Research On Interpretable Methods For Drug-target Interaction Prediction

Predicting drug-target interactions is an important step in drug development.Traditional prediction methods based on biological experiments are time-consuming,risky and costly.In silico methods can effectively improve the efficiency of predicting drug-target interactions and reduce costs.At present,most in silico methods for predicting drug-target interactions lack interpretability.In response to the above problem,this thesis defines drug-target interaction prediction as a binary classification problem,proposes interpretable drugtarget interaction prediction methods,and conducts the following research works:1.In order to improve the interpretability of deep neural network,a drug-target interaction prediction model based on prior knowledge is proposed.For the first time,a deep neural network based on prior knowledge is used to predict drug-target interactions.In addition,a network interpretation method based on the difference on neuron’s activation value is proposed to illustrate the model’s interpretability.2.In order to ensure the interpretability and prediction performance of the model,a drug-target interaction prediction method based on the attention mechanism is proposed.For the first time,the two-way attention mechanism and the multi-head self-attention mechanism are combined to predict drug-target interactions.3.Aiming at the problem that high-performance complex models are difficult to interpret,a drug-target interaction prediction method based on black-box analysis is proposed.It is the first time that knowledge distillation is applied to the field of drugtarget interaction.Knowledge distillation is used to train the global surrogate model of the black-box model,which is interpreted by analyzing the global surrogate model.The experimental results show that using prior knowledge to build a deep neural network can improve its interpretability;the model based on attention mechanism can ensure interpretability and prediction performance at the same time;the global surrogate model trained by knowledge distillation can effectively interpret the black-box model.The above work will provide new ideas for the study of drug-target interaction prediction.