Artificial neural network and classical least-squares methods for neurotransmitter mixture analysis

Area of Science:

• Analytical Chemistry
• Biochemistry
• Spectroscopy

Background:

• Identifying individual components in biological mixtures is analytically challenging.
• Raman spectroscopy offers versatility for biological sample analysis in aqueous media.
• Small-molecule neurotransmitters are crucial biological analytes.

Purpose of the Study:

• To evaluate artificial neural networks (ANNs) for identifying and quantifying neurotransmitters via Raman spectra.
• To compare ANN performance with the classical least-squares (CLS) method.
• To determine the influence of ANN architecture and transfer functions on spectral analysis.

Main Methods:

• Raman spectroscopy was employed to acquire spectra of neurotransmitters and their mixtures.
• Artificial neural networks (ANNs) with varying architectures and transfer functions were trained.
• The classical least-squares (CLS) method was used as a benchmark for comparison.

Main Results:

• ANNs utilizing sigmoid and hyperbolic tangent transfer functions showed superior generalization to novel spectra compared to sine functions.
• Network architectures facilitating local input processing enhanced performance across test datasets.
• ANNs outperformed the CLS method in identifying novel neurotransmitter spectra.
• The CLS method demonstrated robustness with noisy, shifted, and difference spectra.

Conclusions:

• ANNs, particularly with specific architectures and transfer functions, offer a powerful alternative for analyzing complex biological mixtures using Raman spectroscopy.
• The choice of network design is critical for optimizing the identification and quantification of neurotransmitters.
• ANNs provide a valuable tool for advancing the analysis of biological samples, complementing traditional methods like CLS.