Challenges in Computational Drug-Drug Interaction Prediction: A Survey

Drugs are chemical substances capable of inducing biological effects when introduced into a living organism, traditionally derived from medicinal plants and increasingly synthesized through organic chemistry [1, 2]. Advances in pharmaceutical science have led to the creation of comprehensive databases supporting research into drug mechanisms and interactions [3, 4]. A prime example is DrugBank [3], which lists 15,451 drug entries, including approved small molecule drugs, biologics, nutraceuticals, and experimental drugs. In clinical treatment, managing diseases often requires multiple drugs [5], as combination therapies generally yield higher success rates than monotherapies [6]. For instance, drug cocktails like doxorubicin, cyclophosphamide, vincristine, and prednisone are standard in cancer chemotherapy [7], while co-administration of anti-tuberculosis drugs enhances efficacy and delays resistance [8].

Nevertheless, the possibility of harmful drug-drug interactions (DDIs) rises with the number of medications given to a patient [9]. According to the study [10], for example, more than one-third of elderly Americans frequently take five or more medications or supplements, and 15% are at high risk of developing serious DDIs. DDIs include interactions with food, metabolites, endogenous chemicalsand diagnostic agents in addition to interactions between therapeutic medications [11]. They can change the nature, intensity, duration, side effects, and toxicity of drugs by either amplifying (synergistic action) or decreasing (antagonistic action) their efficacy [12]. Reactions resulting from DDIs can be advantageous, negligible, or detrimental [13]. In drug safety, these dangerous DDIs are the main emphasis.

Comprehending the pharmacological actions of each drug in a combination therapy is crucial for optimizing therapeutic efficacy while minimizing adverse reactions [14]. The importance of drug–drug interactions expands even more when aging is considered because the elderly take more than one medication at a time. An investigation performed in 2019 by The Health Insurance Review and Assessment Service (HIRA) found 41.8% of seniors ages 65 and older, who were seen in outpatient clinics, were given five or more prescriptions 14.4% were given 10 or more [15]. The data used in the study was based on the Korean National Health Insurance (NHI). Following the COVID-19 pandemic, the necessity of DDIs has been underscored, particularly for COVID-19 infected individuals with underlying medical comorbidities who are recurrently on multiple drugs. Due to the difficulties triggered by DDI prediction, an increase in computational approaches has been observed to flourish in recent years with revived interest in ML approaches and these methodologies. The one that garnered the most attention was called DeepDDI, the baby of the group as a computational model, being the first of its kind to use DL techniques to predict the possibility of the deadliest hyperport in DDI with any two of the thousand vis-a-vis drugs were to be combined which produces an array of permutations, most deadly [16]. This innovation lead to a brand new era in DDI precision, leading to many new extremely complicated ML models that utilizes much higher bandwidths of data not only new sources but also that are analysed even closer using far more deeper and complicated learning techniques.

The rapid progress in this field makes it a perfect time to complete a comprehensive survey of these modern DL models. Firstly, takes a look at the various dataset sources used by these models, and explains how this data is highly varied and detailed and is not just one type of DDI, which helps the predictions to be more accurate. The second aspect to be considered concerns the featurization approaches of model input data. It demonstrates the ways adopted to process and integrate molecular structures and graphs coding DDIs or biological knowledge into the models. Thirdly, the review will ascertain the methodologies used in utilizing those ML models in predicting DDIs, and the effectiveness of such methodologies in predicting DDIs, as illustrated in Figure 1.

This review demonstrates forthcoming research prospects in DDI predictions. Furthermore, the document includes various gaps that interested researchers can cover in order to advance in this field. Particularly, the paper clearly indicates areas that need to be considered so as there could be a perfect and right DDI predictions, which are basically to increase on the accuracy, reliability and convenience of the DDI predictions towards pharmaceutical and clinical areas. Taking an all-inclusive perspective, the purpose of this review is to guide future research and development of DDI prediction.

1.png

Figure 1. Overview pipeline of DDI prediction based on drug features and deep learning models

Molecular representation is critical in many drug-related tasks, with its importance particularly arising in the DDI prediction. An example of this would be the example of Tranylcypromine, which is a substance known for its function as an inhibitor of the monoamine oxidase enzyme [22]. When it comes to using medication to deal with mood or anxiety disorders a good way to approach this problem is with using Tranylcypromine, which can clear the way for a nonselective, irreversible antidepressant and anxiolytic agent.

There are many different ways to represent a drug molecule, and one of them, is through SMILES notation [23]. An investigation by experimental and computative accounts showed a summary and systematic layout of the arrangement of molecules which scientist employ to view and dissect a compound easily and not in great detail, as demonstrated in Figure 2. The choice of molecular representation is a significant issue, as this is the basis of how the computational techniques especially DL models operate towards the prediction of DDIs also whether it be smiles notation or some other way. The precision of DDI predictions increases and in addition we learn more about to gain more understanding of pharmacological effects, from the DDI given, Molecular representation is important in aspect of its accuracy, to give additional information.

2.png

Figure 2. Different representations of a drug molecule including fingerprint, graph, SMILES, InChI, and 3D electrostatic potential

3.1 SMILES sequence

SMILES, the Simplified Molecular Input Line Entry System, is one of the most useful tools in chemical computing, and is used by every chemist involved with computer information. SMILES strings are character strings that can describe complex chemical structures. Each atom in the SMILES string is represented by a unique ASCII symbol. SMILES strings contain very special symbols for stereochemistry, chemical bonds, and branching patterns.

The fantastic ability of SMILES strings is being able to turn any intricate chemical in the world into a simplified, tree shaped figure that’s easy to understand. The transformation of the molecular information is done by following a tree pattern where it is done longitudinal-first neighbor. A series of characters are going to take as the last result. Models based on DL can provide a well-rounded approach to handling differing measures to the sequence of data that is introduced in given output [24, 25]. Working together, they use what they know to interpret the SMILES Strings into consumable information in the same way that humans read a string of text.

With their compactness, memory efficiency, and ease in searchability, there are many advantages to using sequence-based representations. Since they are so compressed, they don’t waste valuable space in memory while conserving space. They are great for encoding molecular structures because they are so great at being triggered by words in a sentence. SMILES representations can also allow for the translation of chemical context from SMILES sequences, using Mol2Vec and FCS, which are techniques that are built to be able to understand the chemical relationship within the molecule, similar to how NLP have strategies to help with the translation of context [26, 27]. In summary, the remarkable and highly adaptable representation via the SMILES sequences make the entrée into the utilization of DL models for the foreseeing of the drug connections and several other chemical tasks and outcomes.

3.2 2D graph

Molecular representations utilize graph-based structures to make use of the most updated pharmacological compounds that will be very helpful and will be able to be produced by two-dimensional molecular graphs that month RDKit is a computer program that allows a SMILES string to be turned into its 2-D structure.

Following that, each node is given a set of atomic features that was established according to the node’s particular atomic number. Each node in a molecule’s graph has a starting 78-dimensional feature vector consisting of a variety of atomic attributes such as its atomic symbol, implicit value and aromaticity, neighbor atoms, neighbor hydrogens, and also mean values. As a result, we are left with the molecular graph representation, like the one for tranyl-cypromine, which is made up of edge features, atom numbers, and atomic features. It is possible to extract important structural information from the molecular graph thanks to this representation. 2D GNN models often leverage message passing neural networks (MPNN), a classic approach for encoding graph-based methods. Since 2D graphs are commonly stored as adjacency matrices, 2D GNNs facilitate efficient and accurate property combination between adjacent atoms or chemical bonds. Furthermore, they optimize weights during the message-passing procedure. Graph-based representations are advantageous in their ability to extract structural information through graph convolutional operations compared to sequence-based methods. These operations allow for the updating and optimization of bond weights within message-passing networks, enhancing their utility in various computational chemistry tasks.

3.3 3D graph

While 2D molecular graphs effectively capture structural connectivity among atoms, they fall short in representing spatial configurations crucial for understanding real-world drug interactions. The 3D graph representation addresses this limitation by incorporating the spatial coordinates of atoms within a molecule, enabling a more precise depiction of its geometry and conformation. This is especially important for modeling inter-molecular interactions such as ligand-receptor binding, where three-dimensional shape, orientation, and atomic distances determine binding affinity and biological activity. Applications of 3D graphs include the generation of conformer ensembles and the accurate prediction of molecular properties like binding energy, reactivity, or selectivity [28, 29]. By including features such as atomic coordinates, bond angles, and torsional geometry, 3D graph-based models provide richer information and enable deep learning architectures—such as 3D graph neural networks—to learn spatial dependencies that are critical in drug–target or drug–drug interaction prediction.

3.4 Drug-drug interaction network

DDIs are a multifaceted area of research that combines information from areas including biology, chemistry, and other information about drugs to measure the likelihood of interaction between two drugs. A way to take apart the very complex web of DDIs is to create a DDI network. This network is an outline of certain drug molecules and can give a clear account of the possibilities of chemical linkage between them. By doing this we will have a better understand of how the drugs are designed to interact.

Within the context of the DDI prediction task, the problem is often cast as a missing link prediction problem. Drugs are cast as nodes, and the established interactions as the edges connecting them. By putting drugs into a state that allows a predictive model to be created, what’s done is to change drugs into feature vectors. In order to do this, we consider its interaction profiles that have been gotten from interactions known. The use of this representation allows for the building of models that can determine potential DDI occurrences, giving a better look into the very complex interaction of drugs and allowing for better choice-making with regards to drugs.

3.5 Heterogeneous graphs

A repository of extensive information represents a hetergenous graph (HetG) that contains structured relations among diverse types of nodes and unstructured contents related to each node [30]. Hitting the genetic switch proves to have some merit in predicting DDI’s. HetGs signals the function of cancer related gene pairs and is a notable player in determining successes.

When talking about DDIs the format of a typical HetG sometimes looks like a graph, it is usually denoted as G = (V, E, OV, RE), where V is the set of nodes, E is the set of links, OV is the set of object types and RE is the set of relation types. Moreover, each node within this HetG is imbued with heterogeneous content, such as attributes and properties. The graph encodes relationships between several pairs of entities. Among those considered are a drug and the protein that it targets, a drug and the side effects that it elicits, and a drug and the diseases that it treats. To illustrate, consider a biological heterogeneous graph centered around a drug like Fulvestrant. The reassignment of the bar that come in diverse paint illustrates the many solutions and arrow mean the direction of it and it can help for me the view of shift the framework by graping a itself containing more information.

3.6 Knowledge graphs

Knowledge graphs (KGs) have emerged as a valuable resource in the realm of drug discovery, garnering attention from both the academic community and various sectors within the field [31]. KGs offer a structured representation of human knowledge, and their application has proven beneficial in the drug discovery domain. The extraction of high-order semantic characteristics that enhance the estimate of DDIs is made possible by these KGs, which allow the smooth integration of various entity kinds and association interactions among biological entities.

A knowledge graph typically takes the form of G = (V, E, F), where E represents the set of entities, R denotes the set of relations, and F encompasses the set of facts. Facts within the KG are expressed as triples (h, r, t) F, where h and t are entities connected by relation r. Entities are depicted as nodes, each characterized by distinct colors and alphabets, representing real-world biological objects such as drugs, targets, and side-effects. Relationships (edges) illustrate the connections between entities, and they incorporate semantic descriptions, encompassing types and properties with well-defined meanings, including associations like Drug-Disease, Drug-Target Gene, and Drug Brite.

As an illustration of its practical application, KGs have played a pivotal role in addressing challenges posed by the COVID-19 pandemic [32, 33]. Notably, there exist several knowledge graphs tailored to various facets of the drug discovery procedure, such as Clinical Knowledge Graph, DRKG, OpenBioLink, PharmKG, BioKG, and Hetionet. While providing a brief overview, it’s worth noting that a comprehensive review of these KGs goes outside the limits of what the work covers, and interested readers are encouraged to explore devoted reviews on the subject [34].

This rapidly advancing subject has seen a vast number of techniques used as proven by the timetable of results found in Table 2. The table has shown the range of techniques being used as well as their chronological order which can correspond to the date of publication. A comprehensive overview of deep and graph learning techniques that have been developed recently is shown in the table below. In the table, we see that each model has its unique characteristics in a set of columns, i.e., model name, the types of input, the representation method, the architectural frameworks, the classification tasks.

Beginning in 2021 with SumGNN the path has been paved for a variety of new techniques, GNNs with knowledge graph/subgraph representations and attention mechanism, and GCN on molecular graph with contrastive learning. There are a lot of unique methods like MIRACLE and SSI-DDI that are binary classing amongst themselves from SMILES data down to molecular substructures, and others such as AAEs using knowledge graphs with adversarial autoencoders. This trend shows a greater change in this field of work and the way more compound and diverse data representations and also more intricate and multifaceted architectural strategies. Entering 2022, we observe a growing diversification of models and approaches. GNNs on Molecular Graphs such as GNN-DDI, and MFFGNN with multi-type features, continue trending. The increasing complexity of these models can be observed by DeepDrug that introduces a RGCN architecture. Furthermore, this year shows the increasing range of the field, with models employing drug features, biomedical networks, and directed graphs.

In 2023, the trend towards more intricate models continues with DSN-DDI and DGNN-DDI, among others, embracing dual-view encoders and directed MPNNs. This not only reflects the ongoing refinement of existing methodologies but also the introduction of novel approaches to tackle the complexity of DDI prediction. Each model’s contribution to the field is further delineated by their unique approaches to data representation and processing architecture, be it through attention mechanisms, contrastive learning, or capsule networks.

In our comprehensive analysis, we conducted a detailed comparison of various new deep and graph learning models, focusing on their capabilities in binary, multi-label classification, and multi-class tasks for DDI predictions. This comparison, as presented in Tables 3 and 4, evaluates the performance of different models under the binary classification task on two benchmark datasets: DrugBank and TWOSIDES. The evaluation metrics employed to assess the effectiveness of these models include the Area Under the Precision-Recall Curve (AUPRC), Accuracy (ACC), Area Under the Receiver Operating Characteristic (AUROC), and F1-score. These metrics were calculated using a 5-fold cross-validation approach to ensure robustness and reliability in our assessment. Notably, higher values in these metrics correlate with superior predictive performance. An important aspect of our comparison is the recognition that despite the fact that different models may divide training and test data differently, the evaluation remains statistically meaningful. This is because each model’s performance is appraised under consistent, rigorous criteria, providing a fair and comprehensive comparison.

Table 2. Summary of DDI prediction models

Table 3. Performance metrics for DrugBank dataset (in %)

Table 4. Performance metrics for TWOSIDES dataset (in %)

In our observations, particularly on the DrugBank dataset, we noted standout performances by RANEDDI (AUPRC = 98.94%) and KGNN (AUPRC = 98.92%), both of which are network-based methods. These models achieved the highest and second-highest AUPRC performance, respectively, surpassing those of chemical structure-based and hybrid methods. The success of RANEDDI and KGNN can be attributed to their ability to effectively utilize multi-relational information inherent in DDI networks or knowledge graphs. In contrast, graph embedding procedures, such as DeepDDI, GraRep, DeepWalk, and substructure-based methods such as CASTER primarily leverage similar chemical structural information or drug characteristics.

It’s also important to note that R2-DDI performed better in terms of both AUROC and F1-score, while DANN-DDI had the highest ACC result of 0.9962, surpassing all other models.

On the TWOSIDES dataset, the recently published chemical structure-based model DSN-DDI showed remarkable results, outshining other baseline models across all evaluation metrics. Specifically, it achieved an ACC of 0.9883, an AUROC of 0.9990, and an F1-score of 0.9883, indicating its robustness and accuracy in DDI prediction. Another interesting insight derived from the comparative studies is that the network-based approaches, such as DANN-DDI and RANE-DDI, have performance comparable with the chemical structure-based methods (e.g. R2-DDI). In the mean-time, for the binary classification task, the hybrid methods still perform steadily on DrugBank. The value of different methods to predict DDI seems promising, which signifies the ease of selecting suitable models in the dataset that can relate to the task.

In an endeavor to design the multi-class performance evaluating methods. The work we had presented here in this comprehensive analysis deals with the multi-class performance metric, which will mainly enable to investigate various types of DL and graph learning models by taking the benchmark datasets such as DrugBank and TWOSIDES datasets has been intensively evaluated, as depicted in Tables 5 and 6.

Provided in these tables is a thorough comparison of the effectiveness of the models in relation to multiple key evaluation metrics. The experiments lead us to conclude that KG2ECapsule and SumGNN ranked as the first two best models with consistent peak performances regardless of which metric we evaluated. They have demonstrated their extraordinary capability in tackling complicated classification tasks.

KG2ECapsule shows significant improvements over the best baseline models on the DrugBank dataset in particular. It showed an enhancement of 2.71% in PR-AUC, 1.03% in ACC, 2.8% in ROC–AUC, and 2% in F1 score. The reason for this increase in accuracy is due to KG2ECapsule being able to accurately model the triplets and include the connections that are present in the edges into the embedding algorithm. This demonstrates a more discriminate and effective method for integrating and employing KG information.

However, for the TWOSIDES dataset, SumGNN has showed better improvement by at least 2.45% and 2.82% in PR-AUC and ROC-AUC respectively compared to other methods. It is strongly suggested by this that SumGNN must be giving thought to the subgraphs that they are using since they are being so successful. The fact that it is able to exploit these outside information looks like to give it a step ahead against many other models. Moreover, in terms of KG based approaches such as KG-DDI and KGNN, as a comparison, SumGNN and KG2ECapsule are both superior even though the comparison of both methods is only single, while KG2ECapsule and SumGNN consistently are both higher than any other on two datasets. This observation really drives the point home: just using KGEmbed and neighborhood sampling might not be that great an idea to sufficiently leverage KG information for DDI prediction. This implies the requirement for more sophisticated techniques that can handle the complex structures and relationships inherent in KGs in a more comprehensive manner.

In addition, network-based techniques, with all three different approaches, had better performances in the multi-class problem. Evidence of these trends may mean the architectural and computational strategies in certain network-based models could have a high performance on multi-class DDI prediction tasks, because DDI prediction is a complex task involving many aspects of information. This knowledge presents a promising research direction where network-based models can be extrapolated and used to potentially redefine how accurate we can be and how fast we can do predictions for DDI.

Table 5. Multi-class performance metrics for DrugBank dataset (in %)

Table 6. Multi-class performance metrics for TWOSIDES dataset (in %)

In reviewing the progression of deep learning and graph-based models for drug–drug interaction (DDI) prediction from 2021 onward, several key trends and performance patterns emerge. Firstly, network-based models, particularly those leveraging knowledge graphs and biomedical networks—such as RANEDDI, DANN-DDI, and KGNN—demonstrate superior performance in metrics like AUPRC and AUROC. These models excel due to their ability to capture complex, multi-relational patterns inherent in drug interaction networks, as opposed to relying solely on chemical structure or sequence-level data. For example, RANEDDI achieved the highest AUPRC on the DrugBank dataset (98.94%), while DANN-DDI recorded the highest accuracy (99.62%), showcasing the power of embedding techniques and attention mechanisms when applied to relational data. In contrast, structure-based methods using SMILES strings and molecular graphs—such as GMPNN, DeepDrug, and DSN-DDI—have also shown competitive performance, particularly in binary and multi-class classification, thanks to the incorporation of message-passing neural networks (MPNNs), graph convolutional networks (GCNs), and attention modules that enable more expressive molecular feature learning. Hybrid approaches like MFFGNN and DGNN-DDI, which integrate multi-type features from both chemical and relational domains, offer balanced effectiveness across metrics and task types. Another trend is the increasing use of attention mechanisms, contrastive learning, and transformer-like architectures, which enhance the interpretability and generalization of models. Moreover, the recent inclusion of 3D structural information in models such as 3DGT-DDI highlights a growing recognition of spatial configuration’s role in accurate DDI prediction. Overall, the diversity in input types (SMILES, drug IDs, graphs), architectural complexity (from GCNs to capsule networks), and task orientation (binary, multi-class, multi-label) reflects a maturing field, where model choice is often dictated by the specific DDI task, dataset characteristics, and desired performance trade-offs. The following tables offer a detailed breakdown and comparison of these models across different metrics, datasets, and classification types.