Prediction Of Drug-drug Interaction Based On Multi-dimensional Feature Fusion

Objectives:A multi-dimensional feature fusion model integrating 1D sequence features,2D molecular map features,2D substructure features,and 3D molecular geometry features is constructed to predict the specific interaction between drugs.Methods:The current DDIs prediction models can only predict whether there is interaction or the probability of interaction between drugs,and the lack of other information on many drugs makes it difficult to predict a large number of specific types of DDIs between drugs.With the development of the Simplified Molecular Linear Input Specification(SMILES)and the continuous breakthroughs in the field of molecular characterization learning.It is an effective method to predict specific types of DDI events by constructing features of different dimensions based solely on the structure and spatial information of drug molecules themselves.The purpose of predicting specific types of DDIs is achieved by collecting DDIs information of drugs from the Drug Bank database and converting the collected DDIs information into DDI events in the form of "Drug A,Drug B,Mechanism,Action" quaternion in the way of natural language processing.In order to fully represent the different dimensional features of drugs using only SMILES of drug molecules as input to predict specific types of DDI events,we propose a four channel deep learning model.We take SMILES of drug molecules as input,and use FCS algorithm to split SMILES of drugs into smaller subsequences through Transformer to construct 1D sequence features;use the tool RDkit to generate 2D molecular diagram,update the information of neighboring nodes through the message transmission network of adding attention mechanism in the way of atoms as nodes and chemical bonds as edges,and finally aggregate into 2D molecular diagram features;take the molecular graph of a pair of drugs and the specific interaction type as the tuple,gradually extract the substructure of drug molecules through the graph attention network,and input it into the multi-head attention mechanism module to give important substructure a high weight,and finally form a learnable feature matrix to construct 2D substructure features;the best conformation is obtained through the tool RDkit optimization to obtain the atomic position information and atomic type,and calculate the geometric information such as the angle between atoms,bond length,twist angle,and finally input it into the Sch Net model to build 3D geometric features.The four features of different dimensions are spliced into new features through the torch.cat function to achieve multi-dimensional feature fusion.Results:Through multiple parameter adjustments,the accuracy of the model is as high as 0.9516.The accuracy of 3D geometric features is as high as 0.8704 from the single features of ablation experimental results.In the combination of multiple features,four features with different dimensions have the best performance.Compared with other methods,our model is significantly improved from the current deep learning model due to the traditional machine learning method.Finally,the predicted results of the model were externally verified in the Drugs database,which also confirmed the reliability of the predicted results.Conclusions:Our model can effectively predict specific types of drug interactions by fusing the characteristics of four different dimensions.and the reliability of the predicted structure has been proven through external verification,indicating the feasibility of using artificial intelligence technology to predict DDIs.