Author Correction: Optimized model architectures for deep learning on genomic data

Hüseyin Anil Gündüz, René Mreches, Julia Moosbauer, Gary Robertson, Xiao-Yin To, Eric A. Franzosa, Curtis Huttenhower, Mina Rezaei, Alice C. McHardy, Bernd Bischl, Philipp C. Münch, Martin BinderCommunications 2024 · DOI · OpenAlex

This document is a formal Author Correction for the paper "Optimized model architectures for deep learning on genomic data," originally published in Communications Biology in April 2024. The correction addresses a specific clerical error in the labeling of performance data, ensuring that the comparison between baseline and optimized models is accurately represented.

TL;DR

This update corrects a labeling error in the original paper’s comparison of genomic model architectures. Specifically, it clarifies that certain baseline models (DNABERT and RC-LSTM) are "Single" models rather than "Ensemble" models, ensuring the performance benchmarks for pathogenicity detection are correctly categorized.

What Problem Does This Paper Address?

The original research focuses on the challenge of identifying the best neural network architectures for genomic data, specifically for the task of pathogenicity detection (identifying whether a genetic sequence is likely to cause disease).

In the field of bioinformatics, researchers often struggle to determine whether performance gains come from novel architecture optimizations or simply from using "ensembles" (combining multiple models). This correction ensures that the distinction between single-model baselines and ensemble-model baselines is clear, allowing for a fair "apples-to-apples" comparison with the authors' newly optimized architectures.

Methods

While the correction notice does not detail the underlying algorithms, the included figures list the specific architectures evaluated:

  • Baselines: BLAST (alignment-based), PaPrBaG, DNABERT, RC-LSTM, and RC-CNN.
  • Optimized Models: RC-CNN+LSTM, CNN-RNN, CNN-RNN-VC, CNN-GAP, and CNN-GAP-VC.
  • Ensembles: Combinations such as GAP+RNN, GAP+RNN-VC, and a comprehensive "4-model" ensemble.

The primary "method" of this specific document is the visual rectification of Figure 4 to correctly group these models into "Single" and "Ensemble" categories.

Datasets

Not discussed in the paper. (The correction focuses on the performance task of "pathogenicity detection," but specific dataset names or sizes are not listed in this notice).

Key Insights and Ideas

  • Categorization Matters: The core of this correction is the reclassification of DNABERT and RC-LSTM. In the original figure, these were grouped under "Ensemble" baseline models. The corrected figure moves them to the "Single" baseline category. This is a critical distinction for researchers looking to beat the state-of-the-art; it shows that these powerful models are individual units, not combinations of multiple networks.
  • Optimized Architectures Outperform Baselines: Even with the corrected labels, the "Optimized models" (particularly those utilizing Global Average Pooling, or GAP) show a clear downward trend in misclassification rates compared to the baselines.
  • The Power of Ensembling: The results confirm that the "4-model" ensemble achieves the lowest misclassification rate overall, highlighting that while architecture optimization is vital, ensembling remains a superior strategy for maximizing accuracy in genomic tasks.

Key Results

The corrected Figure 4 provides a quantitative breakdown of the Misclassification Rate for the pathogenicity detection task:

  • Traditional Baseline: BLAST shows the highest error rate at 0.335.
  • Single Baseline Models: DNABERT (0.167) and RC-LSTM (0.162) perform significantly better than alignment-based methods.
  • Optimized Single Models: The CNN-GAP-VC architecture reduces the misclassification rate to 0.143.
  • Top Performer: The "4-model" ensemble achieves the best performance with a misclassification rate of 0.136.

Limitations and Future Work

Not discussed in the paper. (As a correction notice, the document is limited to rectifying the specific error identified in the original publication).

Summary updated

BibTeX
@inproceedings{gunduz2024author,
  author = {Hüseyin Anil Gündüz and René Mreches and Julia Moosbauer and Gary Robertson and Xiao-Yin To and Eric A. Franzosa and Curtis Huttenhower and Mina Rezaei and Alice C. McHardy and Bernd Bischl and Philipp C. Münch and Martin Binder},
  title = {{Author Correction: Optimized model architectures for deep learning on genomic data}},
  booktitle = {Communications},
  year = {2024},
  doi = {10.1038/s42003-024-06318-y},
}

Loading related content...

Figure