Predicting Of Ligand Bioactivities Targeting Orphan G Protein-coupled Receptors Through Deep Domain Adaptation

G protein-coupled receptors(GPCRs)are a large protein family with seven transmembrane structures,which include more than 800 members and participate in a variety of human physiological and pathobiological processes.GPCRs are the most widespread and successful targets in modern medicine.More than 30%of drugs on sail in the world have been appointed by GPCRs as their targets.At present,there are still more than 140 GPCRs whose endogenous ligand molecules are not clear.They are orphan GPCRs which are one of the most important sources for drug targets for new drug discovery.Accurately prediction of the biological activity of ligands targeting these orphan GPCRs is essential and important for screening and further optimizing lead compounds targeting these orphan GPCRs.However,due to the extremely lack of experiment data about ligand bioactivities targeting these orphan GPCRs,it is hard to build a stratified model to predict ligand bioactivites.In this paper,we proposed a few shot learning method called UDDA for accurately learning the biological activity of ligands targeting orphan GPCRs.UDDA intends to solve the problem of insuffient data through unsupervised deep domain adaptation.Our UDDA method includes three organically related parts:(1)To design a model for generating target and source domain drug target data sets;(2)To train a virtual screening model by unsupervised domain adaptation;(3)To constructing virtual screening model for small molecule drugs from target domain.We have tested the performance of our UDDA method on 12 important human orphan GPCRs datasets,half of which have less than 100 ligand samples.Experimental results show that UDDA has achieved good performance in modeling the biological activity of ligands,with an average root mean square error(RMSE)of 1.06586 and an average correlation coefficient(r~2)of 0.120977.When compared with traditional supervised machine learning,the performance of UDDA is comparable.When compared with random forest,UDDA achieve an improvement on RMSE with 35.1%on average and on r~2 with 79.0%on average.When compared to support vector regression machine,UDDA has an increase of RMSE with 8.5%on average and of r~2 increase with 67.1%on average.When compared with gradient boosting decision tree,its RMSE increases by an average of 34.9%and its r~2 decreases by an average of 79.2%.When compared with unsupervised algorithms,the performance of our algorithm achieves better performance.When compared with the MSTL-GNN method,the RMSE of our UDDA method decreases by an average of 2.1%and r~2 increases by an average of 82.3%.In addition,we have considered the impact of training sample size and model parameters on model performance.Experimental results show that the bigger size of training sample datas can help improve the performance of the algorithm,while other parameters of the algorithm have no significant impact on performance.