This is the website of Michael Todhunter.

3/24/24: The media optimization project described on this page has been parlayed into a startup venture, called Dragonase (see sidebar), that intends to apply cell culture media optimization to the problem of anti-aging hematopoietic stem cell therapy.

***

I intend to use this space to document a cell culture media automation project recently awarded an ACX Grant.

I am trying to discover how to better control the growth, differentiation, and function of cultured human cells with a robotic system that autonomously formulates and tests cell culture media recipes.

Essentially all laboratory-based human biological research relies on maintaining cells in culture medium that provides an appropriate chemical environment: nutrients, carbon sources, growth factors, and so on. The behaviors of cells - how long they replicate, how quickly they senesce, whether they differentiate, and what functional behaviors they exhibit - are all functions of their culture media environment. Most culture media recipes were designed at least 50 years ago (DMEM was formulated in 1959, Ham’s F12 in 1965, and RPMI in 1966) when the goals of cell culture were much narrower than today. Although the scope and applications of cell biology have vastly increased since then, media formulation has not kept pace.

Many questions in cell biology cannot be answered without appropriate culture media - e.g., there aren’t many ways to study bone differentiation without osteogenic media or lactation without lactogenic media. Most cell types from most human tissues cannot be propagated in vitro, owing to a lack of compatible culture conditions. Culture media formulation is an underdetermined problem.

There has been little progress formulating better culture media, partially because the research is extremely unsexy. The work requires repetitive trial-and-error experimentation, the papers rarely make it into high-impact journals, and the funding is hard to come by - media formulation is not a path that leads a professor to tenure. Aside from that, the design space for media recipes is large and dense. DMEM, a simple medium, has 32 ingredients, and a more complex medium like MEGM can have 70 or more. The relevant concentration range to test for any given ingredient may span several orders of magnitude. Some ingredients, like B27 or FBS, are themselves complex mixtures of dozens (B27) or hundreds (FBS) of constituent ingredients. These ingredients interact with each other - e.g., the uptake of glucose varies with the concentration of insulin, and the bioavailability of steroids varies with the concentration of albumin. This complexity makes the hypothesis-driven research of media recipes very daunting.

If we treat media formulation as a high-dimensional optimization problem instead of a hypothesis-driven problem, we can overcome the challenges of the culture media design space. For a media recipe design space with 70 parameters, where each parameter could take any of 5 values, there are 10⁴⁸ potential media recipes to explore. This greatly exceeds the exploration capacity of factorial Taguchi methods or high-throughput screening, but it is well within the capacity of modern machine learning-based approaches such as gradient descent or quasi-Newton optimization. Compared to state-of-the-art machine learning projects like GPT-3 or AlphaFold that already use similar methods, the parameter space required for media optimization is trivial.

We need robotics to run the numerous experiments required as grist for our high-dimensional optimization. Fortunately, the necessary automation tools already exist. Robotic cell culture and robotic liquid handling are ubiquitous in industrial settings, despite being uncommon in academic labs. Motorized, computer-controlled microscopy makes it feasible to acquire data on the growth and morphology of living cells, and computer vision algorithms, such as random forests or recurrent neural networks, are capable of interpreting massive amounts of this kind of data. The challenge is linking these systems into a closed loop that can operate iteratively and optimize for a specific, defined goal.

I have been working on a prototype system that uses motorized microscopy, computer vision, and a statistics pipeline to automatically evaluate experiments testing several dozen media conditions. The system uses the output of one experiment to suggest conditions to test in the next experiment, allowing the system to operative iteratively. Although we have successfully used this prototype to explore media formulations for human mammary epithelial cells, it is not yet a closed-loop system - human intervention is required for liquid handling (feeding cells and mixing media components) and transferring samples to-and-from the incubator, culture hood, and microscope. Closing the loop will vastly increase the speed and utility of this system.

The next steps for this prototype system are clear. First, I need to incorporate a robotic liquid handling machine, such as an OpenTrons pipetting robot or a solenoid valve manifold, that can mix media components into testable recipes. Second, I need to engineer away the manual transfer of specimen plates between workstations - this could be accomplished by culturing the cells on the microscope itself instead of an incubator, since the microscope can be outfitted with a climate-controlled chamber and already has a robotic arm. Third, I need to refactor the system to make it as cheap and easy to deploy as possible - this way, additional systems can be built beyond the prototype. The system does not need to be perfect in order to be useful - I can simultaneously use the system to formulate media as I build it.

It might be possible to fund this project as a for-profit venture, but doing so would entail either patenting the key technologies or keeping the details secret, and these requirements would impede other researchers from adopting the method. I believe that new technology only matters inasmuch as people adopt it; my foremost desire is accelerating biological research, and I will be better able to spread this approach inasmuch as it is funded in a non-profit capacity. As such, I am soliciting donations to support this project.

In the long run, I think this project helps address a fundamental problem with the pace of biological research. Biology today is a mashup of fast, data-rich omics research and slow, data-poor wet-lab research. Even a lab fully equipped with robotics is limited by scientists needing to formulate and evaluate individual hypotheses before and after each incremental experiment. But, to the extent we can phrase wet-lab biology questions as functions to be maximized instead of hypotheses to be falsified, we can let machines do much more of the work. I envision a future where biologists render research questions in a way that machines can answer unassisted, freeing us to spend more time on the bigger picture. A single biologist could supervise several research automatons, running 24/7, performing myriad experiments in parallel, and iterating towards the biologist’s goals. This is how we accelerate biological research. If we want more headway on biomedical problems such as improving human longevity, designing gene therapies, or growing human organs, we need to change how the scientific discovery process works, and I see no better place to start than changing how we formulate cell culture media.