{
  "abstract": "Background Numerous studies rely on transcriptomic-based annotation to characterize the immune cell-types composition of tumor microenvironment (TME) and blood systemic immunity. However, these annotations remain largely unvalidated, raising reproducibility concerns. Discrepancies between mRNA and protein expression—compounded by scRNA-seq technical limitations—cast doubt on the reliability of transcriptomic-only cell classification. Misannotated immune cells can compromise downstream analyses, including therapeutic target discovery and biomarker development. Establishing ground-truth benchmarks and using them to systematically evaluate scRNA-seq annotation accuracy is therefore a pressing methodological need with potentially broad fundamental implications.Methods We benchmarked seven state-of-the-art scRNA-seq annotation tools—including SingleR (with multiple references), the recent SCimilarity, and CellTypist—against protein-based ground truth, using publicly available CITE-seq datasets. CITE-seq enables simultaneous measurement of gene and surface protein expression, allowing confident cell identity validation and comparison. To this end, we developed a deep learning ensemble framework integrating those multiple SC annotation tools. We additionally generated a new large validation dataset from 40 head and neck cancer (HNC) samples with matched scRNA-seq, CITE-seq, and CyTEK profiles, enabling independent prospective robust cross-modal validation at single-cell resolution.Results Compared to the CITE-seq ground truth established by the original authors of the cohorts investigated, we identified substantial annotation mismatches that challenge the reliability of scRNA-seq transcriptomic immune profiling. While B cells and monocytes were mostly correctly identified (F1 = 1.00 and 0.96, respectively), quite disturbingly, CD8 + T cells and dendritic cells (DCs) were frequently misclassified (F1 = 0.70 and 0.47, respectively) (figure 1). A novel deep learning-based ensemble method we introduced could significantly improve the F1 scores for these populations in held-out test sets (CD8+ T cells and DCs: F1 = 0.87), but could not completely resolve these discrepancies.Conclusions Our findings expose a critical and disturbing annotation mismatch in cancer immunology, which was somehow neglected until now. The systematic misidentification of CD8 + T cells—key effectors in immunotherapy—raises concerns about conclusions drawn from scRNA-seq data alone, particularly regarding T cell exhaustion, activation, and therapeutic response. We demonstrate that transcriptomic data alone is insufficient for reliable annotation and advocate for either proteomic validation or advanced integration strategies. These insights demand reassessment of prior studies and highlight the urgent need for annotation standards to ensure robust, clinically meaningful immune profiling.Abstract 1292 Figure 1Heatmap of the average F1 values of cell types across the different annotation tools",
  "authors": [
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Shai Dulberg"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Tian-Gen Chang"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Dongya Jia"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Chi-Ping Day"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Clint T Allen"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Gregoire Altan-Bonnet"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Eytan Ruppin"
    }
  ],
  "title": "1292 Analysis of single cell transcriptomic profiling uncovers a systematic misidentification of CD8+ T and dendritic cells compared to CITE-seq ground truth annotations",
  "uid": "aac1de4e-6dc7-51e7-97dc-59104b94febd"
}