Abstract
DNA fluorescence in situ hybridization (FISH) is the technique of choice to map the position of genomic loci in three-dimensional (3D) space at the single allele level in the cell nucleus. High-throughput DNA FISH methods have recently been developed using complex libraries of fluorescently labeled synthetic oligonucleotides and automated fluorescence microscopy, enabling large-scale interrogation of genomic organization. Although the FISH signals generated by high-throughput methods can, in principle, be analyzed by traditional spot-detection algorithms, these approaches require user intervention to optimize each interrogated genomic locus, making analysis of tens or hundreds of genomic loci in a single experiment prohibitive. We report here the design and testing of two separate machine learning-based workflows for FISH signal detection in a high-throughput format. The two methods rely on random forest (RF) classification or convolutional neural networks (CNNs), respectively. Both workflows detect DNA FISH signals with high accuracy in three separate fluorescence microscopy channels for tens of independent genomic loci, without the need for manual parameter value setting on a per locus basis. In particular, the CNN workflow, which we named SpotLearn, is highly efficient and accurate in the detection of DNA FISH signals with low signal-to-noise ratio (SNR). We suggest that SpotLearn will be useful to accurately and robustly detect diverse DNA FISH signals in a high-throughput fashion, enabling the visualization and positioning of hundreds of genomic loci in a single experiment.