New Model Predicts Multicellular Development Processes

During early development, tissues and organs arise as thousands of cells shift, divide, and reorganize. A team of MIT engineers has developed a deep-learning approach that predicts, minute by minute, how individual cells will fold, rearrange, and divide in a fruit fly embryo’s earliest developmental stage. The method focuses on gastrulation, an approximately one-hour window in which cells move rapidly and the embryo changes from a smooth structure into one with distinct folds and features. By modeling this early period in detail, the researchers aim to understand how local cell interactions collectively shape tissues and whole organisms.

The model, described recently in Nature Methods, learns how key geometric properties of individual cells evolve over time. It tracks features such as cell position, whether a cell is in contact with specific neighbors, and whether it is folding or dividing at a given moment. The team applied the method to videos of fruit fly embryos, each beginning as a cluster of about 5,000 cells. When tested on unseen data, the model predicted with about 90 percent accuracy how each cell would move and change during the first hour of development, including whether cells maintain or lose shared edges and when such events occur.

To build this predictive framework, the researchers combined two common ways of representing developing tissues. Traditionally, embryos are modeled either as point clouds, where each point is a cell, or as foams, where cells resemble bubbles that slide against one another. Rather than choose between these abstractions, the team treated both as different views of the same underlying graph. This led to a “dual-graph” structure that represents cells simultaneously as moving points and as foam-like compartments, making it possible to capture connections among cells and detailed geometry, such as nuclei location and cell–cell contacts, over time.

The model was trained and tested using high-quality, single-cell–resolution videos of fruit fly gastrulation, which included labeled cell edges and nuclei. These rare datasets provided the submicron spatial detail and fast frame rates needed to follow thousands of cells through rapid shape changes. After learning from three embryo videos, the model could accurately predict cell behaviors in a fourth, including whether individual cells would fold, divide, or remain attached to neighbors, and precisely when these transitions would occur.

The researchers see broad potential for applying this dual-graph approach to other multicellular systems. They hope to extend it to species such as zebrafish and mice, and eventually to certain human tissues and organs. However, they note that the main limitation is not the model itself but the availability of similarly detailed video data. With sufficiently high-quality recordings of specific tissues, the same framework could be used to predict how many more biological structures develop over time.