Technical Summary: PltScanner

Problem Statement

Single-cell RNA sequencing (scRNA-seq) has revolutionized the dissection of cellular composition in complex tissues. However, accurate annotation of specific cell types remains a challenge, particularly for platelets. Due to their anucleate nature and extremely low RNA content, platelets generate weak transcriptional signals that are frequently obscured by technical noise. Consequently, they are often overlooked or misclassified as other cell types in standard scRNA-seq workflows. While automated annotation tools (e.g., SingleR, CellTypist, Garnett) exist, they are primarily optimized for common immune or tissue cell types and lack dedicated support for specialized, low-RNA cells like platelets. Existing methods often require labor-intensive retraining or the construction of custom reference datasets to identify such rare populations, limiting systematic characterization of platelet heterogeneity and function.

Methodology

The authors developed PltScanner, a lightweight, ensemble classifier designed specifically for the automatic annotation of platelets in scRNA-seq data. The methodology proceeds through the following stages:

1. Data Collection and Feature Selection

• Data Sources: The model was trained using a positive dataset of 1,260 platelet samples and a negative dataset of 778 non-platelet samples (immune, endothelial, fibroblast cells) derived from bulk transcriptomic data. Bulk data was preferred over scRNA-seq for feature selection due to higher sequencing depth and signal-to-noise ratios.
• Rank-Based Strategy: To mitigate batch effects inherent in multi-source data, the authors employed a feature selection strategy based on relative gene expression ranks rather than absolute expression values. Genes were ranked within each sample, and the median rank differences between positive and negative groups were calculated.
• Stable Reversal Pairs: The algorithm identified "stable reversal gene pairs" (genes where the rank order flips between platelets and non-platelets). A voting mechanism was applied: if the majority of informative gene pairs supported a positive assignment, the sample was classified as a platelet.
• Optimal Feature Set: Through 5-fold cross-validation, the optimal number of features was determined to be 15: 14 marker genes (7 positive, 7 negative) and one auxiliary feature, nFeature_RNA (total number of genes detected per cell), which accounts for the low transcript complexity of platelets.

2. Model Construction

• Algorithm: The final classifier utilizes XGBoost (Gradient Boosting Decision Trees).
• Handling Class Imbalance: Given the scarcity of platelets in scRNA-seq datasets, the authors addressed class imbalance via random undersampling. They generated 10 stratified subsets of negative cells, each paired with an equal number of platelet cells to create balanced training sets.
• Ensemble Learning: 10 sub-models were trained on these balanced subsets. The final PltScanner model aggregates predictions from these sub-models to generate the final classification.

3. Evaluation and Benchmarking

• Datasets: The model was trained on healthy control cells from the GSE158055 dataset (1.46 million cells) and evaluated on independent COVID-19 samples from the same dataset, as well as external datasets (e.g., Domínguez Conde et al., TCGA, spatial transcriptomics).
• Comparators: Performance was benchmarked against standard machine learning algorithms (LR, SVM, KNN, NB, BP Neural Network) and existing annotation tools (SingleR, Garnett, CellTypist, scBERT).

Key Results

1. Classification Performance

PltScanner demonstrated superior performance compared to existing tools and algorithms:

• Metrics: It achieved a Sensitivity (SE) of 0.9980, Specificity (SP) of 0.9708, and an AUC of 0.9750.
• Efficiency: PltScanner had the shortest prediction time (0.49 seconds) among all tested methods, significantly outperforming tools like SingleR (>1600s) and scBERT (~1178s).
• Comparison: While SingleR and CellTypist showed high specificity, they missed a significant proportion of platelets (lower sensitivity). Garnett showed poor specificity. PltScanner provided the best balance of accuracy and computational efficiency.

2. Discovery of Platelet Heterogeneity

Re-annotation of a pan-tissue scRNA-seq dataset (Domínguez Conde et al.) using PltScanner identified 818 platelets, a substantial increase from the original 242 annotated "Early Megakaryocytes."

• Subpopulations: Unsupervised clustering of these identified platelets revealed five distinct transcriptional subpopulations.
• Characterization:
  • Subpopulations 2 & 4: Co-expressed platelet and endothelial markers.
  • Subpopulation 3: Co-expressed platelet and macrophage markers (Macrophage-associated platelets, MAPs).
  • Subpopulations 0 & 1: Co-expressed platelet and lymphocyte markers (Lymphocyte-associated platelets, LAPs).
• Validation: These subpopulations showed distinct tissue distributions (e.g., LAPs were reduced in tumor tissues compared to normal tissues) and expressed canonical platelet markers (PPBP, PF4, ITGA2B, ITGB3) alongside lineage-specific genes.

3. Tumor Microenvironment Interactions

In tumor scRNA-seq datasets (breast cancer and intrahepatic cholangiocarcinoma), PltScanner identified interactions between platelet subpopulations and tumor cells:

• Signaling Pathways: Distinct platelet subgroups engaged in subtype-specific signaling with tumor cells.
  • MAPK signaling was enriched across all platelet-tumor interactions.
  • PI3K-Akt and Focal Adhesion pathways were enriched in Traditional Platelets (TPs) interacting with tumors, suggesting roles in survival and adhesion.
  • Cellular Senescence and Th17 differentiation pathways were enriched in interactions involving MAPs, suggesting immunoregulatory roles.

4. Cross-Modal Applicability

• Bulk RNA-seq: Using a reference matrix derived from scRNA-seq subpopulations, PltScanner enabled deconvolution of TCGA bulk data, revealing tumor-type-specific variations in platelet subpopulation abundance (e.g., LAPs and TPs decreased in cholangiocarcinoma, while MAPs increased in breast cancer).
• Spatial Transcriptomics: Application to spatial data from ovarian and lung cancers showed that high PltScanner scores localized to vascular-rich regions, consistent with known platelet biology.

Significance and Claims

The paper claims that PltScanner fills a critical methodological gap by providing the first dedicated, lightweight tool for accurate platelet identification in scRNA-seq data. Its significance lies in three main areas:

1. Methodological Innovation: By utilizing a rank-based feature selection strategy and an ensemble learning approach, PltScanner overcomes the challenges of low RNA content and batch effects, offering a robust solution for a cell type previously difficult to annotate automatically.
2. Biological Insight: The tool facilitates the discovery of platelet heterogeneity, revealing that platelets are not a monolithic population but consist of distinct subgroups (e.g., MAPs, LAPs) with unique molecular signatures and potential functions in immune regulation and tumor progression.
3. Broad Applicability: The study demonstrates that PltScanner is not limited to scRNA-seq but can be extended to bulk and spatial transcriptomic data, providing a framework for investigating platelet distribution and function across diverse tissue microenvironments and disease contexts.

The authors position PltScanner as a user-friendly, efficient resource (available as an R package on GitHub) that enables researchers to systematically characterize platelet roles in health and disease without the need for complex, custom reference construction.