NTT constructs model that contributes to the design of artificial tissues derived from iPS cells

NTT has devised a technique for estimating cellular interactions during the formation of tissues made of multiple types of cells. Conventionally, a vast number of experiments and calculations were required for estimating cellular interactions from tissue structures. In this study, the company devised a method for developing an inverse surrogate model by machine learning and quickly estimating cellular interactions directly from structures already formed. NTT will utilize this method in research and development of bio-digital twins for replicating biological systems on computers. Their achievements were introduced at the "NTT Communication Science Laboratories OPEN HOUSE 2026," held from May 20 to 22.

Tissues, which make up organs, contain various types of cells such as blood vessels, muscles and nerves. Proper coordination of these cells is what enables the body to function. Advancement of iPS cell technology has enabled artificial production of various cells that make up tissues.

Many attempts have been made to produce artificial tissues (organoids) using these cells. However, the tissues created with iPS cells do not always function in the same way as the actual tissues. One of the causes is that the artificial tissues fail to reproduce the structures (cellular patterns) of their living counterparts, preventing the cells from properly coordinating with each other.

Identifying the mechanism for achieving the target cellular patterns should lead to the creation of organoids with functions closer to those in living organisms, opening the way for their application in regenerative medicine.

It is known that cellular patterns are generated through interactions (e.g., adhesion, repulsion) of specific types of cells, and experimental control of cellular patterns is becoming more and more feasible.

In this study, the company devised a method for directly estimating the cellular interactions. They combined an "inverse surrogate model," which from a given cellular pattern directly estimates the cellular interactions that achieves its formation, with "multiscale featurization of adjacent structures," which takes into account the stochastic appearance and disappearance of cells.

Cellular interaction is a model that represents cells as agents and simulates the cellular patterns generated through their interactions.

Surrogate models are machine learning models that can be run at relatively low computational costs to serve as alternatives to simulations requiring high computational costs.

The method developed by NTT adopts a model (inverse surrogate model) that runs in the direction opposite to the forward model, i.e., inputs cellular patterns and outputs parameters of cellular interactions responsible for the pattern formation.

The first step in constructing an inverse surrogate model is "training data generation." In this step, simulations by an agent-based model (ABM) that represents individual cells as agents are performed by changing the interaction parameters in various ways. This yields sets of cellular patterns and feature quantity corresponding to individual interaction parameters, which serve as the data set for training the inverse surrogate model.

Next, this data set is used in "machine learning" for training a multilayer perceptron (MLP), an artificial neural network, to construct an inverse surrogate model that estimates interaction parameters based on the simulation results. By adopting these procedures, direct interactions can be achieved without performing repetitive searches.

The developed method was applied to stripe pattern formation in zebrafish. Regarding visual reproducibility, the generated cell patterns showed higher visual similarity (e.g., size of stripes and dots) to the target cell patterns than that achieved with an existing method. The time required for each simulation run in this method is estimated to be a few seconds on a laptop computer, which is much shorter than the time required in a conventional method (about a few hours with 100 simulation iterations).