Medium-aware multi-task learning for robust bacterial identification using Raman spectroscopy.

Area of Science:

• Biotechnology
• Spectroscopy
• Machine Learning

Background:

• Raman spectroscopy shows promise for rapid bacterial identification.
• Complex culture medium backgrounds often interfere with spectral signatures, limiting model generalization.
• Existing deep learning models struggle to separate bacterial signals from environmental noise.

Purpose of the Study:

• To develop a novel deep learning model, the Hybrid Medium-Aware Network (HMA-Net), to address challenges in bacterial identification using Raman spectroscopy.
• To transform culture medium background signals from noise into informative contextual data for precise bacterial feature extraction.
• To enhance the robustness and accuracy of bacterial identification in complex, variable environments.

Main Methods:

• Developed a Residual Cross-Task Attention mechanism to recalibrate bacterial spectral features with environmental context.
• Constructed a Mixture-of-Experts (MoE)-based classifier for adaptive classification based on medium conditions.
• Implemented a hybrid loss function combining contrastive learning and Focal Loss to improve intra-class compactness and handle class imbalance.

Main Results:

• Achieved 97.45% strain identification accuracy and a Macro F1-score of 0.9729 on a mixed dataset of 10 bacterial strains across 4 culture media.
• Demonstrated significant accuracy improvements of 6.39% over single-task 1D-CNN and 2.90% over standard multi-task learning (MTL) models.
• Validated the model's ability to mitigate culture medium interference through synergistic learning of environmental context.

Conclusions:

• The HMA-Net effectively utilizes environmental context to enhance bacterial identification accuracy and robustness.
• The proposed method offers a practically meaningful solution for reliable bacterial identification in complex settings.
• Synergistic learning of spectral and environmental data is crucial for overcoming limitations in Raman spectroscopy-based bacterial identification.