Evaluating Limits of Machine Learning-Assisted Raman Spectroscopy in Classification of Biological Samples

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning Applications

Background:

• Machine learning (ML)-assisted Raman spectroscopy is vital for analyte classification.
• Technical challenges affecting detection accuracy require thorough investigation.

Purpose of the Study:

• To investigate experimental factors impacting ML-assisted Raman spectroscopy classification performance.
• To identify key bottlenecks in achieving robust detection accuracy.

Main Methods:

• Evaluated various ML models and experimental factors (spectral noise, similarity, biological heterogeneity).
• Analyzed single-cell Raman spectra from Saccharomyces cerevisiae strains with gene mutations.
• Assessed transfer learning effectiveness across different Raman spectrometers.

Main Results:

• ML algorithms had minimal impact; spectral similarity and data quality were dominant factors.
• Increased spectral noise and similarity significantly reduced classification accuracy.
• Cell-to-cell variability in biological samples severely impacted single-cell classification accuracy, improved by spectral averaging.

Conclusions:

• Data quality and spectral similarity are primary limitations in ML-assisted Raman spectroscopy.
• Proper sample preparation, data acquisition, and instrument calibration are critical for reliable performance.
• Transfer learning benefits from instrument standardization and calibration.