for each atom of between drugs and the local

OPEN ACCESS
the molecular graph. chemical context in SMILES
sequences.

ll
(Continued on next page)
9
10

Table 4. Continued

OPEN ACCESS
iScience 27, 109148, March 15, 2024

Technology Method Descriptions Advantages Limitations Validate Links

ll
AE Purkayastha The model incorporates different The model uses a GAE to effectively Lack of the exploration of other N –
et al.98 combinations of feature embeddings predict missing DDI links. representations of drugs, such as
from the drug-target interaction textual description of the drugs and
network and chemical structure. side effect-based interaction network
representation of the drugs.
87
Lee et al. Target similarity profiles (TSP), Gene GSP and TSP increase the prediction The DDIs predicted by this model N –
Ontology (GO) term similarity accuracy when using SSP alone, and and their clinical consequences are
profiles (GSP), as well as structural the proposed model identified a mostly unvalidated in DrugBank, and
similarity profiles (SSP) are number of novel DDIs that are the experimental results can be
constructed and combined. supported by medical databases or changed for different settings
existing researches. including different dataset version or
experimental environment.
DDI-MDAE61 A drug representation learning DDI-MDAE can predict potential Only considering the structural N –
method, which can learn unified drug interactions for drugs with topologies of drug feature networks
representations from multiple drug incomplete features even faced with is not enough to learn the more
feature networks simultaneously. large-scale, noisy and sparse data. comprehensive drug
representations.
Ensemble NMDADNN99 NMDADNN extracts the unified drug Five drug-related heterogeneous The quality of drug similarity matrix Y –
mapping features by integrating five information sources are effectively may be improved by utilizing more
drug-related heterogeneous integrated. drug-related sources and suitable
information sources. similarity measures, this is one of its
limitations.
58
DPDDI GCN is utilized to extract the network This is an effective and robust Interactions with the new drugs N https://github.com/NWPU-903PR/
structure features of drugs from DDI method proposed to predict cannot be predicted. DPDDI
network, and DNN as a predictor. potential DDIs by utilizing the DDI
network information without
considering the drug properties.
R 2 -DDI100 The framework encodes drugs and It significantly improves the DDI This model ignores the atom level of Y https://github.com/
relationship embeddings, and builds prediction performance over the pair interaction between drugs, linjc16/R2-DDI
the relation-aware refined features. multiple real-world datasets and the modeling of relation and the
settings, and shows better relation-aware module is relatively
generalization ability. simple, the imbalanced data issue is
also not solved.
83
MDF-SA-DDI The model combines two drugs in Multi-source drug fusion is used to There is room for further N https://github.com/ShenggengLin/

iScience
four different ways and can predict obtain better prediction of DDIs. improvement in the prediction of MDF-SA-DDI

Review
unobserved interactions between interactions between new drugs.
new drugs.

(Continued on next page)

 Review
iScience
Table 4. Continued
Technology Method Descriptions Advantages Limitations Validate Links
AMDE54 AMDE encodes drug features from AMDE integrates drug features from The model can be further optimized Y https://github.com/wan-Ying-Z/
multiple dimensions, including multiple dimensions to enhance the to incorporate more drug features. AMDE-master
information from both Simplified effectiveness of downstream
Molecular-Input Line-Entry System prediction tasks.
sequence and atomic graph of the
drug.
SGRL-DDI84 SGRL-DDI model captures the task- It utilizes the Balance theory and Only enhancive/depressive DDIs and Y https://github.com/NWPU-903PR/
joint information by integrating Status theory to reveal directed DDIs can be predicted, and SGRL-DDI
relation graph convolutional pharmacological interaction patterns there is limitation in predicting
networks with Balance and Status in the directed DDI network. specific events.
patterns and predicts directed DDIs.
Interpretable CASTER69 A sequential pattern mining module The model improves generalizability The model can be further improved N https://github.com/
is developed by using labeled and and interpretable prediction. by extending it to chemical sub- kexinhuang12345/CASTER
unlabeled chemical structure data. graph embedding and incorporating
metric learning.
STNN-DDI101 Mapping drugs into SSI space based The interpretability of the model is The introduction of more features of N https://github.com/zsy-9/STNN-DDI
on a list of predefined substructures improved, and it can also predict drugs may be helpful in improving
with specific chemical meanings, interactions between new drugs. the performance.
which allows STNN-DDI to perform
multiple types of DDI predictions in
both transductive and inductive
scenarios in a unified form with an
explicable manner.
DeSIDE-DDI102 The model uses drug-induced gene It can increase DDI prediction The datasets are very sparse in terms Y https://github.com/GIST-CSBL/
expression signatures followed by accuracy and provide model of side effect type, and the DeSIDE-DDI
gating and translating embedding. interpretability. identification of the side effect
mechanism remains challenging.
DSIL-DDI103 Treat the substructure interactions as DSIL-DDI improves the Effective predictions between new N –
iScience 27, 109148, March 15, 2024

domain-invariant representations of generalization and interpretability of drugs need to be considered.

DDIs. Moreover, a pluggable DDI prediction models.
substructure interaction module and
a practical loss function are
proposed.
GGI-DDI104 A method that employs granular It enhances the interpretability of DDI The rich knowledge contained in the N https://codeocean.com/

OPEN ACCESS
computing to identify key predictions, and offers a consistent different biomedical entities
substructures, drugs are granulated framework for DDI prediction in both associated with drugs (e.g., proteins,

ll
into a set of coarser granules. transductive and inductive scenarios. genes, and targets) is ignored.

(Continued on next page)

11
12

Table 4. Continued
Technology Method Descriptions Advantages Limitations Validate Links

OPEN ACCESS
iScience 27, 109148, March 15, 2024

KG Celebi et al.11 A simple disjoint cross-validation Different KGE methods for DDIs Because the embedding predictors N https://github.com/rcelebi/

ll
scheme is proposed to evaluate DDI prediction are compared. are constructed using the black-box GraphEmbedding4DDI/
predictions in the absence of known model, they are unable to provide
DDIs for drugs. the mechanistic explanations for
predicted potential DDIs.
DDI-BLKG105 A method for discovering potential It uses semantic relations connecting A qualitative analysis of the resulting N https://github.com/kbogas/
DDIs through the generation of the different drugs in the literatures, as KG should be conducted, to gain DDIBLKG
KG from disease-specific literatures. features of DDI. insights on errors generated by the
automatic extraction tools.
BERTKG-DDI106 BERTKG-DDI combines drug This method is in line with the new It is also essential to explore other N –
embeddings obtained from DDIs direction of research of fusing various external drug representation such as
and other biomedical entities with an information to DDI prediction. chemical structure, textual
RC architecture based on domain- description in predicting DDI from
specific BioBERT embeddings. textual corpus.
71
DL and KG Conv-LSTM Using KG integrates 12,000 drug It can integrate multiple sources of There is no interpretation for the N –
features from DrugBank, drugs and related data for predicted DDIs.
PharmaGKB, and KEGG drugs, comprehensive information.
embedding the nodes in the graph
using various embedding
approaches.
107
MDNN A two-pathway framework including MDNN learns the representations The dataset imbalance problem is N –
DKG-based pathway and from multimodal data and mines the ignored.
heterogeneous feature (HF)-based inter-modality similarities from
pathway, which is designed to obtain multiple sources, also exploits the
the multimodal representations of topological structure information
drugs. and semantic relations with DKG.
KGNN55 KGNN is an end-to-end framework KGNN aggregates all topological The introduction of other drug N https://github.com/xzenglab/KGNN
that explores drugs’ topological neighborhood information received features might be helpful to improve
structures in knowledge graph. locally to extract both high-order the prediction performance.
structures and semantic relations.
RANEDDI91 This is a relation-aware network By considering the multi-relational It cannot predict DDIs effectively N https://github.com/DongWenMin/
embedding model, which can information and relation-aware when training sets are very sparse or RANEDDI
embed multi-relational graph. network structure information the drug has no neighbors, and
together, RANEDDI can learn the performs poorly on some low-
more representative entity frequency DDI types.
embeddings.

iScience
Review
(Continued on next page)
 Review
iScience
Table 4. Continued
Technology Method Descriptions Advantages Limitations Validate Links
AAE-FOR-KG108 A new knowledge graph embedding AAE can generate high-quality Compared to small dataset, it is not N https://github.com/dyf0631/
framework is proposed by negative samples, the Gumbel- suitable for large dataset which AAE_FOR_KG
introducing adversarial autoencoder Softmax relaxation and the contains complex relationship types.
(AAE) based on Wasserstein Wasserstein distance help train the
distances and Gumbel-Softmax embedding model steadily.
relaxation.
KG2ECapsule85 This model integrates a probability- High-quality negative samples are Multi-hop is not considered, so the N https://github.com/Blair1213/
based negative sampling strategy to generated, which refrains from the receptive field of each entity cannot KG2ECapsule
generate high-quality negative danger of introducing false negative. be globally determine for accurate
samples, and also utilizes a capsule learning. This is one of its limitations.
network for non-linear
transformation to enrich the
representations of entities under
specified relational space.
MUFFIN52 A bi-level cross strategy is proposed MUFFIN can jointly learn the drug The model lacks interpretability and N https://github.com/xzenglab/
that includes cross- and scalar-level representations based on both the the problem of data redundancy is MUFFIN
components to fuse multi-modal drug-self structure information and also to be resolved.
features well. the KG with rich bio-medical
information.
DDKG77 DDKG fully utilizes the information of Drug attributes are integrated into It is difficult for DDKG to obtain the N https://github.com/Blair1213/DDKG
biomedical KGs, and can learn the the representation learning process global optimal solution.
initial embeddings of drug nodes to improve performance of DDI
from their attributes in the KG, while prediction.
considering both neighboring node
embeddings and triple facts.
BioDKG-DDI109 BioDKG-DDI integrates multi-feature It is a robust, yet simple method and Transportability can be further N –
with biochemical information, and can be used as a benefic supplement improved by changing the way of
predicts potential DDIs through an to the experimental process. extracting the feature of drug
attention machine with superior functional similarity.
iScience 27, 109148, March 15, 2024

performance.
DeepLGF110 DeepLGF is an inductive model that This model is proposed to fully Random selection of negative N https://github.com/MrPhil/
predicts DDIs through aggregating exploit biomedical knowledge graph samples can bring certain noise. DeepLGF
local-global multi-information based (BKG) fusing local-global Moreover, drug sequences as CS
on the BKG. information, which improves the information are too simple and may
performance of DDIs prediction. only provide limited one-
dimensional information.

OPEN ACCESS
(Continued on next page)

ll
13
14

OPEN ACCESS
iScience 27, 109148, March 15, 2024

ll
Table 4. Continued
Technology Method Descriptions Advantages Limitations Validate Links
3WDDI56 A method based on three-way Delay decision is made for objects in Not considering diverse drug N –
decision, which combines the boundary region by integrating features and multi-omics data may
knowledge graph embedding as KG embedding feature, and introduce some limitations to the
supplementary features to enhance improves the accuracy of decision- model.
DDI prediction. making.
DGAT-DDI57 A directed graph attention network DGAT-DDI is the first approach for It cannot accommodate multitype Y https://github.com/F-windyy/
to predict asymmetric DDIs, which predicting asymmetric interactions asymmetric interactions, which is one DGATDDI
learns embeddings of the source among drugs. of its limitations.
roles, the target roles and the self-
roles of drugs.
81
MCFF-MTDDI This model integrates the extra label By using multi-channel feature Extremely unbalanced data may lead N https://github.com/ChendiHan111/
information into KG-based multi- fusion, biomedical KG-based to bad prediction outcomes. MCFF-MTDDI
typed DDI prediction, and features, extra label information and
innovatively proposes a novel KG drug chemical structures are fused
feature learning method and a State more effectively.
Encoder.
Note: ‘Validate’ represents whether the method has been externally validated on independent dataset or in real clinical settings, ‘Y’ represents yes, and ‘N’ represents no, ‘-’ represents no clear explanation in
the original literature.

iScience
Review
iScience ll
Review OPEN ACCESS

Figure 2. The computational methods for predicting DDIs are classified into three main groups: deep learning-based models, KG-based models, models
which combine deep learning and KG
In particular, DL-based methods are categorized based on their underlying deep learning models.

DL-based methods
DNN-based methods
In the real world, it is generally not possible to obtain all the detailed drug information. In order to ease these limitations, Ryu et al.80 proposed
DeepDDI which only used drug names and structure information as inputs. DeepDDI generated a structure similarity profile (SSP) for each
drug, obtained the structurel features of the drug, and then merged two SSPs of a drug pair, input them to DNN for prediction. DeepDDI
could provide important information about drug prescriptions and even dietary recommendations when taking certain drugs, as well as
guidelines during drug development. Although these studies have made significant contributions to DDI prediction, considering more fea-
tures is necessary for comprehensive study and better performances.
To effectively integrate different features of drugs, Deng et al.79 proposed a multi-modal deep learning framework named
DDIMDL. Four drug features including chemical substructure, target, enzyme and pathway, were effectively utilized and each type
of these features was fed into a sub-model with a multilayer neural network, and then were combined using a joint DNN framework
to learn cross-modal representations of drug pairs for DDI events prediction. Liu et al.68 proposed DANN-DDI that utilized SDNE for
learning drug nodes embeddings from five drug-related feature networks, connected the learned drug embeddings and neural
network to represent drug pairs, and then used DNN to predict DDI. However, it is difficult for these methods to preserve
higher-order structure, and they might tend to obtain local optimal solutions. Moreover, integrating drug features without the atten-
tion machine may lead to limited prediction performance. Attention mechanism can be used as a resource allocation scheme, which
is the main means to solve the problem of information overload.127 Ren et al.92 used attention mechanism to integrate drug features,
and proposed BioChemDDI computational method to construct highly representative integrated feature descriptors, which were
input to the DNN for DDI prediction. The experimental results showed the effectiveness of the attention mechanism and also utilized
the deep network structure information.

GNN-based methods
To overcome the current limitations of using single drug compound structures or using only DDI data, Bai et al.93 proposed BI-GNN which
constructed a bi-level graph based on DDIs, in which the highest level contained the interaction graph of DDIs and the representation graphs
of entities. The lower level representation graph neural network generated vector representations for each representation graph. The higher
level interaction graph neural network further propagated information from the lower level graph embeddings to neighboring nodes in the
interaction graph, then provided the final graph representations to a fully connected network to obtain a final link prediction score. This
design of features captured the whole molecular structure of drugs. In fact, DDI usually only depend on a few substructures. Nyamabo
et al.82 have proposed a DDI prediction method named SSI-DDI based on the assumption that DDI is actually caused by chemical substruc-
ture-substructure interactions. SSI-DDI used the GAT layer with shared weights to extract the raw molecular graph representations of drugs
and extract substructure information. The task of DDI prediction was decomposed into substructure-substructure interactions prediction. Li
et al.94 proposed DSN-DDI based on GNN, which iteratively learned drug substructures from the single drug(intra-view) and the drug

iScience 27, 109148, March 15, 2024 15

 ll iScience
OPEN ACCESS Review

Figure 3. The process of using DNN for DDI prediction

The feature vectors of drug a and b are first combined and then fed into the DNN to predict DDI.

pair(inter-view). Each GNN had its own unique receptive field, each one aggregated information from neighboring nodes, which resulted in
updating the nodes and extracting the substructures.
Chen et al.95 utilized GCN to convert the molecular data with irregular structure into the corresponding molecular data in low dimensional
vector spaces. Then, the learned embedding vectors were used to predict whether there has DDI between two drugs or DDI type. Wang
et al.62 proposed a multi-kernel graph convolutional network (GCNMK). GCNMK used two graph cores as the graph convolution layers, in
which an increased DDI graph and a decreased DDI graph were constructed from the ‘increase’-related and ‘decrease’-related DDIs. The
two graphs and the drug feature matrices were fed into the first GCN layer composed of two GCN blocks, then the low-dimensional repre-
sentation vectors of drugs were generated using the second layer composed of one GCN block. Finally, the two drug feature vectors were
concatenated to form a DDI vector, which were then inputted into a block with three fully connected layers to predict potential DDIs. Feeney
et al.86 proposed RS-GCN to model the importance of neighborhood sampling relationship types in the network, in which each relationship

Figure 4. The widely used GNN frameworks

(A) GNN; (B) GCN; (C) GAT.

16 iScience 27, 109148, March 15, 2024

iScience ll
Review OPEN ACCESS

Figure 5. Reconstructing SMILES of a drug using AE

AE involves two processes: encoder and decoder. The encoder compresses the input high-dimensional data into low-dimensional data, while the decoder
returns the low-dimensional data into high-dimensional data, and the goal is sample reconstruction.

type was assigned a learning probability and updated through a reinforcing-based method. The results showed that this model can learn the
right balance: relation-type probabilities that reflect both frequency and importance, and also offered some kind of explanation. However,
these graph-based studies only focus on a single view of the drug, the combination of multi-view information inside the DDI network is often
ignored.
GCN is also a promising way to handle multiple views, Wang et al.96 proposed MIRACLE to simultaneously capture the relationships be-
tween molecular structures across views and the molecular-molecular interactions within views. MIRACLE regarded the DDI network as a
multi-view graph in which each node represented an instance of a drug molecule graph. In the learning stage, GCN was utilized to encode
DDIs, while the key aware attention information propagation method was used to capture the drug molecular structure information. Although
the models mentioned above have taken into account the structure, sequence or interaction information of the drugs, while they neglect the
synergistic effects between them. He et al.97 proposed MFFGNN to efficiently integrate topological information in molecular graphs, DDIs
information and local chemical contexts in SMILES sequences. In MFFGNN, the bi-directional gate recurrent unit extracted local chemical
context information from SMILES sequences, while the graph interactions network with graph wrap unit extracted the topological structure
features of the drugs from the given molecular graphs. Finally, GCN fused the intra-drug features and external DDI features to update the
drug representations.

AE-based methods
Purkayastha et al.98 proposed an effective method to fuse multiple drug features to predict DDIs. The proposed model leveraged the drug-
target interaction (DTI) network to learn drug embeddings. In order to obtain drug representations from rich chemical structures, a variational
autoencoder was also constructed. Finally, the obtained different combinations of feature embeddings of the drugs were incorporated and
input into a Graph Auto Encoder to predict missing DDI links in the network. Lee et al.87 proposed a DDI prediction method based on con-
structed Target Similarity Profile (TSP), Gene Ontology (GO), Term Similarity Profile (GSP) as well as the SSP. An autoencoder was trained to
minimize the difference between inputs and outputs, while training it to minimize the prediction error of DDI labels. However, the relation-
ships between these biomedical events are usually non-linear across all types of drug features. Additionally, some datasets may lack labels or
contain noise, which can potentially have adverse effects on prediction models.

Figure 6. Updating embedding of node D1 in KG

After initializing the embedding of D1, the first-order neighborhood information is recursively propagated and aggregated, the same operation to aggregate the
second-order neighborhood information, and then to update the embedding of node D1.

iScience 27, 109148, March 15, 2024 17

 ll iScience
OPEN ACCESS Review

Figure 7. An example of a model that utilizes both GCN and DNN

In order to predict DDIs from large-scale, noisy and sparse data, Zhang et al.61 proposed a multi-modal deep AE-based method named
DDI-MDAE to learn drug representations. This computational method first utilized an AE to learn uniform drug representations from multiple
drug feature networks simultaneously. Next, four operators were applied to the learned drug embeddings to represent drug-drug pairs.
Finally, a random forest model was trained using the representations to predict DDIs.

Multi-neural network ensemble-based methods

The above proposed prediction methods have achieved excellent performance when using one type of neural network alone. To utilize the
advantages of various deep learning models, researchers have begun to combine different neural networks to predict potential DDIs, thus
developing many high-performance and high-robust computing methods. Figure 7 shows an example of using both GCN and DNN, the GCN
extracts the features from the DDI network and inputs them into the DNN network for prediction.
Yan et al.99 proposed NMDADNN, which constructed similarity networks by integrating five heterogeneous sources of drug-related in-
formation, after unifying these similarities with AE, a DNN was constructed to predict DDI. Feng et al.58 proposed DPDDI to obtain
network structure features of drugs. In DPDDI, GCN was utilized to capture the topological relationships among drugs in the DDI network
and extract their network structural features. DNN acted as a predictor to associate the potential feature vectors with the feature vectors of
the corresponding drug pairs to predict DDIs. Lin et al.100 proposed a novel DDI prediction framework(R 2 -DDI). After encoding drugs
through DeeperGCN,128 the feature refinement module was constructed to obtain mutually aware drugs. Furthermore, the relation em-
beddings could be updated by incorporating the drug information. In this way, the DDI prediction can be improved with more informative
and distinguishable features. Finally, these refined features for drugs and relation were utilized to train and predict the possibility of DDI.
However, the feature fusion methods utilized in these models are simple, and more effective fusion methods can be designed to improve
the performance.
Latent feature fusion via Siamese network is a useful method, Lin et al.83 proposed MDF-SA-DDI computational method, which based on
multi-source drug fusion and multi-source feature fusion with self-attentive mechanism in order to predict interaction events between two
drugs. After combining two drugs in four different ways, the combined drug feature representations were fed into four different drug fusion
networks (Siamese network, convolutional neural network, and two autoencoders) to obtain the potential feature vectors of the drug pairs,
then fused with potential features using a transformation block with self-attentive mechanism. The multi-source drug fusion can provide
diverse information from different views to deep learning models, and accurately predict DDI events. In addition to this method of encoding
drug features, there have also been some models that directly encode SMILES of drugs. Pang et al.54 proposed attention-based multidimen-
sional feature encoder method (AMDE) that effectively utilized this information to enhance the accuracy of DDI prediction. AMDE utilized the
MPAN model as the graphic encoder to process the atomic graphs of drugs and generate 2D atomic graph feature vectors. Meanwhile, a
sequence encoder was used to process the sequence data generated by the SMILES to generate 1D sequence feature vectors. Finally, all
feature vectors were fed into multidimensional decoder to predict whether there is an interaction between two drugs.
Although these methods have achieved inspiring results, they neglected the pharmacological changes that DDI could induce enhance-
ment and inhibition, as well as the different pharmacological roles of two drugs in an interaction. Therefore, Feng et al.84 proposed a new
graph representation learning model SGRL-DDI. This model leveraged Balance theory and Status theory from social networks to characterize
pharmacological patterns of DDI, and organized DDI entries into a signed and directed network that reflects the relational semantics between
drugs. Two-layer embedding and an extra enhancer based on social theory were utilized to represent drugs in DDIs network, which each
embedding layer was constructed by a multi-relation GNN and a two-layer MLP. Finally, the concatenation vectors of two drug latent features
are fed into two dense DNNs to achieve two tasks of predicting enhancive/depressive DDIs and predicting the directed DDIs.

18 iScience 27, 109148, March 15, 2024

iScience ll
Review OPEN ACCESS

Deep learning models with interpretability

Although DL models have shown good performance for DDIs prediction, they usually need a large number of parameters, which are
difficult to interpret.129 DL prediction models with good interpretability can not only help researchers understand the trigger mecha-
nism of DDIs, but also ensure safer medication, so it is especially crucial to improve the interpretability of deep learning models. In
recent years, more and more models have begun to focus on this issue. In this section, we review models that consider the problem
of non-interpretability of deep learning.
Huang et al.69 proposed a computational method named CASTER which utilized a dictionary learning module to help researchers to un-
derstand how this model makes prediction and identify which sub-structures can possibly lead to the interaction. After generating frequent
sub-structures for the input drug SMILES sequences, CASTER generated a functional representation for each frequent sub-structure, then
exploited the encoder to generate a matrix of latent feature vectors, and projected the resulting potential vector into the subspace defined
by the matrix. The basis vector in the matrix was associated with a frequent molecular sub-graph, and its corresponding projection coefficient
revealed the statistical importance of having the sub-graph in the molecular graphs of the drug pairs so that they would interact, thus explain-
ing the rationality behind CASTER’s predictions. Yu et al.101 proposed a novel substructure-aware tensor neural network model STNN-DDI,
which mapped drugs into an SSI space based on a list of predefined substructures with specific chemical meanings, thereby improving the
interpretability of the model. STNN-DDI characterized the SSI space by using the learned substructure 3 substructure 3 interaction tensor
(ST), which was a 3-D tensor and expanded by a series of rank-one tensors. It represented the occurring probability of drug-drug pair as the
sum of the occurring probability of the substructures included in this drug pair, and the types of SSI were defined as the same as the corre-
sponding types of DDI. Potential DDIs were obtained by learning the probability of each SSI under a list of predefined chemical substructures.
This form improved interpretability of the model, because both known drugs and new drugs are mapped into a common SSI space no matter
whether a drug has an interaction or not. Kim et al.102 proposed DeSIDE-DDI, which can provide the interpretable expression level for DDI
prediction with drug-induced gene expression signatures. Specifically speaking, the model engineered dynamic drug features using a gating
mechanism to mimic drug co-administration effects by imposing attention to important genes. The concept of translating embeddings was
also introduced, which considered a side effect as a relationship between two drugs and applied the margin-based loss function, implying
positive pairs of drug combinations are positioned closely on the given side effect space. Via using drug-induced gene expression signatures
followed by gating and translating embedding can provide model interpretability and increase the accuracy of the model.
Tang et al.103 proposed DSIL-DDI, a pluggable substructure interaction module, which not only considered the fine-grained properties of
substructures, but also introduced the attention mechanism to recognize which substructure interaction representations contributed more to
the DDI representation. The substructure representations of the input drugs extracted by GNN were regarded as the property spectrum of
the substructure, which represented the extent to which chemical substructure responds to different properties. In the process of using Ha-
damard product to model the fine-grained attribute interactions, attention weights were introduced to help observe which substructure inter-
action representations contributes more to the DDI representation. Furthermore, experiments were also performed to demonstrate the inter-
pretability of DSIL-DDI. Yu et al.104 proposed GGI-DDI, which utilized granular computing to extract key subgraphs between drug pairs and
enhance the interpretability of DDI prediction models, the key chemical substructures were defined as coarser granules. In detail, after aggre-
gating the node information in the atomic graphs of drugs using GINE, the attention mechanism and bond information were fused to obtain
the embedding of nodes and subgraphs for determining the importance of nodes. Then, a local and global score function were introduced to
evaluate the local significance and long-range dependencies the embedding of each subgraph, respectively, to help identify the key sub-
graphs related to drugs and were reconstructed to create new graphs. Coarser particles (key chemical substructures) were identified via mul-
tiple granulations on graphs of drug pairs, and a cross-attention mechanism was employed to compute the attention score for each coarser
granule, thereby obtaining their final representation. This interpretability provided by GGI-DDI can guide the advancement of novel drug
development and poly-drug therapy strategies.

KG-based methods
In recent years, biological data and knowledge bases have been increasingly built on Semantic Web technologies.130 Many bioinformatics
databases have started to use Semantic Web-related technologies and use these data as Linked Open Data (LOD).131 By using LOD data,
Celebi et al.11 proposed a model to evaluate DDI in the absence of a known DDI. Different knowledge graph embedding methods were
used to extract drug feature vectors from KG, such as RDF2Vec, TransE, and TransD. Finally, some common classifiers were utilized to predict
DDIs. Bougiatiotis et al.105 proposed DDI-BLKG, predicted DDIs by generating disease-specific KG from biomedical publications and manu-
ally curated databases. The human-curated drug database was utilized to train a classifier that identifies patterns of interactions between drug
pairs. The predictive potential and usefulness of the method were demonstrated through a small-scale qualitative evaluation. Based on rep-
resentation learning, Mondal et al.106 proposed a DDI prediction method named BERTKG-DDI, which has used a knowledge graph that con-
tains target-target, drug-drug, drug-disease, disease-target, disease-target interactions, and enhanced entity representations to train the
relational classification model.

Combination of DL with KG
Recent studies have demonstrated the effectiveness of deep learning models in extracting node features from KG and achieved remarkable
results. When using deep learning for DDI prediction, it is common to make predictions based on the features of drug nodes. On the other
hand, KG focuses on prediction based on features of drug nodes, neighbor nodes and relations. As a result, mixing these approaches can

iScience 27, 109148, March 15, 2024 19

 ll iScience
OPEN ACCESS Review

compliment each other, and offer richer information. Additionally, the labels or feature information provided by KG can also enhance the
performance of deep learning models. Thus, the integration of DL and KG offers opportunities for improvement in DDI prediction tasks.
Karim et al.71 have proposed a novel DDI prediction method, Conv-LSTM, which based on a KG including drug-related information infor-
mation from KEGG, DrugBank, TWOSIDES and the literatures. The method utilized RDF2Vec, SimpleIE, TransE, KGloVe, CrossE and PBG for
the KG embeddings. Then input these embeddings into CNN-LSTM layer, where the CNN extracted local relationship values in the drug
features, and the LSTM captured the overall relationship from the features extracted by the CNN. Nevertheless, Conv-LSTM has limitations
on obtaining the rich features of drugs from structural information and semantic relations. Furthermore, it has not considered the drug multi-
modal data coherence and complementarity together. To effectively assist the joint representation learning of multi-modal data related to
DDI events, Lyu et al.107 proposed a multi-modal deep neural network (MDNN) for DDI events prediction. A dual-path framework containing a
drug knowledge graph (DKG)-based pathway and a heterogeneous feature (HF)-based pathway was proposed to extract structural informa-
tion and semantic relations from the DKG to learn drug representations. Moreover, a multi-modal fusion neural layer was utilized to explore
the complementarity among multi-modal representations of drugs.
But one neglected deficiency is that model DDI as an independent data sample and do not consider their multiple related correlations in
knowledge graph. Moreover, directly learning the latent embedding of nodes in KG could also bring some limitations. Therefore, Lin et al.55
proposed KGNN to obtain rich neighborhood information for each entity in KG. The central idea of KGNN was to consider both high-order
structures and semantic relations, by using GNN to encode the drugs and topological neighborhood information to distributed representa-
tions, which facilitated the prediction of DDI events. Nevertheless, Yu et al.91 found that both the multi-relational information and network
structure information can affect the learning of entity embeddings, thereby proposed a relation-aware network embedding model
(RANEDDI) to predict potential DDIs. After embedding entities and relations into the vector space using RotatE, a relation-aware information
propagation mechanism was utilized to extract the relation-aware network structure information of the entities by propagating their neigh-
bor’s information under different relations. Finally, DNN acted as the predictor to predict the probability of the certain interaction between
drug pairs. Experiments have also demonstrated its robust performance even in the case of a scarcity of labeled DDIs.
Several deep learning approaches have been proposed to embed Drug Knowledge Graphs (DKGs) for predicting unknown DDIs. Training
a KG embedding models requires negative samples, but most embedding models have been generating negative triplets via a uniform nega-
tive sampling strategy, and the obtained samples are too simple to train the model effectively. Thus Dai et al.108 proposed AAE-FOR-KG, a
knowledge graph embedding framework. AAE-FOR-KG utilized adversarial autoencoder to generate high-quality negative samples. Based
on positive and negative triples, discriminator could learn drugs and interaction embeddings effectively. Compared to other traditional
knowledge graph embedding methods, it has better performance. Su et al.85 also proposed KG2ECapsule, which generated high-quality
negative samples based on a probability negative sampling strategy. KG2ECapsule constructed a Graph-to-Embedding Layer, which could
recursively propagate embeddings from the neighbors of the entity as well as the relations between them. In addition, a two-layer capsule
network has been innovatively integrated to obtain entity representations in a non-linear form under specified relational space. Finally, these
entity representations were utilized to predict the DDI. Experiments also demonstrated the effectiveness of probability-based sampling strat-
egy and non-linear transformations.
It has been proved that the performance of DDI prediction can be improved by considering KG with rich bio-medical information and drug
molecular structure information or SMILES sequence simultaneously. Chen et al.52 proposed a multi-scale feature fusion deep learning model
named MUFFIN. MUFFIN utilized MPNN to extract molecular structure features from SMILES, and TransE to extract semantic features from
KG. Then, crossed these features to learn local and global features using CNN and flattening operations. Meanwhile, the fine-grained inter-
action features between two different features were obtained using the element product method, the obtained features were stitched
together for prediction. Su et al.77 proposed DDKG, which learned the initial embeddings of drugs from the corresponding attributes of
the nodes in KG using an encoder-decoder layer. To learn accurate global representations of drug nodes, the model recursively propagated
and aggregated the first-order neighborhood information along the top-ranked network paths which determined by the embeddings of
neighbor nodes and triples. Finally, estimated the interaction probability for pairwise drugs with their respective representations.
Considering some unavailable information can lead to inaccuracy of drug features extraction, Ren et al.109 proposed BioDKG-DDI, which
used Mol2Context-vec to extract molecular features, combined ComplEx-DURA with DKG to obtain drug global features, and incorporated
drug functional similarity features as supplementary information. Such approach enabled the combination of drug molecular features, drug
global association information and drug functional similarity features, and then the features were integrated and input to DNN for prediction.
Moreover, Ren et al.110 proposed DeepLGF, which utilized the BFGNN model to construct a heterogeneous network of drugs to obtain the
bio-local information, learned the global feature information of BKG by the KGE method of ComplEx, finally obtained a stable and effective
model to predict potential DDI after integration. Nevertheless, this model relies on obtaining detailed and complete information about the
drug at once. In fact, only the SMILES sequences of the drugs are easy to obtain, while others information needed further experiments to
discover. The uncertainty and incompleteness brought about by drug information might hinder the accuracy of DDI prediction. Three-way
decision-making can solve this problem. Thus Hao et al.56 proposed 3WDDI based on the computational method of three-way decision mak-
ing.132 After embedding KG using ComplEx, CNN was utilized as the decision function to classify the drug pairs into positive region, negative
region and boundary region based on the chemical structure features of the drugs using SCNN. By combining the knowledge graph embed-
ding features, the robustness of the decision could be improved by delay decision for objects in the boundary region.
Many biological experiments have proved pharmacological asymmetry between DDIs, but most of the above models did not consider this
situation. Feng et al.57 designed DGAT-DDI model for asymmetric DDIs prediction by learning the embeddings of source-role, target-role,

20 iScience 27, 109148, March 15, 2024

iScience ll
Review OPEN ACCESS

Table 5. Confusion matrix

Predict
Real Positive Negative
Positive TP FN
Negative FP TN

and self-role of drugs. The method viewed the knowledge graph as directed graph and used GAT to aggregate information of entities and
neighboring nodes to generate source and target embeddings, while using MLP to obtain self-role embeddings. After aligning these em-
beddings, asymmetric DDIs were predicted. However, extra label information has not yet been applied to the task of multi-label DDI predic-
tion. To overcome the problem, Han et al.81 proposed a multi-channel feature fusion model MCFF-MTDDI, which utilized drug chemical struc-
ture features, drug pairs’ extra label features, and KG features of drugs. For the target drug pair, after extracting the structural feature vector,
the three types of KG representations and the extra label vector, MCFF-MTDDI utilized a state encoder to fuse these three KG representations
vectors and obtain two KG fusion representations. In addition, a multi-channel feature fusion framework based on GRU was established, which
comprehensively integrated the information of two KG feature channels, structural feature channels, and additional label feature channels.
Finally, multi-class prediction and multi-label DDIs prediction were performed respectively.

EXPERIMENTS AND COMPARISON

Evaluation metrics
In general, the performance of model is usually evaluated through experiments, therefore, the learning ability of the model needs to be tested
by using a test set. Cross-validation (CV) is commonly utilized to train models, and the basic idea is to split the original data into the train set
and the test set. The train set for training the model and the test set for evaluating the performance of the model. The most popular method is
the K-CV, which the original data is divided into K subsets, one of them is used as the test set while K-1 subsets as the train set. The process is
then repeated K times, with each subset being used as the test set once. The final performance score of the model is obtained by taking the
average of the evaluation scores obtained during these iterations. Normally, the values of K are set to 5 and 10.
Evaluating the performance of a model requires not only feasible experimental methods of evaluation, but also evaluation metrics to mea-
sure the generalization ability of the model. In the classification task, the samples can be grouped into true positive (TP), false positive (FP),
true negative (TN) and false negative (FN) based on the true and prediction labels, which are referred to as ‘confusion matrix’, as shown in
Table 5. Accuracy, precision, recall and F1 are commonly used evaluation metrics in classification problems. Accuracy means the number
of correctly classified samples as a proportion of the total number of samples, and precision means the proportion of samples classified
as TP among all positive samples. The calculation formulas are as follows.

TP+TN
Accuracy = (Equation 1)
TP+FP+FN+TN

TP
Precision = (Equation 2)
TP+FP

TP
Recall = (Equation 3)
TP+FN

2 3 precision 3 recall
F1 = (Equation 4)
precision+recall

Receiver operating characteristic (ROC), by setting out several different continuous values of the continuous variables, thereby calculating
a series of true positive rate and false positive case rate, and plotted as curves for the vertical and horizontal axes respectively, the larger the
area under the curve, the higher the accuracy rate is indicated, and the area under the curve is also called AUC. Precision-recall curve (PR), with
Recall on the horizontal axis and Precision on the vertical axis, shows model performance at different classification thresholds, AUPR is the area
under PR curve. The values of AUC and AUPR range from 0 to 1, and larger values indicate better model performance.

Experiment results
We compare the models in binary classification prediction, multi-class prediction, and multi-label prediction tasks, respectively. The exper-
iment results are shown in Figures 8, 9, and 10.
Figure 8 shows the comparison results of different models under binary classification task on six benchmark datasets. The performance of
DDIs prediction achieved by these models were all measured in terms of ACC, AUC, AUPR and F1. For Dm_l4, we can observe that GNN-
based method SSI-DDI achieves the highest AUC of 96.14. MFFGNN (AUPR = 96.81, F1 = 92.54), which also belongs to GNN-based model,

iScience 27, 109148, March 15, 2024 21

 ll iScience
OPEN ACCESS Review

Figure 8. Performance evaluation under binary classification task

achieves the highest AUPR and F1, respectively. This may be attributed to the fact that the features from drug sequences and molecular
graphs are integrated and comprehensively learned, while some DNN-based models (e.g., DeepDDI and DDIMDL) only consider the struc-
tural information and features of drugs. Interestingly, for Db_3, AMDE, which based on the integration of multi-neural networks shows similar
AUC and ACC as KGNN that belongs to the models of combine DL and KG. KGNN and 3WDDI, which belong to the combined DL and KG
models perform well and obtain the best AUPR and AUC on three different datasets. KGNN obtains AUC of 99.12 and AUPR of 98.92 on Db_4,
also obtains AUC of 99.1 and AUPR of 98.9 on Db_6. For Db_5, the AUC and AUPR result of 3WDDI is 95.82 and 96.14, respectively.
For multi-class prediction task, we can observe from Figure 9 that MUFFIN (F1 = 94.95, ACC = 96.48, Precision = 95.68, Recall = 94.82) and
MCFF-MTDDI (AUPR = 97.57, AUC = 99.81, ACC = 97.74), the models combining DL and KG, achieve the best performance on Dm_c2 and
Dm_c3, respectively. The achievement of high performance means that the combination of KG-based features and drug features or fusion
with additional information is feasible. For Dm_c1, the AUPR and AUC of MDF-SA-DDI that based on multi-neural network ensemble achieve
the best performance, 97.37 and 99.89, respectively. The success of MDF-SA-DDI may be attributed to the utilization of a novel approach to
fuse the information of drug pairs, i.e., four different drug fusion networks are utilized to obtain latent feature vectors of the drug pairs.
Furthermore, we discover that the DNN-based model DeepDDI has the same AUC as the AE-based model Lee et al.87 on Dm_c1, but the
latter has the better performance in general.
Actually, multi-label prediction is more difficult than binary and multi-class prediction. From Figure 10, we can observe that MUFFIN,
KG2ECapsule and MCFF-MTDDI, which belong to the combination of DL and KG methods achieve the best performance on three data-
sets, respectively. For Dm_l2, the AUC and AUPR of KG2ECapsule respectively increased by 23.03% and 31.16% compared to KGNN,
which is a surprising improvement. The success of KG2ECapsule may be attributed to modeling the triplets and integrating the relations
of edges into embedding. On Dm_l3, the AUC of MCFF-MTDDI is 93.15, and AUPR is 74.36, which demonstrate the importance of using
extra label information of drug pairs. It is worth mentioning that DL and KG methods are more often used in multi-label prediction task
with better results.

22 iScience 27, 109148, March 15, 2024

iScience ll
Review OPEN ACCESS

Figure 9. Performance evaluation under multi-class task

CHALLENGES AND OPPORTUNITIES

Although excellent results have been achieved using deep learning and knowledge graph for DDI prediction, there are still some issues that
need to be resolved, which are summarized as follows.

(1) Asymmetric drug-drug interactions prediction. In the field of DDI prediction, relevant computational methods are well developed,
and it is difficult to further improve. In most computational methods related to DDIs prediction, such as the model using DKG as input,
the relationships in DKG are undirected, and the drugs in the triples are symmetric, but that doesn’t fit the real world. When Wicha
et al.133 evaluated their model using a dataset of 200 combination experiments in Saccharomyces cerevisiae, they found that 67%
of the interactions were monodirectional. In addition, asymmetric DDIs prediction can also help patients determine the order of

Figure 10. Performance evaluation under multi-label task

iScience 27, 109148, March 15, 2024 23

 ll iScience
OPEN ACCESS Review

medication, which greatly affects the effectiveness of medication. Bible et al.134 demonstrated that for paclitaxel, cytarabine, topote-
can, doxorubicin and etoposide, synergy was more pronounced when the agents were administered before flavopiridol rather than
concomitant with or following flavopiridol. Therefore, in future researches, we can turn our attention to asymmetric DDIs prediction.
(2) High-order drug-drug interactions prediction. The vast majority of current methods have been developed to predict DDI, and few
methods predict interactions among multiple drugs. Ning et al.135 proposed a purely data-driven fashion for representing, discov-
ering, quantifying, and visualizing high-order DDIs. Peng et al.136 proposed a D3 I model for predicting high-order DDI based on
deep learning techniques. There is still much room for development in predicting high-order DDIs, which should receive more atten-
tion in future research.
(3) Datasets. While utilizing a large amount of complex data can provide the model with rich drug features, it can also introduce a sig-
nificant level of noise. In addition, many datasets only contain known drug pairs with interactions, lacking validated drug pairs without
known interactions. These issues create excessive deviation between experimental results and actual values, which significantly re-
duces the accuracy of the experiment. Therefore, constructing high-quality and available negative samples can further improve the
accuracy of the DDI prediction methods.

DISCUSSION
It’s important to discuss the clinical relevance and practical applications of DDI prediction methods. How to translate these predictions into
actionable insights for healthcare providers and patients is a very important task. Therefore, we can do the following.

(1) The computational model is applied to various drug-drug pairs to predict whether and what kind of interactions exist between these
drug-drug pairs, to generate sentences describing relevant interactions, which suggest specific pharmacological effects (e.g., ‘‘the
decreased therapeutic efficacy’’ and ‘‘the increased anticoagulant activities’’) in addition to ‘‘the increased risk or severity of bleeding’’.
(2) To validate the accuracy of the model’s predictions, outputs need to be compared with the consistent descriptions on the DDIs present
in the Drugs.com (https://www.drugs.com/), which provides additional information regarding DDIs.
(3) Except the drug pairs examined above by comparing with the information presented in the Drugs.com, there may be drug pairs for
which ADEs possible causal mechanisms are not available elsewhere. Therefore, the output sentences describing additional DDI types
for the drug pairs with the reported ADEs can serve as the likely causal mechanisms of DDIs for further validation.

ACKNOWLEDGMENTS
This work was supported by the National Natural Science Foundation of China [grant number 61972134, 61802113, 61802114]; and the Sci-
ence and Technology Development Plan Project of Henan Province [grant number 222102210238, 212102210091].

DECLARATION OF INTERESTS
No competing interest is declared.

REFERENCES
1. Lu, D.-Y., Lu, T.R., Yarla, N.S., Wu, H.Y., Xu, MATEC Web of Conferences, 173, J. Heled prediction in realistic settings. BMC Bioinf.
B., Ding, J., and Zhu, H. (2017). Drug and A. Yuan, eds. (EDP Sciences). 20, 726.
combination in clinical cancer treatments. 7. Karbownik, A., Szałek, E., Soba nska, K., 12. Safdari, R., Ferdousi, R., Aziziheris, K.,
Rev. Recent Clin. Trials 12, 202–211. Grabowski, T., Wolc, A., and Grzeskowiak, Niakan-Kalhori, S.R., and Omidi, Y. (2016).
2. Fisusi, F.A., and Akala, E.O. (2019). Drug E. (2017). Pharmacokinetic drug-drug Computerized techniques pave the way for
combinations in breast cancer therapy. interaction between erlotinib and drug-drug interaction prediction and
Pharm. Nanotechnol. 7, 3–23. paracetamol: a potential risk for clinical interpretation. Bioimpacts 6, 71–78.
3. Asada, M., Miwa, M., and Sasaki, Y. (2018). practice. Eur. J. Pharm. Sci. 102, 55–62. 13. Jordan, M.I., and Mitchell, T.M. (2015).
Enhancing drug-drug interaction extraction 8. Han, K., Cao, P., Wang, Y., Xie, F., Ma, J., Yu, Machine learning: Trends, perspectives, and
from texts by molecular structure M., Wang, J., Xu, Y., Zhang, Y., and Wan, J. prospects. Science 349, 255–260.
information. Preprint at arXiv. https://doi. (2021). A review of approaches for 14. Goodfellow, I., Bengio, Y., and Courville, A.
org/10.48550/arXiv.1805.05593. predicting drug–drug interactions based on (2016). Deep Learning (MIT press).
machine learning. Front. Pharmacol. 12, 15. Mahesh, B. (2020). Machine learning
4. Yu, H., Mao, K.T., Shi, J.Y., Huang, H., Chen, algorithms-a review. Int. J. Sci. Res. 9,
Z., Dong, K., and Yiu, S.M. (2018). Predicting 814858.
381–386.
and understanding comprehensive drug- 9. Gu, Q. (2010). Prescription Drug Use 16. Shinde, P.P., and Shah, S. (2018). A Review of
drug interactions via semi-nonnegative Continues to Increase: US Prescription Drug Machine Learning and Deep Learning
matrix factorization. BMC Syst. Biol. 12, Data for 2007-2008 (US Department of Applications. In 2018 Fourth international
101–110. Health and Human Services, Centers for conference on computing communication
5. Becker, M.L., Kallewaard, M., Caspers, Disease Control and Prevention). control and automation (ICCUBEA) (IEEE).
P.W.J., Visser, L.E., Leufkens, H.G.M., and 10. Giacomini, K.M., Krauss, R.M., Roden, D.M., 17. Cheng, F., and Zhao, Z. (2014). Machine
Stricker, B.H.C. (2007). Hospitalisations and Eichelbaum, M., Hayden, M.R., and learning-based prediction of drug–drug
emergency department visits due to drug– Nakamura, Y. (2007). When good drugs go interactions by integrating drug phenotypic,
drug interactions: a literature review. bad. Nature 446, 975–977. therapeutic, chemical, and genomic
Pharmacoepidemiol. Drug Saf. 16, 641–651. 11. Celebi, R., Uyar, H., Yasar, E., Gumus, O., properties. J. Am. Med. Inform. Assoc. 21,
6. Yan, Z., Zhao, L., Wei, X., and Zhang, Q. Dikenelli, O., and Dumontier, M. (2019). e278–e286.
(2018). Improved label propagation model Evaluation of knowledge graph embedding 18. Li, P., Huang, C., Fu, Y., Wang, J., Wu, Z., Ru,
to predict drug-drug interactions. In approaches for drug-drug interaction J., Zheng, C., Guo, Z., Chen, X., Zhou, W.,

24 iScience 27, 109148, March 15, 2024

iScience ll
Review OPEN ACCESS

et al. (2015). Large-scale exploration and 36. Zeng, X., Tu, X., Liu, Y., Fu, X., and Su, Y. Novartis Foundation Symposium 247 (Wiley
analysis of drug combinations. (2022). Toward better drug discovery with Online Library).
Bioinformatics 31, 2007–2016. knowledge graph. Curr. Opin. Struct. Biol. 51. Ursu, O., Holmes, J., Knockel, J., Bologa,
19. Mei, S., and Zhang, K. (2021). A machine 72, 114–126. C.G., Yang, J.J., Mathias, S.L., Nelson, S.J.,
learning framework for predicting drug– 37. Lin, X., Dai, L., Zhou, Y., Yu, Z.G., Zhang, W., and Oprea, T.I. (2016). DrugCentral: Online
drug interactions. Sci. Rep. 11, 17619. Shi, J.Y., Cao, D.S., Zeng, L., Chen, H., Song, Drug Compendium. Nucleic acids research
20. Song, D., Chen, Y., Min, Q., Sun, Q., Ye, K., B., et al. (2023). Comprehensive evaluation 45, gkw993.
Zhou, C., Yuan, S., Sun, Z., and Liao, J. of deep and graph learning on drug–drug 52. Chen, Y., Ma, T., Yang, X., Wang, J., Song,
(2019). Similarity-based machine learning interactions prediction. Brief. Bioinform. 24, B., and Zeng, X. (2021). MUFFIN: multi-scale
support vector machine predictor of drug- bbad235. feature fusion for drug–drug interaction
drug interactions with improved accuracies. 38. Vo, T.H., Nguyen, N.T.K., Kha, Q.H., and Le, prediction. Bioinformatics 37, 2651–2658.
J. Clin. Pharm. Ther. 44, 268–275. N.Q.K. (2022). On the road to explainable AI 53. Zhang, W., Chen, Y., Liu, F., Luo, F., Tian, G.,
21. Ye, F. (2017). Particle swarm optimization- in drug-drug interactions prediction: A and Li, X. (2017). Predicting potential drug-
based automatic parameter selection for systematic review. Comput. Struct. drug interactions by integrating chemical,
deep neural networks and its applications in Biotechnol. J. 20, 2112–2123. biological, phenotypic and network data.
large-scale and high-dimensional data. 39. Li, J., Zheng, S., Chen, B., Butte, A.J., BMC Bioinf. 18, 18.
PLoS One 12, e0188746. Swamidass, S.J., and Lu, Z. (2016). A survey 54. Pang, S., Zhang, Y., Song, T., Zhang, X.,
22. LeCun, Y., Bengio, Y., and Hinton, G. (2015). of current trends in computational drug Wang, X., and Rodriguez-Patón, A. (2022).
Deep learning. Nature 521, 436–444. repositioning. Brief. Bioinform. 17, 2–12. AMDE: a novel attention-mechanism-based
23. Wu, M., and Chen, L. (2015). Image 40. Wishart, D.S., Feunang, Y.D., Guo, A.C., Lo, multidimensional feature encoder for drug–
recognition based on deep learning. In 2015 E.J., Marcu, A., Grant, J.R., Sajed, T., drug interaction prediction. Brief.
Chinese automation congress (CAC) (IEEE). Johnson, D., Li, C., Sayeeda, Z., et al. (2018). Bioinform. 23, bbab545.
24. Esteva, A., Chou, K., Yeung, S., Naik, N., DrugBank 5.0: a major update to the 55. Lin, X., Quan, Z., Wang, Z.-J., Ma, T., and
Madani, A., Mottaghi, A., Liu, Y., Topol, E., DrugBank database for 2018. Nucleic Acids Zeng, X. (2020). KGNN: Knowledge Graph
Dean, J., and Socher, R. (2021). Deep Res. 46, D1074–D1082. Neural Network for Drug-Drug Interaction
learning-enabled medical computer vision. 41. Ioannidis, V.N., Song, X., Manchanda, S., Li, Prediction. In IJCAI.
NPJ Digit. Med. 4, 5. M., Pan, X., Zheng, D., Ning, X., Zeng, X., 56. Hao, X., Chen, Q., Pan, H., Qiu, J., Zhang, Y.,
25. Lopez, M.M., and Kalita, J. (2017). Deep and Karypis, G. Drkg-drug repurposing Yu, Q., Han, Z., and Du, X. (2023). Enhancing
Learning applied to NLP. Preprint at arXiv. knowledge graph for covid-19. https:// drug–drug interaction prediction by three-
https://doi.org/10.48550/arXiv.1703.0309. github.com/gnn4dr/DRKG.2020. Accessed way decision and knowledge graph
26. Abdel-Hamid, O., Mohamed, A.r., Jiang, H., 01 May 2023. embedding. Granul. Comput. 8, 67–76.
Deng, L., Penn, G., and Yu, D. (2014). 42. Kanehisa, M., Furumichi, M., Tanabe, M., 57. Feng, Y.-Y., Yu, H., Feng, Y.H., and Shi, J.Y.
Convolutional neural networks for speech Sato, Y., and Morishima, K. (2017). KEGG: (2022). Directed graph attention networks
recognition. IEEE/ACM Trans. Audio new perspectives on genomes, pathways, for predicting asymmetric drug–drug
Speech Lang. Process. 22, 1533–1545. diseases and drugs. Nucleic Acids Res. 45, interactions. Brief. Bioinform. 23, bbac151.
27. Kato, Y., Hamada, S., and Goto, H. (2016). D353–D361. 58. Feng, Y.-H., Zhang, S.-W., and Shi, J.-Y.
Molecular activity prediction using deep (2020). DPDDI: a deep predictor for drug-
43. Callahan, A., Cruz-Toledo, J., Ansell, P., and
learning software library. In 2016 drug interactions. BMC Bioinf. 21, 419.
Dumontier, M. (2013). Bio2RDF release 2:
international conference on advanced 59. Liu, Z., Guo, F., Gu, J., Wang, Y., Li, Y.,
improved coverage, interoperability and
informatics: concepts, theory and Wang, D., Lu, L., Li, D., and He, F. (2015).
provenance of life science linked data. In
application (ICAICTA) (IEEE). Similarity-based prediction for anatomical
The Semantic Web: Semantics and Big
28. Feinberg, E.N., Sur, D., Wu, Z., Husic, B.E., therapeutic chemical classification of drugs
Data: 10th International Conference, ESWC
Mai, H., Li, Y., Sun, S., Yang, J., Ramsundar, by integrating multiple data sources.
2013, Montpellier, France, May 26-30, 2013.
B., and Pande, V.S. (2018). PotentialNet for Bioinformatics 31, 1788–1795.
Proceedings 10 (Springer).
molecular property prediction. ACS Cent. 60. Shi, J.-Y., Mao, K.T., Yu, H., and Yiu, S.M.
Sci. 4, 1520–1530. 44. Tatonetti, N.P., Ye, P.P., Daneshjou, R., and
(2019). Detecting drug communities and
Altman, R.B. (2012). Data-driven prediction
29. Zeng, X., Zhu, S., Lu, W., Liu, Z., Huang, J., predicting comprehensive drug–drug
of drug effects and interactions. Sci. Transl.
Zhou, Y., Fang, J., Huang, Y., Guo, H., Li, L., interactions via balance regularized semi-
Med. 4, 125ra31.
et al. (2020). Target identification among nonnegative matrix factorization.
known drugs by deep learning from 45. Kuhn, M., Letunic, I., Jensen, L.J., and Bork, J. Cheminform. 11, 28.
heterogeneous networks. Chem. Sci. 11, P. (2016). The SIDER database of drugs and 61. Zhang, Y., Qiu, Y., Cui, Y., Liu, S., and Zhang,
1775–1797. side effects. Nucleic Acids Res. 44, D1075– W. (2020). Predicting drug-drug interactions
30. Torng, W., and Altman, R.B. (2019). Graph D1079. using multi-modal deep auto-encoders
convolutional neural networks for predicting 46. Kim, S., Chen, J., Cheng, T., Gindulyte, A., based network embedding and positive-
drug-target interactions. J. Chem. Inf. He, J., He, S., Li, Q., Shoemaker, B.A., unlabeled learning. Methods 179, 37–46.
Model. 59, 4131–4149. Thiessen, P.A., Yu, B., et al. (2021). PubChem 62. Wang, F., Lei, X., Liao, B., and Wu, F.X.
31. Liu, Z., and Han, X. (2018). Deep learning in in 2021: new data content and improved (2022). Predicting drug–drug interactions by
knowledge graph. Deep Learning in Natural web interfaces. Nucleic Acids Res. 49, graph convolutional network with multi-
Language Processing, Deng L. and Liu D1388–D1395. kernel. Brief. Bioinform. 23, bbab511.
Y.(eds), Springer, Singapore, pp. 117–145. 47. Avram, S., Bologa, C.G., Holmes, J., Bocci, 63. Davis, A.P., Grondin, C.J., Johnson, R.J.,
32. Zou, X. (2020). A survey on application of G., Wilson, T.B., Nguyen, D.T., Curpan, R., Sciaky, D., Wiegers, J., Wiegers, T.C., and
knowledge graph. In Journal of Physics: Halip, L., Bora, A., Yang, J.J., et al. (2021). Mattingly, C.J. (2021). Comparative
Conference Series (IOP Publishing). DrugCentral 2021 supports drug discovery toxicogenomics database (CTD): update
33. Park, N., Kan, A., Dong, X.L., Zhao, T., and and repositioning. Nucleic Acids Res. 49, 2021. Nucleic Acids Res. 49, D1138–D1143.
Faloutsos, C. (2019). Estimating node D1160–D1169. 64. Iorio, F., Bosotti, R., Scacheri, E., Belcastro,
importance in knowledge graphs using 48. Wishart, D.S., Knox, C., Guo, A.C., V., Mithbaokar, P., Ferriero, R., Murino, L.,
graph neural networks. In Proceedings of Shrivastava, S., Hassanali, M., Stothard, P., Tagliaferri, R., Brunetti-Pierri, N., Isacchi, A.,
the 25th ACM SIGKDD international Chang, Z., and Woolsey, J. (2006). and di Bernardo, D. (2010). Discovery of
conference on knowledge discovery & data DrugBank: a comprehensive resource for drug mode of action and drug repositioning
mining. in silico drug discovery and exploration. from transcriptional responses. Proc. Natl.
34. Qiu, Y., Zhang, Y., Deng, Y., Liu, S., and Nucleic Acids Res. 34, D668–D672. Acad. Sci. USA 107, 14621–14626.
Zhang, W. (2022). A comprehensive review 49. Wishart, D.S., Knox, C., Guo, A.C., Cheng, 65. Subramanian, A., Narayan, R., Corsello,
of computational methods for drug-drug D., Shrivastava, S., Tzur, D., Gautam, B., and S.M., Peck, D.D., Natoli, T.E., Lu, X., Gould,
interaction detection. IEEE/ACM Trans. Hassanali, M. (2008). DrugBank: a J., Davis, J.F., Tubelli, A.A., Asiedu, J.K.,
Comput. Biol. Bioinform. 19, 1968–1985. knowledgebase for drugs, drug actions and et al. (2017). A next generation connectivity
35. Zhang, T., Leng, J., and Liu, Y. (2020). Deep drug targets. Nucleic Acids Res. 36, map: L1000 platform and the first 1,000,000
learning for drug–drug interaction D901–D906. profiles. Cell 171, 1437–1452.e17.
extraction from the literature: a review. Brief. 50. Kanehisa, M. (2002). The KEGG database. In 66. Grover, A., and Leskovec, J. (2016).
Bioinform. 21, 1609–1627. ‘In silico’simulation of biological processes: node2vec: Scalable feature learning for

iScience 27, 109148, March 15, 2024 25

 ll iScience
OPEN ACCESS Review

networks. Proceedings of the 22nd ACM drug interaction prediction. Brief. representation learning for drug-drug
SIGKDD international conference on Bioinform. 24, bbad215. interaction prediction. Proceedings of the
Knowledge discovery and data 82. Nyamabo, A.K., Yu, H., and Shi, J.-Y. (2021). Web Conference 2021; ACM.
mining; ACM. SSI–DDI: substructure–substructure 97. He, C., Liu, Y., Li, H., Zhang, H., Mao, Y., Qin,
67. Huang, K., Xiao, C., Glass, L.M., Zitnik, M., interactions for drug–drug interaction X., Liu, L., and Zhang, X. (2022). Multi-type
and Sun, J. (2020). SkipGNN: predicting prediction. Brief. Bioinform. 22, bbab133. feature fusion based on graph neural
molecular interactions with skip-graph 83. Lin, S., Wang, Y., Zhang, L., Chu, Y., Liu, Y., network for drug-drug interaction
networks. Sci. Rep. 10, 21092. Fang, Y., Jiang, M., Wang, Q., Zhao, B., prediction. BMC Bioinf. 23, 224.
68. Liu, S., Zhang, Y., Cui, Y., Qiu, Y., Deng, Y., Xiong, Y., and Wei, D.Q. (2022). MDF-SA- 98. Purkayastha, S., Mondal, I., Sarkar, S., Goyal,
Zhang, Z., and Zhang, W. (2023). Enhancing DDI: predicting drug–drug interaction P., and Pillai, J.K. (2019). Drug-drug
drug-drug interaction prediction using events based on multi-source drug fusion, interactions prediction based on drug
deep attention neural networks. IEEE/ACM multi-source feature fusion and transformer embedding and graph auto-encoder. In
Trans. Comput. Biol. Bioinform. 20, 976–985. self-attention mechanism. Brief. Bioinform. 2019 IEEE 19th International Conference on
69. Huang, K., Xiao, C., Hoang, T., Glass, L., and 23, bbab421. Bioinformatics and Bioengineering
Sun, J. (2020). Caster: Predicting drug 84. Feng, Y.-H., Zhang, S.W., Feng, Y.Y., Zhang, (BIBE) (IEEE).
interactions with chemical substructure Q.Q., Shi, M.H., and Shi, J.Y. (2023). A social 99. Yan, X.-Y., Yin, P.W., Wu, X.M., and Han, J.X.
representation. In Proceedings of the AAAI theory-enhanced graph representation (2021). Prediction of the drug–drug
conference on artificial intelligence, learning framework for multitask prediction interaction types with the unified
34Proceedings of the AAAI conference on of drug–drug interactions. Brief. Bioinform. embedding features from drug similarity
artificial intelligence, pp. 702–709. 24, bbac602. networks. Front. Pharmacol. 12, 794205.
70. Zitnik, M., Sosic, R., and Leskovec, J. (2018). 85. Su, X., You, Z.-H., Huang, D.S., and Wang, L. 100. Lin, J., Wu, L., Zhu, J., Liang, X., Xia, Y., Xie,
BioSNAP Datasets: Stanford Biomedical (2022). Biomedical knowledge graph S., Qin, T., and Liu, T.Y. (2023). R2-DDI:
Network Dataset Collection. http://snap. embedding with capsule network for multi- relation-aware feature refinement for drug–
stanford.edu/biodata. label drug-drug interaction prediction. IEEE drug interaction prediction. Brief. Bioinform.
71. Karim, M.R., Cochez, M., Jares, J., Uddin, Trans. Knowl. Data Eng 35, 5640–5651. 24, bbac576.
M., Beyan, O., and Decker, S. (2019). Drug- 86. Feeney, A., Gupta, R., Thost, V., Angell, R., 101. Yu, H., Zhao, S., and Shi, J. (2022). STNN-
drug interaction prediction based on Chandu, G., Adhikari, Y., and Ma, T. (2021). DDI: a substructure-aware tensor neural
knowledge graph embeddings and Relation matters in sampling: a scalable network to predict drug–drug interactions.
convolutional-LSTM network. Proceedings multi-relational graph neural network for Brief. Bioinform. 23, bbac209.
of the 10th ACM international conference drug-drug interaction prediction. Preprint at 102. Kim, E., and Nam, H. (2022). DeSIDE-DDI:
on bioinformatics, computational biology arXiv. https://doi.org/10.48550/arXiv.2105. interpretable prediction of drug-drug
and health informatics; ACM. 13975. interactions using drug-induced gene
72. Herrero-Zazo, M., Segura-Bedmar, I., 87. Lee, G., Park, C., and Ahn, J. (2019). Novel expressions. J. Cheminform. 14, 9–12.
Martı́nez, P., and Declerck, T. (2013). The deep learning model for more accurate 103. Tang, Z., Chen, G., Yang, H., Zhong, W., and
DDI corpus: An annotated corpus with prediction of drug-drug interaction effects. Chen, C.Y.C. (2023). DSIL-DDI: A Domain-
pharmacological substances and drug–drug BMC Bioinf. 20, 415–418. Invariant Substructure Interaction Learning
interactions. J. Biomed. Inform. 46, 914–920. 88. Chatr-Aryamontri, A., Oughtred, R., for Generalizable Drug–Drug Interaction
73. Whirl-Carrillo, M., McDonagh, E.M., Hebert, Boucher, L., Rust, J., Chang, C., Kolas, N.K., Prediction. IEEE Trans. Neural Netw. Learn.
J.M., Gong, L., Sangkuhl, K., Thorn, C.F., O’Donnell, L., Oster, S., Theesfeld, C., Syst. 1–9.
Altman, R.B., and Klein, T.E. (2012). Sellam, A., et al. (2017). The BioGRID 104. Yu, H., Wang, J., Zhao, S.Y., Silver, O., Liu, Z.,
Pharmacogenomics knowledge for interaction database: 2017 update. Nucleic Yao, J., and Shi, J.Y. (2024). GGI-DDI:
personalized medicine. Clin. Pharmacol. Acids Res. 45, D369–D379. Identification for key molecular
Ther. 92, 414–417. 89. Ashburner, M., Ball, C.A., Blake, J.A., substructures by granule learning to
74. Dhami, D.S., Kunapuli, G., Das, M., Page, D., Botstein, D., Butler, H., Cherry, J.M., Davis, interpret predicted drug-drug interactions.
and Natarajan, S. (2018). Drug-drug A.P., Dolinski, K., Dwight, S.S., Eppig, J.T., Expert Syst. Appl. 240, 122500.
interaction discovery: kernel learning from et al. (2000). Gene ontology: tool for the 105. Bougiatiotis, K., Aisopos, F., Nentidis, A.,
heterogeneous similarities. Smart Health 9– unification of biology. Nat. Genet. 25, 25–29. Krithara, A., and Paliouras, G. (2020). Drug-
10, 88–100. 90. The Gene Ontology Consortium (2017). drug interaction prediction on a biomedical
75. Sridhar, D., Fakhraei, S., and Getoor, L. Expansion of the Gene Ontology literature knowledge graph. In Artificial
(2016). A probabilistic approach for knowledgebase and resources. Nucleic Intelligence in Medicine: 18th International
collective similarity-based drug–drug Acids Res. 45, D331–D338. Conference on Artificial Intelligence in
interaction prediction. Bioinformatics 32, 91. Yu, H., Dong, W., and Shi, J. (2022). Medicine, AIME 2020, Minneapolis, MN,
3175–3182. RANEDDI: Relation-aware network USA, August 25–28, 2020, Proceedings 18
76. Zhang, P., Wang, F., Hu, J., and Sorrentino, embedding for drug-drug interaction (Springer).
R. (2015). Label propagation prediction of prediction. Inf. Sci. 582, 167–180. 106. Mondal, I. (2020). Towards Incorporating
drug-drug interactions based on clinical 92. Ren, Z.-H., Yu, C.Q., Li, L.P., You, Z.H., Pan, Entity-specific Knowledge Graph
side effects. Sci. Rep. 5, 12339. J., Guan, Y.J., and Guo, L.X. (2022). Information in Predicting Drug-Drug
77. Su, X., Hu, L., You, Z., Hu, P., and Zhao, B. BioChemDDI: Predicting Drug–Drug Interactions. Preprint at arXiv. https://doi.
(2022). Attention-based knowledge graph Interactions by Fusing Biochemical and org/10.48550/arXiv.2012.11142.
representation learning for predicting drug- Structural Information through a Self- 107. Lyu, T., Gao, J., Tian, L., Li, Z., Zhang, P., and
drug interactions. Brief. Bioinform. 23, Attention Mechanism. Biology 11, 758. Zhang, J. (2021). MDNN: A Multimodal
bbac140. 93. Bai, Y., Gu, K., Sun, Y., and Wang, W. (2020). Deep Neural Network for Predicting Drug-
78. Hu, W., Fey, M., Zitnik, M., Dong, Y., Ren, H., Bi-level graph neural networks for drug- Drug Interaction Events. In IJCAI.
Liu, B., Catasta, M., and Leskovec, J. (2020). drug interaction prediction. Preprint at 108. Dai, Y., Guo, C., Guo, W., and Eickhoff, C.
Open graph benchmark: Datasets for arXiv. https://doi.org/10.48550/arXiv.2006. (2021). Drug–drug interaction prediction
machine learning on graphs. Adv. Neural 14002. with Wasserstein Adversarial Autoencoder-
Inf. Process. Syst. 33, 22118–22133. 94. Li, Z., Zhu, S., Shao, B., Zeng, X., Wang, T., based knowledge graph embeddings. Brief.
79. Deng, Y., Xu, X., Qiu, Y., Xia, J., Zhang, W., and Liu, T.Y. (2023). DSN-DDI: an accurate Bioinform. 22, bbaa256.
and Liu, S. (2020). A multimodal deep and generalized framework for drug–drug 109. Ren, Z.-H., Yu, C.Q., Li, L.P., You, Z.H., Guan,
learning framework for predicting drug– interaction prediction by dual-view Y.J., Wang, X.F., and Pan, J. (2022). BioDKG–
drug interaction events. Bioinformatics 36, representation learning. Brief. Bioinform. DDI: predicting drug–drug interactions
4316–4322. 24, bbac597. based on drug knowledge graph fusing
80. Ryu, J.Y., Kim, H.U., and Lee, S.Y. (2018). 95. Chen, X., Liu, X., and Wu, J. (2019). Drug- biochemical information. Brief. Funct.
Deep learning improves prediction of drug– drug interaction prediction with graph Genomics 21, 216–229.
drug and drug–food interactions. Proc. Natl. representation learning. In 2019 IEEE 110. Ren, Z.-H., You, Z.H., Yu, C.Q., Li, L.P., Guan,
Acad. Sci. USA 115, E4304–E4311. International conference on bioinformatics Y.J., Guo, L.X., and Pan, J. (2022). A
81. Han, C.-D., Wang, C.C., Huang, L., and and biomedicine (BIBM) (IEEE). biomedical knowledge graph-based
Chen, X. (2023). MCFF-MTDDI: multi- 96. Wang, Y., Min, Y., Chen, X., and Wu, J. method for drug–drug interactions
channel feature fusion for multi-typed drug– (2021). Multi-view graph contrastive prediction through combining local and

26 iScience 27, 109148, March 15, 2024

iScience ll
Review OPEN ACCESS

global features with deep neural networks. 120. Zhang, S., Tong, H., Xu, J., and Maciejewski, driven continuous representation of
Brief. Bioinform. 23, bbac363. R. (2019). Graph convolutional networks: a molecules. ACS Cent. Sci. 4, 268–276.
111. Najafabadi, M.M., Villanustre, F., comprehensive review. Comput. Soc. Netw.
130. Katayama, T., Wilkinson, M.D., Aoki-
Khoshgoftaar, T.M., Seliya, N., Wald, R., and 6, 1–23.
Kinoshita, K.F., Kawashima, S., Yamamoto,
Muharemagic, E. (2015). Deep learning 121. Velickovic, P., Casanova, A., Lio, P., Cucurull,
Y., Yamaguchi, A., Okamoto, S., Kawano, S.,
applications and challenges in big data G., Romero, A., and Bengio, Y. (2017). Graph Kim, J.D., Wang, Y., et al. (2014).
analytics. J. Big Data 2, 1–21. attention networks. stat 1050, 10–48550. BioHackathon series in 2011 and 2012:
112. Sze, V., Chen, Y.H., Yang, T.J., and Emer, J.S. 122. Rumelhart, D.E., Hinton, G.E., and Williams, penetration of ontology and linked data in
(2017). Efficient processing of deep neural R.J. (1985). Learning Internal life science domains. J. Biomed. Semantics
networks: A tutorial and survey. Proc. IEEE Representations by Error Propagation 5, 1–13.
105, 2295–2329. (Institute for Cognitive Science, University of
113. Larochelle, H., Bengio, Y., Louradour, J., and California, San Diego La). 131. Smith, B., Ceusters, W., Klagges, B., Köhler,
Lamblin, P. (2009). Exploring strategies for 123. Tschannen, M., Bachem, O., and Lucic, M. J., Kumar, A., Lomax, J., Mungall, C.,
training deep neural networks. J. Mach. (2018). Recent advances in autoencoder- Neuhaus, F., Rector, A.L., and Rosse, C.
Learn. Res. 10. based representation learning. Preprint at (2005). Relations in biomedical ontologies.
114. Washio, T., and Motoda, H. (2003). State of arXiv. https://doi.org/10.48550/arXiv.1812. Genome Biol. 6, 1–15.
the art of graph-based data mining. SIGKDD 05069. 132. Hu, B.Q. (2014). Three-way decisions space
Explor. Newsl. 5, 59–68. 124. Kasun, L.L.C., Yang, Y., Huang, G.B., and and three-way decisions. Inf. Sci. 281, 21–52.
115. Bunke, H., and Riesen, K. (2011). Recent Zhang, Z. (2016). Dimension reduction with
advances in graph-based pattern extreme learning machine. IEEE Trans. 133. Wicha, S.G., Chen, C., Clewe, O., and
recognition with applications in document Image Process. 25, 3906–3918. Simonsson, U.S.H. (2017). A general
analysis. Pattern Recogn. 44, 1057–1067. 125. Dai, Y., Wang, S., Xiong, N.N., and Guo, W. pharmacodynamic interaction model
116. Jiao, L., Chen, J., Liu, F., Yang, S., You, C., (2020). A survey on knowledge graph identifies perpetrators and victims in drug
Liu, X., Li, L., and Hou, B. (2023). Graph embedding: Approaches, applications and interactions. Nat. Commun. 8, 2129.
representation learning meets computer benchmarks. Electronics 9, 750. 134. Bible, K.C., and Kaufmann, S.H. (1997).
vision: A survey. IEEE Trans. Artif. Intell. 126. Mohamed, S.K., Nounu, A., and Novácek, V. Cytotoxic synergy between flavopiridol
4, 2–22. (2021). Biological applications of knowledge (NSC 649890, L86-8275) and various
117. Zhang, Y., Lei, X., Fang, Z., and Pan, Y. graph embedding models. Brief. Bioinform. antineoplastic agents: the importance of
(2020). CircRNA-disease associations 22, 1679–1693. sequence of administration. Cancer Res. 57,
prediction based on metapath2vec++ and 127. Niu, Z., Zhong, G., and Yu, H. (2021). A 3375–3380.
matrix factorization. Big Data Min. Anal. 3, review on the attention mechanism of deep
280–291. learning. Neurocomputing 452, 48–62. 135. Ning, X., Schleyer, T., Shen, L., and Li, L.
118. Fan, C., Lei, X., Guo, L., and Zhang, A. (2019). 128. Li, G., Xiong, C., Thabet, A., and Ghanem, B. (2017). Pattern discovery from directional
high-order drug-drug interaction relations.
Predicting the associations between (2020). Deepergcn: All you need to train
microbes and diseases by integrating deeper gcns. Preprint at arXiv. https://doi. In 2017 IEEE International Conference on
Healthcare Informatics (ICHI) (IEEE).
multiple data sources and path-based org/10.48550/arXiv.2006.07739.
HeteSim scores. Neurocomputing 129. Gómez-Bombarelli, R., Wei, J.N., 136. Peng, B., and Ning, X. (2019). Deep learning
323, 76–85. Duvenaud, D., Hernández-Lobato, J.M., for high-order drug-drug interaction
119. Zhou, J., Cui, G., Hu, S., Zhang, Z., Yang, C., Sánchez-Lengeling, B., Sheberla, D., prediction. Proceedings of the 10th ACM
Liu, Z., Wang, L., Li, C., and Sun, M. (2020). Aguilera-Iparraguirre, J., Hirzel, T.D., international conference on bioinformatics,
Graph neural networks: A review of methods Adams, R.P., and Aspuru-Guzik, A. (2018). computational biology and health
and applications. AI Open 1, 57–81. Automatic chemical design using a data- informatics; ACM.

iScience 27, 109148, March 15, 2024 27