Comparing machine learning methods on Raman spectra from eight different spectrometers | JoVE Visualize

Area of Science:

Biotechnology
Analytical Chemistry
Chemometrics
Machine Learning

Background:

Raman spectroscopy is a valuable Process Analytical Technology (PAT) in biotechnology for non-invasive molecular analysis.
Machine learning models are essential for translating complex spectral data into quantifiable concentrations.
Partial least squares (PLS) models, assuming linear relationships, are commonly used but may be limited in complex biological systems.

Purpose of the Study:

To evaluate the performance of a single convolutional neural network (CNN) trained on data from multiple Raman spectrometers.
To compare the efficacy of CNNs against traditional PLS models for quantitative analysis in biotechnology.
To assess the potential for a unified CNN model to streamline calibration across diverse spectral datasets.

Main Methods:

Preparation of samples with known concentrations of glucose, sodium acetate, and magnesium sulfate.
Acquisition of over 2200 Raman spectra from these samples using eight different spectrometers.
Training a single CNN model on the combined spectral data from all eight spectrometers.
Comparison of CNN model performance against conventional PLS models.

Main Results:

The single CNN model trained on data from all eight spectrometers significantly outperformed individual PLS models.
The joint CNN approach demonstrated improved overall accuracy and reduced calibration effort for laboratories with multiple spectrometers.
Analysis identified three specific spectrometers as being better suited for the accurate quantification of glucose, sodium acetate, and magnesium sulfate.

Conclusions:

A unified CNN model offers a powerful and efficient alternative to traditional PLS calibration for Raman spectroscopy in biotechnology.
This multi-spectrometer CNN approach enhances accuracy and simplifies workflow, particularly beneficial for labs utilizing diverse instrumentation.
The findings highlight the potential for advanced machine learning to overcome limitations of conventional methods in complex analytical challenges.