Multimodal deep learning-driven Raman spectroscopy for rapid identification of closely related pathogenic Bacillus species.

Abstract

Rapid and fine-grained identification of closely related pathogenic Bacillus species is a critical challenge in optical biosensing and biomolecular spectroscopy, hindered by their highly similar biochemical compositions and severe spectral overlap. Existing single-modality spectroscopic analysis and computational frameworks frequently fail to resolve these subtle spectral shifts, particularly when confronted with instrumental noise, background interference, and limited training data. To overcome these spectroscopic bottlenecks, this study proposes the Spectral Continuous Wavelet Transform Transformer Network (Spec-CWT-TransNet), a novel multimodal deep learning framework designed for the precise identification of closely related microbial strains. By transforming 1D Raman spectra into 2D wavenumber-scale scalograms utilizing the Continuous Wavelet Transform (CWT), the proposed method synergizes the robust local topological texture extraction of Convolutional Neural Networks (CNNs) with the global dependency modeling capabilities of Transformers. Furthermore, a Fisher Score-based scale optimization strategy is introduced to adaptively select discriminative frequency bands, effectively suppressing redundant instrumental and background noise. Experimental results demonstrate that Spec-CWT-TransNet achieves a superior overall identification accuracy of 99.38% on four Bacillus species. Notably, under limited data conditions (using only 40% of training data), the framework maintains a robust accuracy of 92.50%, quantifiably surpassing the performance of traditional 1D baselines trained on the full dataset. Moreover, under severe spectral interference, the model retains 90.63% accuracy, outperforming unimodal baseline models by margins exceeding 22.5%. Overall, this multimodal Raman data analysis methodology demonstrates great potential for rapid, robust, and reproducible spectral discrimination in non-destructive optical sensing and bioanalytical applications.