Amino-acid classification based on terahertz absorption spectroscopy with Gaussian process and maximum likelihood

Abstract

The rotational and vibrational energy levels of numerous biomolecules lie in the terahertz (THz) band, which makes THz spectroscopy a viable option for biosensing and monitoring the chemical processes such as the formation of hydrogen bonds. However, classifying the fingerprint THz spectra of various molecules has constantly challenged real-world applications. Here, a machine learning model based on a discrete Gaussian process and maximum likelihood estimation (DGPML) for classifying an ensemble of 20 amino acids is proposed. The model demonstrates a prediction accuracy of 99.5% in the presence of scattering noise and various experimental conditions and is able to distinguish spectra without significant absorption peaks. Compared to the conventional Gaussian process classifier, which generally exhibits less time and space complexity than neural-network models, DGPML requires 99% less training time and achieves 1530 times faster processing speed with around 0.29 million parameters. Regarding the DGPML, a database that encloses the THz absorption spectrum and its covariance matrix, which describes the deviation of the absorption spectrum under various experimental conditions, is designed. Due to the compact size and high efficiency of the proposed model, it has the potential to power the THz chemical sensor in the future.