MSTune: A Data-Driven Approach to Parameter Tuning Using Grid Search and Differential Evolution for Gas...

Area of Science:

• Analytical Chemistry
• Computational Chemistry

Background:

• Spectrum preprocessing parameters critically impact GC-MS library-based compound identification.
• Conventional grid search optimization is computationally inefficient and limited by discretization.

Purpose of the Study:

• Compare the performance of differential evolution (DE) and grid search for optimizing GC-MS compound identification parameters.
• Evaluate DE's flexibility and efficiency in exploring parameter spaces.

Main Methods:

• Applied cosine similarity to the NIST GC-MS library.
• Utilized DE to maximize cross-validated accuracy or mean reciprocal rank (MRR).
• Compared DE results with grid search over five parameter values, evaluating performance using accuracy, MRR, and AUC.

Main Results:

• DE achieved slightly higher accuracy and MRR than grid search when optimizing four parameters simultaneously.
• More significant differences were noted in unidimensional tuning, especially for the intensity weight factor.
• Simultaneous multidimensional parameter optimization outperformed isolated tuning.

Conclusions:

• DE offers a flexible approach for high-dimensional or unknown parameter spaces, while grid search is suitable for limited, known spaces.
• DE demonstrated comparable identification performance to grid search with modest improvements in certain settings.
• A publicly available tool, MSTune, was developed for spectrum preprocessing parameter optimization.