Deep Transfer Learning of Drug Responses by Integrating Bulk and Single-cell RNA-seq data

Abstract

Massively bulk RNA sequencing databases incorporating drug screening have opened up an avenue to inform the optimal clinical application of cancer drugs. Meanwhile, the growing single-cell RNA sequencing data contributes to improving therapeutic effectiveness by studying the heterogeneity of drug responses for cancer cell subpopulations. Yet, the drug response information for single-cell data is scarcely obtained. Thus, there is an urgent need to develop computational pipelines to infer and interpret cancer drug responses in single cells. Here, we developed scDEAL, a deep transfer learning framework integrating large-scale bulk and single-cell RNA sequencing drug response datasets. We benchmarked scDEAL on six single-cell RNA sequencing datasets and indicate its model interpretability by several case studies. scDEAL not only achieves accurate and robust performance in single-cell drug response predictions, but also can infer signature genes to reveal potential drug resistance mechanisms based on integrated gradient feature interpretation. This work may help study cell reprogramming, drug selection, and repurposing for improving therapeutic efficacy.

Inferring therapeutic vulnerability within tumors through integration of pan-cancer cell line and single-cell transcriptomic profiles

“…To date, a few other methods have been proposed to computationally predict single cell drug response using drug screen data on CCLs, such as Beyondcell, CaDRRes-sc, SCAD, and scDEAL [21][22][23]26 . Each of these methods employ a unique transfer learning-like approach, utilizing relationships between CCL expression data and drug response to predict single cell level drug sensitivity.…”

“…While computational tools utilizing relationships between CCL gene expression from RNA-seq and drug response have demonstrated practicality in predicting efficacious treatments 13,[17][18][19] , such relationships cannot be directly applied to generate predictions of drug response at the cellular level, as RNA-seq is limited to measuring average expression across a diverse set of cells, which obscures cell type and composition, as well as temporal and spatial distributions. Thus, inferring cellular drug response requires specialized tools to transfer current bulk-learned drug-gene information to single-cell RNA sequencing (scRNA-seq) data that encapsulate cell level expression patterns [20][21][22][23][24] .…”

Enabling Single‐Cell Drug Response Annotations from Bulk RNA‐Seq Using SCAD

“…First, we adopted IntegratedGradients [41] method to infer the subset of genes that contributes the most to drug sensitivity prediction [42] (see Experimental Section). The IntegratedGradients is a method to infer how much an independent variable affects the value of the prediction output from the model, which is one of the most widely used ML model explainer to interpret the association between input data and machine learning model outputs.…”

Deep learning analysis of single‐cell data in empowering clinical implementation

“…15 scDEAL, a deep transfer learning framework integrating large-scale bulk and scRNA-seq data, adapts a domain-adaptive neural network to predict single-cell drug responses from scRNA-seq data by integrating and harmonizing large-scale drug response data of bulk cancer cell lines; it does not depend on predefined single-cell labels. 16 It can further predict critical genes that significantly contribute to drug sensitivity and resistance prediction. In another study, a convolutional neural network (CNN)-based model was designed to predict antitumor drugs for CTCs at the single-cell level.…”

“…Deep transfer learning can transfer knowledge and relationship patterns from bulk data to single‐cell data to overcome the issue of limited training data 15 . scDEAL, a deep transfer learning framework integrating large‐scale bulk and scRNA‐seq data, adapts a domain‐adaptive neural network to predict single‐cell drug responses from scRNA‐seq data by integrating and harmonizing large‐scale drug response data of bulk cancer cell lines; it does not depend on predefined single‐cell labels 16 . It can further predict critical genes that significantly contribute to drug sensitivity and resistance prediction.…”