Deep Learning Outperforms Descriptor-Based Classification of Food Items Using Chromatography-Mass Spectrometry Data.

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Food Science

Background:

• Food authentication and adulteration detection face challenges due to complex chemical matrices.
• Traditional methods struggle to capture intricate interactions within chromatography-mass spectrometry (CMS) data.
• Deep learning offers a promising alternative for objective and scalable food analysis.

Purpose of the Study:

• To compare the performance of a convolutional neural network (CNN) against a random forest (RF) model for food classification.
• To demonstrate the advantage of deep learning in analyzing CMS data for food authentication.
• To evaluate the CNN's capability in detecting food adulteration.

Main Methods:

• Simulated 3000 CMS spectra based on FooDB compound profiles for 15 food classes.
• Engineered six descriptors for RF training; developed a SimpleCNN with two convolutional layers.
• Utilized t-SNE and K-means for chemical space visualization and simulated adulteration by mixing corn syrup into honey spectra.

Main Results:

• CNN achieved 93.1% classification accuracy, outperforming RF (87.3%).
• t-SNE visualization revealed five coherent clusters, indicating good data separation.
• Adulteration detection demonstrated high sensitivity (94.2%) and specificity (96.8%).

Conclusions:

• End-to-end deep learning on CMS fingerprints provides a robust pipeline for food authentication and adulteration screening.
• The study highlights the superiority of deep learning over descriptor-based methods for CMS data analysis.
• Further validation with experimentally acquired spectra is recommended for practical application.