Correction to: scBFA: modeling detection patterns to mitigate technical noise in large-scale single-cell genomics data

Ruoxin Li, Gerald QuonGenome 2019 · DOI · OpenAlex

Abstract

Following the publication of the original article "scBFA: modeling detection patterns to mitigate technical noise in large-scale single-cell genomics data" (2019), two typographical errors were identified in the mathematical formulae within the Methods section. This correction provides the accurate notation for the Matthews Correlation Coefficient (MCC) and the batch effect correction generative model. These changes do not affect the results, conclusions, or the underlying scBFA algorithm.

Introduction

The scBFA (single-cell Binary Factor Analysis) method was developed to address the challenges of technical noise and "dropout" events in large-scale single-cell RNA sequencing (scRNA-seq) data. By modeling gene detection as a binary process, scBFA provides a robust framework for dimensionality reduction and batch effect correction. This document serves to correct two specific mathematical misprints found in the original publication to ensure clarity for researchers implementing the method.

Correction of Benchmarking Metrics

In the original section titled "Benchmarking dimensionality reduction methods for scRNA-seq," the formula for the Matthews Correlation Coefficient (MCC) contained a notation error in the denominator. The MCC is a critical metric for evaluating binary classification performance, particularly in the context of imbalanced single-cell datasets where non-detections (zeros) are prevalent. Unlike simple accuracy, MCC accounts for all four quadrants of a confusion matrix (True Positives, True Negatives, False Positives, and False Negatives).

The corrected denominator for the MCC calculation should be represented as the product of four distinct square root terms: $$\sqrt{TP + FP} \cdot \sqrt{TP + FN} \cdot \sqrt{TN + FP} \cdot \sqrt{TN + FN}$$ The original version incorrectly included brackets inside the square roots that could lead to misinterpretation of the calculation.

Correction of the Batch Effect Generative Model

In the "Batch effect correction" section, the positioning of superscripts and subscripts within the generative model equation was corrected. This equation describes how the model accounts for covariates (like batch effects) when predicting the probability of detection for a specific gene in a specific cell.

The corrected generative model is expressed as: $$M_{sub(i', j)} \sim P\left(g^{-1}\left(x^T_{sub(i')} \beta_j + z^T_{sub(i')} a_j + u_{sub(i')} + v_j\right)\right)$$ This correction ensures that the indices for cells ($i'$) and genes ($j$) are properly aligned, accurately reflecting the matrix operations intended by the authors. The model treats gene detection as a probabilistic event influenced by biological factors ($\beta_j$), latent factors ($z$), and technical offsets ($u, v$).

Conclusion

Precision in mathematical notation is vital for the implementation of bioinformatics tools. These typographical corrections ensure that the mathematical "source code" of the paper matches the intended implementation. While the underlying logic and the findings of the original study remain unchanged, the corrected formulae provide the necessary mathematical accuracy for the bioinformatics community.

Summary updated

Abstract

Following publication of the original article [1], the following two errors were found in formulae.
BibTeX
@inproceedings{li2019correction,
  author = {Ruoxin Li and Gerald Quon},
  title = {{Correction to: scBFA: modeling detection patterns to mitigate technical noise in large-scale single-cell genomics data}},
  booktitle = {Genome},
  year = {2019},
  doi = {10.1186/s13059-019-1822-0},
}
Figure