Pseudo-grading of tumor subpopulations from single-cell transcriptomic data using Phenotype Algebra

Reviewer #1 (Public Review):

Summary:

This review evaluates the SCellBOW framework, which applies phenotype algebra to obtain vectors from cancer subclusters or user-defined subclusters.

Strengths:

SCellBOW employs an innovative application of NLP-inspired techniques to analyze scRNA-seq data, facilitating the identification and visualization of phenotypically divergent cell subpopulations.

The framework demonstrates robustness in accurately representing various cell types across multiple datasets, highlighting its versatility and utility in different biological contexts.

By simulating the impact of specific malignant subpopulations on disease prognosis, SCellBOW provides valuable insights into the relative risk and aggressiveness of cancer subpopulations, which is crucial for personalized therapeutic strategies.

The identification of a previously unknown and aggressive AR−/NElow subpopulation in metastatic prostate cancer underscores the potential of SCellBOW in uncovering clinically significant findings.

Weaknesses:

The reliance on bulk RNA-seq data as a reference raises concerns about potentially misleading results due to the presence of RNA expression from immune cells in the TME. It is unclear if SCellBOW adequately addresses this issue, which could affect the accuracy of the cancer subcluster vectors.

The method of extracting vectors in phenotype algebra appears to be a straightforward subtraction operation. This simplicity might limit its efficiency in excluding associations with phenotypes from specific subpopulations, potentially leading to inaccurate interpretations of the data.

The review would benefit from additional validation studies to assess the effectiveness of SCellBOW in distinguishing between cancerous and non-cancerous signals, particularly in heterogeneous tumor environments.

Further clarification on how SCellBOW handles mixed-cell populations within bulk RNA-seq data would strengthen the evaluation of its applicability and reliability in diverse research settings.

Reviewer #2 (Public Review):

Summary:

The authors developed a novel tool, SCellBOW, to perform cell clustering and infer survival risks on individual cancer cell clusters from the single-cell RNA seq dataset. The key ideas/techniques used in the tool include transfer learning, bag of words (BOW), and phenotype algebra which is similar to word algebra from natural language processing (NLP). Comparisons with existing methods demonstrated that SCellBOW provides superior clustering results and exhibits robust performance across a wide range of datasets. Importantly, a distinguishing feature of SCellBOW compared to other tools is its ability to assign risk scores to specific cancer cell clusters. Using SCellBOW, the authors identified a new group of prostate cancer cells characterized by a highly aggressive and dedifferentiated phenotype.

Strengths:

The application of natural language processing (NLP) to single-cell RNA sequencing (scRNA-seq) datasets is both smart and insightful. Encoding gene expression levels as word frequencies is a creative way to apply text analysis techniques to biological data. When combined with transfer learning, this approach enhances our ability to describe the heterogeneity of different cells, offering a novel method for understanding the biological behavior of individual cells and surpassing the capabilities of existing cell clustering methods. Moreover, the ability of the package to predict risk, particularly within cancer datasets, significantly expands the potential applications.

Weaknesses:

Given the promising nature of this tool, it would be beneficial for the authors to test the risk-stratification functionality on other types of tumors with high heterogeneity, such as liver and pancreatic cancers, which currently lack clinically relevant and well-recognized stratification methods. Additionally, it would be worthwhile to investigate how the tool could be applied to spatial transcriptomics by analyzing cell embeddings from different layers within these tissues.

Author response:

Reviewer #1:

This review evaluates the SCellBOW framework, which applies phenotype algebra to obtain vectors from cancer subclusters or user-defined subclusters.

Strengths:

SCellBOW employs an innovative application of NLP-inspired techniques to analyze scRNA-seq data, facilitating the identification and visualization of phenotypically divergent cell subpopulations. The framework demonstrates robustness in accurately representing various cell types across multiple datasets, highlighting its versatility and utility in different biological contexts. By simulating the impact of specific malignant subpopulations on disease prognosis, SCellBOW provides valuable insights into the relative risk and aggressiveness of cancer subpopulations, which is crucial for personalized therapeutic strategies. The identification of a previously unknown and aggressive AR−/NElow subpopulation in metastatic prostate cancer underscores the potential of SCellBOW in uncovering clinically significant findings.

Major concerns:

The reliance on bulk RNA-seq data as a reference raises concerns about potentially misleading results due to the presence of RNA expression from immune cells in the TME. It is unclear if SCellBOW adequately addresses this issue, which could affect the accuracy of the cancer subcluster vectors.

To address the concern about potentially misleading results due to the TME when using bulk RNA-seq data as a reference:

a. We account for systematic biases between the single-cell and bulk transcriptomics readouts by creating pseudo-bulk profiles for single-cell clusters, enabling more accurate comparisons.

b. We encode expressions into word vectors and co-embed them together. By doing this, we mitigate any possibility of systematic differences in the embedding.

c. It is imperative that we subject both single-cell and bulk data through the same treatments because otherwise, it will be difficult to perform algebraic operations on them.

d. We rely on tumor bulk transcriptomics data from TCGA due to its high sample size and patient meta-data such as information pertaining to patient survival.

We will discuss this in the revised manuscript.

The method of extracting vectors in phenotype algebra appears to be a straightforward subtraction operation. This simplicity might limit its efficiency in excluding associations with phenotypes from specific subpopulations, potentially leading to inaccurate interpretations of the data.

Vector algebra operations are not done in the gene expression space (i.e., gene expression vectors associated with tumor samples), rather we process the single cell and bulk expression profiles through multiple steps (pseudo-bulk vector generation for single cell clusters, mapping gene expression values to word frequencies as better understood by the Doc2vec neural networks etc.) to ensure their embeddings are consistent and capture intricate phenotypic information. We have demonstrated this through rigorous validation of the clusters yielded on various types of healthy and diseased samples. Furthermore, we have demonstrated the consistency of the vector algebra operations on known cancer subtypes in breast cancer, glioblastoma, and prostate cancer.

We will discuss this in the revised manuscript.

The review would benefit from additional validation studies to assess the effectiveness of SCellBOW in distinguishing between cancerous and non-cancerous signals, particularly in heterogeneous tumor environments.

In our study, we are primarily interested in signals from malignant cells. However, we may consider scRNA-seq data with stromal cells and test whether SCellBOW can identify the influence of different stromal cell types on cancer aggressiveness.

Further clarification on how SCellBOW handles mixed-cell populations within bulk RNA-seq data would strengthen the evaluation of its applicability and reliability in diverse research settings.

We will elaborate on our discussion in the Result as well as Discussion sections.

Reviewer #2:

The authors developed a novel tool, SCellBOW, to perform cell clustering and infer survival risks on individual cancer cell clusters from the single-cell RNA seq dataset. The key ideas/techniques used in the tool include transfer learning, bag of words (BOW), and phenotype algebra which is similar to word algebra from natural language processing (NLP). Comparisons with existing methods demonstrated that SCellBOW provides superior clustering results and exhibits robust performance across a wide range of datasets. Importantly, a distinguishing feature of SCellBOW compared to other tools is its ability to assign risk scores to specific cancer cell clusters. Using SCellBOW, the authors identified a new group of prostate cancer cells characterized by a highly aggressive and dedifferentiated phenotype.

Strengths:

The application of natural language processing (NLP) to single-cell RNA sequencing (scRNA-seq) datasets is both smart and insightful. Encoding gene expression levels as word frequencies is a creative way to apply text analysis techniques to biological data. When combined with transfer learning, this approach enhances our ability to describe the heterogeneity of different cells, offering a novel method for understanding the biological behavior of individual cells and surpassing the capabilities of existing cell clustering methods. Moreover, the ability of the package to predict risk, particularly within cancer datasets, significantly expands the potential applications.

Major concerns:

Given the promising nature of this tool, it would be beneficial for the authors to test the risk-stratification functionality on other types of tumors with high heterogeneity, such as liver and pancreatic cancers, which currently lack clinically relevant and well-recognized stratification methods. Additionally, it would be worthwhile to investigate how the tool could be applied to spatial transcriptomics by analyzing cell embeddings from different layers within these tissue

(1) Our selection of glioblastoma and breast cancer for this study was primarily driven by the focus on extensively studied and well-defined cancer types. To demonstrate the effectiveness of our model, we tested it on advanced prostate cancer, which currently lacks clinically relevant and well-recognized stratification methods. This application to metastatic prostate cancer serves as a proof of concept, illustrating our model's potential to provide valuable insights into cancer types where established stratification approaches are limited or absent. However, as suggested by the Reviewer, we will try to incorporate results for liver cancer, subject to the availability of adequate data for model building.

(2) Regarding the application of our tool to spatial transcriptomics, we have already analyzed data from Digital Spatial Profiling (DSP). The article is already quite complex and involved, and we are afraid the inclusion of spatial transcriptomics may amount to a significant extension of the method. To this end, although we will discuss the future possibilities, we will skip the method validity check on spatial transcriptomics data.