DeSpecNet: A decomposition spectral network for Raman-based ratio quantification of the mixed Lactobacillus crispatus and Lactobacillus iners culture
Author Identifier (ORCID)
Liang Wang: https://orcid.org/0000-0001-5339-7484
Abstract
The vaginal microbiota is vital to female reproductive health, with Lactobacillus crispatus and Lactobacillus iners dominating in healthy states yet differing markedly in ecological function. Accurate discrimination and quantification of these closely related species are crucial for monitoring vaginal health. However, current clinical methods, reliant on bacterial culture, molecular testing, or empirical interpretation, suffer from long processing times and limited resolution. Surface-enhanced Raman spectroscopy (SERS), a rapid and label-free technique capturing molecular fingerprints, shows promise for microbial detection, but in mixed communities, its performance is hindered by spectral overlap and nonlinear effects, complicating precise unmixing. In this study, we propose a novel spectral deconstruction neural network, DeSpecNet, designed to directly reconstruct the pure spectra. The workflow integrates β-VAE data augmentation to enhance spectral diversity and authenticity, employs preprocessing to improve training stability, and combines convolutional feature extraction, channel attention mechanisms, and gated modulation units to achieve layer-by-layer parsing and component separation of mixed spectra, thereby effectively separating overlapping signals from different species. The external testing demonstrate that DeSpecNet consistently extracts pure spectra highly concordant with experimental references across various mixing ratios, with similarity correlation coefficients exceeding 0.95 and outperforming classical algorithms in unmixing fidelity. Ablation experiments further confirm that each module of the model contributes positively to performance enhancement. In sum, this study presents a new strategy for quantifying mixed systems of vaginal Lactobacilli. With continued data accumulation and model iteration, DeSpecNet holds promise as a new paradigm for quantitative spectral analysis of complex microbial communities.
Keywords
Lactobacillus crispatus, Lactobacillus iners, machine learning algorithm, Raman spectroscopy, ratio quantification
Document Type
Journal Article
Date of Publication
6-30-2026
Volume
280
Publication Title
Measurement
Publisher
Elsevier
School
Nutrition and Health Innovation Research Institute / School of Medical and Health Sciences
Funding Information
This study was financially supported by the Research Foundation for Advanced Talents of Guangdong Provincial People’s Hospital (grant no. KY012023293) and National Natural Science Foundation of China (grant no. 82372258). JWT acknowledges the support of the Research Training Program scholarship by the Australian Commonwealth Government (no grant number is assigned for the scholarship).
Copyright
subscription content
Comments
Tang, J., Luan, L., Xie, Y., Chen, J., Zhou, Z., Wang, Y., Chen, H., & Wang, L. (2026). DeSpecNet: A decomposition spectral network for Raman-based ratio quantification of the mixed Lactobacillus crispatus and Lactobacillus iners culture. Measurement, 280, 121845. https://doi.org/10.1016/j.measurement.2026.121845