U a goal in personalized medicine. Thus, estimating how

SING the right drug in the right dose at the right time is each patient responds to a drug based on their biological characteristics (e.g., omics data) is important in biomedical research. However, patients’ drug response data is very limited and not well-structured. Indeed, there have been only a few studies on drug response for cancer patients gathered in TCGA [ 1 ]. This has formed a barrier to largescale research on this topic.

Fortunately, large-scale projects on drug response for “artificial patients” (i.e., cell line), such as GDSC [ 2 ], CCLE [ 3 ] and NCI60 [ 4 ] have facilitated the development of computational methods for drug response prediction [ 5 ], [ 6 ], [ 7 ], [ 8 ]. Indeed, a DREAM challenge for drug sensitivity prediction was launched and with methods proposed by many research groups [ 9 ]. Most of them are machine learningbased, where different strategies for data and model integration were introduced. For example, multiple-kernel and multiple-task learning techniques were proposed to integrate various types of –omics data of cell lines and response data [ 10 ], [ 11 ]. Besides, ensemble learning strategies were used to integrate individual models [ 12 ], [ 13 ], [ 14 ]. In parallel, network-based methods relying on similarity networks (e.g., structural similarity between drugs and biological similarity between cell lines) and known drugcell line responses [ 15 ], [ 16 ], [ 17 ] have been proposed. In addition, protein interaction and gene regulatory networks have also been used also used to predict drug response [ 18 ].

The machine learning-based methods have shown their ability in the data and model integration, thus drug response prediction was generally systematically approached. However, drugs and cell lines are often represented by predefined features such as structural features of drugs and -omics profiles of cell lines. As the number of cell lines is much smaller than the number of genes in -omics profiles of cell lines, thus some of the traditional machine learningbased methods often face the “small n, large p” problem. Consequently, this limits the prediction performance of traditional machine learning-based methods.

Deep learning is a state-of-the-art branch of machine learning for extracting a feature from complex data and making accurate predictions [ 19 ]. Recently, deep learning has been applied to drug discovery [ 20 ], [ 21 ]. It has achieved superior performance compared to traditional machine learning techniques in many problems in drug development such as visual screening [ 22 ], [ 23 ], drug-target profiling [ 24 ], [ 25 ], [ 26 ], [ 27 ], drug repositioning [ 28 ], [ 29 ]. Especially in drug response, deep learning is utilized to automatically learn genomic features of cell lines and the structural features of drugs to predict anticancer drug responsiveness [ 30 ], [ 31 ], [ 32 ], [ 33 ]. For example, the deep neural network is used in DeepDR [ 31 ] to predict the half-maximal inhibitory concentrations (IC50) or the convolutional neural network is utilized in tCNNS [ 33 ] and CDRScan [ 30 ] to extract the features of cell lines and drugs. In addition, in DeepDSC [ 32 ], a pre-trained stacked deep autoencoder is used to extract genomic features of cell lines from gene expression data and then combine with chemical features of compounds to predict response data. However, in these deep learning models, drugs are represented as strings which are not a natural presentation, thus the structural information of drugs may be lost.

Graph convolutional networks can learn representations of compound structures represented as molecular graphs [ 34 ]. For example, in GraphDTA [ 35 ], the drugs are presented as graphs where the edges are the bonding of atoms and the model achieves the best performance compared to other deep learning-based methods which represent drugs as strings in the task of drug-target binding affinity prediction. However, the graph neural network has not been employed yet [ 34 ] for drug response prediction. So it is promising to apply graph neural network to drug response prediction. In addition, although deep learning-based methods often achieve better prediction performance when compared to traditional machine learning-based methods, it is considered as a black-box approach because of not being interpretable. The saliency map [ 36 ] was introduced to visualize image features in classification task at first, now it plays an important role in various practical applications right from video surveillance [ 37 ] to traffic light detection [ 38 ]. In this research, this strategy can help evaluate the degree of genomics features such as aberration attributes to drug response prediction.

tCNNS was recently published and shown to be the state-of-the-art method among other deep learning-based methods [ 33 ]. We also tested traditional machine learning methods such as decision tree, gradient boosting but even tCNNS surpassed these methods. So in our work, we only compared our result directly with tCNNS.

In this study, we propose GraphDRP (Graph convolutional network for drug response prediction), a new neural network architecture capable of modeling drugs as molecular graphs to predict drug response on cell-line. We compared our method with the state-of-the-art, tCNNS [ 33 ], where drug molecules were represented as SMILES [ 39 ] strings. Experimental results indicate that our method achieves better performance in terms of root mean square error (RMSE) and Pearson correlation coefficient for all experiments. Also, by visualizing the resulting networks through saliency maps, we can discover the most significant genomic aberrations for the prediction of the response value. This suggests a novel way to interpret the result of deep learning models for drug response prediction. 2

The proposed model of drug response prediction is shown in Fig 1. The input data includes chemical information of drugs and genomic features of cell lines including mutations and copy number alternations (i.e., genomic aberration).

For the drug features, the drugs represented in SMILES format [ 39 ] were downloaded from PubChem [ 40 ]. Then, RDKit, an open-source chemical informatics software [ 41 ], was used to construct a molecular graph reflecting interactions between the atoms inside the drug. Atom feature design from DeepChem [ 42 ] was used to describe a node in the graph. Each node contains five types of atom features: atom symbol, atom degree calculated by the number of bonded neighbors and Hydrogen, the total number of Hydrogen, implicit value of the atom, and whether the atom is aromatic. These atom features constituted a multidimensional binary feature vector [ 35 ]. If there exists a bond among a pair of atoms, an edge is set. As a result, an indirect, binary graph with attributed nodes was built for each input SMILES string. Several graph convolutional network models, including GCN [ 43 ], GAT [ 44 ], GIN [ 45 ] and combined GAT-GCN architecture [ 35 ], were used to learn the features of drugs. Following the graph neural network, a fully connected layer (FC layer) was also used to convert the result to 128 dimensions.

In deep learning models, 1D convolutions are normally used for genomic features. We used the same approach as other models because 1D convolution with a large kernel has the ability to combine genomic abbreviations in the genomic features to make good predictions. In addition, 1D pooling was also used to reduce the size of input feature then 1D convolutions can learn abstract features from genomic features. The genomic features of cell lines were represented in one-hot encoding. 1D convolutional neural network (CNN) layers were used to learn latent features on those data. Then the output was flattened to 128 dimension vector of cell line representation.

Finally, the 256-dimension vector, the combination of drug’s feature and cell line’s feature was put through two fully-connected layers with the number of nodes 1024 and 256 respectively, before predicting the response.

CNNs have recently achieved success in computer vision [ 46 ], [ 47 ] and natural language processing [ 48 ], [ 49 ], which motivates the use of convolutional neural networks to graph structures. Similar to the use of CNN with image, CNN in the graph also has two main layers: convolutional layer and pooling layer. While convolutional layer is for learning receptive fields in graphs whose data points are not arranged as Euclidean grids, pooling layer is for down-sampling a graph [ 35 ]. Graph convolutional network (GCN) is wellfitted for the drug response because the drug molecular itself is represented in the form of a graph. In order to evaluate the effectiveness of graph-based models, we investigated several graph convolutional models, including GCN [ 43 ], GAT [ 44 ], GIN [ 45 ] and combined GAT-GCN architecture [ 35 ]. The details of each GCN architecture are described as follows. 2.1

Given a drug-cell line pair, saliency value was defined by using the idea of the saliency map [ 36 ] to measure the importance of each genomic aberration to the prediction of response value (Y). In our proposed model, each drug (D) was represented by a graph, meanwhile, each cell-line (C) was displayed by a binary vector of 735 dimensions, with each value indicates whether or not the cell-line had a specific genomic aberration.

f is the whole deep learning model function.

Y = f (C; D) (6)

Then saliency value (S) was defined as the gradient of cell-line with respect to predicted response as follows: @Y S = (7)

@C

This saliency value has the same size as cell-line vector. The higher value indicates the more important of genomic aberration that was encoded in this position. 4 4.1

In this study, we proposed a novel method for drug response prediction, called GraphDRP. In our model, drug molecules were presented as graphs instead of strings, celllines were encoded into one-hot vector format. Then graph convolutional layers were used to learn the features of compounds and 1D convolutional layers were used to learn cell-line representation. After that the combination of drug and cell-line representation was used to predict IC50 value. Four variants of graph neural networks including GCN, GAT, GIN and combination of GAT&GCN were used for learning features of drugs. We compared our method with state-of-the-art one, tCNNS [ 33 ], where drug molecules were represented as SMILES strings.

Experimental results indicate that our method achieves better performance in terms of both root mean square error and Pearson correlation coefficient. The performance suggests that representing drugs in graphs is more suitable than in strings since it conserves the nature of chemical structures of drugs. Furthermore, the responses of missing drug-cell line pairs in GDSC dataset were predicted and analyzed. We figured out that Bortezomib and Epothilone B have the lowest IC50 values and we found the evidence showing that some types of cancer are sensitive to these drugs. Similarly, we also found that cancers are less sensitive to drugs having the highest IC50 values. It means that the model actually learns from data and has a potential to predict the response of new drug-cell line pairs. Also, through saliency maps, we discovered ten most important genomic aberrations of the three cell-lines having lowest IC50s to that drug and seek their contribution to the sensitivity of that drug. This technique suggests a novel way to interpret the result of deep learning model in drug response.

Since cell lines in the same tissue types will share similar genetic information, in the future, we will split the cell lines based on different tissue types to demonstrate how the performance varies as cell line similarity decreases.

Because all drugs are screened at a certain concentration, if this concentration is low, that does not necessarily make it a good cancer drug. Rather, a drug that is particularly toxic to a cancer cell line compared to non-cancer cells is a good drug for this cell line. However, this non-cancer toxicity is not measured in the GDSC panel, and hence only looking at a low IC50 value is not sufficient. In this study, we only focused on improving the prediction of IC50 values by using GCNs. Also, in our study, we concentrated on improving drug response prediction by extracting drug features from their graph representation by GCN; thus, we used the same dataset with the tCNNS study, which learned drug’s features from their string format. In addition, only genomic data of cell lines was used in our study. Therefore, in future work, we will additionally use other -omics data such as gene expression, since it is proven to support drug response prediction [ 58 ] [ 59 ] [ 60 ].

Tuan Nguyen is a data scientist at Big Data Institute, Vingroup. He received his MSc degree from University of Bristol in 2019 in the area of machine learning. His current research topic includes application of deep learning in medicine and the application of GAN.

Giang T.T. Nguyen is a PhD Candidate in Information System at Posts and Telecommunications Institute of Technology. She received BSc and MSc in Computer Science at Hanoi University of Science and Technology. Her current research interests include Bioinformatics and Computational biology.

Thin Nguyen is a Senior Research Fellow with the Applied Artificial Intelligence Institute (A2I2), Deakin University, Australia. He graduated with a PhD in Computer Science from Curtin University, Australia in the area of machine learning and social media analytics. His current research topic is strongly inter-disciplinary, bridging large-scale data analytics, pattern recognition, genetics and medicine. His research direction is to utilize artificial intelligence to discover functional connections between drugs, genes and diseases.

Duc-Hau Le received his Ph.D. degree in 2012 with research interests in bioinformatics and computational biology. So far, he has published many papers on prestigious journals. He is now leading the Department of Computational Biomedicine, Vingroup Big Data Institute in Vietnam, which aims to develop/apply computational methods to build and analyze high-throughput biomedical data in order to improve the screen, diagnosis and treatment of complex diseases.

Specially, he is the principal investigator of the biggest genome project in Vietnam (i.e., building databases of genomic variants for Vietnamese population).