Tokyo, Japan – Researchers from Tokyo Metropolitan University have developed a suite of algorithms to automate the counting of sister chromatid exchanges (SCE) in chromosomes under the microscope. Conventional analysis requires trained personnel and time, with variability between different people. The team's machine-learning-based algorithm boasts an accuracy of 84% and gives a more objective measurement. This could be a game changer for diagnosing disorders tied to abnormal numbers of SCEs, like Bloom syndrome.
DNA, the blueprint of life for all living organisms, is found packaged inside complex structures called chromosomes. When DNA is replicated, two identical strands known as sister chromatids, each carrying exactly the same genetic information, are formed. Unlike in meiosis, sister chromatids do not need to undergo recombination during mitosis, and in most cases they are transmitted intact to the daughter cells. However, when some form of damage occurs in DNA, the organism attempts to repair the lesion by using the remaining undamaged DNA as a template. During this repair process, it often happens that specific segments of the sister chromatids are exchanged with each other. During this repair process, it often happens that specific segments of the sister chromatids are exchanged with each other. This "sister chromatic exchange" (SCE) is not harmful itself, but too many can be a good indicator for some serious disorders. Examples include Bloom syndrome: affected people can have a predisposition to cancer.
To count SCEs, normal methods involve experienced clinicians looking at stained chromosomes under the microscope, trying to identify the telltale "swapped" segments of sister chromatids. Not only is this labor intensive and slow, but it can also be subjective, dependent on how the human eye perceives features. A fully automated analysis of microscope images would save time and give objective measures of the number of SCEs, for more consistent diagnoses across different clinical environments.
Now, a team led by Professors Kiyoshi Nishikawa and Kan Okubo from Tokyo Metropolitan University have developed a suite of algorithms using machine learning to count SCEs in images. They combined separate methods, one to identify individual chromosomes, another to tell whether there are SCEs, and finally, another to cluster and count them, giving an objective, fully automated measurement of the number of SCEs in a microscope image. They found an accuracy of 84.1%, a level which is enough for practical applications. To see how it performed with real data, they collected images of chromosomes from cells with an artificially knocked out BLM gene, the kind of suppression seen in Bloom syndrome patients. The team's algorithm was able to give counts for SCEs which were consistent with those given by human counters.
Work is currently under way to use the vast amounts of available clinical data to train the algorithm, with more refinements to come. The team believes that replacing manual counting with full automation will help realize faster, more objective clinical analysis than ever before, and that this is only the beginning for what AI can bring to medical research.
This work was supported by JSPS KAKENHI Grant Numbers 22H05072, 25K09513, and 22K12170.
