Researchers have created an artificial intelligence (AI) model that can forecast the early development of organoids with greater speed and precision than human experts. This innovation has the potential to enhance efficiency and reduce costs in organoid cultivation.
Organoids, which are small, lab-grown tissues that replicate the functions and structures of organs, are paving the way for advancements in biomedical research. They offer opportunities for personalized transplants, better understanding of diseases such as Alzheimer’s and cancer, and more accurate insights into how medical drugs affect the body.
Recently, scientists from Kyushu University and Nagoya University in Japan introduced a model that employs AI to predict how organoids develop in their initial stages. This model provides faster and more precise predictions than those made by experienced researchers, potentially streamlining the organoid cultivation process and cutting costs.
The study, published in Communications Biology on December 6, 2024, examined the development of hypothalamic-pituitary organoids, which serve to emulate the functions of the pituitary gland, such as producing adrenocorticotropic hormone (ACTH). ACTH plays a critical role in governing stress, metabolism, blood pressure, and inflammation, with a deficiency possibly resulting in severe issues like fatigue or anorexia.
“Our experiments with mice in the lab suggest that transplanting hypothalamic-pituitary organoids could address ACTH deficiency in humans,” states Hidetaka Suga, Associate Professor at Nagoya University’s Graduate School of Medicine and the study’s corresponding author.
One of the major hurdles for researchers is verifying the correct development of the organoids. As they are derived from stem cells suspended in a liquid medium, even slight environmental changes can affect their growth and quality.
The team discovered that the wide presence of a protein called RAX during the early stages indicates proper development, often leading to organoids that later secrete high levels of ACTH.
“We can monitor development by genetically modifying the organoids to make the RAX protein glow,” Suga explains. “However, organoids intended for clinical applications, like transplants, cannot be altered genetically to fluoresce. This forces our researchers to assess them based solely on visual observations, which is a slow and unreliable method.”
To tackle this, Suga and his team collaborated with Hirohiko Niioka, a Professor at Kyushu University’s Data-Driven Innovation Initiative, to develop deep-learning models for this analysis.
“Deep-learning models function similarly to the human brain, processing information and recognizing patterns to categorize vast datasets,” elaborates Niioka.
The Nagoya research team gathered both fluorescent and bright-field images of organoids with fluorescent RAX proteins after 30 days of development. They categorized 1,500 bright-field images into three quality levels based on the RAX expression: A (high quality), B (medium quality), and C (low quality).
Niioka then utilized two sophisticated image recognition models, EfficientNetV2-S and Vision Transformer, created by Google, to predict the quality of the organoids. He trained the models using 1,200 bright-field images, with 400 from each category.
Once training was complete, Niioka amalgamated the two deep-learning models into a single ensemble model for enhanced performance. The research team then evaluated this optimized ensemble model using 300 remaining images (100 from each category), which resulted in a classification accuracy of 70% for the organoids based on bright-field images. In comparison, experienced researchers could only achieve below 60% accuracy in categorizing the same images.
“The deep-learning models surpassed expert predictions in accuracy, sensitivity, and speed,” highlighted Niioka.
The next phase involved testing whether the ensemble model could accurately categorize bright-field images of organoids that were not genetically modified to exhibit fluorescent RAX proteins.
The researchers assessed the trained ensemble model on bright-field images of hypothalamic-pituitary organoids without fluorescence after 30 days of development. Through staining techniques, they confirmed that the organoids labeled as A (high quality) by the model indeed exhibited high levels of RAX after 30 days. These organoids later also demonstrated significant ACTH secretion. Conversely, the organoids classified as C (low quality) exhibited low levels of RAX and subsequently, low ACTH production.
“Thus, our model can reliably predict the future quality of an organoid based solely on its visible characteristics at an early development stage,” Niioka states. “As far as we know, this marks the first use of deep-learning technology to forecast the trajectory of organoid development globally.”
The researchers aim to boost the accuracy of their deep-learning model by utilizing a larger dataset in the future. Nevertheless, even at its current accuracy, the model holds significant promise for advancing organoid research.
“We can swiftly and efficiently identify high-quality organoids for transplantation and disease modeling while simultaneously reducing time and costs by discarding those that are developing poorly,” concludes Suga. “This is a transformative advancement.”
