Artificial Intelligence Surpasses Experts in Forecasting Quality of Laboratory-Grown Organoids

Recent advancements in biomedical research have seen organoids--miniaturized, lab-cultivated tissues that replicate the functions and structures of real organs--emerge as powerful tools for various applications, including personalized transplants, enhanced disease modeling for conditions such as Alzheimer's and cancer, and better assessments of drug efficacy.

A new study conducted by researchers from Kyushu University and Nagoya University in Japan has introduced an innovative model utilizing artificial intelligence (AI) to predict the development of organoids at an early stage. This AI-driven model has demonstrated superior speed and accuracy compared to traditional expert assessments, which could significantly streamline organoid culturing processes and reduce associated costs.

The focus of this research, published in Communications Biology, was on hypothalamic-pituitary organoids, which are designed to emulate the functions of the pituitary gland, particularly its role in producing adrenocorticotropic hormone (ACTH). This hormone is essential for regulating stress responses, metabolism, blood pressure, and inflammation. A deficiency in ACTH can result in severe health issues, including fatigue and anorexia.

Hidetaka Suga, an associate professor at Nagoya University, noted the potential for hypothalamic-pituitary organoids to treat ACTH deficiencies in humans based on laboratory studies involving murine models. However, one of the primary challenges faced by researchers is determining the correct development of these organoids, which are derived from stem cells that are sensitive to environmental fluctuations. This sensitivity can lead to inconsistencies in their growth and overall quality.

In their investigations, researchers identified that a broad expression of a protein known as RAX during early developmental stages is indicative of favorable progression. Organoids exhibiting wide RAX expression are more likely to achieve robust ACTH secretion later in their development.

Utilizing advanced imaging techniques, the researchers captured both fluorescent and bright-field images of organoids at 30 days of growth. The fluorescent images served as a benchmark for categorizing the bright-field images into three distinct quality classifications: A (high quality with wide RAX expression), B (medium quality with moderate RAX expression), and C (low quality with narrow RAX expression).

To enhance the categorization process, the team collaborated with Hirohiko Niioka, a professor specializing in data-driven innovation at Kyushu University, to develop deep-learning models capable of performing this classification. Deep-learning technologies simulate human cognitive functions, enabling the analysis and recognition of patterns across extensive datasets.

Using a training set of 1200 bright-field images, comprising 400 images from each quality category, two deep-learning models--EfficientNetV2-S and Vision Transformer, both designed by Google for image recognition--were trained. The ensemble of these models achieved a classification accuracy of approximately 70% when tested against a separate set of 300 images, significantly surpassing the less than 60% accuracy achieved by experienced researchers.

This study marks a pioneering application of deep-learning technology in predicting organoid development trajectories based solely on visual data. The next steps involve refining the deep-learning model by expanding the dataset used for training, aiming to enhance its predictive accuracy further.

Ultimately, the implications of this research are substantial, as it allows for the rapid identification of high-quality organoids suitable for transplantation and disease modeling while simultaneously minimizing the time and resources spent on less viable specimens. This advancement is poised to revolutionize organoid research and its applications in clinical settings.