Machine Learning Classification of 53BP1 Foci

Abstract

AbstractBackground53BP1 foci are reflective of DNA double-strand break formation and have been used as radiation markers. Manual focus counting, while prone to bias and time constraints, remains the most accurate mode of detecting 53BP1 foci. Several studies have pursued automated focus detection to replace manual methods. Deep learning, spatial 3D images, and segmentation techniques are main components of the highest performing automated methods. While these approaches have achieved promising results regarding accurate focus detection and cell classification, they are not compatible with time-sensitive large-scale applications due to their demand for long run times, advanced microscopy, and computational resources. Further, segmentation of overlapping foci in 2D images has the potential to represent focus morphologies inaccurately.ResultsTo overcome these limitations, we developed a novel method to classify 2D fluorescence microscopy images of 53BP1 foci. Our approach consisted of three key features: (1) general 53BP1 focus classes, (2) varied parameter space composed of properties from individual foci and their respective Fourier transform, and (3) widely-available machine learning classifiers. We identified four main focus classes, which consisted of blurred foci and three levels of overlapping foci. Our parameter space for the training focus library, composed of foci formed by fluorescently-tagged BP1-2, showed a wide correlation range between variables which was validated using a publicly-available library of immunostained 53BP1 foci. Random forest achieved one of the highest and most stable performances for binary and multiclass problems, followed by a support vector machine and k-nearest neighbors. Specific metrics impacted the classification of blurred and low overlap foci for both train and test sets.ConclusionsOur method classified 53BP1 foci across separate fluorescent markers, resolutions, and damage-inducing methods, using off-the-shelf machine learning classifiers, a diverse parameter space, and well-defined focus classes.