Table 1 Classification of methods and models for predicting drug-drug interactions
Taxonomy | Method | Description | Reference |
|---|---|---|---|
Method based on learning paradigms | Position-aware deep multi-task learning | A position-aware deep multi-task learning approach for extracting drug-drug interactions from biomedical texts, combining word and position embeddings with an attention-based BiLSTM network, achieving state-of-the-art performance in both binary DDI classification and interaction type identification. | |
DEML | DEML is an ensemble-based multi-task neural network that optimizes multiple tasks, including synergy prediction and DDI classification, using chemical and transcriptomics data, and achieves superior performance in synergy prediction while alleviating the multi-task learning “seesaw effect” for cancer treatment strategies. | ||
LCM-DS | A local classification-based model (LCM-DS) for predicting potential DDIs of new drugs, combining it with a Dempster-Shafer fusion algorithm to improve training efficiency, memory usage, and prediction performance. | ||
link prediction as a binary classification task on networks | This method uses link prediction techniques, framed as a binary classification task, to predict unknown drug-drug interactions (DDIs) in large-scale databases. | ||
Method based on representation-level | Similarity-based machine learning support vector machine predictor | A machine learning model using support vector machines (SVMs) to predict drug-drug interactions by integrating five similarity measures (2D molecular, 3D pharmacophoric, interaction profile fingerprint, target, and adverse drug effect) from established databases, using a pairwise kernel for SVM input. | |
A Probabilistic Approach for Collective Similarity | A probabilistic method for using Probabilistic Soft Logic to infer unknown drug-drug interactions from a network of drug similarities and known interactions, outperforming existing methods by over 50% in AUPR and identifying five novel interactions validated by external sources. | ||
DDI-IS-SL | DDI-IS-SL is a method for predicting drug-drug interactions by integrating drug chemical, biological, and phenotype similarities with semi-supervised learning. | ||
DDI-SSL | DDI-SSL is a drug-drug interaction prediction method that leverages substructure signature learning, deep clustering, and a collaborative attention mechanism to improve prediction accuracy by aggregating similar substructures and mitigating noise from drug heterogeneity. | ||
SSI–DDI | SSI-DDI is a deep learning framework that directly operates on raw molecular graphs to extract richer features and predicts drug-drug interactions by identifying pairwise interactions between drug substructures. | ||
MSResG | MSResG is a deep learning framework that integrates multi-source drug features with a Graph Auto-Encoder and residual graph convolution networks to predict drug-drug interactions, demonstrating advanced performance and high accuracy. | ||
StructNet-DDI | StructNet-DDI is a deep learning framework using SMILES representations and a modified ResNet18 architecture to extract key molecular features for accurate drug-drug interaction prediction, addressing training challenges and achieving superior performance. | ||
Method based on specific model families | BRSNMF | BRSNMF is a method that predicts drug-drug interactions by detecting drug communities and using drug-binding protein features to handle cold start issues. | |
MRMF | MRMF is a drug-drug interaction prediction method that treats DDI prediction as a matrix completion task, utilizing manifold regularization to incorporate various drug features, improving prediction accuracy. | ||
ISCMF | ISCMF is a method that integrates multiple drug similarity data types using nonlinear fusion and Gaussian interaction profiles to predict drug-drug interactions. | ||
TMFUF | TMFUF is a unified framework based on triple matrix factorization that improves drug-drug interaction prediction by capturing pharmacological changes and enhancing the prediction of enhanced and weakened DDIs, offering support for clinical applications. | ||
GRPMF | GRPMF is a method that predicts drug-drug interactions by incorporating expert knowledge through a graph-based regularization term, enhancing matrix factorization for more accurate predictions. | ||
Wasserstein Adversarial Autoencoder-based knowledge graph embeddings | A new KG embedding framework using adversarial autoencoders (AAEs) with Wasserstein distances and Gumbel-Softmax relaxation for drug-drug interaction (DDI) tasks, achieving improved performance by generating high-quality negative samples and addressing vanishing gradient issues. | ||
Predicting rich DDI information through graph embedding | It uses graph embedding and biomedical text integration to predict drug-drug interactions (DDIs), overcoming data incompleteness and sparsity through a link prediction process. | ||
RaGSEs | RaGSECo is a novel DDI prediction method that combines relation-aware graph structure embedding with co-contrastive learning, using multi-relational DDI and drug–drug similarity graphs to learn and propagate effective drug embeddings for more accurate predictions. | ||
SmileGNN | SmileGNN is a drug–drug interaction prediction model that integrates SMILES-based structural features and knowledge graph topological features through graph neural networks, demonstrating superior prediction performance and credibility. | ||
ACDGNN | ACDGNN is an attention-based cross-domain graph neural network for DDI prediction that integrates drug-related biomedical entities and uses cross-domain transformation to handle entity heterogeneity, achieving superior performance in both transductive and inductive settings. | ||
reverse GNN | Reverse GNN is a model that learns high-quality graphs from the intrinsic space of original data points and addresses out-of-sample extension, improving feature learning and enabling both supervised and semi-supervised learning. | ||
SGFNNs | By introducing signed GNNs to model assortative and disassortative relationships in drug pairs, using two spectral filters on signed graphs and an end-to-end framework with SGFNNs and a discriminators are used to improve DDI prediction. | ||
AutoDDI | AutoDDI is an automated DDI prediction method that leverages reinforcement learning to optimize GNN architectures, adapting the depth of layers and capturing interaction information for improved prediction accuracy. | ||
Method based on modeling strategies and training paradigms | MMADL | MMADL model integrates multi-source drug features and a multi-view, multichannel attention mechanism to enhance drug-drug interaction prediction by adaptively learning the importance of different drug attributes and entity information. | |
MDF-SA-DDI | MDF-SA-DDI is a novel method that predicts drug-drug interactions by fusing multi-source drug and feature data through Siamese networks, convolutional neural networks, and autoencoders, with transformer self-attention for enhanced feature integration and prediction accuracy. | ||
CNN-Siam | CNN-Siam is a novel convolutional neural network based on a Siamese architecture that learns feature representations of drug pairs from multimodal data to predict drug interactions, optimized with RAdam and LookAhead algorithms. |
