Sparse Low-rank Multi-task Learning And Its Application In The Prediction Of Bioactivity Of Ligand Molecules

With the development of machine learning,multi-task learning has become one of the hot issues.In many applications,multiple tasks are often correlated with each other.Multi-task learning can use information to build more robust models by exploiting the correlations among tasks.The related features of tasks make the models of multiple tasks share a low-rank structure.Each task has its own unique features,which constitute a sparse structure.Therefore,sparse low-rank multi-task learning has been used to deal with many real cases in life.Sparse low-rank multi-task learning is mainly studied to predict the binding activity value of drug targets and ligand molecules in this paper.This method includes the following steps: first,to construct the required datasets.Then,to process the datasets to extract features using the Pub Chem molecular fingerprint method,and to construct a sparse low-rank multi-task learning model.Finally,perform regression prediction and evaluate the performance of the model.In addition,the number of training samples is changed to evaluate the performance of the methods on small-size datasets with insufficient information.The method of sparse low-rank multi-task learning not only takes advantage of the commonness of many related tasks,but also considers the uniqueness of each task.When applied to small-size datasets,it can overcome the shortcomings of insufficient sample information and the inability to build robust models.Finally,feature selection was performed.The features are re-ranked according to the value of the weight,and a new regression model is constructed by using the filtered relevant features,which can reduce the difficulty and complexity of the model.In this paper,the experiments of regression prediction and feature selection are conducted on 46 datasets,and the results of regression prediction are measured by two commonly-used evaluation criterions.The experiment chooses the traditional single-task learning methods and other multi-task learning methods for comparison.Experiments on 9 datasets are conducted to evaluate the effect of the training samples number.The final experimental results show that correlation coefficient(r2)of the regression prediction model is 30% higher than single-task learning on average,50% higher than deep multi-task learning,and 10% higher than other multi-task learning,and the feature selection model achieved the best performance on two thirds of the datasets.