One of the most fundamental processes in all of biology is the spontaneous organization of cells into clusters that divide and eventually turn into shapes – be they organs, wings or limbs.
Scientists have long explored this enormously complex process to make artificial organs or understand cancer growth – but precisely engineering single cells to achieve a desired collective outcome is often a trial-and-error process.
Harvard applied physicists consider the control of cellular organization and morphogenesis to be an optimization problem that can be solved with powerful new machine learning tools. In new research published in Nature Computational Science , researchers in the John A. Paulson School of Engineering and Applied Sciences (SEAS) have created a computational framework that can extract the rules that cells need to follow as they grow, in order for a collective function to emerge from the whole.
The computer learns these "rules" in the form of genetic networks that guide a cell's behavior, influencing the many ways cells chemically signal to each other, or the physical forces that make them stick together or pull apart.
Currently a proof of concept, the new methods could be combined with experiments to allow scientists to understand and control how organisms develop from the cellular level.
The research was co-led by graduate student Ramya Deshpande and postdoctoral researcher Francesco Mottes. The senior author was Michael Brenner , Catalyst Professor of Applied Mathematics and Applied Physics at SEAS.
Automatic differentiation
The search for rules that cells must follow was enabled by a computational technique called automatic differentiation. This method, which forms the backbone of training deep learning models in artificial intelligence, consists of algorithms designed to efficiently compute highly complex functions. Automatic differentiation allows the computer to detect the precise effect that a small change in any part of the gene network would have on the behavior of the whole cell collective.
For the last several years, Brenner's team has been applying such algorithms to problems beyond neural networks, including designing self-assembling colloid materials, improving fluid dynamics simulations, or engineering certain types of proteins.
Deshpande said the principles from the paper could help guide follow-up experiments on physical systems of cells. "Once you have a model that can predict what happens when you have a certain combination of cells, genes or molecules that interact, can we then invert that model and say, 'We want these cells to come together and do this particular thing. How do we program them to do that?'
Mottes said that by enabling the scaling of physics-based systems biology models, automatic differentiation offers a promising path toward achieving the predictive control needed to, in the distant future, engineer the growth of organs — the holy grail of computational bioengineering.
"If you have a model that is predictive enough and calibrated enough on experimental data, the hope is that you can just say, for example, 'I want a spheroid with these characteristics. How should I engineer my cells to achieve this?'" Mottes said.
The research was supported by the Office of Naval Research and the NSF AI institute of Dynamic Systems. The paper is dedicated to the memory of former Harvard postdoctoral researcher Alma Dal Co.
