Motivation Computational approaches for predicting drug-target interactions (DTIs) can provide valuable insights into the drug mechanism of action. DTI predictions can help to quickly identify new promising (on-target) or unintended (off-target) effects of drugs. However, existing models face several challenges. Many can only process a limited number of drugs and/or have poor proteome coverage. The current approaches also often suffer from high false positive prediction rates. Results We propose a novel computational approach for predicting drug target proteins. The approach is based on formulating the problem as a link prediction in knowledge graphs (robust, machine-readable representations of networked knowledge). We use biomedical knowledge bases to create a knowledge graph of entities connected to both drugs and their potential targets. We propose a specific knowledge graph embedding model, TriModel, to learn vector representaions (i.e. embeddings) for all drugs and targets in the created knowledge graph. These representations are consequently used to infer candidate drug target interactions based on their scores computed by the trained TriModel model. We have experimentally evaluated our method using computer simulations and compared it to five existing models. This has shown that our approach outperforms all previous ones in terms of both area under ROC and precision-recall curves in standard benchmark tests. Availability The data, predictions, and models are available at: drugtargets.insight-centre.org

Complex biological systems are traditionally modelled as graphs of interconnected biological entities. These graphs, i.e. biological knowledge graphs, are then processed using graph exploratory approaches to perform different types of analytical and predictive tasks. Despite the high predictive accuracy of these approaches, they have limited scalability due to their dependency on time-consuming path exploratory procedures. In recent years, owing to the rapid advances of computational technologies, new approaches for modelling graphs and mining them with high accuracy and scalability have emerged. These approaches, i.e. knowledge graph embedding (KGE) models, operate by learning low-rank vector representations of graph nodes and edges that preserve the graph’s inherent structure. These approaches were used to analyse knowledge graphs from different domains where they showed superior performance and accuracy compared to previous graph exploratory approaches. In this work, we study this class of models in the context of biological knowledge graphs and their different applications. We then show how KGE models can be a natural fit for representing complex biological knowledge modelled as graphs. We also discuss their predictive and analytical capabilities in different biology applications. In this regard, we present two example case studies that demonstrate the capabilities of KGE models: prediction of drug–target interactions and polypharmacy side effects. Finally, we analyse different practical considerations for KGEs, and we discuss possible opportunities and challenges related to adopting them for modelling biological systems.

Phosphorylation of specific substrates by protein kinases is a key control mechanism for vital cell-fate decisions and other cellular processes. However, discovering specific kinase-substrate relationships is time-consuming and often rather serendipitous. Computational predictions alleviate these challenges, but the current approaches suffer from limitations like restricted kinome coverage and inaccuracy. They also typically utilise only local features without reflecting broader interaction context. To address these limitations, we have developed an alternative predictive model. It uses statistical relational learning on top of phosphorylation networks interpreted as knowledge graphs, a simple yet robust model for representing networked knowledge. Compared to a representative selection of six existing systems, our model has the highest kinome coverage and produces biologically valid high-confidence predictions not possible with the other tools. Specifically, we have experimentally validated predictions of previously unknown phosphorylations by the LATS1, AKT1, PKA and MST2 kinases in human. Thus, our tool is useful for focusing phosphoproteomic experiments, and facilitates the discovery of new phosphorylation reactions. Our model can be accessed publicly via an easy-to-use web interface (LinkPhinder).

Knowledge graph embedding (KGE) models have become popular means for making discoveries in knowledge graphs (e.g., RDF graphs) in an efficient and scalable manner. The key to success of these models is their ability to learn low-rank vector representations for knowledge graph entities and relations. Despite the rapid development of KGE models, state-of-the-art approaches have mostly focused on new ways to represent embeddings interaction functions (i.e., scoring functions). In this paper, we argue that the choice of other training components such as the loss function, hyperparameters and negative sampling strategies can also have substantial impact on the model efficiency. This area has been rather neglected by previous works so far and our contribution is towards closing this gap by a thorough analysis of possible choices of training loss functions, hyperparameters and negative sampling techniques. We finally investigate the effects of specific choices on the scalability and accuracy of knowledge graph embedding models.

Phosphorylation of specific substrates by protein kinases is a key control mechanism for vital cell-fate decisions and other cellular processes. However, discovering specific kinase-substrate relationships is time-consuming and often rather serendipitous. Computational predictions alleviate these challenges, but the current approaches suffer from limitations like restricted kinome coverage and inaccuracy. They also typically utilise only local features without reflecting broader interaction context. To address these limitations, we have developed an alternative predictive model. It uses statistical relational learning on top of phosphorylation networks interpreted as knowledge graphs, a simple yet robust model for representing networked knowledge. Compared to a representative selection of six existing systems, our model has the highest kinome coverage and produces biologically valid highconfidence predictions not possible with the other tools. Specifically, we have experimentally validated predictions of previously unknown phosphorylations by the LATS1, AKT1, PKA and MST2 kinases in human. Thus, our tool is useful for focusing phosphoproteomic experiments, and facilitates the discovery of new phosphorylation reactions.Our model can be accessed publicly via an easy-to-use web interface (LinkPhinder).

Knowledge graphs became a popular means for modelling complex biological systems where they model the interactions between biological entities and their effects on the biological system. They also provide support for relational learning models which are known to provide highly scalable and accurate predictions of associations between biological entities. Despite the success of the combination of biological knowledge graph and relation learning models in biological predictive tasks, there is a lack of unified biological knowledge graph resources. This forced all current efforts and studies for applying a relational learning model on biological data to compile and build biological knowledge graphs from open biological databases. This process is often performed inconsistently across such efforts, especially in terms of choosing the original resources, aligning identifiers of the different databases and assessing the quality of included data. To make relational learning on biomedical data more standardised and reproducible, we propose a new biological knowledge graph which provides a compilation of curated relational data from open biological databases in a unified format with common, interlinked identifiers. We also provide a new module for mapping identifiers and labels from different databases which can be used to align our knowledge graph with biological data from other heterogeneous sources. Finally, to illustrate practical relevance of our work, we provide a set of benchmarks based on the presented data that can be used to train and assess the relational learning models in various tasks related to pathway and drug discovery.

The field of drug discovery has entered a plateau stage lately. It is increasingly more expensive and time-demanding to introduce new drugs into the market. One of the main reasons is the slow progress in finding novel targets for drug candidates and the lack of insight in terms of the associated mechanisms of action. Current works in this area mainly utilise different chemical, genetic and proteomic methods, which are limited in terms of the scalability of experimentation and the scope of studied drugs and targets per experiment. This is mainly due to their dependency on laboratory experiments and available physical resource. This has led to an increasing importance of computational methods for the identification of candidate drug targets. In this work, we introduce a novel computational approach for predicting drug target proteins. We approach the problem as a link prediction task on knowledge graphs. We process drug and target information as a knowledge graph of interconnected drugs, proteins, disease, pathways and other relevant entities. We then apply knowledge graph embedding (KGE) models over this data to enable scoring drug-target associations, where we employ a customised version of state-of-the-art KGE model ComplEx. We generate a benchmarking dataset based on KEGG database to train and evaluate our method. Our experiments show that our method achieves best results in comparison to other traditional KGE models. Specifically, the method predicts drug target links with mean reciprocal rank (MRR) of 0.78 and Hits@10 of 0.88. This provides a promising basis for further experimentation and comparisons with domain-specific predictive models.

PURPOSE Stratifying patients with cancer according to risk of relapse can personalize their care. In this work, we provide an answer to the following research question: How to use machine learning to estimate probability of relapse in patients with early-stage non–small-cell lung cancer (NSCLC)? MATERIALS AND METHODS For predicting relapse in 1,387 patients with early-stage (I-II) NSCLC from the Spanish Lung Cancer Group data (average age 65.7 years, female 24.8%, male 75.2%), we train tabular and graph machine learning models. We generate automatic explanations for the predictions of such models. For models trained on tabular data, we adopt SHapley Additive exPlanations local explanations to gauge how each patient feature contributes to the predicted outcome. We explain graph machine learning predictions with an example-based method that highlights influential past patients. RESULTS Machine learning models trained on tabular data exhibit a 76% accuracy for the random forest model at predicting relapse evaluated with a 10-fold cross-validation (the model was trained 10 times with different independent sets of patients in test, train, and validation sets, and the reported metrics are averaged over these 10 test sets). Graph machine learning reaches 68% accuracy over a held-out test set of 200 patients, calibrated on a held-out set of 100 patients. CONCLUSION Our results show that machine learning models trained on tabular and graph data can enable objective, personalized, and reproducible prediction of relapse and, therefore, disease outcome in patients with early-stage NSCLC. With further prospective and multisite validation, and additional radiological and molecular data, this prognostic model could potentially serve as a predictive decision support tool for deciding the use of adjuvant treatments in early-stage lung cancer.