Similarity Metric Learning of Molecular Structure Features based Method for Drug Bioactivity Prediction

Drug bioactivity prediction entails estimating drugs' biological activity by analyzing their chemical formulas' structural characteristics. This technique establishes a functional relationship that maps the structural features of chemical molecules to their corresponding bioactivities, allowing for the estimation of the bioactivity of potential drugs, thereby significantly reducing the cost and cycle of drug development. An artificial neural network (ANN) prediction model based on principal component features is proposed for this task. The model filters through feature variable information to eliminate invalid features and then extracts significant features. Finally, an ANN model is used to establish the connection between molecular structural features and bioactivity, enabling the prediction of the bioactivity performance of different chemical molecules. In addition, to address the overfitting problem in forecasts, this paper designs a semi-supervised classification-based method for bioactivity prediction. This method first classifies samples through the k-means algorithm and then learns the projection subspace through the LDA method. Next, the projection subspace is improved through semi-supervised learning. Based on the similarity of samples, the estimated value of the test sample's bioactivity is obtained through the average bioactivity of the most similar s samples. Finally, we tested the algorithm proposed on a dataset of 1974 samples. The model's fit degree reached 0.88, and the mean square error on the test data reached 0.64, fully proving the accuracy of the algorithm presented in this paper.