Machine learning and deep learning-based drug-drug interactions prediction: a systematic review focused on anticancer drugs

Abstract

Cancer patients are particularly susceptible to Drug–Drug Interactions (DDIs) due to frequent polypharmacy in oncology care. Co-administered drugs can increase toxicity or reduce effectiveness, potentially causing serious adverse events—for example, QTc-prolonging Tyrosine Kinase Inhibitors with CYP3A4 inhibitors can lead to torsade de pointes. Traditional DDI identification methods are time-consuming and costly, relying mainly on in vitro and in vivo wet lab experiments, clinical studies, or post-marketing surveillance. Many Machine Learning (ML) and Deep Learning (DL)-based DDI prediction models have been developed in recent decades to accelerate the identification of DDIs. We systematically reviewed ML- and DL-based DDI prediction models involving anticancer drugs. Key features of anticancer drugs involved and details of prediction models, such as the prediction tasks (existence or types of DDI) and performance, were summarised, as well as a list of newly predicted DDIs. Additionally, verification through up-to-date DrugBank and Drugs.com confirmed 22 of 96 newly predicted potential DDI drug pairs, demonstrating the practical value of these techniques. By understanding the current DDI prediction studies from both methodological and clinical standpoints, novel approaches may be tailored to the unique characteristics of oncology drugs, thereby enhancing the clinical relevance and applicability of DDI predictions.

Introduction

Cancer is a global disease that poses a significant public health and economic burden^1,2, and a leading cause of death³. In 2022, there were approximately 20 million new cases of cancer and 9.7 million cancer-related deaths. According to statistics from the International Agency for Research on Cancer (IARC), cases are estimated to reach 35 million by 2050¹. Polypharmacy, defined as the use of multiple medications simultaneously⁴, is more common among cancer patients due to the complexity of treatment regimens and the presence of comorbidities^5,6, particularly in older individuals⁷. One study found a median number of six medications among 350 cancer patients; 47.1% were taking six or more medications, and 11.4% were taking eleven or more medications⁸. The increasing number of concomitant medicines poses a potential threat of drug–drug interactions (DDI)⁹.

Drug–drug interactions occur when the effect of one drug is changed by the co-administration of another drug¹⁰. This can be categorised as a Pharmacokinetic (PK) and Pharmacodynamic (PD) interaction, an event where the delivery of a drug (PK) or the response of the drug target (PD) is altered by another drug¹¹. Figure 1 summarises the mechanisms of PK and PD DDIs according to the pharmacological actions of therapeutic agents^11–22.

Fig. 1. Mechanisms of pharmacokinetic and pharmacodynamic drug–drug interactions.

CYP450 cytochrome P450, DDI drug–drug interaction, PD pharmacodynamic, OATP organic anion transporting polypeptide, P-gp P-glycoprotein, PK pharmacokinetic.

A drug (Drug 1) administered at a specific dose via a predefined administration route reaches the systemic circulation after undergoing Absorption, Distribution, Metabolism, and Excretion (ADME) processes (the PK process). Once the therapeutic concentration is achieved at the target site, the drug binds to the target and activates the subsequent signalling pathway, stimulating a pharmacological effect that can lead to both clinical toxicity and therapeutic effectiveness^11,14. During these processes, a co-administered drug (Drug 2) may inhibit or activate transporters and/or enzymes^18,19, or modify gastric acid²¹, thereby changing the ADME behaviour of Drug 1, causing PK DDIs. Drug 2 may also affect the binding between Drug 1 and the target, or influence the signalling pathway, resulting in PD DDIs¹⁷. The structure of a drug determines its PK and PD behaviours¹¹, while patient characteristics such as age, gender, and gene polymorphisms have also been shown to be significantly associated with DDIs^12,13,22.

DDIs may lead to reduced efficacy or increased toxicity^11,23 and are important causes of adverse drug reactions (ADRs). It has been reported that 9.69% of ADRs are attributed to DDIs²⁴. These can be life-threatening or fatal, accounting for 11% and 8%, respectively, of all serious ADRs caused by DDIs²⁵. DDI severity can be stratified into three levels: minor, moderate, and major¹⁵.

The related risk of reduced efficacy or increased toxicity of anticancer drugs rises particularly for those with a Narrow Therapeutic Index (NTI). A minor difference in dose or systemic concentration, determined by PK and PD processes, may substantially impact therapeutic effectiveness and/or toxicity^26,27. For example, 5-Fluorouracil (5-FU), a cytotoxic anticancer drug with an NTI, has been reported to cause fatal toxicity in patients receiving concurrent treatment of sorivudine, which inhibits the enzyme responsible for metabolising 5-‍FU²¹. Another widely used anticancer drug, Methotrexate, often used in the treatment of haematological cancers, has been reported to cause DDIs when administered simultaneously with Non-Steroidal Anti-Inflammatory Drugs (NSAIDs). NSAIDs can block the elimination of Methotrexate, leading to a significantly increased risk of serious adverse events, even when Methotrexate is given at low doses^21,28. Besides small molecular anticancer drugs, the approval of biologics for cancer treatment has increased in recent years²⁹, drawing attention to the potential for related DDIs³⁰. Although DDIs are less likely to occur with biologics than with small molecules³¹, the risk warrants attention, particularly for cytokine modulators, which may affect Cytochromes P450 (CYP450) enzymes and transporters³².

Thus, it is crucial for healthcare professionals to be aware of potential DDIs before prescribing medications, thereby reducing safety risks and improving patients’ quality of life. Traditional DDI identification is mainly based on in vitro and in vivo wet lab experiments, clinical studies, or post-marketing surveillance^33,34, but these methods are time-consuming and costly³⁵. To address these challenges, Artificial Intelligence (AI)-based DDI prediction models have been developed³⁶. By using publicly available biomedical and drug databases containing drug information on structures, target, transporters, enzymes, pathways, or disease-related literature, self-supervised methods can effectively extract valuable information to predict potential DDIs among unlabelled drug pairs, and determine DDI types³⁷. These DDI prediction outputs can serve as a valuable reference for healthcare professionals, aiding the identification of potential DDIs and mitigating associated risks. The prediction enhances clinical decision support by alerting clinicians to potential harmful combinations³⁸, suggesting alternative agents with lower interaction risk³⁹, and enabling regimen simplification to minimise long-term toxicity⁴⁰.

Although several reviews on AI-based DDI prediction models have been published (Supplementary Table 1), some limit searches to single database, leading to incomplete bibliometric overviews⁴¹, while others adopt a primarily methodological perspective and lack clinical focus⁴². In this study, we aimed to systematically synthesise ML or DL-based DDI prediction models on anticancer drugs by summarising key information, including model algorithms, drug data sources, and features, prediction objects (DDI existence or type classification), evaluation methods (evaluate known and unknown interacting drug pairs), negative sampling methods for unreported interactions, and prediction performance. This framework seeks to facilitate the development and application of robust DDI prediction tools that fit the clinical needs of anticancer agents, which could improve clinical decision-making and patient outcomes.

Results

DDI study searching and quality assessment

Studies on DDI prediction involving anticancer drugs were identified and screened manually according to the eligibility criteria outlined in the ‘Methods’ section, along with detailed study selection procedures (Fig. 2). An initial total of 2457 records were identified. After removing duplicates and a comprehensive review of titles and abstracts, 237 were selected for full-text screening and assessed for eligibility. Ultimately, a total of 59 studies were filtered and reviewed in detail (Supplementary Table 2).

Fig. 2. PRISMA 2020 flow diagram for literature screening process.

DDI = drug-drug interaction.

Most of the included studies were published as journal articles (n = 48), with the number of publications increasing over time, especially from 2022 onward (Supplementary Table 2). One study⁴³ used lung cancer as the clinical case to illustrate the DDI model development process, and another study⁴⁴ specifically collected actual anticancer drugs to develop its prediction model. In the remaining studies (n = 57), anticancer drugs were incorporated into the model evaluation, validation, prediction application, case studies, or presented in the reviewed literature or DDI prediction visualisation figures.

Reporting quality of each included study was assessed through the Clinical Artificial Intelligence Research (CAIR) checklist⁴⁵ (Supplementary Table 3). Except for three items that were not applicable to DDI prediction studies noted in the methodology section, all included studies satisfied the other items 1–12. One study⁴⁶ primarily demonstrated the architecture of the DDI prediction approach and the presentation of predicted DDIs, and did not provide empirical results or evaluation, which violates item 13 in CAIR. Additional variation in reporting quality was observed in item 14 (“State if and how the AI model/data are accessible, including any restrictions to access or reuse”) and item 15 (“Describe ethical considerations and implications, biases, and limitations”). These differences contributed to the verified CAIR scores, resulting in an average score of 11.3 out of 12 (Supplementary Table 3).

Classification of anticancer agents

The anticancer drugs identified from the included studies were classified as small molecules or biologics according to DrugBank (https://go.drugbank.com/). Among the included 130 specified anticancer drugs, and only three drugs, Cetuximab, Denileukin diftitox, and Trastuzumab, are biologics (Supplementary Table 2).

Algorithm for DDI prediction

Most of the included models (72.9%, 43/59) are based on DL (Fig. 3), showing an increasing trend over the past five years compared to ML. These studies often integrated multiple technologies, such as Graph Neural Network (GNN, including Graph Convolutional Network [GCN]), attention mechanisms (e.g., self-attention and multi-head attention), multi-modal or multi-scale learning, transformers and autoencoder, to develop DDI prediction models. For example, the Deep Multimodal Fusion for Drug–Drug Interaction (DMFDDI) model combined a deep multimodal feature fusion framework with a GCN⁴⁷. Another study applied a Graph Auto-Encoder (GAE) in combination with residual GCN on multiple sources of drug features⁴⁸. Additionally, a deep learning-based DDI prediction model was specifically designed for predicting interactions between small molecules and biologic agents⁴⁹. Among the hybrid models that combine DL and Knowledge Graphs (KGs), a study by Xu et al.⁵⁰ enhanced DDI prediction by integrating biomedical KGs with the semantic capabilities of large language models.

Fig. 3. Classification and trends of deep learning and machine learning models for drug–drug interaction prediction involving anticancer drugs.

Updated as of May 2025. DL deep learning, KG knowledge graph, ML machine learning.

Drug information resource

A total of 22 drug information databases were utilised across the 59 included studies during DDI prediction model development (Supplementary Table 4). DrugBank (https://go.drugbank.com/) was the most frequently used source, appearing in approximately 67.8% (40/59) of studies. From DrugBank, researchers collected DDIs and a variety of drug features, including structures, targets, transporters, enzymes, pathways, and Morgan fingerprints. Utilising amino acid sequences of biologics from the DrugBank, one study⁴⁹ also predicts DDIs between small molecules and biologics. TWOSIDES (https://tatonettilab.org/offsides/) was the second most used database for DDIs collection. Databases such as the Kyoto Encyclopedia of Genes and Genomes (KEGG, https://www.genome.jp/kegg/), Uniprot (https://www.uniprot.org/), and PubChem (https://pubchem.ncbi.nlm.nih.gov/) were typically used to collect common drug features. The Drug Repurposing Knowledge Graph (DRKG, https://github.com/gnn4dr/DRKG) was generally used for KG construction. For side effect data, OffSIDES (https://tatonettilab.org/offsides/) and SIDER (http://sideeffects.embl.de/) databases were commonly used.

In addition to extracting DDIs and drug features directly from databases, many studies derived their datasets from previously existing datasets (Supplementary Tables 4 and 5). The most commonly used datasets were those collected by Zhang et al.⁵¹ and Deng et al.⁵², which include 548 drugs with 48,584 DDIs and 572 drugs with 74,528 DDIs, respectively. These datasets also provide various drug features such as chemical structures, substructures, targets, enzymes, and pathways. A third frequently used larger dataset collected by Ryu et al.⁵³ contains 1704 drugs and 192,284 DDIs⁴⁷.

Many studies utilised multiple databases and/or datasets, often incorporating both small-scale databases and large-scale databases (Supplementary Table 5). For instance, Chen et al.⁵⁴ used the small-scale dataset derived from Zhang et al.⁵¹ for binary classification and a large-scale database derived from Ryu et al.⁵³ for multi-label classification, to evaluate model performance on different tasks. Similarly, Gan et al.⁴⁷ evaluated the prediction performance across three datasets of varying sizes. Detailed information on dataset usage for each study, including the number of drugs, the number of DDIs, and the types of drug features used, is provided in Supplementary Table 5.

Drug features

The drug features used in the included studies can be broadly classified into four categories⁵¹: (1) chemical features, such as structures, fingerprints, and substructures; (2) biological features, such as targets, enzymes, transporters, pathways and gene ontology; (3) phenotypic features, such as indications, adverse effects and off-side effects; (4) network features, such as semantic relationships, or knowledge graphs connecting entities (e.g. drugs, targets, diseases). Initially, studies primarily focused on chemical structures, biological features, and phenotypic features. However, there is a growing trend toward integration with biomedical network features to enhance model prediction performance (Supplementary Table 5).

DDI prediction objects

Most of the included models (94.9%, 56/59) are designed to predict DDI existence (binary classification) or type (multiple classification), with the number of DDI types based on the dataset used (Supplementary Table 6). These DDI types are typically related to changes in PK or PD, or associated adverse effects, which help explain DDI mechanisms. Among these studies, two models by Kang et al.⁵⁵ and Zhang et al.⁵⁶ specifically focused on predicting the type of side effect resulting from DDIs, while two models by Allahgholi et al.⁴⁰ and Ryu et al.⁵³ are also able to recommend alternative drugs for DDIs with negative health effects. Additionally, one model by Feng et al.⁵⁷ provides suggestions on whether a DDI is asymmetric (i.e., monodirectional interaction). Three models were constructed to predict whether one drug impacts the efficacy or activity of another drug^58–60. The other two models are related to CYP450 enzymes; one predicts whether a DDI involves a substrate or inhibitor of CYP450 enzyme⁶¹ and the other predicts whether a DDI is related to CYP450 enzyme⁶².

Among multi-label classification models, in addition to predicting DDI types, one study by Pham et al.⁶³ aimed to predict the severity of DDIs based on clinical relevance between antiretroviral therapies and co-medications. For binary classification beyond simply predicting the existence of a DDI, a model by Peng et al.⁶⁴ was also designed to predict whether the drug combination, including those with arbitrary numbers of drugs (including high-order DDIs), would cause a specific adverse event, i.e., myopathy. Another model developed by Liu et al.⁶⁵ focused on predicting whether a DDI should be classified as high or low priority for alerting purposes, based on the severity of DDIs.

DDI prediction task and prediction performance

Across the reviewed studies, DDI prediction models generally demonstrated strong predictive performance, with results varying by prediction task. For DDI existence prediction, most models achieved high performance, with accuracies commonly above 0.80, Area Under the Receiver Operator Characteristic curve (AUROC) values typically exceeding 0.90, and Area Under the Precision-Recall Curve (AUPRC) values frequently above 0.80. Performance remained strong overall for DDI type prediction, even in multi-label tasks where most studies predict more than 65 interaction types. Reported accuracies generally ranged from 0.85 to 0.98, with AUROC values consistently high (commonly >0.95) (Supplementary Table 6).

The primary function of DDI prediction is to forecast unknown interactions between drug pairs, thereby better reflecting real-world scenarios such as providing clinicians with early warning signals and helping to prioritise experimental validation efforts before clinical confirmation. To evaluate a model’s ability in realistic clinical scenarios, DDI prediction tasks are typically stratified into three categories, which were adopted in 12^{52,55,59,66–74} included studies. Specifically, task 1 involves known–known drug pair, where both drugs are from the training set; task 2 involves known–new drug pair, where one drug is from the test set and the other from the training set; task 3 involves new–new drug pair, in which both drugs were not used during the construction of the DDI models. The first task, also referred to as a transductive setting, is used for internal evaluation. In contrast, the latter two tasks are inductive settings, which pose a greater challenge for DDI prediction, but provide a more realistic assessment of a model’s applicability to predict interactions involving novel, unseen drugs.

Among the reviewed studies, DDI prediction models’ performance varies depending on the prediction task, the drug features incorporated into each model, and the method employed for handling negative samples.

Features vs. performance

The combination of multiple features has been shown to improve model performance in 12 included studies^{44,47–49,52,57,66,75–79}. For example, Lee et al.⁷⁵ demonstrated that adding Gene Ontology (GO) terms and target genes increased the prediction accuracy compared to using only structure features. Similarly, Deng et al.⁵² found that combining features improved performance (Accuracy: 0.8725 [structure + target + pathway + enzyme] vs. 0.8623 [structure alone]), although chemical structure remained the most informative feature, which was also reported by Asfand et al.⁷⁶ and Liang et al.⁴⁴. However, these latter two studies also identified contradictory results: combining all four features (substructures, targets, enzymes and pathways) did not outperform the combination of three features (substructures, targets and enzymes)⁷⁶, and no improvement was observed when combining targets and pathways if a drug targeted only a single pathway⁴⁴. To address this, Liang et al. developed the Multi-View Semi-supervised Graph-based framework to automatically assign different weights to each feature according to their contributions for different drugs⁴⁴.

Performance on prediction tasks 1, 2, and 3

Prediction models usually performed better in task 1, with a significant decline observed in tasks 2 and 3, especially for metrics such as Accuracy, F1 score, and AUPRC (Supplementary Table 6). For instance, the DDI prediction model under the multimodal deep learning framework proposed by Deng et al.⁵² achieved the following results in task 1: 0.8852 (Accuracy), 0.7585 (F1 score), 0.9976 (AUROC), and 0.9208 (AUPRC). In contrast, for tasks 2 & 3, the corresponding values dropped to 0.6415 & 0.4075 (Accuracy), 0.4460 & 0.1590 (F1 score), 0.9799 & 0.9512 (AUROC), and 0.6558 & 0.3635 (AUPRC), respectively. These results indicate that there is still room for improvement in model generalisability to better align with real-world DDI prediction scenarios.

Performance on the negative sample strategy

In DDI prediction studies, experimentally unverified interactions are often mixed with pairs that do not have any interaction, which makes the labelling of negatives particularly challenging. Most included studies (48 out of 59) either did not specify their definition of negative samples or treated drug pairs without known DDI annotations as negative samples through random sampling, and in some cases did not report the sampling method at all (Supplementary Table 6).

Among studies that took extra steps to improve the quality of negative samples, one study⁵⁰ further confirmed the negative samples using literature evidence. Others defined negative samples by evaluating DDI confidence^49,80 or calculated the Average Euclidean Distance (AED)⁵⁸ among drug pairs and selected drug pairs with large distances as negative. Two studies combined random selection with the previously mentioned AED⁵⁸ or with autoencoder-based error screening⁶⁵. Other studies selected negative samples based on frequency⁶⁴, corrupted sampling^43,74, under-sampling methods based on positive-unlabelled learning⁴⁹, ensemble methods and class weight balance⁶³, drug similarity⁸¹, or interaction type replacement⁸².

The performance in Huang et al.’s study⁴⁹ using random sampling shows the Accuracy, AUROC, AUPRC, and F1 score as 0.9101, 0.9991, 0.9586, and 0.8345, respectively. These metrics increased to 0.9676, 0.9997, 0.9892, and 0.8673 with positive-unlabelled learning sampling. Additionally, two studies^64,81 found that although model performance slightly decreased with novel negative sampling strategies, the model maintained good prediction ability.

Novel predicted DDIs

We extracted the anticancer drugs from model predictions and compiled a list of newly predicted DDIs involving these agents, cross-referencing them with up-to-date DDI databases (DrugBank and Drugs.com). A total of 96 newly predicted DDIs involving anticancer drugs were identified in these studies, although these have not been corroborated with the public drug database in the original studies. This involves 112 drugs (42 anticancer and 70 non-anticancer drugs) and includes the biologics trastuzumab and denileukin diftitox, while others were small-molecule drugs (Supplementary Table 7).

Among all newly identified DDIs, 22 interactions were subsequently confirmed using the drug interaction checker of DrugBank (https://go.drugbank.com/drug-interaction-checker) or Drugs.com (https://www.drugs.com/drug_interactions.html) as of March 2026, of which 10 were classified as major in severity (Table 1). As pharmacovigilance data continue to accumulate over time, additional predicted interactions may be confirmed, further demonstrating the predictive capability of AI‑based DDI models.

Table 1.

Newly predicted anti-cancer drugs with DDI confirmed by DrugBank and Drugs.com

Author & Year	Anticancer drug	Interacting drug	Is the interacting drug an anticancer agent?	Evidence source	Interaction severity
Schwarz et al., 2021111	Erlotinib	Epirubicin	Yes	Drugs.com	Moderate
Chen et al., 2021112	Neratinib	Quinidine	No	DrugBank	Major
Ren et al., 202277	Cisplatin	Sparfloxacin	No	Drugs.com	Minor
	Cisplatin	Pitavastatin	No	Drugs.com	Moderate
Ren et al., 202283	Regorafenib	Paclitaxel	Yes	DrugBank	Major
Masumshah et al., 202378	Tamoxifen	Orphenadrine	No	DrugBank	Major
Liang et al., 202444	Ibrutinib	Fidaxomicin	No	Drugs.com	Moderate
	Ibrutinib	Phenylbutazone	No	DrugBank	Moderate
	Ibrutinib	Rimegepant	No	Drugs.com	Moderate
	Ibrutinib	Tazemetostat	Yes	Drugs.com	Major
Nejati et al., 202487	Rucaparib	Prasterone	No	DrugBank	Moderate
Kaliappan et al., 2024113	Abemaciclib	Mitotane	Yes	DrugBank	Major
	Abemaciclib	Lumacaftor	No	DrugBank	Moderate
Xu et al., 202450	Acalabrutinib	Selumetinib	Yes	DrugBank	Minor
	Acalabrutinib	Crizotinib	Yes	DrugBank	Moderate
	Acalabrutinib	Doxorubicin	Yes	DrugBank	Moderate
	Acalabrutinib	Lapatinib	Yes	DrugBank	Major
	Acalabrutinib	Olaparib	Yes	DrugBank	Major
	Acalabrutinib	Palbociclib	Yes	DrugBank	Major
	Acalabrutinib	Vandetanib	Yes	DrugBank	Major
	Acalabrutinib	Venetoclax	Yes	DrugBank	Major
Zhao et al., 202559	Amsacrine	Warfarin	No	DrugBank	Minor

Interaction confirmed according to DrugBank (https://go.drugbank.com/drug-interaction-checker) or Drugs.com (https://www.drugs.com/drug_interactions.html) as of March 2026.

Figure 4 visualised the DDI relationships among these drugs. In the network, nodes represent anticancer drugs and non-anticancer drugs using distinct colours. Confirmed interactions were presented as solid edges, with edge thickness indicating interaction severity (minor, moderate, or major). Among the remaining 74 DDIs, 5 were confirmed by evidence found in the DrugBank drug homepages (but not by the DDI checker); these are displayed as solid edges and labelled with a severity of “NA” (not available, as the DDI checker did not confirm them). The other 69 unconfirmed DDIs are shown as dotted edges, indicating that further validation through pre‑clinical studies, clinical investigations, or post‑marketing surveillance is required.

Fig. 4. Novel predicted DDIs of anticancer drugs by reviewed studies.

Interactions checked according to DrugBank (https://go.drugbank.com/drug-interaction-checker) or Drugs.com (https://www.drugs.com/drug_interactions.html) as of March 2026. Confirmed interactions were presented as solid edges, with edge thickness indicating interaction severity (minor, moderate, or major); interactions confirmed by evidence found in the DrugBank drug homepages (but not by the DDI checker) are displayed as solid edges and labelled with a severity of “NA”; unconfirmed interactions are shown as dotted edges.

Notably, the DDI between Regorafenib and Cisplatin was predicted by two models^77,83, indicating a higher possibility of DDI between these two drugs. Both Regorafenib (tyrosine kinase inhibitor for treatment of advanced colon, liver, and gastrointestinal cancer) and Cisplatin (platinum-based chemotherapy for treatment of various cancers, including sarcomas, carcinomas, germ cell tumours, and lymphomas) were both reported to cause hepatic toxicity^84–86. Therefore, the co-administration of these two drugs may increase hepatic toxicity. To illustrate the potential for unconfirmed DDIs, we further analysed one additional predicted drug pair, Niraparib and Phenytoin, which was predicted by the graph attention-based deep neural network model⁸⁷. Niraparib, a Poly (ADP‑ribose) polymerase (PARP) inhibitor indicated for ovarian cancer⁸⁸, is primarily metabolised by carboxylesterases and to a lesser extent by CYP3A4⁸⁹. Phenytoin, used for treatment of seizures which are common in cancer patients, is a strong CYP3A4 inducer^90–92. Consequently, co-administration of Phenytoin could increase the CYP3A4 metabolisation of Niraparib, potentially reducing its serum concentration and diminishing therapeutic efficacy.

Discussion

This study systematically reviewed the ML and DL-based DDI prediction models involving anticancer drugs. Most existing models focus on predicting DDI existence or DDI types, such as interaction-associated adverse effects. This, thereby, addresses the clinical need for early identification of potentially harmful anticancer drug combinations, particularly in patients exposed to complex multi‑drug regimens. Overall, the included studies showed high reporting quality, with deep learning approaches dominating and increasingly enhanced by advanced techniques, such as GNNs, attention mechanisms, multimodal learning, transformers, and autoencoders (Supplementary Table 2). DrugBank was the most commonly used data source, alongside self-compiled datasets and shared code, supporting model reproducibility (Supplementary Table 4).

Given their generally strong evaluation performance, these models can benefit clinicians and pharmacists by supporting safer prescribing decisions and helping to identify high‑risk drug combinations⁹³, ultimately improving patient safety. As a demonstration of potential clinical value, we provide a list of newly predicted DDIs involving anticancer drugs that had not yet been clinically validated at the time of model construction, illustrating the usefulness of proactively identifying potential DDIs in advance (Fig. 4). In practice, the model outputs can be integrated into clinical decision support systems⁹⁴ to generate interaction alerts, guide drug selection, suggest potential alternative therapies, or prioritise high‑risk DDIs for closer monitoring.

Despite the promising performance achieved by current ML or DL-based DDI prediction models, challenges remain in translating accurate predictions into real-world practice, highlighting the need for future methodological and translational research exploration.

First of all, almost all current input drug features focus on the properties of the drug itself. However, to enhance the accuracy of the DDI prediction, additional factors should be considered, such as the drug administration route and patient characteristics, especially for cancer patients. Older cancer patients are at high risk of ADRs from DDIs because of the greater frequency of polypharmacy compared with younger patients, alongside age-related changes in liver and kidney function^95,96. The drug administration route can impact the PK and PD of the co-administered drugs, and therefore potentially influence their effect. Besides the above-mentioned factors, demographic and genetic factors such as age, gender, and gene polymorphism have also been demonstrated to be significantly associated with drug–drug interactions^12,13,22,97. Furthermore, gene polymorphism is associated with CYP450 enzyme, which can impact DDIs related to metabolism by CYP450 enzymes.

Real-world routine data, such as Electronic Health Records (EHRs) containing patient information and drug administration details, can be used to train DDI prediction models. Wolff et al. developed a machine learning model that used features such as patient age, gender, somatic and psychiatric diagnoses, treatment setting, and functional and disease scores documented in the electronic medical records to predict DDI risk in psychiatric hospitals. Jeon et al. developed a machine learning model using data from laboratory measurements, prescriptions and patient information (including age, gender, race, and disease diagnose codes) recorded in EHRs to predict the physiological effects of two or more drugs. However, as both studies only considered real-world routine data, combining this data with drug features needs further exploration.

Secondly, existing studies commonly employ 5- or 10-fold cross-validation to validate the performance of the DDI model. While DDI prediction performance on the known drug pairs is generally strong, this tends to overestimate real-world generalisation as newly introduced drugs in practice have no prior pairs in the training data. For prediction tasks involving drug pairs with drugs outside the training set, there is still considerable room for improvement in model performance, particularly regarding external evaluation.

In addition to combining various drug features with real-world routine data, collecting reliable negative samples should also be considered to minimise mislabelling of experimentally unverified drug pairs as negatives and enhance model generalisability. Several approaches have been proposed to enhance negative-sample reliability, including using confidence scores from existing resources⁵² and selecting negatives based on distance derived from molecular representations⁵⁸. However, this strategy may fail to account for important pharmacological factors when calculating distance. For example, a drug pair that is distant from the cluster of known interacting pairs in molecular fingerprint space (i.e., the ‘positive sample centre’) may still exhibit a DDI if both drugs act on the same biological target, as illustrated by Nifedipine and Verapamil⁹⁸. This underscores that chemical fingerprints alone may fail to capture target-mediated interactions. Improving the accuracy of negative-sample construction may therefore require complementary strategies, such as cross-checking labels against public databases⁵⁰ or incorporating expert pharmacological assessment to validate non-interacting pairs.

Thirdly, most current DDI prediction models focus on small molecules, as evidenced by the systematic review finding that 110 out of 112 drugs in newly identified DDIs involving anti-cancer drugs are associated with small molecule agents. However, DDIs resulting from the co-administration of biologics with other biologics or with small molecules should also be considered, given the increasing trend in biologics approval⁹⁹. Since the first approval of interferon (IFN)-α, biologics have shown effectiveness in the treatment of multiple cancers^100–102. Unlike small molecules, biologics are mainly administered parenterally, and their metabolism and elimination primarily occur through lysosomal degradation, neonatal Fc receptor (FcRn)-mediated recycling, and target-mediated elimination, among others¹⁰³. Therefore, biologics are not considered to directly impact the PK processes of small molecules, but may do so indirectly¹⁰⁴. The most well-documented PK mechanism of DDIs involving biologics is the impact of cytokine modulators on CYP450 enzymes and transporters³². For example, Tocilizumab, an inhibitor of interleukin (IL)-6, which suppresses CYP3A4 activity, can increase the CYP3A4 metabolism of Simvastatin¹⁰⁵. In addition, adverse PD DDIs between two biologics with the same targets are also possible.

Among the reviewed studies, limited models consider amino acid sequences when predicting DDI. Huang et al.⁴⁹ developed a DL-based model that utilised structure (molecular structure for small molecules, and amino acid sequence for biologics) as well as network topology features of drugs. Ru et al. developed a dual-modal graph learning framework that used SMILES strings to represent small molecules and amino acid sequence-based residue contact maps for biologics.

However, several challenges remain in developing DDI prediction models for biologics. One is the limited number of positive DDI pairs, as those involving biologics occur less frequently than those involving small molecules³¹. Another challenge is that the complex structure of biologics makes it difficult to accurately characterise the small molecule–biologic drug pairs⁴⁹. Despite challenges such as cytokine release or stimulation impacting pharmacological function, biologics with their higher-order structure can also be incorporated into models.

The review has certain limitations that should be acknowledged. By design, it focused on studies employing ML and DL approaches for DDI prediction involving cancer drugs, thereby excluding traditional statistical, rule-based, and mechanistic approaches. As a result, the review may underrepresent the broader landscape of anticancer drug interaction research. Nevertheless, the primary aim of this work is to critically examine contemporary AI‑based DDI prediction models. By identifying current gaps and challenges, this review seeks to inform future methodological advances and enhance the clinical applicability of DDI prediction models in real‑world decision‑making contexts.

To conclude, the application of machine learning, deep learning, and knowledge graph-based AI techniques in DDI prediction models involving anticancer drugs is on the rise, particularly with the combination of deep learning technologies with knowledge graphs. Although significant progress has been achieved in improving DDI prediction accuracy, including algorithm combinations, feature integration, and new strategies for negative sample collection, further efforts are necessary. These could include incorporating drug features, administration routes, and patient characteristics, as well as involving pharmacological experts during DDI prediction construction, to help translate accurate prediction outputs into real-world practice. Furthermore, there is a pressing need to develop more DDI prediction models, especially for biologics.

Methods

This study conducts a systematic review aimed at evaluating the applications of machine learning or deep learning for DDI prediction, with a focus on anticancer drugs. This review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines¹⁰⁶, and the PRISMA checklist was utilised to ensure comprehensive and transparent reporting¹⁰⁷ (Supplementary Table 8).

Search strategy and eligibility criteria

Following PRISMA guidelines¹⁰⁶, a systematic review was performed across multiple databases, including PubMed (https://pubmed.ncbi.nlm.nih.gov/), Embase (https://www.embase.com/), Cochrane Library (https://www.cochranelibrary.com/), Web of Science (https://www.webofscience.com/wos/woscc/smart-search), and IEEE Xplore databases (https://ieeexplore.ieee.org/). The search strategy employed the keywords “machine learning”, “deep learning”, and “drug–drug interaction” within the title and abstract without publication date restrictions. The initial search was completed on 18 November 2024 and updated to 18 March 2025. Both Medical Subject Headings (MeSH) terms and free-text words were used, with the search terms combined using Boolean operators (AND/OR).

The detailed inclusion and exclusion criteria were shown in Table 2. To ensure comprehensive coverage of DDI prediction models, the keywords “anticancer” or “cancer” were not included in the search strategy but were used as exclusion criteria. Instead, we manually screened all retrieved studies to identify whether anticancer or cancer drugs were involved and documented where applicable. This approach was adopted because the general DDI prediction model can also be applied to the anticancer drug, and many studies involve detailed anticancer drugs in the case study or application demonstration without explicitly mentioning the terms “anticancer” or “cancer”. For example, an anticancer drug such as Cisplatin may be used in a case study for a machine learning or deep learning-based DDI prediction model, which does not include the words “anticancer” or “cancer” in the literature.

Table 2.

Inclusion and exclusion criteria for study selection

	Inclusion criteria:	Exclusion criteria:
Study design	Studies employing machine learning and/or deep learning based drug–drug interaction prediction	Studies not using machine learning or deep learning technologies for drug–drug interaction prediction
Application	Studies involving anticancer drugs	Studies not involving anticancer drugs
Publication type	Original research articles	Review articles, conference abstracts, editorials, or comments
Language	English publications	Non-English publications
Availability	NA	Studies without accessible full text

Quality assessment

The quality of the included studies was assessed using the CAIR checklist⁴⁵. Each item was scored 1 if the relevant content was reported (“Yes”), and 0 if it was not reported (“No”). Items 4 (participant inclusion/exclusion criteria), 6 (handling of input data by human or AI operators), and 7 (management of missing or low-quality data) were not applicable to the DDI prediction studies and marked as “NA”.

Data extraction and synthesis

Data were extracted based on a structured set of criteria, including information on bibliography, algorithm, data source, number of drugs and DDIs, features, prediction objects, evaluation methods, performance, negative sampling methods, and anticancer drug information. In addition, to provide healthcare professionals with a list of potential DDIs involving anticancer drugs, we separately extracted DDIs of anticancer drugs newly predicted by the included models with high probabilities (i.e., top predicted DDIs highlighted in the original studies) that have not been previously confirmed by public data sources (e.g., DrugBank, https://go.drugbank.com/; Drugs.com, https://www.drugs.com/; or published literature). These potential DDIs were further checked against the DrugBank and Drugs.com databases using the Drug Interaction Checker in May 2025. The included models were also categorised according to their underlying algorithms.

Two independent researchers (Yingying Zhao and Jiaqi Wang) performed literature screening and quality assessment. Data were extracted by one researcher (Yingying Zhao) and cross-checked by another researcher (Jiaqi Wang). All discrepancies were resolved in consensus meetings.

Algorithm categorisation

In this article, we broadly classify these models as: (1) Machine Learning (ML), such as logistic regression, Support Vector Machine (SVM), and ensemble methods (e.g. XGBoost), which mainly uses handcrafted features and domain knowledge^108,109; (2) Deep Learning (DL), which uses multiple processing layers to learn representations of data at varying levels of abstraction³⁷, such as Convolutional Neural Network (CNN), Deep Neural Network (DNN), Graph Convolutional Network (GCN)^37,110; (3) Hybrid models of Knowledge Graph (KG) and ML or DL. A KG is a multi-relational graph where nodes represent entities (e.g. drugs, enzymes, targets, transporters) and edges define their relations³⁷. These hybrid models integrate KG-based representations with ML or DL techniques to leverage both structured biomedical knowledge and advanced predictive algorithms.

Author contributions

Yingying Zhao: Formal analysis, investigation, methodology, visualization, writing—original draft; Jiaqi Wang: Methodology, validation, writing—review and editing; Jiyeong Kim: Writing—review and editing; Fatima Rodriguez: Writing—review and editing; Eleni Linos: Writing—review and editing; Rong Na: Writing—review and editing; Khuloud T. Al-Jamal: Writing—review and editing; Xue Li: Conceptualization, methodology, supervision, writing—review and editing.

Data availability

All data generated or analysed during this study are included in this published article and its supplementary materials. Additional datasets referenced, such as DrugBank (https://go.drugbank.com/), Drugs.com (https://www.drugs.com/), are publicly accessible through their respective platforms.

Competing interests

F.R. reports consulting fees from Novartis, Esperion Therapeutics, Edwards, Arrowhead Pharmaceuticals, HeartFlow, iRhythm, Amgen, and Cleerly Health outside the submitted work. X.L. received research grants or contracts from the Health and Medical Research Fund (HMRF) Main Scheme and HMRF Fellowship Scheme of the Hong Kong Special Administrative Region(SAR), China, and from the Research Grants Council Early Career Scheme, Research Impact Fund (Hong Kong SAR); is also the former nonexecutive director of ADAMS Hong Kong SAR; received commission grants from Department of Health, Government of Hong Kong SAR, China; commission grants from Hospital Authority of Hong Kong SAR, China, internal funding from the University of Hong Kong, Hong Kong SAR, China, and research or education grants from Pfizer, Janssen, Bristol Myers Squibb (BMS) and Novartis; received consultancy fees from Merck Sharp & Dohme, Pfizer, Open Health and The Office of Health Economics ; and received honoraria for associate editorship from Springer Nature. None are related to the work reported in the current manuscript.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41698-026-01512-z.