Transformer-Based Classification of Bacterial Raman Spectra with LOOCV

This study investigates whether transformer-based deep learning models can improve the classification of bacterial Raman spectra compared to traditional chemometric pipelines. A critical focus is the use of a 'replicate-aware' validation framework, which addresses the common pitfall of over-optimistic performance estimates in spectral analysis by ensuring that training and testing data do not share the same measurement replicates.

The authors utilized a dataset of 5,417 single-cell Raman spectra across six bacterial species and nine independent measurement replicates. They compared a transformer architecture—which processes spectral data as sequences using self-attention—against six conventional pipelines (PCA or ICA combined with LDA, SVM, or Random Forest). To ensure a realistic evaluation, they employed a nested leave-one-replicate-out cross-validation strategy, where the model is tested on entirely unseen measurement sessions.

The transformer model achieved the highest classification sensitivity across all independent test replicates, significantly outperforming all conventional machine learning combinations. Statistical analysis via Wilcoxon signed-rank tests confirmed that these performance gains were significant. Furthermore, the transformer demonstrated robust performance even when applied directly to raw spectral data without extensive preprocessing, suggesting that the model effectively learns discriminative features from the raw input. Visualization of the latent feature space indicated that the transformer achieves superior class separation compared to PCA- and ICA-based representations.

In Raman spectroscopy, models often fail to generalize to new experimental conditions due to replicate-specific variations. By demonstrating that transformers can handle these variations more effectively than traditional methods, this study provides a pathway for more reliable, automated bacterial identification in clinical and microbiological settings. It also underscores the necessity of using replicate-aware validation to ensure that machine learning models are truly ready for real-world application.