Deep learning aided quantitative analysis of anti-tuberculosis fixed-dose combinatorial formulation by terahertz spectroscopy

Abstract

Anti-tuberculosis fixed-dose combinatorial formulation (FDCs) is an effective drug for the treatment of tuberculosis. As a compound medicine, its efficacy is based on the comprehensive action of multiple main ingredients. If the content of an active ingredient is insufficient, not only will it reduce the curative effect, but it also causes patients to develop drug resistance and leads to the evolution of drug-resistant strains of tuberculosis, which hamper the treatment of the disease. Thus accurate detection of the contents of active components in the anti-tuberculosis FDC is the key link of its quality control. For the first time, convolutional neural networks (CNN), one of the most popular deep learning methods, is adopted to develop a quantitative calibration model based on terahertz time-domain spectroscopy (THz-TDS) for the accurate detection of the content of each active component in the anti-tuberculosis FDCs. For comparison with CNN, partial least squares regression (PLSR) was also introduced to build a reference quantitative calibration model. For CNN modeling, the raw THz spectral is fed to the model directly; While for PLSR, prior to the spectrum feeding to the model, the raw spectral data are processed by multiple different combinations of preprocessing. Experimental and simulation results demonstrate that although preprocessing techniques can improve the prediction performance of PLSR, its prediction capabilities is still inferior to CNN based on raw spectrum. Therefore, for the quantitative analysis of the content of each active component in the anti-tuberculosis FDCs, CNN appears to be an ideal modeling method.

Keywords: Anti-tuberculosis fixed-dose combination formulations (FDCs); Convolutional neural networks (CNN); Quantitative analysis; Terahertz time-domain spectroscopy (THz-TDS).