Peer Review #1 of "SC-JNMF: single-cell clustering integrating multiple quantification methods based on joint non-negative matrix factorization (v0.2)"

Mikio Shiga, Shigeto Seno, Makoto Onizuka, Hideo Matsuda, N Bray, H Pimentel, P Melsted, L Pachter, J.-p Brunet, P Tamayo, T Golub, J Mesirov, Chin Lin, J Costa-Silva, D Domingues, F Lopes, A Dobin, C Davis, F Schlesinger, J Drenkow, C Zaleski, S Jha, P Batut, M Chaisson, T Gingeras, Z Duren, X Chen, M Zamanighomi, W Zeng, A Satpathy, H Chang, Y Wang, W Wong, S Freytag, L Tian, I Lnnstedt, M Ng, M Bahlo, N Fujita, S Mizuarai, K Murakami, K Nakai, S Jin, L Zhang, Q Nie, V Kiselev, K Kirschner, M Schaub, T Andrews, A Yiu, T Chandra, K Natarajan, W Reik, M Barahona, A Green, M Hemberg, B Langmead, S Salzberg, D Lee, H Seung, B Li, V Ruotti, R Stewart, J Thomson, C Dewey, G Monaco, B Lee, W Xu, S Mustafah, Y Hwang, C Carre, N Burdin, L Visan, M Ceccarelli, M Poidinger, A Zippelius, J Pedro De Magalhes, A Larbi, S Nik-Zainal, L Alexandrov, D Wedge, P Van Loo, C Greenman, K Raine, D Jones, J Hinton, J Marshall, L Stebbings, A Menzies, S Martin, K Leung, L Chen, C Leroy, M Ramakrishna, R Rance, K Lau, L Mudie, I Varela, D Mcbride, G BignellOpenalex Unknown 2021 · DOI · OpenAlex

Abstract

Single-cell RNA-sequencing (scRNA-seq) provides unprecedented resolution for understanding biological processes. However, the standard analysis pipeline—comprising quantification (converting sequences to gene-cell matrices) and downstream analysis (clustering)—is susceptible to "quantification bias." Different quantification tools, such as STAR or Salmon, often produce divergent gene expression profiles from the same raw data, leading to inconsistent cell clustering results. We present SC-JNMF, a framework that utilizes joint non-negative matrix factorization (joint-NMF) to integrate multiple gene expression profiles. By enforcing a shared coefficient matrix across different quantification methods, SC-JNMF extracts robust common biological factors, improving clustering accuracy and stability.

Introduction: The Challenge of Quantification Bias

The transition from raw scRNA-seq reads to biological discovery involves complex computational pipelines. A pivotal step is quantification, where tools map reads to a reference genome or transcriptome to generate a gene expression matrix. Researchers typically choose between alignment-based methods (e.g., STAR) and alignment-free "pseudo-alignment" methods (e.g., Salmon, kallisto).

The problem arises when these tools disagree. Even with identical input data, different software packages introduce unique technical artifacts and biases. Consequently, the resulting cell clusters—essential for identifying cell types and trajectories—can vary significantly depending on the chosen upstream tool. This dependency undermines the reproducibility and reliability of single-cell studies.

Methodology: Joint Non-negative Matrix Factorization

To address this inconsistency, SC-JNMF leverages the strengths of multiple quantification pipelines simultaneously. Instead of choosing one "best" method, it finds a consensus across several.

Standard Non-negative Matrix Factorization (NMF) decomposes a single expression matrix $D$ into a basis matrix $W$ (representing gene features) and a coefficient matrix $H$ (representing cell factors), such that $D \approx WH$. SC-JNMF extends this to a joint framework. Given multiple matrices ($D_1, D_2, \dots$) from the same sample processed by different methods, the model performs simultaneous factorization with a critical constraint: while each method maintains its own basis matrix ($W_1, W_2, \dots$), they must all share the same coefficient matrix ($H$).

The SC-JNMF workflow integrates multiple quantification pipelines into a single shared cell-factor matrix (H) for robust clustering.

By forcing a shared $H$ matrix, the algorithm effectively filters out method-specific noise and technical artifacts, capturing the underlying biological signal common to all pipelines. The inclusion of $L_1$ regularization further ensures sparsity in the shared matrix, resulting in more distinct and interpretable cell clusters.

Experimental Evaluation

The authors validated SC-JNMF across six diverse datasets, ranging from small-scale lung epithelium samples to large-scale benchmarking datasets:

  • Treutlein & Pollen: Lung epithelium and developing brain datasets for initial validation.
  • Segerstolpe & Xin: Pancreatic datasets with varying cell counts and classes.
  • Monaco: Bulk RNA-seq data used to test the method's versatility.
  • CellBench: A large-scale benchmarking dataset of lung adenocarcinoma cell lines (3,918 cells).

Performance was measured using the Adjusted Rand Index (ARI), which quantifies the similarity between the predicted clusters and the ground-truth cell types.

Key Insights and Results

Superior Accuracy and Stability

SC-JNMF consistently outperformed established single-method tools like SC3 and Seurat, as well as standard NMF approaches, particularly on the Pollen, Xin, and CellBench datasets. The joint approach proved more stable, showing less sensitivity to initial conditions than conventional NMF.

The Power of Diversity

A crucial finding was that the method's effectiveness depends on the diversity of the input pipelines. Combining an alignment-based method (STAR) with an alignment-free method (Salmon) yielded better results than combining two similar alignment-free methods (Salmon and kallisto). When the underlying logic of the quantification tools differs, the "joint" constraint is more effective at isolating biological signal from technical noise.

Scalability and Marker Discovery

The performance advantage of SC-JNMF became more pronounced as dataset size increased, making it highly suitable for modern high-throughput experiments. Furthermore, the method-specific $W$ matrices remain useful for identifying marker genes, demonstrating that the integration process does not sacrifice high-resolution biological detail.

Conclusion and Future Directions

SC-JNMF provides a robust solution to the problem of quantification bias in single-cell analysis. By integrating multiple data-processing perspectives, it offers a more reliable foundation for downstream biological interpretation. While the method requires more computational resources than single-pipeline approaches, the gain in clustering accuracy and stability is significant. Future work may explore extending this joint framework to multi-omic data, such as integrating transcriptomics with spatial data or chromatin accessibility (ATAC-seq), to further refine our understanding of cellular diversity.

Summary updated

Abstract

Single-cell RNA-sequencing is a rapidly evolving technology that enables us to understand biological processes at unprecedented resolution.Single-cell expression analysis requires a complex data processing pipeline, and the pipeline is divided into two main parts: The quantification part, which converts the sequence information into gene-cell matrix data; the analysis part, which analyzes the matrix data using statistics and/or machine learning techniques.In the analysis part, unsupervised cell clustering plays an important role in identifying cell types and discovering cell diversity and subpopulations.Identified cell clusters are also used for subsequent analysis, such as finding differentially expressed genes and inferring cell trajectories.However, single-cell clustering using gene expression profiles shows different results depending on the quantification methods.Clustering results are greatly affected by the quantification method used in the upstream process.In other words, even if the original RNA-sequence data is the same, gene expression profiles processed by different quantification methods will produce different clusters.In this article, we propose a robust and highly accurate clustering method based on joint non-negative matrix factorization (joint-NMF) by utilizing the information from multiple gene expression profiles quantified using different methods from the same RNA-sequence data.Our joint-NMF can extract common factors among multiple gene expression profiles by applying each NMF under the constraint that one of the factorized matrices is shared among multiple NMFs.The joint-NMF determines more robust and accurate cell clustering results by leveraging multiple quantification methods compared to conventional clustering methods, which use only a single gene expression profile.Additionally, we showed the usefulness of discovering marker genes with the extracted factors using our method.seq).This technology has reached a surprising level of resolution that reveals the regulation of gene 35 expression within cells.scRNA-seq measures gene expression on a cell-by-cell basis and allows the 36 analysis of the functions and properties of cells using this information.Many experimental protocols 37 and computational analyses exist for scRNA-seq, and they have different goals such as differential ex-38 pression analysis, cell clustering, cell classification, and trajectory reconstruction.Therefore, single-cell 39 analysis forms a pipeline (a series of procedures) that mainly consists of two parts, quantification of 40 gene expression and downstream analysis depends on the goals.In the quantification part of a pipeline, 41 gene expression levels measured for each cell (raw sequence data) are converted to a matrix called the 42 "gene expression profile".The quantification part also forms a pipeline including read quality control, 43 adaptor trimming, demultiplexing, deduplicating using barcodes, mapping to genome, counting transcripts.
BibTeX
@inproceedings{shiga2021peer,
  author = {Mikio Shiga and Shigeto Seno and Makoto Onizuka and Hideo Matsuda and N Bray and H Pimentel and P Melsted and L Pachter and J.-P Brunet and P Tamayo and T Golub and J Mesirov and Chin Lin and J Costa-Silva and D Domingues and F Lopes and A Dobin and C Davis and F Schlesinger and J Drenkow and C Zaleski and S Jha and P Batut and M Chaisson and T Gingeras and Z Duren and X Chen and M Zamanighomi and W Zeng and A Satpathy and H Chang and Y Wang and W Wong and S Freytag and L Tian and I Lnnstedt and M Ng and M Bahlo and N Fujita and S Mizuarai and K Murakami and K Nakai and S Jin and L Zhang and Q Nie and V Kiselev and K Kirschner and M Schaub and T Andrews and A Yiu and T Chandra and K Natarajan and W Reik and M Barahona and A Green and M Hemberg and B Langmead and S Salzberg and D Lee and H Seung and B Li and V Ruotti and R Stewart and J Thomson and C Dewey and G Monaco and B Lee and W Xu and S Mustafah and Y Hwang and C Carre and N Burdin and L Visan and M Ceccarelli and M Poidinger and A Zippelius and J Pedro De Magalhes and A Larbi and S Nik-Zainal and L Alexandrov and D Wedge and P Van Loo and C Greenman and K Raine and D Jones and J Hinton and J Marshall and L Stebbings and A Menzies and S Martin and K Leung and L Chen and C Leroy and M Ramakrishna and R Rance and K Lau and L Mudie and I Varela and D Mcbride and G Bignell},
  title = {{Peer Review \#1 of "SC-JNMF: single-cell clustering integrating multiple quantification methods based on joint non-negative matrix factorization (v0.2)"}},
  booktitle = {Openalex Unknown},
  year = {2021},
  doi = {10.7287/peerj.12087v0.2/reviews/1},
}
Figure