I Wanted AI to Predict Cell Changes—Then the Workflow Got in the Way
Last Wednesday night, I played a “virtual cell” interview for a biotech friend. We were at a café near Tsinghua, and she laughed and interrupted me: “No matter how powerful the model is, someone still has to put the cells in a dish, photograph them and check the results.” That was when I realized AI biology is farther from everyday workflows than I imagined—and even more dependent on real experimental data.
Fabian Theis is a researcher at Helmholtz Munich in Germany and a key advocate of the “Virtual Cell” concept. His work goes beyond designing a single protein; he wants to understand how cells change over time and across layers: how genes are expressed, how cells transform and how those changes influence one another. AI x Bio attention used to center on molecular tools such as AlphaFold. Now Biohub, NIH, DOE, Google and Meta, among others, are committing roughly $1.8B in total to “virtual cells.”
Here is a concrete snapshot: in March 2025, Biohub introduced its joint funding program in San Francisco with $1.8B on the table. Theis’s team had already helped develop laser phase-contrast microscopy, light-sheet microscopy and post-processing tools for conventional microscopes. AI is not conjuring answers from thin air here. It first finds patterns in increasingly large sets of images and measurements, then automated labs test those patterns in the physical world.
That also explains why funding is arriving now. New imaging equipment, automated lab facilities and data-sharing mechanisms are maturing. For the first time, the field is not merely piling up models; it is filling the gap that lets models keep receiving real-world feedback. Still, the question of whether simulating a cell is “good enough” has not gone away. A larger scale does not automatically make the conclusions more reliable.
What “Virtual Cell” Actually Is
I’d think of it as a continuously updated cell-simulation system. Feed it cell images, molecular measurements and experimental results, and it tries to predict how a cell will respond to an intervention. Robots in the lab then run the experiment, and fresh data flows back into the system. The core goal for Theis and Biohub is to turn a loop once completed by hand into a scalable closed loop.
Who benefits? For now, mainly research institutions, pharmaceutical companies and teams with automated lab access. Theis has also worked with multiple startups and on Biohub-related efforts, so he is not merely offering commentary from the sidelines; he is helping push this infrastructure into practice. In September 2024, Biohub announced a partnership with Boston-based Schmidt Futures to build a “Virtual Cell” lab, bringing AI prediction, automated experimentation and biological data under one direction.
What It Would Take Us to Replicate It
Replicating the full “virtual cell” stack is not as abstract as saying “an $1.8B budget.” According to public industry estimates, equipping and retrofitting a medium-sized automated cell lab typically costs several million dollars. Running it for a year, including staff, consumables and data maintenance, can add hundreds of thousands of dollars or more. On the timeline, procuring equipment, building workflows and generating stable training data takes at least a year. The technical barrier is straightforward too: the team needs expertise in cell experiments, image analysis, machine learning and automation hardware. Prompting alone cannot bridge that gap.
If I only wanted a low-barrier way to explore the field, my first step would not be buying equipment. I would open Biohub’s Virtual Cell Challenge page and browse its public problems and datasets. That can cost $0 and take one to three hours, though it assumes I can read papers and organize experimental data. The first click would be “Participate” or “Datasets,” followed by completing one public baseline task. I might also watch Theis’s full interview first, but I would not rush to order lab equipment.
This tool is not for everyone, nor is it a practical productivity tool for building a personal brand or working as a freelancer. There is nothing wrong with waiting. Without lab access especially, forcing it would only add unnecessary cost.
How I’d Approach It at Different Stages
If I’m just starting: I would read public cases, learn basic biology and AI evaluation, and avoid hardware purchases. I would run one small experiment with $0 and a week of time to see whether I genuinely enjoy this long-cycle work.
If I already have one or two client partners: I would package the capability as an “image analysis or experimental-data organization service” and look for a small paid pilot with a biology research team. I would not promise cell prediction on day one; I would start with verifiable data cleaning and reporting.
If I’m scaling up: I would seriously evaluate lab and data partners, model equipment utilization, repeated experiment costs and compliance requirements, then decide whether to form a dedicated team. “Virtual cell” is hot right now, but reliable results—and revenue—will take longer.