Source-linked AI summary
Drug-Drug Interaction Prediction Based on Knowledge Graph Embeddings and Convolutional-LSTM Network
Md. Rezaul Karim, Michael Cochez, Joao Bosco Jares, Mamtaz Uddin, Oya Beyan, Stefan Decker
TL;DR
DDI prediction is difficult because available drug knowledge is limited and existing approaches often rely on single data sources. The paper integrates multiple biomedical sources with knowledge graphs, embeds the graph, and combines Conv-LSTM with machine-learning classifiers, achieving strong cross-validated results while lacking explanations for predicted interactions.
Problem
DDI prediction is important for reducing medical injuries, but available drug knowledge is limited and many existing approaches rely on a single information source.
Method
The paper integrates drug-related data into a knowledge graph, applies graph embeddings, and evaluates Conv-LSTM and classic machine-learning prediction models.
Results
Up to 0.94 AUPR, 0.92 F1-score, and 0.80 MCC were obtained during 5-fold cross-validation tests.
Takeaways & Limitations
The proposed architecture consistently outperformed baseline classifiers and earlier state-of-the-art models, while PBG produced the best embedding results among those evaluated.
Takeaways & Limitations
The latent features produced by the embeddings do not provide explanations for predicted DDIs, making the approach a black-box model.
Abstract
from arXiv · showhide
Interference between pharmacological substances can cause serious medical injuries. Correctly predicting so-called drug-drug interactions (DDI) does not only reduce these cases but can also result in a reduction of drug development cost. Presently, most drug-related knowledge is the result of clinical evaluations and post-marketing surveillance; resulting in a limited amount of information. Existing data-driven prediction approaches for DDIs typically rely on a single source of information, while using information from multiple sources would help improve predictions. Machine learning (ML) techniques are used, but the techniques are often unable to deal with skewness in the data. Hence, we propose a new ML approach for predicting DDIs based on multiple data sources. For this task, we use 12,000 drug features from DrugBank, PharmGKB, and KEGG drugs, which are integrated using Knowledge Graphs (KGs). To train our prediction model, we first embed the nodes in the graph using various embedding approaches. We found that the best performing combination was a ComplEx embedding method creating using PyTorch-BigGraph (PBG) with a Convolutional-LSTM network and classic machine learning-based prediction models. The model averaging ensemble method of three best classifiers yields up to 0.94, 0.92, 0.80 for AUPR, F1-score, and MCC, respectively during 5-fold cross-validation tests.
1 INTRODUCTION
The paper frames DDI prediction as a medical-safety problem and proposes integrating multiple biomedical data sources with knowledge graphs and machine learning. It evaluates graph embeddings and a Conv-LSTM architecture alongside baseline classifiers.
- DDIs can cause preventable medical injuries and are associated with adverse drug reactions and increased healthcare costs.
- Deep learning is underexplored for DDI prediction, while CNNs and LSTMs can capture complementary local and sequential drug features.
- Knowledge graphs integrate drugs with diseases, protein targets, substructures, side effects, and pathways before embedding entities into dense feature vectors.
- The study combines DDI information from DrugBank, KEGG, TWOSIDES, and scientific literature with an integrated graph using DrugBank, KEGG drug, PharmGKB, and OFFSIDES.
- The dataset contains 2,898,937 DDI pairs, and the integrated knowledge graph contains 1.2 billion triples.
- The paper evaluates embedding techniques and machine-learning models, reporting Conv-LSTM as the highest-accuracy approach among those tested.
2 RELATED WORK
Prior DDI research uses diverse similarity, text-mining, and knowledge-graph approaches, but biomedical knowledge graphs can remain incomplete or inaccurate. Translation-based embeddings also have limited capacity for modeling complex relation types, motivating broader embedding and data-integration strategies.
- Earlier DDI prediction used in vitro and in vivo experiments, similarity features, statistical models, text mining, and graph-based methods.
- Biomedical knowledge graphs support large-scale integration of drug-related facts, but incomplete and inaccurate data can impede safe-medicine applications.
- Graph embeddings are used to address data incompleteness and sparsity, with learned representations commonly applied to DDI prediction.
- Translation-based embeddings model triples through a relation-dependent transformation from head entity h to tail entity t.
- Tiresias integrates drug-related data into a knowledge graph, computes drug similarities, and uses logistic regression to predict DDIs.
- Most translation-based embeddings have limited capacity to represent symmetric, transitive, one-many, many-to-one, and many-many relations.
3 MATERIALS AND METHODS
The methods section covers problem formulation, data collection and integration, knowledge-graph embeddings, Conv-LSTM construction, training with hyperparameter optimization, and DDI inference. Figure 1 presents the workflow of the proposed approach.
- The proposed workflow includes problem formulation, data collection and integration, KG embeddings, Conv-LSTM construction, network training, and DDI inference.
- Training includes hyperparameter optimization before the model performs DDI predictions.
3.1 Problem formulation
The paper formulates DDI prediction as link prediction on a directed drug-interaction graph. Its binary matrix marks observed interactions, while zero can also represent an interaction that has not yet been discovered.
- DDI prediction is treated as a link-prediction problem on a directed graph whose nodes are drugs and whose edges represent potential interactions.
- The interaction matrix Y is binary, with y_u,v equal to 1 when an interaction between drugs u and v exists and 0 otherwise.
- A zero matrix entry does not prove that no interaction exists because the interaction may not yet have been discovered.
3.2 DDIs extraction and KG construction
The study combines multiple drug and interaction sources to construct a DDI dataset and an integrated knowledge graph, while excluding explicit DDI relations from the background graph. Sources are mapped and consolidated to address incomplete and distributed DDI information.
- DrugBank, KEGG drug, and PharmGKB provide drug and drug-target data for the integrated knowledge graph.
- OFFSIDES, TWOSIDES, and MEDLINE literature provide DDI evidence for extracting interaction pairs.
- 2,641,889 pairwise DDIs and 2,630,796 unique DDIs were extracted from DrugBank across 12,112 drugs.
- 63,473 distinct pairwise DDIs between 645 drugs were obtained from literature-based mapping of TWOSIDES interactions.
- The integrated knowledge graph excludes explicit drug-drug interaction relations and represents knowledge as subject–predicate–object triples.
3.3 Knowledge graph embeddings
The authors encode integrated-graph information as dense entity and relation representations using several knowledge-graph embedding methods, including random-walk, translation-based, crossover, and global-pattern approaches. ComplEx embeddings from PBG are additionally trained for scalable DDI prediction.
- KG embedding converts graph information into dense vectors suitable for fixed-length machine-learning inputs.
- RDF2Vec generates edge-labeled random walks and learns node embeddings with skip-gram word2vec using fixed-size context windows.
- TransE represents relations as translations between head and tail entities, while CrossE models crossover interactions.
- KGloVe creates embeddings from normalized co-occurrence matrices built using personalized PageRank over original and reversed graph edges.
- PBG trains ComplEx embeddings to improve speed, scalability, and parallelization for large graphs before DDI prediction.
3.4 Network construction
The proposed Conv-LSTM combines convolutional filtering with recurrent modeling to capture local and overall relationships in drug features. Embedded drug sequences pass through convolutional, pooling, recurrent, and fully connected components for DDI prediction.
- The Conv-LSTM combines CNN layers for local feature relationships with LSTM layers for overall relationships.
- Each input, cell output, hidden state, and gate is represented as a 2D tensor over spatial drug-feature dimensions.
- The embedding layer maps a drug sequence into a real vector domain, producing a 100x300 representation for convolution.
- The recurrent output is flattened and passed through fully connected layers to predict DDIs and provide input at the next time step.
- Convolutional outputs are regularized with dropout, transformed into a 100x100 representation, and down-sampled by max pooling to 25x100.
3.5 Network training
Training addresses class imbalance in unknown interactions and compares multiple embedding and machine-learning configurations. Hyperparameters are selected through random search and cross-validation using held-out evaluation and validation splits.
- Treating all unknown interactions as negative samples creates imbalance that influences AUPR and F1-score.
- PBG, CrossE, TransE, SimpleIE, KGloVe, and RDF2Vec embeddings are evaluated with 300-dimensional feature vectors and varying negative samples.
- Adam, AdaGrad, RMSprop, and AdaMax are tested with varying learning rates and batch sizes to optimize binary cross-entropy loss.
- Random-search hyperparameter optimization uses cross-validation with batch size 128, 70% training data, 30% network evaluation, and 10% of training data for validation.
4 EXPERIMENTS
The experiments compare classifiers and graph embeddings for DDI prediction, finding strong performance from Conv-LSTM and ComplEx-based features. Performance also varies with training-set size, negative-sample ratio, and evaluation setting.
- 0.93 AUPR was the best Conv-LSTM result across the evaluated embedding methods, outperforming the baseline models.
- 1.5% higher F1-score was achieved by averaging GBT, RF, and Conv-LSTM than by the best Conv-LSTM model alone.
- 0.70 Pearson correlation was obtained between predictions and ground truth, while Conv-LSTM achieved the highest AUC, at least 3% above RF.
- ComplEx features produced consistently strong F1, MCC, and AUPR results for GBT, RF, and Conv-LSTM, whereas RDF2Vec produced the worst prediction accuracy.
- 0.82 to 0.93 probability values were produced by Conv-LSTM for predictions generated from PBG embeddings.
- Adding more training samples increased generalization for RF, GBT, and Conv-LSTM, while SVM scores converged to a low value.
- 1% AUPR improvement occurred with 20 negative triples, but increasing the number to 25 caused an AUPR decrease.
- Up to 0.94 AUPR, 0.92 F1-score, and 0.80 MCC were obtained in 5-fold cross-validation, comparable to current state-of-the-art methods.
5 CONCLUSION AND OUTLOOK
The paper integrates drug-related knowledge from multiple sources into a knowledge graph, embeds its nodes, and uses classifiers including a convolutional-LSTM to predict DDIs. The approach outperformed baseline and earlier state-of-the-art models, while its latent embeddings cannot explain individual predictions.
- 5 CONCLUSION AND OUTLOOK: Knowledge graphs integrate background information about drugs, diseases, pathways, proteins, enzymes, and chemical structures for DDI prediction.Node embeddings convert this background knowledge into dense vectors suitable for traditional classifiers and a convolutional-LSTM network.
- 5 CONCLUSION AND OUTLOOK: The proposed architecture consistently outperformed baseline classifiers and earlier state-of-the-art models.
- 5 CONCLUSION AND OUTLOOK: 0.94, 0.92, and 0.80 were obtained for AUPR, F1-score, and MCC, respectively, during 5-fold cross-validation.The paper describes these results as showing high confidence in predicting potential DDIs.
- 5 CONCLUSION AND OUTLOOK: PBG performed best among the evaluated embedding models, while SimpleE and KGloVe also produced reasonable results.
- 5 CONCLUSION AND OUTLOOK: The approach cannot provide explanations for predicted DDIs because its embeddings create latent features resembling a black-box model.Future directions include adding data, modeling negative interactions and interaction types, and investigating explanation methods.