Efficient and accurate, Zhengzhou University team develops new AI tool to identify drug-target interactions

Editor | Dry Leaf Butterfly

Accurate identification of drug-target interactions (DTIs) is one of the key steps in the drug discovery and drug repositioning process.

Currently, many computational-based models have been proposed for predicting DTI, and some significant progress has been achieved.

However, these methods rarely focus on how to fuse multi-view similarity networks related to drugs and targets in an appropriate manner. Furthermore, how to fully incorporate known interaction relationships to accurately represent drugs and targets has not been well studied. Therefore, improving the accuracy of DTI prediction models remains necessary.

In the latest research, teams from Zhengzhou University and University of Electronic Science and Technology of China proposed a new method, MIDTI. This method adopts a multi-view similarity network fusion strategy and a deep interactive attention mechanism to predict drug-target interactions.

The results show that MIDTI performs significantly better than other baseline methods on the DTI prediction task. The results of the ablation experiment also confirm the effectiveness of the attention mechanism and the deep interactive attention mechanism in the multi-view similarity network fusion strategy.

The study is titled "Drug–target interaction predictions with multi-view similarity network fusion strategy and deep interactive attention mechanism" and was published in "Bioinformatics" on June 6, 2024.

Drug target interaction (DTI) prediction

occupies a core position in the process of new drug development and reuse. Traditional wet experiment methods are costly and time-consuming, prompting researchers to turn to computationally assisted drug screening methods. Speed ​​up the process.

Computational DTI prediction methods

are mainly divided into:

1. Structure-based methods: Rely on drug molecules and target structures and binding sites, but are limited by structural information of certain targets such as membrane proteins lack of.
2. Ligand-based method: Build a model based on known active small molecules, but it does not work well when the number of target binding ligands is limited.
3. Machine learning-based method: Predict potential DTI by extracting drug chemical structure and target gene sequence features for binary classification.

Limitations of machine learning methods

Current methods only learn representations based on the structure of the drug and target itself, ignoring the interaction between DTI pairs.

Heterogeneous network construction

The relationships between biological entities contain rich semantic information. Building a network that integrates heterogeneous information helps the system understand DTI.

MIDTI method

The Zhengzhou University team proposed MIDTI, a new method for predicting DTI, based on:

• Multi-view similarity network fusion strategy
• Deep interactive attention mechanism

  

  MIDTI as a whole Framework

Illustration: The overall framework of MIDTI (Source: paper)

Steps:

1. Building a similarity network: MIDTI builds a drug similarity network based on drug association information and adopts a fusion strategy Obtain an integrated drug similarity network; similarly build an integrated target similarity network.
2. Learning embeddings: MIDTI adopts GCN to learn drug and target embeddings from integrated drug similarity networks, integrated target similarity networks, drug-target bipartite networks, and drug-target heterogeneous networks.
3. Discriminative embedding: MIDTI utilizes an interactive attention mechanism to learn discriminative embeddings based on known DTI relationships.
4. Predict DTI: The learned drug-target pair representations are fed into the MLP to predict DTI.

   

   Illustration: Four steps of multi-perspective drug similarity network fusion strategy. (Source: paper)

To evaluate the performance of MIDTI, researchers used a variety of evaluation metrics, including accuracy (ACC), area under the curve (AUC), area under the precision-recall curve (AUPR), F1 score and Matthews correlation coefficient (MCC). The researchers compared MIDTI with ten other competing methods, including random forests, graph convolutional networks, graph attention networks, MMGCN, GraphCDA, and DTINet, among others.

MIDTI achieved scores of 0.9340, 0.9787, and 0.9701 on the ACC, AUC, and AUPR metrics, respectively, which are 2.55%, 2.31%, and 2.30% higher than the highest scores of MMGCN and GraphCDA. This indicates that MIDTI is one of the most competitive methods in predicting drug-target interactions. In experiments with different positive and negative sample ratios, MIDTI also showed excellent performance.

Illustration: Visualizing the drug target embeddings learned by MIDTI at different times. (Source: paper)

The study also shows the visualization results of the drug-target embeddings learned by MIDTI, using the t-SNE tool to map the embeddings into two-dimensional space. As the number of training rounds increases, positive examples and negative examples are gradually distinguished, which proves that the embeddings learned by MIDTI have good discriminability and interpretability, thereby improving the accuracy of DTI predictions.

The core contribution of MIDTI is:

1. It proposes a new multi-view similar network fusion strategy that can integrate different similar networks in an unsupervised manner;
2. Using a deep interactive attention mechanism, based on known DTI information learns discriminative representations of drugs and targets;
3. A large number of experiments prove that MIDTI outperforms other advanced methods in DTI prediction tasks.

In short, MIDTI is an efficient and accurate drug-target interaction prediction method. Its innovation lies in the use of multi-view information and deep attention mechanisms to enhance prediction capabilities.

The researchers said that the next work will be carried out in the following two aspects. First, embedding learning is performed using other relevant data sources of drugs and targets. Second, MIDTI can be applied to other link prediction problems, such as miRNA-disease association prediction.

Related reports: https://academic.oup.com/bioinformatics/article/40/6/btae346/7688335

The above is the detailed content of Efficient and accurate, Zhengzhou University team develops new AI tool to identify drug-target interactions. For more information, please follow other related articles on the PHP Chinese website!