Convolutional neural network-based regression analysis to predict subnuclear chromatin organization from...

Area of Science:

• Biophysics
• Computational Biology
• Cell Biology

Background:

• Azimuth-resolved optical scattering signals from cell nuclei are sensitive to internal refractive index profiles.
• These signals offer insights into chromatin organization and subnuclear refractive index fluctuations.

Purpose of the Study:

• To determine if 2D scattering signals can quantitatively extract spatial correlation length and extent of subnuclear refractive index fluctuations.
• To provide quantitative information on chromatin distribution using an inverse scheme.

Main Methods:

• A data-driven approach using machine learning (convolutional neural network - CNN) for regression analysis.
• Numerical computation of 198 scattering signals for nuclear models with varying parameters.
• Five-fold cross-validation to quantify prediction performance.

Main Results:

• CNN-based regression accurately predicted spatial correlation length ($$) and extent ($$) of refractive index fluctuations.
• Mean absolute percent errors were 8.5% for $$ and 13.5% for $$.
• Prediction performance was excellent, with errors smaller than model parameter increments.

Conclusions:

• CNN-based regression is a powerful method for analyzing 2D optical scattering signals.
• This approach enables quantitative monitoring of chromatin organization.
• The study demonstrates the potential of machine learning in biophysical analysis.