In vivo patterning with in silico signaling, or how we put the cell in the loop

We substituted chemical cell-to-cell signaling with computer-controlled light inputs to generate checkerboard patterns of gene expression, quantitatively predicted by mathematical theory.

Like Comment
Read the paper

Synthetic biologists have recently made great strides in engineering gene expression patterning into multicellular systems, with prospective applications including biomaterials and regenerative medicine.  Patterning is particularly challenging for researchers from both a theoretical and a practical standpoint.  For one, patterning systems often involve many interacting cells, each containing a complex genetic network; such high-dimensional nonlinear systems can be difficult to model, let alone analyze.  From an experimental perspective, finding biological components that satisfy parameter requirements is not straightforward, nor is monitoring synthetic circuits to ensure they are functioning properly.  Cells also come "preprogrammed" with intricate circuitry intended to help the cell survive, and synthetic additions can interact in complex, unexpected ways with host components.  Future research in synthetic patterning would benefit from theory that makes accurate predictions for system behavior, as well as from scalable and efficient platforms to validate these predictions in practice.

What might these platforms look like?  We took a hint from "conventional" engineering systems that already involve living organisms: humans.  In applications such as airplane flight or lunar landing, humans act as system components that react to outputs from—and provide new inputs to—other system components.  Especially for applications where safety is critical, a system must be carefully tested in simulation before it is deployed.  But individual humans are complex, unpredictable, and consequently quite difficult to model with much accuracy.  So instead of simulating humans, engineers simulate the remaining system components digitally and allow real humans to interact with them through an interface.  The combination of humans and digitally simulated components forms a human-in-the-loop simulation of the full system, which is more accurate than a completely digital model, and more cost effective (and safe) than testing directly on a physical prototype.

Inspired by human-in-the-loop simulations, we propose a cell-in-the-loop approach in which computationally calculated inputs substitute for physical cell-to-cell signaling in a synthetic biological system.  Incorporating live cells into the simulation eliminates the need to make assumptions about many aspects of cell behavior, while permitting users to easily prototype synthetic designs by programming them directly into the computer before going through the effort and expense to implement them fully in vivo.  We expect the benefits of cell-in-the-loop to be particularly clear in multicellular patterning systems, which produce appealing visual results that depend crucially on the parameter values involved in the process of cell-to-cell signaling.

Spontaneous checkerboard patterning with cell-in-the-loop.
Optogenetically responsive cells signal to each other through computer-controlled light inputs that vary in intensity based on the gene expression levels of other cells. Here, we implemented lateral inhibition, in which cells expressing high levels of gene inhibit their neighbors' gene expression by reducing the light input they receive.

Our project was a close collaboration between two labs: Murat Arcak's networked systems lab at the University of California, Berkeley and Mustafa Khammash's Control Theory and Systems Biology lab at ETH Zürich (in Basel).  The Arcak lab developed theory to predict the emergence of checkerboard patterns through lateral inhibition, which occurs when neighboring cells inhibit each other's gene expression.  Lateral inhibition was postulated as early as 1940 to form regularly spaced bristles on insect cuticles, and has since been found in diverse developmental contexts from the inner ear to the intestine.  Thus, lateral inhibition is interesting both for its relevance to natural patterning systems and for its relatively intuitive interpretation.  Our theory reduces the problem of predicting patterning to an easy-to-use graphical test for bistability.  Furthermore, our theory allows us to clearly delineate in vivo from in silico elements, since cell behavior is represented separately from the network structure (which cells signal to which).

The Khammash lab established the perfect experimental setup to implement cell-in-the-loop patterning.  In this setup, optogenetically responsive yeast cells are placed under a microscope and individually targeted with light to induce gene expression.  The computer can automatically image, segment, and track cells, as well as score their gene expression levels based on fluorescence measurements.  The computer can then use these scores to calculate and administer light inputs to cells in real time based on user-defined parameters.  Thus, we can program into the computer which cells interact with which and how strongly they inhibit each other, thereby implementing lateral inhibition-like behavior among cells that are not chemically communicating or even necessarily physically neighboring each other.  Unlike previous research using optogenetics for synthetic multicellular patterning, in our system the light inputs do not encode a pattern a priori.  Rather, light is used to implement cell-to-cell signaling, such that patterns emerge spontaneously without external control.

In our experiments, cells that signaled each other were visualized as neighbors on a virtual grid.  Boundary conditions on the grid were periodic, so that each cell had four virtual neighbors.  Lateral inhibition was then emulated by reducing the light input to cells whose virtual neighbors expressed high levels of gene.  We estimated gene expression from measurements of a fast-acting nuclear translocation reporter system (i.e., yeast responded to light by pulling fluorescent protein from the cytoplasm into the nucleus).  At each point in time, we assigned to each cell a nuclear localization score that increased with increasing gene expression.

Cells illuminated with light of constant intensity reached a steady-state response in about 20 minutes.  We tested several different intensities of light to construct a dose response curve from input light intensity to output cell score.  On average, cells had a graded response with higher expression levels for higher input intensity, although individual cell responses proved to be quite variable with regard to both intensity and time.  To reduce such stochastic influences, in the final patterning experiments lateral inhibition took place between patches of four or six cells whose scores were averaged to determine the patch score.  All cells in the same patch received the same input.  In our visualizations of patterning, the brightness of a patch on the virtual grid increased with increasing score.

Our theory combines a dose response curve with the strength of inhibition and each cell's number of neighbors to predict whether a given parameter set will produce steady-state checkerboard patterns.  We chose a fixed form for inhibition among cells, and varied one parameter to tune the inhibition strength.  For systems of 16 patches, our theory correctly predicted which parameter values were expected to produce patterns and which were not.  Although individual patterning experiments produced highly stochastic outcomes, the theory also quantitatively predicted the average contrast level in systems with patterning (checkerboard) and overall brightness is systems with no patterning (homogeneous).

Interestingly, although predictions for final patterning outcome were the same regardless of the number of patches under consideration, when we tested systems of 36 patches they produced localized patterns, but would not resolve into a global pattern before experiments ended.  In fact, they appeared not to reach a steady state at all.  We hypothesized this was because increasing the dimensionality of the grid also increases the possible space of system behaviors, and our experiments could not run long enough for such large systems to converge to steady state.  Fortunately, our setup also provided us with a way to test this prediction.  Because we calculated light inputs on the computer, we effectively had measurements not only for gene expression levels but also for signaling levels among cells (in fully biochemical systems, usually only the gene expression level is measured).  We also calculated new inputs more quickly than individual cells could reach a steady state under constant illumination, meaning cells would only report a constant-in-time score if the light input (and therefore the gene expression levels of other cells) were also constant for a few cycles beforehand.  This means we could assess whether the full lateral inhibition system had reached steady state by comparing the dose response curve to a plot of the relationship between average input level and average patch score during the experiment.  If the curves matched, then the system had probably reached steady state.  If the curves did not match, the system was probably still evolving.  We found that at the end of experiments the curves did indeed match for the 16-patch systems but not for the 36-patch systems, supporting our hypothesis that the larger systems would need more time to produce a pattern.

There are many potential steps for further research with cell-in-the-loop.  Although we chose an arbitrary deterministic function to represent inhibition strength, one could also model a specific biochemical mechanism of cell-to-cell signaling to compare with theoretical predictions.  Our experimental results suggested a sizeable role for variability both in individual cell responses and in patterning outcomes, so one could also consider simulating stochastic communication, and developing theory to understand stochastic influences on large-scale patterning.  Most of our discussion centers on using cell-in-the-loop as a simulation/prototyping tool for fully biochemical implementations, but we think there could also be "final applications" for cell-in-the-loop systems themselves—for example, by coupling cell-in-the-loop with a downstream pathway to achieve real-world results, or by implementing cell-in-the-loop signaling alongside chemical signaling.  In principle, there is no reason why cells even need to be on the same microscope slide to interact.  It might be hard to imagine the benefits of cells signaling each other over the internet, but then, our ancestors of a hundred years ago might have asked the same question of our human-in-the-loop selves.

Part of what we enjoyed about this project were the boundaries we crossed to make it happen: from in vivo to in silico, theory to experiment, engineering to biology, Basel to Berkeley.  As much for this process as for the results, we hope that our work will inspire future collaboration and interdisciplinary research in multicellular patterning, cybergenetics, and beyond.

Mindy Liu Perkins

Ph.D. student, University of California, Berkeley