Towards generalization of drug response prediction to single cells and patients utilizing importance-aware multi-source domain transfer learning

TL;DR

This work tackles single-cell drug response prediction by transferring knowledge from bulk-cell-line drug sensitivity data to single-cell and patient contexts through scAdaDrug, a multi-source domain adaptation framework. It introduces an adaptive importance-aware weight generator that assigns sample- and feature-level transfer weights, along with a conditional independence constraint to reduce redundancy, and couples this with adversarial domain alignment and supervised prediction. Across multiple datasets including cell lines, PDXs, and TCGA patients, the approach achieves state-of-the-art performance and demonstrates that increasing the number of source domains and using gene- and pathway-informed inputs enhance generalization. The study highlights the potential of bulk-to-single-cell transfer learning for precision oncology while acknowledging domain gaps and data sparsity challenges, and points toward domain generalization as a future direction. Overall, scAdaDrug provides a robust framework for integrating diverse data sources to predict drug response at the single-cell and patient levels, with practical implications for identifying resistant subpopulations and informing therapy choices.

Abstract

The advancement of single-cell sequencing technology has promoted the generation of a large amount of single-cell transcriptional profiles, providing unprecedented opportunities to identify drug-resistant cell subpopulations within a tumor. However, few studies have focused on drug response prediction at single-cell level, and their performance remains suboptimal. This paper proposed scAdaDrug, a novel multi-source domain adaptation model powered by adaptive importance-aware representation learning to predict drug response of individual cells. We used a shared encoder to extract domain-invariant features related to drug response from multiple source domains by utilizing adversarial domain adaptation. Particularly, we introduced a plug-and-play module to generate importance-aware and mutually independent weights, which could adaptively modulate the latent representation of each sample in element-wise manner between source and target domains. Extensive experimental results showed that our model achieved state-of-the-art performance in predicting drug response on multiple independent datasets, including single-cell datasets derived from both cell lines and patient-derived xenografts (PDX) models, as well as clinical tumor patient cohorts. Moreover, the ablation experiments demonstrated our model effectively captured the underlying patterns determining drug response from multiple source domains.

Figures (6)

• Figure 1: Illustrative diagram of the proposed scAdaDrug architecture for predicting single-cell drug response. It consists of four components: autoencoder-based feature extractor, adaptive weight generator, adversarial domain discriminator and drug sensitivity predictor.
• Figure 2: Performance evaluation on single-cell drug response datasets of cell lines. a The imbalanced number of drug-sensitive and drug-resistant cells in both source and target domain. b-c The scatter plots between the predicted and true log(IC50) values of cell lines treated by Etoposide, PLX4720, respectively. d ROC curve of scAdaDrug for predicting single-cell drug response of PC9 cell line treated by Etoposide; e-f ROC curves of scAdaDrug for predicting single-cell drug response of PLX4720-treated 451Lu and A375 cell lines, respectively. g Performance comparison of scAdaDrug to four competing methods on three single-cell datasets. h UMAP feature plots of single-cell transcriptional profiles and corresponding embeddings output by the trained encoder for Etoposide-treated PC9, PLX4720-treated 451Lu and PLX4720-treated A375 cell lines, respectively.
• Figure 3: Performance evaluation in predicting single-cell drug response of PDX samples of hepatocellular carcinoma (GSE175716). a Volcano plot of differentially expressed genes between sensitive and resistant cells upon sorafenib treatments. b The scAdaDrug predicts labels for drug-resistant and drug-sensitive cells from PDX samples. c-d Violin plots and ROC curves achieved by baseline model and scAdaDrug (2- and 3-source domains) in predicting single-cell drug response. e Normalized expression profiles of top-ranked up-regulated and down-regulated genes. f UMAP plots of the transcriptional profiles with the individual cells colored by actual response labels, binarized predicted labels, continuous predicted probabilities, as well as the learned embeddings of individual cells colored by actual response labels.g-hEnrichment analysis of differentially expressed genes. g-h GO enrichment analysis using the top 600 up-regulated genes in the drug-resistant cells.
• Figure 4: Performance evaluation in predicting single-cell drug response of patient tissues of lung adenocarcinoma (GSE223779). a Volcano plot of differentially expressed genes between sensitive and resistant cells to crizotinib treatments. b The scAdaDrug predicts labels for drug-resistant and drug-sensitive cells derived from patient tissues. c-d Violin plots and ROC curves achieved by baseline model and scAdaDrug (2- and 3-source domains) in predicting single-cell drug responses. e Normalized expression profiles of top-ranked up-regulated and down-regulated genes. f UMAP plots of the transcriptional profiles with the individual cells colored by actual response labels, binarized predicted labels, continuous predicted probabilities, as well as the learned embeddings of individual cells colored by actual response labels.g-hEnrichment analysis of differentially expressed genes.
• Figure 5: Comparison of AUROC and AUPR values between scAdaDrug and scDEAL on five single-cell drug sensitivity datasets.