Computer vision techniques for high-speed atomic force microscopy of DNA molecules

Abstract High-speed atomic force microscopy (HSAFM) can produce thousands of topographic, nanoscale images over a small area. One emerging application of this technique is the detection and sizing of single DNA molecules derived from biological experiments and genetic testing. Using HSAFM images, researchers can visually categorize healthy and mutant DNA based on size and sequence-specific labeling, leading to rapid, high precision diagnostics. However, manually sifting through large numbers of images is time consuming and labor intensive, presenting a bottleneck for sample analysis. In this paper we look at how deep learning methods like object detection and image classification can assist in streamlining the process. To instantly assess image quality on data collected from trinucleotide repeat expansion disease samples, a fully convolutional network (FCN) was compared against Laplacian of Gaussian and fast Fourier transform methods. The FCN performed the best, reproducing human categorizations with an accuracy of 96% and an AUC of 0.990. Additionally, using the YOLOv8 architecture, we have developed an object detection model capable of detecting marked DNA from patients with Fragile X syndrome with an average precision of 0.966. The object detection model searched through 20 000 images containing DNA molecules and identified 248 marked molecules, of which 33 were real targets, greatly reducing the time taken to find target molecules for diagnostics. The integration of machine learning techniques in HSAFM systems shows promise to enhance the data collection and analysis process for genomics-based disease diagnosis.