TL;DR
This paper introduces DrugReSC, a novel computational method that leverages single-cell RNA sequencing (scRNA-seq) data to identify specific, disease-critical cell subpopulations within tumors. By targeting these key cells, DrugReSC aims to improve drug repurposing strategies for cancer, outperforming existing methods that rely on bulk data or less precise single-cell analyses.
What Problem Does This Paper Address?
Cancer is notoriously complex due to intra-tumor heterogeneity, meaning tumors are not uniform but composed of diverse cell types with varying characteristics and responses to treatment. This heterogeneity is a major hurdle in developing effective cancer therapies. While single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to study this heterogeneity at the cellular level, existing drug repurposing methods often fail to fully exploit this rich data. They either rely on bulk RNA sequencing (which averages out cellular differences) or don't adequately pinpoint the specific, often rare, cell populations that drive disease progression and drug resistance. This gap means potentially effective drugs might be overlooked because they target the wrong cells or because their efficacy is masked by the bulk analysis.
Methods
DrugReSC tackles this by integrating scRNA-seq data with drug information in a multi-step process:
- Identification of Disease-Associated Cells: DrugReSC first uses a previously developed method called PACSI, which combines both single-cell and bulk RNA sequencing data, to identify cells that are specifically linked to the disease phenotype (e.g., cancer metastasis).
- Drug Signature Construction: Drug signatures are created by analyzing gene expression changes induced by drugs in cell lines from the LINCS L1000 project. These signatures represent how a drug perturbs cellular gene expression, distinguishing between genes that are upregulated (up-signature) and downregulated (down-signature).
- Drug-by-Cell Matrix Construction: A crucial step is creating a "drug-by-cell" (D2C) matrix. This matrix quantifies the "reversal" effect of a drug on the transcriptional state of each individual cell. This is achieved using the single-sample gene set enrichment analysis (ssGSEA) algorithm to calculate enrichment scores for drug signatures within each cell's gene expression profile. A higher D2C score suggests a drug is effectively reversing the disease-associated transcriptional state of a cell.
- Drug Contribution Assessment: To determine which drugs are most likely to be effective, DrugReSC employs a perturbation-based random forest model. This model uses the D2C matrix to predict cell phenotypes (disease-related vs. disease-unrelated). By permuting drug instance values and observing the impact on prediction accuracy, the model assesses the "importance" or contribution of each drug to distinguishing between these cell phenotypes. Drugs with higher importance scores are considered more promising candidates.
Datasets
The study utilized several publicly available datasets:
- Single-cell RNA sequencing (scRNA-seq) data:
- Melanoma: GSE189889
- Non-small cell lung cancer (NSCLC): GSE140819
- Renal cell carcinoma (RCC): GSE152938
- Normal skin: GSE183047
- Normal lung: GSE122960
- Bulk RNA sequencing data:
- Melanoma: TCGA-SKCM
- NSCLC: TCGA-LUAD
- RCC: TCGA-KIRC
- Drug data:
- DrugBank database for FDA-approved drugs.
- LINCS L1000 project for drug-induced gene expression profiles.
The study also generated "ground truth" lists of positive and unlabeled drugs for each indication by querying DrugBank and randomly selecting negative drug instances from LINCS, with multiple samplings to ensure robustness.
Key Insights and Ideas
- Targeting "Disease-Critical" Cells is Key: The core idea is that focusing on specific, disease-driving cell subpopulations, rather than the entire tumor bulk or all cell types, can reveal more effective drug repurposing candidates.
- Single-Cell Data Unlocks Nuance: DrugReSC demonstrates that the detailed resolution of scRNA-seq is essential for capturing the subtle transcriptional differences that indicate a cell's role in disease and its potential response to drugs.
- Random Forest for Drug Importance: The use of a random forest model with permutation-based feature importance is a robust way to quantify a drug's contribution to distinguishing disease-associated cells, providing a ranked list of candidates.
- Flexibility in Cell Identification: While PACSI is the default, DrugReSC is designed to be adaptable, allowing for the integration of other cell identification methods (like Scissor or DEGAS), showcasing its modularity.
Key Results
- Superior Performance: DrugReSC, particularly when combined with the random forest model for feature importance, consistently outperformed other methods (including bulk-based methods like DrInsight and iLINCS, and single-cell-based ASGARD) across melanoma, NSCLC, and RCC.
- For example, DrugReSC achieved average AUROC values of 0.702, 0.718, and 0.709 for melanoma, NSCLC, and RCC, respectively, significantly higher than ASGARD (0.502, 0.497, 0.457).
- Validation in Cancer Types: The method was successfully applied to melanoma, NSCLC, and RCC, identifying candidate drugs with strong literature support.
- For melanoma, top candidates included FDA-approved drugs like Vemurafenib and Dabrafenib, as well as Wortmannin (under investigation).
- For NSCLC, candidates included Afatinib and Crizotinib, with Doxorubicin identified as a promising non-FDA-approved option.
- For RCC, Tivozanib and Pazopanib were highlighted, alongside Irinotecan (under investigation).
- Correlation with Literature: The predicted drug importance scores showed a significant positive correlation with the number of published studies related to the drug and disease, providing external validation.
Limitations and Future Work
- Directionality of Drug Effects: The current method quantifies drug importance but doesn't explicitly determine if a drug's effect is a "promotion" or "inhibition" of a cellular state. Future work aims to address this.
- Imputation Impact: Imputing missing values in scRNA-seq data (due to dropout rates) actually decreased DrugReSC's performance, suggesting that imputation might introduce noise or smooth out critical biological signals. This highlights the need for careful consideration of imputation strategies.
- Experimental Validation Needed: As a computational method, DrugReSC's predictions require validation through in vitro experiments, animal models, and clinical trials before clinical application.
- Microenvironment Cell Roles: While malignant cells are the primary identified population, the role of other microenvironmental cells (immune, fibroblasts, endothelial) in drug therapy warrants further investigation.
- Adaptability: The authors explored various methods for cell identification and drug-by-cell matrix construction, finding PACSI and ssGSEA to be optimal, but the framework is designed to accommodate alternatives.
Summary updated