• What is CytoTRACE 2?
• Why should I use CytoTRACE 2?
• How does CytoTRACE 2 work?
• What do I need to run CytoTRACE 2?
• Should I normalize the data before running CytoTRACE 2?
• What if my dataset includes rare cell types?
• My data appears to be too large to analyze on the website. What is the file size limit to run CytoTRACE 2 on the website?
• What if I have multiple batches of data? Should I perform any integration?
• Which genes can I explore in the “Explore genes” –> “Custom genes” section?
• Does CytoTRACE 2 work with different platforms, tissues, and species?
• Are there limitations to CytoTRACE 2?
• What are the CytoTRACE 2 potency categories?
• What organism can my data be from?
• I still have so many questions, is there someone I can reach out to for help?

What is CytoTRACE 2?

CytoTRACE 2 is a computational method for predicting cellular potency categories and absolute developmental potential from single-cell RNA-sequencing data, across various tissue types, scRNA-seq platforms, and species data. This framework learns multivariate gene expression programs for each potency category and calibrates outputs across the full range of cellular ontogeny, facilitating direct cross-dataset comparison of developmental potential in an absolute space. Underlying this method is a novel, interpretable deep learning framework trained and validated across 34 human and mouse scRNA-seq datasets encompassing 24 tissue types, collectively spanning the developmental spectrum.

More information on why you should use CytoTRACE 2 and how it works is provided below.

Why should I use CytoTRACE 2?

CytoTRACE 2 was developed to predict differentiaton states in scRNA-seq data without any prior information. Some examples where CytoTRACE 2 may be useful, include:

• Validating differentiation states in tissues with functional evidence of developmental states

• Assessing absolute developmental potential of cells and comparing these across different datasets

• Predicting differentiation states in samples where a continuous developmental process is not captured

• Predicting differentiation states in human tissues where developmental hierarchies are poorly understood

• Predicting differentiation states in diseased tissues, such as cancer

• Identifying cells and genes associated with stemness and differentiation

• Associating cancer cell differentiation states with survival, metastasis, response to therapies, and other clinical outcomes

How does CytoTRACE 2 work?

CytoTRACE 2 is designed to assess the absolute developmental potential of cells using a three-step process in its core model architecture. Here’s a simplified breakdown:

1. Preprocessing:

   • Ortholog Mapping and Normalization: CytoTRACE 2 first aligns and filters input genes against a unified dictionary of 14,271 mouse/human orthologs, accounting for alias gene names, mapping unused genes to zero. It then transforms the gene expression data into two representations:
     • Log2-adjusted CPM/TPM to capture detailed transcriptomic signals.
     • Rank space to dampen extreme values and reduce noise. This dual representation provides a robust foundation for subsequent computations.

2. Prediction:

   • Gene Set Binary Networks: The core model features 19 modules each containing a geneset binary weight matrix for each of the 6 potency categories. Within each module, through numerous matrix operations with the exact weights and parameters learned by training an ensemble model on our training cohort, rank-based and log2 expression-based scores are computed and combined to capture distinct features of gene set activity.
   • Integration of Scores: Scores from these operations across potency categories for each member model of the ensemble, across all models, are obtained. For each model both continuous potency scores and softmax output prediction probabilities (to be later mapped to discrete potency categories) are then aggregated to form a consensus prediction of the ensmeble.

3. Postprocessing: The raw predictions of the previous step are then passed through 3 layers of post-processing:

   Markov Diffusion Smoothing:

   • Utilizes a Markov matrix based on transcriptional similarities to smooth raw potency scores (RPS), resulting in less noisy and more biologically accurate smoothed potency scores (SPS).

   Binning Procedure:

   • Cells are ranked within their predicted potency categories and uniformly partitioned across the potency spectrum. This ensures that each category spans a specific segment of the range from 0 to 1, refining the smoothed scores (SPS^B) to maintain relative order.

   Adaptive Nearest Neighbor Smoothing:

   • Further refines the binned smooth potency scores (SPS^B) for datasets with more than 100 cells by adaptively selecting a neighborhood for each cell, and recalculates each cell’s potency score as a distance-weighted average of its neighbors including itself.

The resulting values are continuous scores ranging from 0 and 1 and discrete labels, representing the predicted absolute developmental potential of cells.

What do I need to run CytoTRACE 2?

All you need to run CytoTRACE 2 is a gene expression matrix generated by single-cell RNA-sequencing, where columns are cells and rows are genes/transcripts. For the website, we require this file to be a text (txt), tab separated value (tsv), or comma separated value (csv) file containing < 5,000 cells. For datasets containing > 5,000 cells, please use the R or Python package implementation. You’ll need to specify your input species (“human” or “mouse”). Phenotype tables are optional, but if included, should be formatted so that the first column contains single cell IDs matching the columns of the gene expression matrix and the second column must contain the labels.

For more details on how to format and upload your data to the website, please refer to the Tutorial page (see sidebar for link).

Should I normalize the data before running CytoTRACE 2?

There is no need to normalize data prior to running CytoTRACE 2, provided there are no missing values and all values are non-negative. The input needs to be raw counts or CPM/TPM data, and should not be log-transformed, as the model employs normalization methods internally.

What if my dataset includes rare cell types?

When analyzing phenotypes expected to have five or fewer cells, we recommend bypassing the KNN smoothing step so that predictions for these rare cells are not forced toward more abundant phenotypes. In practice, you can simply use the preKNN score output (preKNN_CytoTRACE2_Score) instead of the final KNN-smoothed value (CytoTRACE2_Score). This preserves the original predictions for rare cell types while still benefiting from the other postprocessing steps.

My data appears to be too large to analyze on the website. What is the file size limit to run CytoTRACE 2 on the website?

To manage website traffic, we have limited file size uploads to 800 MB. For all datasets greater than 800 MB, please use the R or Python package implementation. You can visit our Github site for detailed instructions on how to install and run these packages.

I have data from multiple batches. Can I still run CytoTRACE 2?

We recommend running CytoTRACE 2 separately over each dataset. While raw predictions are made per cell without regard to the broader dataset, the postprocessing step to refine predictions adjusts predictions using information from other cells in the dataset, and so may be impacted by batch effects. Note that CytoTRACE 2 outputs, except for Relative Order, are calibrated to be comparable across datasets without further adjustment. Therefore, no integration is recommended over the predictions either. As for Relative Order, it should not be compared across different runs, as it is based on the ranking and scaling of CytoTRACE2 Scores within the context of the specific input dataset.

Which genes can I explore in the “Explore genes” –> “Custom genes” section?

For this section utilize Single Sample GSEA (ssGSEA) to estimate gene set enrichment scores across various cell types, tissues, and potency categories within our 34 gold-standard datasets. These include both training and test datasets employed for model development and validation.You can select one or multiple genes from the dropdown menu to explore their enrichment across dataset cohorts and ground truth potency categories. Of note, the genes currently available for exploration are limited to the feature space of CytoTRACE 2, which is comprised of 14,271 total genes, including mouse gene symbols and human gene symbols mapped to mouse orthologs. Accordingly, all queries should use mouse gene symbols. For a list of supported gene symbols, download here.

Does CytoTRACE 2 work with different platforms, tissues, and species?

Although there are some caveats to note (see section ‘Are there limitations to CytoTRACE 2?’ below), CytoTRACE 2 can be applied to scRNA-seq data from any platform, tissue, and species. We show in our original manuscript (Kang et al., 2022) that CytoTRACE 2 is robust to variation in dataset characteristics. Just to note, CytoTRACE 2 was developed over mouse and human data, so datasets supplied to the website should have human or mouse gene names, following HGNC and MGI gene naming conventions.

Are there limitations to CytoTRACE 2?

There are some situations where direct application of CytoTRACE 2 to the dataset would be suboptimal. These include:

Dependence on Feature Overlap with Model Features: CytoTRACE 2’s performance is closely tied to the overlap between the features in the input data and those used by the model. If the overlap in gene features is less than 10,000 genes, the model may show poorer performance, due to not observing the needed level of signal for accurate predictions.

Sensitivity to Low UMI/Gene Counts: Although generally insensitive to variation in gene/UMI counts per cell, CytoTRACE 2 requires further optimization for cells that have exceedingly low gene expression levels, particularly those with fewer than 500 genes expressed per cell, as its performance can become less reliable in these cases. For best results, a minimum gene count of 500-1000 per cell is recommended.

Limitation in Species Generalization: While CytoTRACE 2 is designed for human and mouse data, you can attempt to map gene names from other species to mouse orthologs to utilize the model. However, this approach is experimental and does not guarantee reliable performance without further testing and optimization.

Verification of Predictions: Although the differentiation states and molecular signatures predicted by CytoTRACE 2 are biologically aware and promising, they are preliminary and intended to generate hypotheses. These predictions need to be independently confirmed and functionally validated in future studies.

What are the CytoTRACE 2 potency categories?

CytoTRACE 2 classifies cells into six potency categories:

• Totipotent: Stem cells capable of generating an entire multicellular organism
• Pluripotent: Stem cells with the capacity to differentiate into all adult cell types
• Multipotent: Lineage-restricted multipotent cells capable of producing >3 downstream cell types
• Oligopotent: Lineage-restricted immature cells capable of producing 2-3 downstream cell types
• Unipotent: Lineage-restricted immature cells capable of producing a single downstream cell type
• Differentiated: Mature cells, including cells with no developmental potential

I still have so many questions, is there someone I can reach out to for help?

If you have further questions, please email us at cytotrace2team@gmail.com.