Author Correction: Employing fingerprinting of medicinal plants by means of LC-MS and machine learning for species identification task

Pavel Kharyuk, Dmitry Nazarenko, Ivan Oseledets, И. А. Родин, О. А. Шпигун, Andrey Tsitsilin, Mikhail LavrentyevScientific 2020 · DOI · OpenAlex

Introduction

The identification of medicinal plants is a cornerstone of botanical quality control, ensuring the safety and efficacy of herbal products. A previously published study proposed a pipeline using Liquid Chromatography-Mass Spectrometry (LC-MS) fingerprints and machine learning to automate this process. However, a subsequent Author Correction was issued to address a critical oversight: data leakage. This correction serves as a significant case study in the importance of rigorous data partitioning in biological machine learning.

The Challenge of Data Leakage

The original study reported exceptionally high accuracies (~96%). Upon re-evaluation, the authors discovered "train-test leakage" stemming from the presence of highly correlated LC-MS repletions—multiple measurements of the same physical sample—across both training and validation sets. In effect, the models were partially "memorizing" specific samples rather than learning the generalizable chemical characteristics of the species.

Figure 4: Accuracy vs. Variable Number across different test sets. Note the performance gap between the leaked Test 1 and the independent Test 2.

Methodology and Revised Evaluation

The researchers utilized LC-MS to generate chemical "barcodes" for various plant species and organs. To handle the high-dimensional data, they employed feature extraction techniques such as Autoencoders and Tucker decomposition (specifically Sparse NTD). These compressed features were then fed into classifiers including Logistic Regression, Naive Bayes, and Hybrid Bayesian Networks.

Following the correction, the models were re-evaluated using a strict partitioning strategy where all repletions of a single physical sample were kept within the same set. This ensures that the model is tested on truly unseen biological material.

Figure 3: Accuracy and computational time vs. Rank for Sparse NMF and Sparse NTD methods.

Key Results and Insights

The corrected results show a significant but realistic decrease in performance metrics:

  • Overall Accuracy: Species and organ identification accuracies dropped from over 92% to approximately 85%.
  • Decomposition Performance: Accuracy for Tucker decomposition on the primary test set fell from 93% to 78%.
  • TopN Robustness: Despite the drop in "Top 1" accuracy, the "Top 5" accuracy remained high at 90% on independent test data. This suggests that machine learning is highly effective as a decision-support tool, narrowing down candidate species for expert verification.

Conclusion

This correction highlights a vital lesson for the metabolomics and machine learning communities: data must be partitioned at the physical sample level to ensure true generalizability. While the revised accuracies are lower, the study confirms that LC-MS fingerprinting combined with Bayesian networks remains a powerful tool for botanical identification. Future work will focus on resolving "very similar pairs," such as distinguishing between different organs of the same plant (e.g., roots vs. rhizomes), which remains a more complex task than inter-species classification.

Summary updated

Abstract

An amendment to this paper has been published and can be accessed via a link at the top of the paper.
BibTeX
@inproceedings{kharyuk2020author,
  author = {Pavel Kharyuk and Dmitry Nazarenko and Ivan Oseledets and И. А. Родин and О. А. Шпигун and Andrey Tsitsilin and Mikhail Lavrentyev},
  title = {{Author Correction: Employing fingerprinting of medicinal plants by means of LC-MS and machine learning for species identification task}},
  booktitle = {Scientific},
  year = {2020},
  doi = {10.1038/s41598-020-67201-4},
}

Loading related content...

Figure