Optimum molecular descriptors based on 89 machine learning methods for predicting the recovery rate of pesticides in crops by GC-MS

Importance of the Topic

Residual pesticide residues in crops represent a critical food safety challenge. Gas chromatography–mass spectrometry (GC-MS) is extensively used for routine analysis, yet recovery rates can fluctuate due to matrix effects. Optimizing molecular descriptors for predictive modeling enhances accuracy and efficiency in pesticide quantification.

Objectives and Study Overview

This work aimed to identify the optimal set of molecular descriptors (MDs) for predicting pesticide recovery rates in fruits and vegetables by GC-MS. A total of 178 descriptors derived via the rcdk package were considered, and 89 machine learning regression methods were evaluated to build robust predictive models.

Methodology

• Correlation analysis of all 178 MDs using Pearson coefficient (threshold r=0.7) to flag highly correlated pairs.
• Graph clustering via DPClus to visualize descriptor relationships and select representative, weakly correlated MDs.
• Descriptor groups MD-r1a, MD-r1b and MD-r1c defined through sequential selection steps to balance correlation reduction and information retention.
• Regression models trained with selected MD subsets and compared against full descriptor set using prediction error metrics.

Instrumentation Used

• GC-MS system for pesticide residue analysis in crop matrices.
• R packages: rcdk for MD calculation, corrr for correlation analysis, strech function, and DPClus for clustering.

Main Results and Discussion

• Correlation analysis identified 118 strongly correlated MDs, leading to 28 clusters; 83 descriptors selected for modeling.
• Of 89 machine learning methods, 57 showed improved prediction error with descriptor selection, while 32 performed worse.
• Notable improvements observed with bagEarthGCV, projection pursuit regression (ppr), and several sparse modeling algorithms.
• Some simple linear models exhibited degraded performance after descriptor reduction, highlighting method-specific sensitivity.

Benefits and Practical Applications

• Streamlined descriptor set reduces multicollinearity and speeds up model development.
• Enhanced predictive accuracy supports more reliable pesticide residue quantification in QA/QC workflows.
• Generalizable selection protocol can be adapted for other analytes and analytical platforms.

Future Trends and Opportunities

• Integration of deep learning and automated feature engineering for descriptor optimization.
• Extension of the workflow to liquid chromatography–mass spectrometry and other food matrices.
• Development of web-based tools for real-time descriptor selection and model deployment.

Conclusion

A systematic workflow combining correlation analysis and graph clustering effectively reduced descriptor redundancy while preserving predictive power. The selected MDs improved machine learning performance for pesticide recovery prediction by GC-MS, offering a reproducible strategy for analytical method development.

References

• Serino T, Nakamura S, Takigawa Y, Anumol T, Altaf-UI-Amin M, Kanaya S. Comprehensive Machine Learning Prediction of GC/MS Pesticide Recovery Based on the Molecular Fingerprinting for Food QA/QC; Poster TP-298, ASMS; June 2019.
• Garg A, Tai K. Comparison of Statistical and Machine Learning Methods in Modelling of Data with Multicollinearity. Int J Model Ident Control. 2013;18:295-312.
• Altaf-Ul-Amin M, Shibo Y, Mihara K, Kurokawa K, Kanaya S. Development and Implementation of an Algorithm for Detection of Protein Complexes in Large Interaction Networks. BMC Bioinformatics. 2006;7:207.