Today, I would like to tell you about an article on drug-target interaction published in Briefings in Bioinformatics in March 2023, in which the authors propose a metapath-based layered transformer and transformer method (MHTAN-DTI) for drug-target interaction prediction, which uses the attention mechanism to mainly capture the importance of different metapath types by introducing different weights of different metapath instances to obtain node embeddings with specific semantics. And perform weighted ensemble to obtain the final node embedding representation. Compared with other DTI prediction methods, the prediction performance of MHTAN-DTI has significant advantages. In addition, ablation experiments were performed. The results show that a variety of heterogeneous information is important in predicting DTI and has a certain degree of interpretability, which accelerates the process of drug design.
Background of drug target effect research
Target selection and identification is particularly important in early drug discovery. The prediction of drug-target interactions is important for identifying interactions between novel ligands and specific targets, multi-pharmacological studies, and drug resistance studies. Since December 2019, the highly contagious and latent coronavirus 2019 (COVID-19) has spread rapidly and poses a serious threat to human health. Therefore, understanding drug target information can help predict off-target toxicity and its efficacy, and also provide new insights for studying the mechanism of drug side effects. Among them, the evaluation indicators of drug target interaction include maximum semi-inhibitory concentration (IC50), inhibitory constant (Ki) and dissociation constant (Kd). Existing benchmark methods include the introduction of drug-drug interactions, drug-disease interactions, drug-side effects, protein-protein, and protein-disease interactions to improve prediction, and they contain potential information related to drug target effects. However, how to effectively use the topology between nodes to improve the prediction of drug target effect is difficult to solve. The authors propose a metapath-based Transformer model and attention network (MHTAN-DTI) drug target prediction method. First, a heterogeneous bioinformatics network containing 708 drugs, 1512 proteins, 5603 diseases, and 4192 side effects was constructed. The characteristics of the network edge include drug-drug interaction, drug-protein interaction, drug-disease interaction, drug-side effect interaction, protein-protein interaction, and protein-disease interaction. Based on heterogeneous bioinformatics networks, the model generates low-dimensional vectors of drugs and proteins to achieve accurate drug target prediction.
Introduction to the MHTAN-DTI model
2.1 Model Architecture
The authors propose a Transformer and attention network approach to predict drug-target interactions. As shown in Figure 1, first, a heterogeneous bioinformatics network with four node types was constructed: drug, target, disease, and side effect. Then, using Transformer and attention network, the structural and semantic information of the biological network is encoded into low-dimensional potential embedding of drugs and targets. Among them, the meta-path instance Transformer internally aggregates the meta-path instance, single-semantic attention combines the meta-path instance of the same meta-path type with the node itself, and multi-semantic attention integrates multiple single-semantic node embeddings to obtain the final node embedding. Finally, a binary classification result is output, and the weight coefficient of the model is updated by the cross-entropy loss function.
Figure 1 MHTAN-DTI design flow
Experimental results
3.1 Evaluation and comparison of benchmark methods
As shown in Figure 2, the authors verified the performance of MHTAN-DTI on the constructed metapath dataset and compared MHTAN-DTI with 7 state-of-the-art DTI prediction methods (BLMNII, CMF, NetLapRLS, NRLMF, DTINet, NeoDTI, and EEG-DTI). Experimental results show that MHTAN-DTI performs best in DTI prediction performance. Compared to the best results of the 7 control experimental methods, the AUROC of MHTAN-DTI increased by 3.01% and AUPR by 2.71%.
The rest of the methods are less effective, probably due to the relatively limited number of relevant features they capture. In addition, embedding representations learned from unsupervised learning are likely not suitable for DTI prediction. Therefore, MHTAN-DTI can better capture the key characteristics of drug-protein interactions by combining metapath instance-level transformation, single-semantic attention, and multi-semantic attention, resulting in significant performance improvements.
Figure 2 Comparison of benchmark methods
3.2 Ablation experimental analysis
The authors introduced the Transformer encoder as an aggregator for metapath instances. In order to evaluate the influence of different metapath instance encoders on MHTAN-DTI, bidirectional long short-term memory network (BiLSTM) encoders, average encoders, and maximum pooled encoders are also used as benchmark comparisons. In addition, an additional experiment was conducted that only metapath-based neighbor information was utilized, without using any metapath instance encoder. As shown in Figure 3, AUROC and AUPR indicators drop significantly without aggregation of intermediate nodes, indicating that the final embedding of drugs and proteins loses important information. In addition, apart from the Transformer encoder, the BiLSTM encoder performs best in terms of performance. The average encoder and the maximum pooling encoder also achieved good experimental results. The Transformer encoder achieves a significant performance improvement over considering only metapath-based neighbor information, indicating the effectiveness of the encoder on metapath instances.
Figure 3 Ablation experiment of metapath instance encoder
conclusion
The authors constructed a metapathway-based Transformer and a drug-protein interaction (DTI) prediction method combined with attention networks. Firstly, a heterogeneous bioinformatics network based on metapathway is constructed. The method of characterizing data based on metapath type makes the model interpretable. In addition, the Transformer encoder and the two-layer attention mechanism are used to obtain the semantic information of protein and compound sequences to generate low-dimensional vectors of both, which are finally used for DTI prediction. It can be found that the Transformer encoder helps capture dependencies between metapath instances, while the two-layer attention network also mitigates the impact of noise on predictive performance.
Further, to verify the effectiveness and robustness of MHTAN-DTI, the authors compared it with 7 different DTI prediction methods. Experimental results show that MHTAN-DTI shows competitive advantages in performance. In addition, ablation studies were carried out to demonstrate the contribution of different metapath types and different components of MHTAN-DTI to improve DTI prediction performance. These findings highlight the potential and applicability of MHTAN-DTI as a powerful DTI prediction method.
bibliography
- Cummings, M. D. & Sekharan, S. Structure-based macrocycle design in small-molecule drug discovery and simple metrics to identify opportunities for macrocyclization of small-molecule ligands. J. Med. Chem. 62, 6843–6853 (2019)
Copyright Information
This article is an external submission accepted by AIDD Pro, the views stated in the article only represent the author's own views, not on behalf of the AIDD Pro platform, if you find that there is any copyright infringement or other information misinterpretation of the published content, please contact AIDD Pro (please add WeChat sixiali_fox59) for deletion and modification.
The original content is prohibited from being reproduced to other platforms without authorization. If you have any questions, you can email [email protected]