Weakly Supervised Identification and Localization of Drug Fingerprints Based on Label-Free Hyperspectral CARS...

Area of Science:

• Chemical imaging
• Deep learning
• Biomedical optics

Background:

• Characterizing drug fingerprints in biological samples is crucial for drug development.
• Hyperspectral coherent anti-Stokes Raman scattering (HS-CARS) microscopy offers label-free chemical imaging based on endogenous vibrational contrast.
• Existing methods lack the resolution and specificity for precise drug localization in complex tissues.

Purpose of the Study:

• To develop a deep learning-assisted HS-CARS imaging approach for identifying and localizing drug fingerprints.
• To achieve single-cell resolution for drug profiling in complex biological systems.
• To investigate drug fingerprints of a hepatitis B virus therapy in murine liver tissues.

Main Methods:

• An attention-based deep neural network, hyperspectral attention net (HAN), was developed for drug fingerprint identification and localization.
• The task was formulated as a multiple instance learning problem, enabling the network to highlight informative regions using an attention mechanism.
• HS-CARS microscopy was employed for label-free chemical imaging of murine liver tissues.

Main Results:

• The hyperspectral attention net (HAN) achieved high classification accuracy, with an average area under the curve (AUC) of 0.942 for the high-dose drug group.
• Increased drug dosage correlated with improved classification accuracy.
• Predicted informative tissue structures by HAN closely matched drug localization visualized by in situ hybridization staining.

Conclusions:

• The proposed deep learning-assisted HS-CARS technique enables label-free profiling, identification, and localization of drug fingerprints in biological samples.
• This method offers potential for nonperturbative investigations of complex biological systems.
• The approach demonstrates significant promise for advancing drug development and understanding therapeutic mechanisms.