مينا طباطبايي

The possibility of treating complex disease using combination of drugs with increasing therapeutic efficacy and reducing toxicity and side effects has made this a vital strategy. Finding the synergistic combinations is usually done by trial and error, which can be time-consuming and costly due to the vast combinatorial search space, and it may cause unnecessary or even harmful treatment. Computational techniques can reduce the number of combinations that should be eva‎luated in an experimental environment. In this paper, we propose two neural network based methods to predict potential synergistic combinations using a varied range of features including chemical structure, biological, biomedical literature embedding, and biological networks interactions features of drugs. In the first method, which is based on the feedforward neural network, we tried to overcome the challenges of datasets for existing combinations, which include small and sparse data and imbalance positive and negative samples. We have applied negative sampling to create balance data and used different regularization methods, including activity regularization, dropout, and early stopping to decrease the chance of overfitting the model due to small training data. In addition, autoencoders and principal component analysis (PCA) are used as dimension reduction approaches to obtain latent representation from sparse data. Using ten-fold cross-validation, we succeed to achieve an AUC of 98 percent, outperforming previous state-of-the-art approaches by an average of 8%. In the second method, we combine multiple Graph Auto Encoder (GAE) models that use Graph Neural Network (GNN) to find new synergistic drug pairs. The comparison of this method with eight of the previous state-of-the-art models indicates its superiority, which outperforms them by an average of 7% w.r.t AUC score (AUC=0.974). In addition to numeral eva‎luation, we constructed a graph of newly predicted drug combinations that are biologically interpreted with interesting patterns. We successfully found drug combinations that were not available in DCDB, but are mentioned and discussed to be synergistic in recent medical papers. Network analysis based on anatomical chemotherapy (ATC) codes revealed that the network constructed of suggested synergestic combinations had a significant degree of homophily.

تركيبات دارويي , تاثيرات هم‌افزا , روش‌هاي محاسباتي , يادگيري عميق

Drug Combinations , Synergistic effects , Computational techniques , Deep Learning

Mina Tabatabaei

Dr. Hossein Rahmani