Research On The Prediction Method Of Anticancer Drug Combination Based On Deep Learnin

Compared to single-drug treatments,drug combinations have demonstrated significant potential in cancer therapy.Combination therapy is a well-established concept in cancer treatment;however,identifying synergistic drug combinations remains challenging due to uncertainties in the combination space.Computational methods based on high-throughput screening data have emerged as efficient and time-saving approaches for predicting drug combinations with synergistic effects.Traditional machine learning methods have exhibited some predictive capabilities in discovering anticancer drug combinations,but there is still room for improvement in performance.Moreover,due to the rapid expansion of data volume and increasing data complexity,traditional methods face limitations in computation time and accuracy.In contrast,deep learning models possess distinct advantages in processing large amounts of complex data.This study explores a new anticancer drug combination prediction model using various biological data and deep learning techniques,aiming to enhance the model’s predictive capabilities and elucidate the biological significance of the results.The specific content is as follows:(1)Considering the implicit relationships between different drug target proteins,we propose a network embedding prediction model that utilizes the topology of protein interaction networks.This model combines corresponding protein modules of drug-cancer cell lines with protein interaction network information.By employing struc2 vec,the hidden topological features of each protein node within the protein interaction network are extracted,exploring targets that are distant in the network and share the same local network structure.Synergistic relationships between drug combinations and cancer cell lines are predicted using Extreme Gradient Boosting.The model demonstrates good predictive performance,further validating the effectiveness of network information in identifying combination therapies for cancer and other complex diseases.(2)Considering the interpretability of biological pathway information in processes such as human metabolism,we propose a deep model for drug combination prediction based on cancer biological pathways.This model employs Net PEA to calculate the relationships between genes in the protein interaction network and pathway genes,combining the positional information of drugs and cancer cell lines to generate corresponding features.The model uses a Graph Autoencoder and deep neural network for prediction.The results indicate that calculating the relationships between target genes and pathway genes can improve predictive capabilities and,to some extent,enhance the model’s interpretability.(3)Considering the importance of local network features and the impact of data diversity,we introduce multiple types of biological information,including chemical information,network information,and pathway information,to propose a deep prediction model based on the self-attention mechanism.In addition to protein interaction networks and cancer pathways,the model incorporates drug molecular descriptor information.Building on network topology and biological pathway proximity,the model further obtains hidden features using a Graph Attention Network.Molecular descriptors are calculated using the RDKit Descriptors module and input into a deep neural network to predict drug combination synergy values.The results indicate that combining various biological information with the self-attention mechanism effectively enhances predictive capabilities.The models proposed in this thesis have been applied to drug combination datasets,and the results demonstrate their outstanding predictive capabilities for drug combinations in specific cancer tissues.Furthermore,these models allow for a certain degree of interpretation of the results.Not only do they exhibit good performance metrics,but they also outperform existing prediction models.