Skip to main content
NCBI home page
AMIA Annual Symposium Proceedings logoLink to AMIA Annual Symposium Proceedings
. 2021 Jan 25;2020:1364–1372.

Predicting Tumor Cell Response to Synergistic Drug Combinations Using a Novel Simplified Deep Learning Model

Heming Zhang 1,2, Jiarui Feng 1,3, Amanda Zeng 1,3, Philip Payne 1, Fuhai Li 1,4,*
PMCID: PMC8075535  PMID: 33936513

Abstract

Drug combinations targeting multiple targets/pathways are believed to be able to reduce drug resistance. Computational models are essential for novel drug combination discovery. In this study, we proposed a new simplified deep learning model, DeepSignalingSynergy, for drug combination prediction. Compared with existing models that use a large number of chemical-structure and genomics features in densely connected layers, we built the model on a small set of cancer signaling pathways, which can mimic the integration of multi-omics data and drug target/mechanism in a more biological meaningful and explainable manner. The evaluation results of the model using the NCI ALMANAC drug combination screening data indicated the feasibility of drug combination prediction using a small set of signaling pathways. Interestingly, the model analysis suggested the importance of heterogeneity of the 46 signaling pathways, which indicates that some new signaling pathways should be targeted to discover novel synergistic drug combinations.

1. Introduction

Acquired and innate drug resistance is one major challenge in cancer therapy, due to the complex signaling pathways of cancer. Drug combinations targeting multiple targets or multiple signaling pathways are believed to be one possibility to reduce drug resistance. Many studies have attempted to identify potentially effective and synergistic drug combinations for cancer treatment in experimental laboratories. For example, RAS and ERK inhibitors were recently reported to be synergistic with autophagy inhibitors in RAS-driven cancers1,2. The mechanism of synergy is that the inhibition of RAS signaling causes the activation of autophagy signaling, which prevents cancer cell death1,2. In BRAF inhibitor resistant Melanoma, vemurafenib (BRAF inhibitor) + tretinoin (retinoic acid receptor agonist) were found to be effective and synergistic in cell assays and mouse models3. However, there are a few effective drug combinations for clinical use in cancer therapy. Novel and effective drug combinations are needed for personalized treatment to reduce drug resistance.

Many cancer cell lines and mouse models are available to experimentally screen drugs and drug combinations. However, the experimental screening approaches are limited, considering the numerous possible combinations of thousands of FDA approved drugs and thousands of investigational agents. For example, there are currently about 4 available datasets of experimental screening drug combinations: 1) NCI-ALMANAC Drug Combination Data Set4 (~5,232 combinations from ~100 drugs on NCI60 cell-lines); 2) the AstraZeneca-Sanger Drug Combination Prediction DREAM Challenge Data Set5 (900 combinations from 118 compounds on 85 cancer cell lines); 3) the Yale-Stern Melanoma DataSet6 (~7000 combinations from 145 drugs/compounds on 19 melanoma cancer cell lines); and 4) the Merck-2016 DataSet7 (583 combinations from 38 drugs/compounds on 39 cancer cell lines). These datasets provide a valuable basis to build machine-learning and deep learning-based models.

Computational models that integrate diverse pharmacogenomics datasets with multi-omics data of cancer patients to prioritize drug combinations are essential for novel drug combination discovery. The combination of computational and experimental models can facilitate drug combination discovery in a fast manner. Though a set of prediction models have been reported for drug combination prediction, it remains an open problem. For example, the network-based and connectivity map8,9 based drug combination models10,11,12 based on the gene-gene interaction network have been proposed. In addition, a semi-supervised learning model integrating diverse pharmacogenomics datasets was proposed to predict drug combination13. Network message propagation-based models developed using drug-target interactions and multi-omics data were also proposed to predict combinations14,15. Deep learning models have also been proposed for drug combination prediction. For example, A deep belief network (DBN) model, DeepSynergy, that integrates a large number of chemical structure and genomics features on the Merck-2016 DataSet7 was recently proposed16 as a prediction drug combination method. In DeepSynergy, about 8300 numerical features (chemical descriptors and genomic features) were used for the prediction. The other deep learning model, AuDNNsynergy17 (Deep Neural Network Synergy model with Autoencoders), integrates the multi-omics data of over 10,000 cancer genome atlas (TCGA) cancer samples. One limitation of the existing deep learning models of drug combination prediction is the use of a large number of chemical and omics features (>10 thousand features) and fully connected dense layers (a huge number of parameters in the model to be trained) relative to the small number (30~100 drugs on 30~60 cancer cell lines) of drug combination synergy scores experimentally obtained. Though the model with a large number of parameters can fit/predict the data, the model parameters cannot be well trained, and cannot be explained.

To reduce the complicity of the deep learning model and make the models more explainable, in this study, we propose a novel simplified deep learning model,DeepSignalingSynergy, for drug combination prediction. Compared with existing models that make use of a large number chemical and genomics features, we built the model on a small set of cancer signaling pathways, with the aim of investigating the importance of individual signaling pathways. Moreover, the model can mimic the integration of multi-omics data and drug target/mechanism in a relatively more biological meaningful and understandable manner. The results from evaluating the model on the NCI ALMANAC drug combination screening data indicated the feasibility of using a small set of signaling pathways and showed the importance of signaling pathways that affect the drug combination response.

2. Materials and Methodology

2.1. Drug combination screening data in NCI ALMANAC dataset

The drug pair data was obtained from the NCI ALMANAC database, which is a resource created in 2017. The NCI Almanac dataset includes a score assigned to each of the drug pairs was assigned a score, termed the NCI “ComboScore”1 to indicate the synergy scores of drug combinations. In summary, the synergistic effects of combinations of 104 FDA approved drugs in terms of cancer cell growth inhibition were evaluated on NCI 60 cancer cell lines. The average comboScore of two drugs with different doses on a given cancer cell lines was used to indicate the synergy score of two drugs on the cancer cell line, with a 4-element tuple: <DA, DB, CC, SABC>.

2. 2 RNA-seq gene expression and copy number data of NCI-60 Cancer Cell Lines from Cancer cell line encyclopedia (CCLE)

Cancer cell line encyclopedia (CCLE) database18 provides the multi-omics data of more than 1000 cancer cell lines, e.g., RNA-seq (gene expression), copy number variation, metabolomics, miRNA, RPPA. The large panel of cancer cell lines with comprehensive genetic characterization provides a data source to investigate the associations between molecular features and cancer phenotypes, including drug responses. For this study, the RNA-sequencing gene expression values, using TPM (transcripts per million), and copy number values of genes of 1019 cancer cell lines were downloaded from the cancer cell line encyclopedia (CCLE) website: https://portals.broadinstitute.org/ccle. The CCLE dataset collected the omics data of 45 of the 60 NCI-60 cancer cell lines, as shown in Table I.

Table I:

NCI-60 cancer cell lines included in CCLE with RNAseq data. Orange text indicates the cancer cell lines that are not included in CCLE.

786-0 HCC-2998 KM12_LARGE_INT ESTINE NCIH23_LUNG RPMI8226_HAEMA TOPOIETIC_AND_L YMPHOID_TISSUE
A498_KIDNEY HCT116_LARGE_IN TESTINE LOXIMVI_SKIN NCIH322_LUNG RXF 393
A549_LUNG HCT15_LARGE_INT ESTINE M14 NCIH460_LUNG SF268_CENTRAL_N ERVOUS_SYSTEM
ACHN_KIDNEY HL60_HAEMATOPO IETIC_AND_LYMP HOID_TISSUE MALME3M_SKIN NCIH522_LUNG SF295_CENTRAL_N ERVOUS_SYSTEM
BT549_BREAST HOP62_LUNG MCF7_BREAST NCI/ADR-RES SF539_CENTRAL_N ERVOUS_SYSTEM
CAKI1_KIDNEY HOP92_LUNG MDAMB231_BREAS T NIHOVCAR3_OVAR Y UO31_KIDNEY
CCRF-CEM HS578T_BREAST MDAMB435S_SKIN OVCAR4_OVARY SK-MEL-2
COLO 205 HT29_LARGE_INTE STINE MDAMB468_BREAS T OVCAR-5 SKMEL28_SKIN
DU145_PROSTATE IGROV1_OVARY MOLT-4 OVCAR8_OVARY SKMEL5_SKIN
EKVX_LUNG K562_HAEMATOPO IETIC_AND_LYMP HOID_TISSUE NCIH226_LUNG PC3_PROSTATE SKOV3_OVARY
SN12C SNB-75 SW620_LARGE_INT ESTINE TK-10 UACC257_SKIN
SNB-19 SR-almanac T47D_BREAST U251MG_CENTRAL_NERVOUS SYSTEM UACC62_SKIN
Open in a new tab

2.3. KEGG signaling pathways and cellular process

KEGG (Kyoto Encyclopedia of Genes and Genomes)19 is a database for the systematic understanding of gene functions. The KEGG signaling pathways provide the knowledge of signaling transduction and cellular processes. There are 303 pathways in KEGG database, and 45 of them are annotated as “signaling pathways”. Many of the signaling pathways are important oncogenic signaling pathways20, e.g., EGFR, WNT, Hippo, Notch, PI3K-Akt, RAS, TGFβ, p53. The ‘cell cycle’ cellular process is also included. For simplification, the ‘cell cycle’ is also viewed as one ‘signaling’ pathway. In total, 46 signaling pathways (45 signaling pathways + cell cycle) are selected (see Table II). Among these 46 signaling pathways, there are 1684 genes with both gene expression and copy number variation data. In summary, there are gene expression (TPM) and copy number variation data of 1684 genes in 46 signaling pathways of 45 cancer cell lines, which was used as the input of the deep learning model.

Table II:

The 46 signaling pathways used in the proposed model.

MAPK FoxO TGF-beta T cell receptor Adipocytokine
ErbB Sphingolipid VEGF B cell receptor Oxytocin
Ras Phospholipase D Apelin Fc epsilon RI Glucagon
Rap1 p53 Hippo TNF Relaxin
Calcium mTOR Toll-like receptor Neurotrophin AGE-RAGE
cGMP-PKG PI3K-Akt NOD-like receptor Insulin Cell cycle
cAMP AMPK RIG-I-like receptor GnRH
Chemokine Wnt C-type lectin receptor Estrogen
NF-kappa B Notch JAK-STAT Prolactin
HIF-1 Hedgehog IL-17 Thyroid hormone
Open in a new tab

2.3. Drug-Target interactions from DrugBank database

DrugBank21 is a widely used database to retrieve the information of drugs, such as drug name, chemo-structure, drug mechanism as well as comprehensive drug target information. There are more than 13,000 drug entries in the latest release of DrugBank (version 5.1.5, released 2020-01-03). Among these entries, 2,630 are FDA approved small molecule drugs, and about 6,355 are investigational agents (not approved yet). In total, there are 15263 drug-target interactions between 5435 drugs/investigational agents and 2775 targets. Among the drugs in NCI ALMANAC, 67 drugs are included in DrugBank with targets; and 21 (see Table III) drugs with targets on the 1684 signaling pathways were kept as the input of the model.

Table III:

The 21 drugs used in the proposed model.

Celecoxib Gefitinib Quinacrine hydrochloride Tretinoin
Cladribine Imatinib mesylate Romidepsin Vinblastine sulfate (hydrate)
Dasatinib Lenalidomide Sirolimus Vorinostat
Docetaxel Mitotane Sorafenib tosylate Thalidomide
Everolimus Nilotinib Tamoxifen citrate Paclitaxel
Fulvestrant
Open in a new tab

2.4. Architecture of DeepSignalingSynergy

Fig. 1 shows the schematic architecture of the proposed DeepSignalingSynergy model. In the ‘input layer’, there are 4 input features, i.e., gene expression (RNA-seq TPM values), copy number, is_target_of_DA (0: this gene is not a target of Drug_A; 1: this gene is a target of Drug_A), and is_target_of_DB, for each of 1684 genes on cancer cell line CC. For the connections between the ‘gene’ and ‘pathway’ layers, the 1684 genes are connected the 46 signaling pathways, only if a gene is included in a signaling pathway (not dense connections). The output of the ‘46 signaling pathway’ is used as the input of the deep belief network (DBN) (densely connected). The ‘output’ layer is the synergy score of a drug combination < DA, DB> on cancer cell line CC. The mean square error (MSE) is used as the loss function. For the DBN, there are 3 hidden layers: first hidden layer has 256 nodes with the relu activation function; the second hidden layer has 128 nodes with the relu activation function; the third hidden layer has 32 nodes with the relu activation function. The linear activation function is used in the output layer.

Figure 1:

Figure 1:

Open in a new tab

Schematic architecture of the proposed DeepSignalingSynergy model.

3. Results

3.1. Evaluation of drug combination prediction of DeepSignalingSynergy

There were about 5658 synergy scores of 21 drugs on 45 cancer cell lines, i.e., <DA, DB, CC, SABC>. To evaluate the performance of the DeepSignalingSynergy model, 5-fold cross validation was used. Specifically, we divided the dataset into 5 folds. Then 4 folds were used as training datasets (80%) and one fold was used as validation datasets (20%) 5 times respectively. We empirically tested a few hyperparameters, like training epochs and dropout rates, of the model, which have the stable performance on the validation datasets. The following results were obtained with the model trained with 30 epochs. The Pearson correlation was used as the metric for the model performance evaluation. Fig. 2 shows the evaluation results on the training and validation datasets. As seen, the proposed model had the average Pearson correlation coefficients, 0.795326 and 0.6414188 on the 5 training datasets and validation datasets respectively (see Table IV). This result indicated that the performance of the proposed model, using a small set of signaling pathways, is relatively low but were comparable with other existing deep learning models using a large number of chemical-structure and genomics features reported17, like AuDNNsynergy17 and DeepSynergy16 (which had the Pearson correlation coefficients of 0.74 and 0.73 respectively).

Figure 2:

Figure 2:

Open in a new tab

Scatter plots of the predicted and experimental synergy scores on the 5-fold training and test datasets respectively.

Table IV:

MSEs and Pearson correlation coefficients on the 5-fold cross validation training and validation datasets.

Data set Epoch number MSE Pearson Correlation
Training-1 30 33.44123 0.788073
Validation-1 45.45195 0.655341
Training-2 30 30.961385 0.806522
Validation-2 49.404432 0.643726
Training-3 30 32.126069 0.797911
Validation-3 51.863341 0.650089
Training-4 30 35.2655924 0.772891
Validation-4 46.7580787 0.636611
Training-5 30 31.876059 0.811233
Validation-5 41.687952 0.621327
Open in a new tab

3.2. Importance of signaling pathways analysis for understanding potential mechanism of synergy

To investigate the potential mechanism of synergy in terms of the contributions of individual signaling pathways to the prediction of synergy of drug combination, we employed the Layer-Wise Relevance Propagation (LRP) approach implemented in the “iNNvestigate” package22, which can be used to visualize the importance of individual inputs at different layers. Fig. 3 shows the density distribution maps of importance scores of 46 signaling pathways on the first testing dataset. The results of 5 testing datasets indicated that the importance of the individual signaling pathways is relatively stable in drug combination prediction. Though the importance scores are positive or negative in different validation datasets, the rough range and values of absolute importance scores are consistent. First, some of the 46 signaling pathways, e.g., the MAPK, TGF-β, cell cycle, AMPK, RAS, JAK-STAT, NOTCH, HIF-1a signaling pathways have much more important than other oncogenic signaling pathways. Second, the Apelin, Adipocytokine, Fc epsilon RI, Neurotrophin, Insulin, T-Cell receptor, and IL-17 signaling pathways surprisingly contribute to the drug combination response prediction. Third, some signaling pathways showed the similar interesting distributions, e.g., the MAPK and RAS signaling pathways, the FoxO and cAMP signaling pathways, as well as Apelin and Neurotrophin signaling pathways. Fourth, other oncogenic signaling pathways, like the mTOR, ER, Hippo, Rap1, ERbB signaling pathways, can only contribute to the drug combination synergy prediction moderately. Though results are interesting, further investigations are needed to understand and explain the roles of individual signaling pathways and their associations with drug combination response.

Figure 3:

Figure 3:

Open in a new tab

Distribution of importance of 46 signaling pathways on the first validation dataset.

3.3. Importance of individual genes

We conducted a similar analysis to investigate the importance of individual genes. Fig. 4 shows the top 50 genes with the largest absolute importance scores of 1684 genes on the first 5-flod validation dataset. The selected top 50 genes (out of 1684 genes, which are much more than 46 signaling pathways) are not so consistent. The common genes selected in all the 5 validation datasets are: 'PRKCG', 'FLT1', 'CSF1R', 'JUN', 'BCL2', which indicate the potential synergy among these targets. It can be possible to understand the mechanism of synergy further by investigating the importance scores of individual genes and pathways for a specific synergy drug combination on a specific cancer cell line. However, it is still challenging to associate the importance scores to specific synergy mechanism of drug combinations.

Figure 4:

Figure 4:

Open in a new tab

Importance of individual genes on the first testing dataset.

4. Discussion and conclusion

Synergistic drug combinations are important factors in reducing drug resistance in cancer therapy. Computational models that can integrate multi-omics data of cancer patients with pharmacogenomics data of drugs and investigational agents are needed to predict potential synergistic drug combinations (to narrow down the search space of drug combination). The combination of computational and experimental models can facilitate the discovery of synergistic drug combinations in a fast manner.

Deep learning models have been widely used and outperform the traditional machine learning models in image analysis, natural language processing, healthcare data analysis, and drug combination prediction. However, it is challenging to make the model explainable, especially the models with a large number of features and parameters. In the existing deep learning models of drug combination prediction, a large number of chemical-structure and genomics features are used via the densely connected layers, which requires the training of a large number of parameters. However, only small sets of drug combination experimental validation results that can be used as training labels are available. Thus, it is hard to train a large number of parameters well, and it is also hard to explain the model to investigate the potential mechanism of drug combination synergy.

In this study, we propose to reduce the number of parameters in deep learning models by using a simplified deep learning model built based on a set of biological meaningful signaling pathways. In the model, we can integrate multi-omics data of individual genes and drug-target information, and link the genes to 46 pathways in a sparse manner with a much fewer number of parameters (compared to densely connected layers). The evaluation results showed that the proposed simplified model can achieve good prediction results in terms of the Pearson correlation coefficient between the predicted and experimental synergy scores. Moreover, the explainable analysis of the deep learning model identified some interesting results in terms of the importance of individual signaling pathways that contribute to the drug combination synergy. Further analyses are needed to investigate the unclear mechanisms of synergy using these signaling pathways.

This is our first exploratory study to investigate and prediction drug combination synergy with a simplified deep learning model with increased possibility of model explanation. There are some limitations of the proposed model that need to be further addressed. First, STITCH23 and PharmGKB24,25 database can provide much more drug-target or drug-genetic biomarker interactions, in addition to drug-target interactions obtained from DrugBank. With more drug-target interactions, more drugs can be included in the model, and the prediction accuracy could be better. Second, in addition to the 46 signaling pathways, other KEGG pathways, like metabolism pathways, will be further evaluated. Also, other signaling pathway database, Reactome26, can be helpful to include more pathways and biological processes. Third, Gene oncology27 (GO) terms provide alternative biological meaningful biological processes (BP) (gene sets), which can cover many more genes (drug targets) and can be used for drug combination prediction. Third, other omics data, like protein, methylation, genetic mutation can be integrated conveniently to the model in addition to the copy number, gene expression data. Fourth, the signaling crosstalk and interactions among multiple pathways are important for drug combination discovery, and should be further investigated in the computational models. We will investigate these possible directions in the future work, and we will develop novel graph neural network (GNN) models to uncover the explainable mechanisms of synergy of effective drug combinations, e.g., the synergy mechanism of RAS/ERK inhibitors and Autophagy inhibitors recently reported1,2, which can provide clues to discover novel synergistic drug combinations to reduce drug resistance in cancer therapy. Moreover, it is an interesting idea to study the knowledge graph engineering-oriented approaches, e.g., Resource Description Framework (RDF) and Web Ontology Language (OWL), to integrate the diverse and heterogeneous data resources. The existing inference models on the knowledge graph representation can be useful for drug and drug combination prediction.

Figures & Table

References

  • 1.Kinsey C. G., et al. Protective autophagy elicited by RAF→MEK→ERK inhibition suggests a treatment strategy for RAS-driven cancers. Nat. Med. 2019. doi:10.1038/s41591-019-0367-9. [DOI] [PMC free article] [PubMed]
  • 2.Bryant K. L., et al. Combination of ERK and autophagy inhibition as a treatment approach for pancreatic cancer. Nat. Med. 2019. doi:10.1038/s41591-019-0368-8. [DOI] [PMC free article] [PubMed]
  • 3.Regan-Fendt K. E., et al. Synergy from gene expression and network mining (SynGeNet) method predicts synergistic drug combinations for diverse melanoma genomic subtypes. npj Syst. Biol. Appl. 2019;5 doi: 10.1038/s41540-019-0085-4. doi.org/10.1038/s41540-019-0085-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Holbeck S. L., et al. The National Cancer Institute ALMANAC: A comprehensive screening resource for the detection of anticancer drug pairs with enhanced therapeutic activity. Cancer Res. 2017;77:3564–3576. doi: 10.1158/0008-5472.CAN-17-0489. doi: 10.1158/0008-5472.CAN-17-0489. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Menden M. P., et al. Community assessment of cancer drug combination screens identifies strategies for synergy prediction. bioRxiv bioRxiv. 2017;200451 doi:10.1101/200451. [Google Scholar]
  • 6.Held M. A., et al. Genotype-selective combination therapies for melanoma identified by high-throughput drug screening. Cancer Discov. 2013;3:52–67. doi: 10.1158/2159-8290.CD-12-0408. doi:10.1158/2159-8290.CD-12-0408. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.O’Neil J., et al. An Unbiased Oncology Compound Screen to Identify Novel Combination Strategies. Mol. Cancer Ther. 2016;15:1155–62. doi: 10.1158/1535-7163.MCT-15-0843. doi:10.1158/1535-7163.MCT-15-0843. [DOI] [PubMed] [Google Scholar]
  • 8.Lamb J., et al. The connectivity map: Using gene-expression signatures to connect small molecules, genes, and disease. Science (80-.) 2006;313:1929–1935. doi: 10.1126/science.1132939. doi:10.1126/science.1132939. [DOI] [PubMed] [Google Scholar]
  • 9.Subramanian A., et al. A Next Generation Connectivity Map: L1000 Platform and the First 1,000,000 Profiles. Cell. 2017;171:1437–1452. doi: 10.1016/j.cell.2017.10.049. doi:10.1016/j.cell.2017.10.049. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Regan K. E., Payne P. R. O., Li F. Integrative network and transcriptomics-based approach predicts genotype- specific drug combinations for melanoma. AMIA Jt. Summits Transl. Sci. proceedings. AMIA Jt. Summits Transl. Sci. 2017. pp. 247–256. [PMC free article] [PubMed]
  • 11.Huang L., et al. DrugComboRanker: Drug combination discovery based on target network analysis. Bioinformatics. 2014;30:i228–i236. doi: 10.1093/bioinformatics/btu278. doi: 10.1093/bioinformatics/btu278. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Shukuya T., et al. The Effect of LKB1 Activity on the Sensitivity to PI3K/mTOR Inhibition in Non-Small Cell Lung Cancer. J. Thorac. Oncol. 2019;14:1061–1076. doi: 10.1016/j.jtho.2019.02.019. doi.org/10.1016/j.jtho.2019.02.019. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Chen X., et al. NLLSS: Predicting Synergistic Drug Combinations Based on Semi-supervised Learning. PLoS Comput. Biol. 2016. doi:10.1371/journal.pcbi.1004975. [DOI] [PMC free article] [PubMed]
  • 14.Xu J., et al. Diffusion mapping of drug targets on disease signaling network elements reveals drug combination strategies. in Proceedings of tpacific symposium on biocomputing. 2018. pp. 92–103. doi: 10.1142/9789813235533_0009. [PubMed]
  • 15.Li H., Li T., Quang D., Guan Y. Network propagation predicts drug synergy in cancers. Cancer Res. 2018. doi: 10.1158/0008-5472.CAN-18-0740. [DOI] [PubMed]
  • 16.Preuer K., et al. DeepSynergy: Predicting anti-cancer drug synergy with Deep Learning. Bioinformatics. 2018 doi: 10.1093/bioinformatics/btx806. doi: 10.1093/bioinformatics/btx806. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Zhang T., Zhang L., Payne P., Li F. Synergistic Drug Combination Prediction by Integrating Multiomics Data in Deep Learning Models. arXiv Prepr arXiv1811.07054. 2018 doi: 10.1007/978-1-0716-0849-4_12. [DOI] [PubMed] [Google Scholar]
  • 18.Barretina J., et al. The Cancer Cell Line Encyclopedia enables predictive modelling of anticancer drug sensitivity. Nature. 2012;483:603–7. doi: 10.1038/nature11003. doi.org/10.1038/nature11003. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Ogata H., et al. KEGG: Kyoto encyclopedia of genes and genomes. Nucleic Acids Research. 1999;28 doi: 10.1093/nar/27.1.29. doi: 10.1093/nar/27.1.29. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Sanchez-Vega F., et al. Oncogenic Signaling Pathways in The Cancer Genome Atlas. Cell. 2018;173:321–337. doi: 10.1016/j.cell.2018.03.035. e10. doi:10.1016/j.cell.2018.03.035. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Wishart D. S., et al. DrugBank 5.0: A major update to the DrugBank database for 2018. Nucleic Acids Res. 2018;46:D1074–D1082. doi: 10.1093/nar/gkx1037. doi: 10.1093/nar/gkx1037. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Alber Maximilian, Lapuschkin Sebastian, Seegerer Philipp, Hägele Miriam, Schütt Kristof T., Montavon Grégoire, Samek Wojciech, Müller Klaus-Robert, Sven Dähne P.-J. K. iNNvestigate neural networks! ArXiv. 2018.
  • 23.Szklarczyk D., et al. STITCH 5: augmenting protein-chemical interaction networks with tissue and affinity data. Nucleic Acids Res. 2015;44:D380–D384. doi: 10.1093/nar/gkv1277. doi: 10.1093/nar/gkv1277. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Barbarino J. M., Whirl-Carrillo M., Altman R. B., Klein T. E. PharmGKB: A worldwide resource for pharmacogenomic information. Wiley Interdiscip. Rev. Syst. Biol. Med. 2018;10:e1417–e1417. doi: 10.1002/wsbm.1417. doi: 10.1002/wsbm.1417. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25.Thorn C. F., Klein T. E., Altman R. B. PharmGKB: the Pharmacogenomics Knowledge Base. Methods Mol. Biol. 2013;1015:311–320. doi: 10.1007/978-1-62703-435-7_20. doi: 10.1007/978-1-62703-435-7_20. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Croft D., et al. Reactome: a database of reactions, pathways and biological processes. Nucleic Acids Res. 2011;39:D691–D697. doi: 10.1093/nar/gkq1018. doi: 10.1093/nar/gkq1018. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Gene Ontology Consortium T., et al. Gene Ontology: tool for the unification of biology NIH Public Access Author Manuscript. Nat Genet. 2000;25:25–29. doi: 10.1038/75556. doi: 10.1038/75556. [DOI] [PMC free article] [PubMed] [Google Scholar]

Articles from AMIA Annual Symposium Proceedings are provided here courtesy of American Medical Informatics Association

RESOURCES