Related Experiment Video
Updated: Jun 15, 2025
07:37
An Integrated Raman Spectroscopy and Mass Spectrometry Platform to Study Single-Cell Drug Uptake, Metabolism, and Effects
Published on: January 9, 2020
9.4K
Raman spectroscopic deep learning with signal aggregated representations for enhanced cell phenotype and signature
Songlin Lu1,2, Yuanfang Huang1, Wan Xiang Shen3
1The State Key Laboratory of Chemical Oncogenomics, Key Laboratory of Chemical Biology, Tsinghua Shenzhen International Graduate School, Tsinghua University, 2279 Lishui Road, Nanshan District, Shenzhen 518055, Guangdong, P. R. China.
PNAS Nexus
|August 28, 2024
Summary
Novel 2D representations improve deep learning for Raman spectroscopy, enhancing cell identification. New DSCANets models significantly outperform existing methods for accurate, label-free cell analysis.
Area of Science:
- Spectroscopy
- Machine Learning
- Deep Learning
Background:
- Machine learning (ML) and deep learning (DL) are used for label-free cell identification using Raman spectra.
- Challenges include high-dimensional, unordered, and low-sample spectroscopic data.
- Accurate cell phenotype identification is crucial for diagnostics, therapeutics, and microbiology.
Purpose of the Study:
- To develop novel feature representations for spectroscopic data.
- To enhance deep learning models for cell phenotype and signature identification.
- To improve the accuracy and performance of Raman spectroscopy analysis.
Main Methods:
- Introduced novel 2D image-like dual signal and component aggregated representations by restructuring Raman spectra and principal components.
- Developed new Convolutional Neural Network (ConvNet) models named DSCARNets.
- Evaluated DSCARNets on six benchmark datasets and four additional datasets.
Main Results:
- DSCARNets significantly outperformed state-of-the-art (SOTA) ML and DL models on six datasets, achieving >2% improvement over 85-97% accuracies.
- Demonstrated strong performance on four additional datasets against SOTA models with >98% accuracy.
- Successfully applied to datasets lacking published supervised phenotype classification models.
- Explainable DSCARNets identified Raman signatures consistent with experimental findings.
Conclusions:
- Novel 2D representations and DSCARNets enhance deep learning for Raman spectroscopy.
- Achieved superior performance in cell phenotype and signature identification.
- Provides a powerful, explainable tool for label-free cell analysis in various applications.
