Rice University researchers have created CLASSIC, a technique that lets artificial intelligence help design complex genetic circuits at unprecedented scale. The work could speed up the development of smarter cell-based therapies for cancer and other diseases.
Designing DNA instructions that tell cells what to do has long been one of synthetic biology’s biggest bottlenecks. A team at Rice University now says artificial intelligence can finally help crack that problem.
In a study published in the journal Nature, the researchers report the first successful use of AI to design complex genetic circuits — DNA programs that control how cells behave. Their approach, called CLASSIC, lets researchers build and test enormous numbers of DNA designs at once and then feed that data into machine-learning models.
For a field that has often advanced one painstaking circuit at a time, the shift to high-throughput, AI-guided design marks a major turning point.
The challenge has always been that the number of possible DNA designs explodes as circuits get more complex. There are countless ways to wire together genes, switches and control elements to make a cell glow, sense a tumor or release a drug at the right moment.
“There are many possible designs for any given function, and finding the right one can be like looking for a needle in a haystack,” senior author Caleb Bashor, an assistant professor of bioengineering and biosciences at Rice, deputy director for the Rice Synthetic Biology Institute and member of the Ken Kennedy Institute, said in a news release.
CLASSIC, which stands for combining long- and short-range sequencing to investigate genetic complexity, is the team’s answer to that search problem.
“We created a new technique that makes hundreds of thousands to millions of DNA designs all at once — more than ever before,” Bashor added.
Instead of building genetic circuits one by one, the Rice group learned how to assemble them in huge batches. Co-first authors Kshitij Rai and Ronan O’Connell, who worked on the project when they were doctoral students in the Bashor lab, described the lab work behind that leap as a lot of molecular cloning — cutting DNA into pieces and pasting it together in new ways.
“We invented a way to do this in large batches, which allowed us to make really large sets — known as ‘libraries’ — of circuits,” Rai, who will be pursuing a postdoctoral fellowship at the University of Washington, said in the news release.
Those libraries are the raw material that makes AI useful. Machine-learning models need massive, well-labeled datasets: for each DNA design, they must know exactly what sequence was built and how it performed inside a cell. CLASSIC was engineered to generate that information at a scale synthetic biology has not seen before.
To do that, the team combined two flavors of next-generation sequencing, or NGS. One, known as long-read sequencing, can scan thousands of DNA bases in a single pass, capturing an entire genetic circuit in one go. The other, called short-read sequencing, reads much smaller stretches of DNA but with high accuracy and throughput.
“Most people do one or the other, but we found using both together unlocked our ability to build and test the libraries,” added O’Connell, who is now a postdoctoral researcher at Baylor College of Medicine.
The researchers first built a proof-of-concept library of genetic circuits that controlled reporter genes designed to make cells glow. Using long-read sequencing, they recorded the full DNA sequence of each circuit and tagged it with a short, unique barcode.
They then pooled the circuits and inserted them into human embryonic kidney cells. Inside the cells, each circuit drove a different level of glow — some bright, some dim. By sorting the cells into groups based on brightness and using short-read sequencing to count the barcodes in each group, the team could link every circuit’s full DNA blueprint to its performance.
That map of genotype (DNA sequence) to phenotype (cell behavior) is what powers the AI step.
“We end up with measurements for a lot of the possible designs but not all of them, and that is where building the ML model comes in,” O’Connell added. “We use the data to train a model that can understand this landscape and predict things we were not able to generate data on, and then we kind of go back to the start: We have all of these predictions — let’s see if they’re correct.”
To check whether CLASSIC’s measurements and the model’s predictions were trustworthy, Rai and O’Connell manually tested a smaller, random set of circuit variants. They compared those results to what CLASSIC and the AI had reported.
“We started lining them up, and first one worked, then another … and then they just started hitting,” added Rai. “All 40 of them matched perfectly. That’s when we knew we had something.”
According to the team, this was the first time AI/ML could be used to analyze circuits and make accurate predictions for untested ones, because up to this point, nobody could build libraries as large as ours.
With that validation in hand, the researchers could begin to extract deeper insights. By testing vast numbers of complete circuits at once, CLASSIC reveals the underlying rules that govern how genetic parts behave when combined. Those patterns, in turn, help machine-learning models design new circuits that are more likely to work as intended.
The study also suggests that there is rarely a single perfect solution to a design problem. Many different DNA circuits can achieve the same functional outcome, much like different driving routes can get you to the same destination.
“This is akin to navigation apps: There are multiple routes to reach your destination, some highways, some backroads, but all get you to your destination,” O’Connell added.
Another finding was that medium-strength genetic components often performed better than very strong or very weak ones. Rai likened this to a biological sweet spot.
“Call it biology’s version of ‘Goldilocks zones’ where everything is just right,” Rai said.
The work has immediate implications for cell-based therapies, which use engineered cells as living drugs to fight cancer, repair tissues or modulate the immune system. Because the Rice team demonstrated CLASSIC in a human cell line, the platform is already working in a context similar to many therapeutic applications.
In the longer term, the researchers see their approach as a foundation for more predictable, programmable biology. As they and others generate larger datasets, AI models could learn to design circuits that, for example, sense multiple disease signals at once, respond only under safe conditions or coordinate the behavior of many cells working together.
“We think AI/ML-driven design is the future of synthetic biology,” added Bashor. “As we collect more data using CLASSIC, we can train more complex models to make predictions for how to design even more sophisticated and useful cellular biotechnology.”
The project itself was a case study in combining different strengths. Rai said the science mirrored the circuits they were building.
“I think a big part of what we did in this project is to show that while the same individual parts might not have any spectacular function by themselves, if you put the right combination of things together, it gives you these dramatically better genetic circuits,” Rai added. “That holds true for the science as well. And that was the best part of this project behind the scenes — all of us bringing our different skill sets together.”
Leaders in the field see the Rice team’s work as a milestone. One early pioneer, James Collins, a biomedical engineer at MIT, noted that two classic synthetic circuits — the genetic toggle switch and the repressilator — proved decades ago that cells could be programmed, but each required months of trial-and-error tuning and was built essentially by hand.
“Twenty-five years ago, those early circuits showed that we could program living cells, but they were built one at a time, each requiring months of tuning,” said Collins, who was one of the inventors of the toggle switch. “Bashor and colleagues have now delivered a transformative leap: CLASSIC brings high-throughput engineering to gene circuit design, allowing exploration of combinatorial spaces that were previously out of reach. Their platform doesn’t just accelerate the design-build-test-learn cycle; it redefines its scale, marking a new era of data-driven synthetic biology.”
Another leading synthetic biologist, Michael Elowitz, emphasized that for years, synthetic biologists have dreamed of programming cells by snapping together biological circuits from interacting genes and proteins. Approaches like CLASSIC, which systematically explore the vast biological design space and feed it into AI, move that dream closer to reality.
If that vision pans out, future bioengineers may spend less time hunting for a single working DNA design and more time choosing among many viable options — letting AI help chart the best route through biology’s maze of possibilities.
Source: Rice University