Deriving accurate molecular indicators of protein synthesis through Raman-based sparse classification

Area of Science:

• Biophysics
• Spectroscopy
• Machine Learning

Background:

• Raman spectroscopy provides non-invasive molecular insights from live biological samples.
• Machine learning can build predictive models from complex spectroscopic data.
• Linear models offer interpretable insights into feature contributions for classification.

Purpose of the Study:

• To compare the performance of principal component analysis/linear discriminant analysis (PCA/LDA) with regularized logistic regression (Lasso) for single-cell Raman spectroscopy analysis.
• To evaluate the stability and interpretability of linear models for detecting macrophage activation.
• To investigate the utility of sparse separation vectors derived from Lasso for identifying key molecular features.

Main Methods:

• Application of PCA/LDA and Lasso to single-cell Raman spectroscopy data.
• Analysis of macrophage activation models.
• Evaluation of model stability and sample size requirements.
• Investigation of feature selection using Lasso on a protein synthesis inhibition model.

Main Results:

• Lasso demonstrated superior classification performance over PCA/LDA for macrophage activation detection.
• Lasso models were more stable and required smaller sample sizes.
• Sparse separation vectors from Lasso identified relevant Raman bands, including side bands, for classification.
• Feature selection highlighted RNA accumulation and protein depletion under inhibited protein synthesis.

Conclusions:

• Lasso is a more effective and stable method than PCA/LDA for analyzing single-cell Raman spectroscopy data.
• Sparse features identified by Lasso provide biologically relevant insights into cellular states.
• The prevalence of side bands in discriminating features suggests their specificity for accurate classification over broad Raman peaks.