Exploring Anti-Tumor Drug Combinations With Multiple Models Based On Deep Learning

In recent years,new use of old drugs has become another focus of drug development due to the difficulty of new drug development,and drug synergy has received much attention as a method of new use of old drugs.The traditional approach of drug synergy research is to experiment one by one based on experience or laboratory results until a usable drug combination is found.This approach is not only resource intensive and requires a certain level of expertise to support.It is very difficult for the common researcher to discover an effective new combination.With the development of computer technology,more and more researchers employ computer technology to assist in the discovery of drug combinations,but the traditional computer method is no longer applicable in the era of big data because the processing ability of data decreases with the increase of data volume and data dimensionality,and the computing time is longer and the computational accuracy decreases.Deep learning methods based on neural network processing data are excellent for big data,therefore,researchers introduced deep learning into drug synergy combination discovery.However,current models still have problems in practical applications,such as insufficient amount of data,overfitting problem of models,low generalization ability and inability to fully extract higher-order relationships.To address these problems,this paper discusses deep learning drug synergy prediction models in two categories,feature-based drug synergy prediction models and interaction-based drug synergy prediction models.For feature-based drug synergy prediction models,this paper proposed a convolution-based siamese neural network model called SCADS.SCADS extracted features using a siamese neural network composed of convolutional modules with shared weights,and adds the attention mechanism,Batch Norm,and dropout to the convolutional modules and used fully connected layers for feature prediction.In this paper,several features(ECFP,mol2 vec,graph2vec)are evaluated and the optimal feature input is selected by analyzing the impact of features on the model.The prediction of novel drug combinations was performed using SCADS,and the combination with the highest prediction score(Mitoxantrone and Fludarabine)was experimented,and the experimental results showed that the results predicted by SCADS were indeed valid.The comprehensive experimental results show that the SCADS proposed in this paper as a drug prediction model has improved in generalization ability as well as overfitting problem handling,and the model has great application prospects.For the interaction-based drug synergy prediction model,a model GAECDS based on graph auto-encoder and convolutional neural network is proposed in this paper.GAECDS extracted the higher-order relationships between drugs based on graph neural network and predicted drug synergy using convolutional neural network.The testing of the model stability shows that the model has a good improvement in generalization ability,and the problem for the lack of data volume is solved by increasing the data dimension.The drug combinations of laboratory interest were predicted using GAECDS,and four of the top ten scored drug pairs were validated,indicating that the model is indeed applicable.In conclusion,in this paper,we divide the deep learning drug synergy prediction models into two types and propose related models based on these two types,namely,feature-based drug synergy prediction model SCADS and interaction-based drug synergy prediction model GAECDS.We confirm through experiments that the models proposed in this paper are practical and usable,and we also propose effective strategies to solve the problems of current models such as lack of data and overfitting.