Search

Application of artificial intelligence (AI) in precision oncology typically involves predicting whether the cancer cells of a patient (previously unseen by AI models) will respond to any of a set of existing anticancer drugs, based on responses of previous training cell samples to those drugs. To expand the repertoire of anticancer drugs, AI has also been used to repurpose drugs that have not been tested in an anticancer setting, i.e., predicting the anticancer effects of a new drug on previously unseen cancer cells de novo. Here, we report a computational model that addresses both of the above tasks in a unified AI framework. Our model, referred to as deep learning-based graph regularized matrix factorization (DeepGRMF), integrates neural networks, graph models, and matrix-factorization techniques to utilize diverse information from drug chemical structures, their impact on cellular signaling systems, and cancer cell cellular states to predict cell response to drugs. DeepGRMF learns embeddings of drugs so that drugs sharing similar structures and mechanisms of action (MOAs) are closely related in the embedding space. Similarly, DeepGRMF also learns representation embeddings of cells such that cells sharing similar cellular states and drug responses are closely related. Evaluation of DeepGRMF and competing models on Genomics of Drug Sensitivity in Cancer (GDSC) and Cancer Cell Line Encyclopedia (CCLE) datasets show its superiority in prediction performance. Finally, we show that the model is capable of predicting effectiveness of a chemotherapy regimen on patient outcomes for the lung cancer patients in The Cancer Genome Atlas (TCGA) dataset*.

BackgroundDrug sensitivity prediction and drug responsive biomarker selection on high-throughput genomic data is a critical step in drug discovery. Many computational methods have been developed to serve this purpose including several deep neural network models. However, the modular relations among genomic features have been largely ignored in these methods. To overcome this limitation, the role of the gene co-expression network on drug sensitivity prediction is investigated in this study.MethodsIn this paper, we first introduce a network-based method to identify representative features for drug response prediction by using the gene co-expression network. Then, two graph-based neural network models are proposed and both models integrate gene network information directly into neural network for outcome prediction. Next, we present a large-scale comparative study among the proposed network-based methods, canonical prediction algorithms (i.e., Elastic Net, Random Forest, Partial Least Squares Regression, and Support Vector Regression), and deep neural network models for drug sensitivity prediction. All the source code and processed datasets in this study are available athttps://github.com/compbiolabucf/drug-sensitivity-prediction.ResultsIn the comparison of different feature selection methods and prediction methods on a non-small cell lung cancer (NSCLC) cell line RNA-seq gene expression dataset with 50 different drug treatments, we found that (1) the network-based feature selection method improves the prediction performance compared to Pearson correlation coefficients; (2) Random Forest outperforms all the other canonical prediction algorithms and deep neural network models; (3) the proposed graph-based neural network models show better prediction performance compared to deep neural network model; (4) the prediction performance is drug dependent and it may relate to the drug’s mechanism of action.ConclusionsNetwork-based feature selection method and prediction models improve the performance of the drug response prediction. The relations between the genomic features are more robust and stable compared to the correlation between each individual genomic feature and the drug response in high dimension and low sample size genomic datasets.

In line with recent advances in neural drug design and sensitivity prediction, we propose a novel architecture for interpretable prediction of anticancer compound sensitivity using a multimodal attention-based convolutional encoder. Our model is based on the three key pillars of drug sensitivity: compounds' structure in the form of a SMILES sequence, gene expression profiles of tumors and prior knowledge on intracellular interactions from protein-protein interaction networks. We demonstrate that our multiscale convolutional attention-based (MCA) encoder significantly outperforms a baseline model trained on Morgan fingerprints, a selection of encoders based on SMILES as well as previously reported state of the art for multimodal drug sensitivity prediction (R2 = 0.86 and RMSE = 0.89). Moreover, the explainability of our approach is demonstrated by a thorough analysis of the attention weights. We show that the attended genes significantly enrich apoptotic processes and that the drug attention is strongly correlated with a standard chemical structure similarity index. Finally, we report a case study of two receptor tyrosine kinase (RTK) inhibitors acting on a leukemia cell line, showcasing the ability of the model to focus on informative genes and submolecular regions of the two compounds. The demonstrated generalizability and the interpretability of our model testify its potential for in-silico prediction of anticancer compound efficacy on unseen cancer cells, positioning it as a valid solution for the development of personalized therapies as well as for the evaluation of candidate compounds in de novo drug design., 11 pages, 5 figures, 1 table, Workshop on Computational Biology at the International Conference on Machine Learning (ICML), Long Beach, CA, 2019

A requirement for precision medicine is the accurate prediction of the sensitivity of a given drug on an individual patient. A very common method for this prediction is the use of Random Forest built on Genomic Features such as gene expression. However, effective drug sensitivity prediction requires the use of multiple heterogeneous dataset but it is rare that all such information is available for all drug and cell line combinations. To effectively incorporate all available data into a drug sensitivity prediction problem requires the use of stacking multiple models, built using a variety of different methods and data. In this article we investigate stacking in the context of drug sensitivity prediction. First, we examine basic models utilized in drug sensitivity prediction including the Random Forest, Neural Network, and K-Nearest Neighbor approach. To stack individual models we utilize a linear stacking method built utilizing a held-out validation set. We then investigate which form of stacking is most effective in improving accuracy. The two forms of stacking investigate are vertical stacking, in which our training samples are split and then stacked, and horizontal stacking, where our features are split before stacking. From a theoretical standpoint we have shown that horizontal stacking outperforms vertical stacking, especially as the sample size becomes large. Our theory is then proven utilizing both synthetically generated data, as well as data extracted from the Cancer Cell Line Encyclopedia. Despite having the best individual predictor, vertical stacking consistently underperformed compared to horizontal stacking. Finally, we build a stacked model utilizing data extracted from a variety of sources. Gene Expression for individual cell lines are taken from the GDSC database, drug target data is data mined from Pubchem and physical features are extracted utilizing PaDEL-Descriptor software. Using this data we aim to predict the Area Under the Curve value for individual drug-cell line combinations. We examine the performance of individual and stacked models. We note that stacking models built on two heterogeneous datasets provide superior performance to stacking different models built on the same dataset. It is also noted that stacking provides a noticeable reduction in the bias of our predictors when the dominant eigenvalue of the principle axis of variation in the residuals is significantly higher than the remaining eigenvalues.