Use the DeepDR model to predict a query sample’s response to 265 anti-cancer drugs. Three prebuilt models are implemented for mutation data alone, gene expression data alone, and both. DeepDR has a transfer learning design that incorporates features of tumors and cell lines, and is applicable to predict both sample types. Drug responses are represented by IC50 (in log(uM)) values, where a more negative value denotes a stronger inhibitory effect.Learn more about input data format and download examples
Search for cell lines across the Cancer Cell Line Encyclopedia (CCLE; n = 704) or tumors of the Cancer Genome Atlas (TCGA; n = 9,059) that have similar genomic features to those of a query sample. Given mutation or gene expression data (or both) of a sample, our model calculates its similarity with samples in our database by the Jaccard index (mutations) and/or the Pearson correlation (expression). Drug responses are represented by IC50 (in log(uM)), where a more negative value denotes a stronger inhibitory effect.Learn more about input data format and download examples
This supplemental module provides access to pre-calculated predictions of drug responses across all TCGA samples as documented in the original DeepDR paper. The module explores the relationships between drug sensitivities and gene mutations or aberrant expressions. Users can interact with the tool by selecting the symbol of a specific gene and the type of alteration of interest (mutated vs. wild-type; top 25% high expression vs. others; or bottom 25% low expression vs. others). The module identifies drugs achieving significantly enhanced responses in tumors harboring specific gene alterations in a pan-cancer or cancer type-specific manner.
DeepDR is a deep learning model that predicts drug sensitivity using mutation and/or gene expression profiles in a cancer sample (cell line or tumor) (Chiu et al. BMC Medical Genomics 2019). shinyDeepDR is an R Shiny app that provides a user-friendly interface to DeepDR with no requirement of programming skills. shinyDeepDR runs two main analysis modules with user's uploaded genomic data, Module 1: Find Drug (core module) and Module 2: Find Sample. Here we provide an overview of our tool. For details regarding the input data format or each module, please visit corresponding help pages or click on the question mark at each input/output section.
DeepDR: Deep Learning-Based Predictiion of drug response of tumors from integrated genomic profiles
Given a pair of mutation and expression profiles, DeepDR predicts IC50 values of 265 FDA-approved or investigational anti-cancer compounds. It contains three deep neural networks:
In the publication of DeepDR, we have systematically evaluated the prediction performance using hold-out cancer cell lines by multiple measures, including mean squared error in drug response and per-cancer cell line correlation coefficients between real and predicted data (Pearson and Spearman correlation coefficients, 0.70-0.96 and 0.62-0.95, respectively). DeepDR achieved marked improvement over conventional methods, including linear regression, support vector machine, and alternative deep learning models trained either with cell lines alone without transferring features learned from tumors, or using principal components to replace encoder outputs. Furthermore, we validated the predictions by real-world clinical data of corresponding patients, such as an approved estrogen receptor agonist (tamoxifen) for breast cancer, approved drugs targeting the EGFR mutations (afatinib and gefitinib) for non-small cell lung cancer, and an investigational compound, CX-5461, for treating hematopoietic malignancies. Please refer to the publication of DeepDR (Chiu et al. BMC Medical Genomics 2019) for more details regarding model design, validation, and applications.
We also built mutation-alone and expression alone models for sampls with only one omics data available.
shinyDeepDR: Implementation of DeepDR by a User-Friendly R Shiny Framework
Module 1: Find Drug that is predicted to be effective in inhibiting the query sample
Module 1 runs DeepDR to predict the query sample’s response to 265 anti-cancer compounds. The app provides an intuitive user interface to interactively visualize, search, and filter all prediction results, as well as detailed annotations of individual compounds.
Module 2: Find Sample that shares similar genomic features to the query sample
Module 2 searches for cell lines in the CCLE or tumors of TCGA that have similar genomic features to those of the query sample and examines their real or predicted drug responses. We incorporated real drug response data of pan-cancer cell lines from the Genomics of Drug Sensitivity in Cancer (GDSC; 265 compounds x 704 cell lines) and prediction by our DeepDR model, as well as predicted drug responses of pan-cancer tumors from our publication of DeepDR (265 compounds x 9,059 tumors).
If you have any questions or comments about shinyDeepDR, please contact:
NIH/NCI - K99/R00 Pathway to Independence Award (R00CA248944)
Deep learning of drug sensitivity and genetic dependency of pediatric cancer cells
NIH/OD - Common Fund R03 (R03OD036494)
In silico screening for immune surveillance adaptation in cancer using Common Fund data resources
NIH/OD – Administrative Supplement (3R00CA248944-04S1)
Enhancing AI-readiness of multi-omics data for cancer pharmacogenomics
UPMC Hillman Cancer Center Developmental Pilot Program (P30CA047904)
Pittsburgh Liver Research Center (PLRC) Pilot and Feasibility Grant (P30DK120531)
Leukemia Research Foundation New Investigator Research Grant Program