Selection of organoids
Machine learning will help grow artificial organs
Phys Tech blog, Naked Science
Scientists from MIPT together with colleagues from the Institute of System Programming and the Charles Schepens Institute of Eye Research at Harvard School of Medicine (USA) To do this, the algorithm, unlike humans, does not require additional modification of cells. This allows the method to be used when growing the retina for transplantation.
The results are published in the journal Frontiers in Cellular Neuroscience (Kegeles et al., Convolutional Neural Networks Can Predict Retinal Differentiation in Retinal Organoids).
In multicellular organisms, each organ and tissue type consists of cells with different functions and properties. They acquire these functions in the process of development. At the very beginning, all cells are the same, but they are potentially capable of creating all types of cells of a mature organism, during this period of time such cells are called stem cells.
When proteins working in certain tissues begin to be synthesized in some of them, differentiation and specialization of cells occurs. After that, groups of cells form different tissues and organs. The most modern approach for reproducing the process of development of various tissues in a test tube is the technology of differentiation in three–dimensional cellular aggregates - organoids. This technology has already shown its effectiveness for studying the development of the retina, brain, inner ear, intestines, pancreas and many other tissues.
Due to the fact that the differentiation process according to this technology is based on natural mechanisms of development, the resulting tissue has a significant similarity with a natural organ. The nature of some stages of differentiation is random, which leads to a significant change in the number of cells with a certain function, even among artificial organs in one batch, not to mention different cell lines.
This means that for reproducibility of experiments and, as a consequence, for the greatest reliability in clinical applications, it is necessary to be able to determine which cells have specialized and which have not. To determine differentiated cells when working with tissues, specialists use fluorescent proteins – the gene of the luminous protein is added to the DNA of cells, as a result of which the latter begin to synthesize it when they pass the desired stage of development.
Unfortunately, this sensitive, specific and quantifiable method is not suitable for the production of cells for transplantation or modeling hereditary diseases of a genetic nature. That is why scientists in this paper have proposed an alternative approach for analysis – based on the structure of the tissue itself.
To date, there are no reliable and objective criteria to predict the quality of cell differentiation. To solve the problem of selecting the best retinal tissues for further transplantation, drug screening or disease modeling, scientists decided to use neural network and artificial intelligence methods.
"One of the main activities of our laboratory is the application of bioinformatics, machine learning and artificial intelligence methods to solve applied problems in the field of genetics and molecular biology. This development is just at the junction of sciences. In it, the classic tools of neural networks for Physical Technology are used for a very significant applied biomedical problem – the prediction of differentiation into the retina from stem cells. The human retina has an extremely limited potential for regeneration.
This means that any progressive loss of neurons, for example, in glaucoma, inevitably leads to complete blindness. Now doctors have practically nothing to offer such patients, except to start learning Braille tables. Our work brings biomedicine one step closer to the creation of cell therapy for retinal diseases, which will not only prevent the progression of the disease, but also restore to patients already lost vision," explains Pavel Volchkov, head of the Laboratory of Genomic Engineering at MIPT.
The authors of the article trained a neural network (a computer algorithm named so by analogy with the work of human neurons in the brain) to find the tissues of the developing retina based on photographs from a simple light microscope. First, they asked experts to identify differentiated cells in 1200 images using an accurate method using a fluorescent reporter.
The neural network was trained on 750 images, another 150 were used for validation and 250 for tests. After checking all the predictions, it turned out that people identified differentiated cells with an accuracy of about 67 percent, while the neural network had an accuracy of 84 percent.
"Our results show that the criteria for selecting retinal tissues at an early stage are subjective and depend on the expert who makes the decision. At the same time, the morphology (that is, the structure) of the tissue itself, even at a very early stage, makes it possible to predict the differentiation of the retina. And a program, unlike a human, can extract this information.
Given that this approach does not require complex images, fluorescent reporters or dyes for analysis, it is easy to implement. This allows us to take another step towards the creation of cellular therapies for retinal diseases such as glaucoma and macular degeneration, which now almost inevitably lead to blindness. In addition, this approach can be transferred not only to other cell lines, but also to human artificial organs," adds Evgeny Kegeles, an employee of the MIPT Orphan Disease Therapy Laboratory.
Portal "Eternal youth" http://vechnayamolodost.ru