Datasets and Training

Classification

Datasets

Training

Source Cell States Organ/Tissue Cells No. Paper DOI Author
E-MTAB-3929 Pluripotent-like Embryonic 1356 https://doi.org/10.1016%2Fj.cell.2016.03.023 Petropoulos et al.
E-MTAB-6819 Pluripotent-like Embryonic 1462 https://doi.org/10.1016%2Fj.celrep.2018.12.099 Messmer et al.
GSM5519457 Multipotent-like Dermis 15563 https://doi.org/10.1002/ctm2.650 Wang et al.
GSM5519466 Multipotent-like Adipose 9039 https://doi.org/10.1002/ctm2.650 Wang et al.
GSE221853 Multipotent-like Neuron 5859 https://doi.org/10.1038/s41467-024-47945-7 Guerrero et al.
GSE147482 Multipotent-like Skin 486 https://doi.org/10.1038/s41467-020-18075-7 Wang et al.
GSE248995 Bi/Unipotent-like Neuron 63 https://doi.org/10.1016/j.isci.2024.109342 Baig et al.
GSE136831 Bi/Unipotent-like Lung 1154 https://doi.org/10.1126/sciadv.aba1983 Adams et al.
GSE143704 Bi/Unipotent-like Muscle 2757 https://doi.org/10.1186/s13395-020-00236-3 Micheli et al.

Validation

Source Cell States Organ/Tissue Cells No. Paper DOI Author
GSE36552 Pluripotent-like Embryonic 124 https://doi.org/10.1038/nsmb.2660 Yan et al.
GSM3370006 Pluripotent-like Embryonic 370 https://doi.org/10.1038/s41467-020-16214-8 Lau et al.
GSM5519456 Multipotent-like Dermis 24241 https://doi.org/10.1002/ctm2.650 Wang et al.
GSM5519465 Multipotent-like Adipose 12740 https://doi.org/10.1002/ctm2.650 Wang et al.
Dryad Bi/Unipotent-like Muscle 65085 https://doi.org/10.7554/elife.51576 Barruet et al.

Workflow

  1. Aggregation of Experimentally Annotated Datasets

    • Datasets were aggregated from multiple publicly available sources to train the classifier model.
    • The aggregation included ensuring consistency and quality control measures.

  2. SCENT Score Calculation

    • The SCENT score was calculated for all samples to assess their biological consistency.
    • Samples with original annotations misaligned with their SCENT score were removed, improving dataset accuracy.

  3. Handling Class Imbalance

    • A class imbalance was identified: Multipotent samples were approximately 8 times more abundant than Pluripotent and Unipotent samples.
    • To address this, we employed an ensemble learning strategy.
    • Eight distinct ensemble models were trained, each with shared Pluripotent and Unipotent samples but unique subsets of Multipotent samples, balancing the dataset across models.

  4. Conversion to Rank Space

    • All gene expression data was transformed to rank space to:
      • Remove batch effects.
      • Mitigate the influence of extreme values and outliers.
      • Reduce the risk of overfitting.
    • This transformation normalized gene expression differences across samples for consistent downstream analysis.

  5. Log2 Transformation

    • Rank-transformed data was then log2-transformed to compress the scale of gene expression data.

  6. Standardization to Z-Scores

    • The log2-transformed data was standardized by converting to z-scores, ensuring all variables had a mean of zero and a standard deviation of one.
    • This final step facilitated uniformity and improved model performance during training.

  7. Model Validation on Independent Datasets

    • To ensure robustness and prevent overfitting, traditional k-fold cross-validation was avoided as it can introduce bias when biological replicates are present in both training and validation sets, leading to double-dipping.
    • Instead, models were validated on completely independent datasets from different sources to ensure an unbiased evaluation of model performance.
    • The trained models were evaluated for their ability to generalize to novel, unseen data.

  8. SCENT-Based Smoothening and Test Data Integrity

    • SCENT-based smoothening was not applied to the test data to maintain the integrity of the unseen test data during prediction, ensuring accurate evaluation.

  9. Model Selection

    • Five different models were developed, and the one demonstrating the highest accuracy on the independent test dataset was selected for subsequent analysis.

Deconvolution

Datasets

The single-cell RNA sequencing (scRNA-seq) gene expression matrix for various cancer types was obtained from the publicly available Weizmann Institute's 3CA dataset, which contains high-resolution data on cancer-associated cell types. We validated our deconvolution model using the given datasets.

Source Type Use Organism Source URL Author
GSE157329 Developmental Training/Validation Human https://doi.org/10.1038/s41556-023-01108-w Xu et al.
PBMC 20k Immune Cells Training/Validation Human https://www.10xgenomics.com/datasets 10X Genomics
Tabula Muris Tissue Training/Validation Mouse https://doi.org/10.1038/s41586-018-0590-4 The Tabula Muris Consortium
PBMC 10k Immune Cells Training/Validation Human https://www.10xgenomics.com/datasets 10X Genomics
E-MTAB-5061 Pancreas Training Human https://doi.org/10.1016/j.cmet.2016.08.020 Segerstolpe et al.
GSE81608 Pancreas Validation Human https://doi.org/10.1016/j.cmet.2016.08.018 Xin et al.
SDY67 (Immport NIAID) Validation Human https://doi.org/10.1371/journal.pone.0152034 Zimmermann et al.
GSE107019 PBMC Validation Human https://doi.org/10.1016/j.celrep.2019.01.041 Monaco et al.
GSE65133 PBMC Validation Human https://doi.org/10.1038/nmeth.3337 Newman et al.
GSE93722 Metastatic Melanoma Validation Human https://doi.org/10.7554/elife.26476 Racle et al.
GSE77940, GSE72056 Melanoma Validation Human https://doi.org/10.1126/science.aad0501 Tirosh et al.

Workflow

Data Preprocessing

Preprocessing was performed using the Seurat package (v5.1.0) in R to ensure stringent quality control. The following steps were applied:

  1. Initial Filtering

    • Cells with fewer than 200 genes were excluded to avoid low-quality cells.
    • Cells with more than 6,000 genes were removed to eliminate potential doublets.
    • Cells with a mitochondrial gene fraction >10% were excluded to remove dying or stressed cells.

  2. Doublet Removal

    • The DoubletFinder package (v2.0) was used for doublet detection and removal.
    • Optimal pK parameter was determined using the paramSweep function, and find.pK was used to summarize results.
    • Adjustments were made for homotypic doublets using the nExp_poi.adj parameter.
    • Classified doublets were removed, leaving only high-confidence singlets for downstream analysis.

Malignant and Stromal Cell Annotation

  • Annotations for non-immune populations (malignant cells, fibroblasts, epithelial cells, endothelial cells) were preserved from the Weizmann Institute’s 3CA dataset.

Immune Cell Annotation

  • The SCINA (Semi-supervised Category Identification and Assignment) package was employed for immune cell classification.
  • The LM22 matrix (from the study by Newman et al.) was used as the reference for major immune populations (T cells, B cells, macrophages, NK cells).
  • Custom MDSC markers were curated from external studies (cite: [Science paper on MDSCs]) and combined with the LM22 matrix for MDSC identification.

Cancer Stem Cell (CSC) Annotation

  1. Isolation of Malignant Cells

    • The annotation of CSCs was restricted to malignant cells only to distinguish them from normal stem cells in cancer tissues.

  2. Data Normalization

    • The NormalizeData function was used to normalize gene expression using the LogNormalize method (scales gene expression levels by a factor of 10,000, followed by log transformation).

  3. Detection of Variable Features

    • FindVariableFeatures detected the top 2,000 highly variable genes using the variance-stabilizing transformation (vst) method.

  4. Data Scaling

    • ScaleData was used to standardize expression levels, reducing bias from overrepresented genes.

  5. Dimensionality Reduction

    • Principal Component Analysis (PCA) was performed using the top 2,000 highly variable genes, retaining the first 20 principal components (PCs).

  6. Batch Effect Correction

    • RunHarmony from the Harmony package was employed to correct batch effects, preserving biological variation.

  7. Clustering

    • The FindNeighbors function constructed a nearest neighbor graph.
    • FindClusters applied the Louvain algorithm (resolution = 0.3) to partition cells into clusters.

  8. Marker Expression Visualization

    • Violin plots were generated to visualize known cancer stem cell markers in each cancer type across clusters.

  9. UCell Analysis

    • UCell was used to calculate enrichment scores based on stem cell marker genes tailored to each cancer type.
    • Clusters with high UCell scores and elevated stem cell marker expression were identified as potential CSC clusters.