Deep learning accelerates research on early pregnancies

Researchers at King Abdullah University of Science and Technology (KAUST; Saudi Arabia) have announced the development of a new deep learning tool, deepBlastoid, to study models of human embryo development in artificial laboratory conditions. The KAUST scientists showed that deepBlastoid can evaluate images of the models equally to expert scientists but 1000 times faster.  

The earliest stages of the human embryo are crucial for understanding fertility, pregnancy complications, and the origins of developmental disorders. However, direct research on human embryos is limited by ethical considerations.  

Blastoids are cellular models that represent the embryo at a period known as the blastocyst stage. This stage begins about five days after fertilization and continues until the embryo has implanted itself into the mother's uterine wall (i.e. the moment of pregnancy). Importantly, the human blastoids in the KAUST study are made of stem cells but not any embryonic tissue, and since they were first discovered in 2021 have quickly become a preferred human model for scientists to studying early embryo development.  

In the new study, the KAUST researchers trained deepBlastoid on over 2000 microscopic images of blastoids and then used it to judge the effects of chemicals on blastoid development by examining over another 10 000 images. Understanding how these chemicals can disrupt blastoids have profound implications for women who are taking prescription medicine or other drugs but seek to become pregnant.  

"Little is known about the very early stages of embryo development. With deepBlastoid, we can scale up blastoid research to study embryo development and the effects of chemicals on the embryo and pregnancy," said KAUST Associate Professor Mo Li, an expert of stem cell biology and whose lab pioneered embryo models using human blastoids.  

He added that deepBlastoid will help advance reproductive technologies like in vitro fertilization too.  

Generally, scientists evaluate blastoids manually by systematically reviewing a library worth of images taken under a microscope. This approach not only takes time, it is also sensitive to the knowledge of the scientist and to the method used to produce the blastoid, which can vary across laboratories. On the other hand, deepBlastoid can process 273 images per second, offering scientists a tool to assess tens of thousands of blastoids in just a few minutes. 

"deepBlastoid not only matches human performance in accuracy, it delivers an unparalleled increase in throughput. This efficiency allows scientists to analyze vast amounts of data in a short time, enabling experiments that were previously unfeasible," said KAUST Professor Peter Wonka, an expert in deep learning and computer vision and whose research team developed deepBlastoid.  

While Li, Wonka, and their colleagues used deepBlastoid to study the blastoid, they said that by adapting the deep learning algorithm, their deep learning approach could be used to other stem cell models for other embryo stages and organs.