Application of neural network-based image analysis to detect sister chromatid cohesion defects

Sister chromatid cohesion (SCC) is mediated by the cohesin complex and its regulatory proteins. To evaluate the involvement of a protein in cohesin regulation, preparations of metaphase chromosome spreads and classifications of chromosome shapes after depletion of the target protein are commonly employed. Although this is a convenient and approved method, the evaluation and classification of each chromosome shape has to be performed manually by researchers. Therefore, this method is time consuming, and the results might be affected by the subjectivity of researchers. In this study, we developed neural network-based image recognition models to judge the positional relationship of sister chromatids, and thereby detect SCC defects. Transfer learning models based on SqueeezeNet or ResNet-18 were trained with more than 600 chromosome images labeled with the type of chromosome, which were classified according to the positional relationship between sister chromatids. The SqueezeNet-based trained model achieved a concordance rate of 73.1% with the sample answers given by a researcher. Importantly, the model successfully detected the SCC defect in the CTF18 deficient cell line, which was used as an SCC-defective model. These results indicate that neural-network-based image recognition models are valuable tools for examining SCC defects in different genetic backgrounds.