Feynman Returns: Controlling Biological Systems
By: Filippo Menolascina
Environmental Microfluidics Group
Dept. of Civil & Environmental Engineering
Massachusetts Institute of Technology
“What I cannot create I do not understand” – Richard Feynman
Over the past several decades of research in biology a great deal of effort has been expended to understand the inner mechanics of cells. How do cells respond to stimuli? How do they process information and make decisions? The detail we gained on fundamental processes underlying life has been unprecedented. With all this new information in our hands, embracing the challenge Richard Feynman launched to all disciplines from his blackboard seemed the natural “next step:” creating new biological systems was the ultimate demonstration that we really understood what we learnt from them. A challenge that remain largely unanswered until the early 2000s when Synthetic Biology made its appearance and changed the game: by combining principles from engineering, math, physics and biology, scientists were able to build biological systems ex-novo. In less than fifteen years we went from the genetic toggle switches in bacteria to mind-controlled gene expression in mammalian cells. We had won the challenge.
The question was then, “what’s next?” What if Richard Feynman could formulate a new challenge, in light of these new discoveries? What would that be? A natural answer can be found looking at the history of engineering: we build systems starting from detailed understanding of their dynamics, then, in order to make real use of them, we need to control them. An example will clarify what I mean. In order to build the first plane we observed birds for a long time, studying how the interplay between the air and their wings allowed them to fly. When we were confident with our understanding of the physics behind flights we built a “flying machine”. The first fight was a great success, but it only lasted 12 seconds. Orville Wright, the pilot, commented: “I found the control […] quite difficult”. Without the ability to control the plane, that machine, a great proof of our understanding of physics, was useless. Controlling the plane, i.e. letting an automatic system such as a mechanical/electrical device identify the right commands (inputs) to, for example, keep a steady altitude (output) is still at the heart of any flight. Someone might argue that this automation is the very essence of what makes flights, as we know them, possible. What if the next step in biology is controlling living systems? Being able to force living systems to display a desired behavior (just like we would do with cruise control on planes) would enable potentially groundbreaking advancements. For example we could interpret cancer as an undesirable working point of a dynamic system, just like a wrong altitude in the previous example, and control the cell to revert to a physiological state via the application of an appropriate stimulus. Intriguing, but how do we go about it?
Inspired by this question, my co-authors and I  started looking for a compelling way to demonstrate we could use principles from control theory to force a biological system to behave like we wanted. To this aim we devised a simple yet effective experiment: we took two networks in S. cerevisiae and we set out to precisely control the concentration of one protein in each of them. The two networks we focused on were very different. The GAL1 system features a galactose inducible promoter driving the transcription of the fusion between GAL1 and GFP genes (Fig. 1); a very essential design indeed from yeast genome. IRMA, in contrast, is one of the most complex synthetic gene networks built in a eukaryotic cell and features five genes, one of them fused to GFP, regulating each other with both activation and repression. Interestingly enough, both systems had a similar behavior: when exposed to galactose, cells would glow green when observed under the microscope due to the GFP fluorescence. Glucose, on the other hand, would repress GFP expression, making cells dim. We had two very different gene circuits that worked similarly, with the same input (glucose/galactose), same output (green fluorescence) and same behavior (switches). The formulation of the control problem followed naturally: we aimed at precisely regulating the concentration of GFP produced in response to the administration of glucose/galactose.
To accomplish this goal we analyzed the mathematical models of the two circuits and devised a simple control strategy that, in essence, compared the actual fluorescence cells were producing with the level of fluorescence we desired. Based on the difference between the two the algorithm decided whether the input should have been galactose (if the actual fluorescence was smaller than the desired one) or glucose (Fig. 2). We tested this algorithm on the mathematical models we had started from. Indeed, the results were exciting: the simulations confirmed we could achieve gene network control, in-silico.
We were one step away from the results we were after, as we had to prove our control strategy worked on live cells. Achieving this goal was not trivial. Our simulations suggested we had to move cells from glucose to galactose (or vice versa) in as little as 5 minutes, too quickly for any classical experimental approach based on large volumes (e.g. flasks). Going “micro” was the solution to this problem. We used a microfluidic device to trap cells and observe them while they were exposed to a continuous flow of nutrients controlled by a computer running our control algorithm. To close the loop we then connected this computer to the microscope the microfluidic device was mounted on. We also programmed the system to take, at regular time intervals, micrographs of the cells both in phase contrast and fluorescence. The control algorithm then identified the cells, estimated their fluorescence and used this information to decide whether to push galactose or glucose, all in real-time. Thanks to this approach we were able to make our gene circuits produce steady as well as time-varying amounts of GFP, demonstrating in-vivo control of gene circuits (Fig. 3).
What if Richard Feynman could launch one last challenge? I always thought it would have been: “What I cannot control I do not understand”. This is the challenge we envisioned, we embraced and eventually won. What’s next? Control of behavioral traits, development of engineered cell lines serving as controllers for tissue engineering; these are only few of the many unchartered territories open for exploration. For sure we will have plenty to choose from, until the next big question.
 Menolascina F, Fiore G, Orabona E, De Stefano L, Ferry M, et al. (2014) In-Vivo Real-Time Control of Protein Expression from Endogenous and Synthetic Gene Networks. PLoS Comput Biol 10(5): e1003625. doi:10.1371/journal.pcbi.1003625