Researchers at Stanford University have developed a new protein engineering method called MIDAS, short for Microbe-Independent Deep Assembly and Screening, that dramatically accelerates the process of designing, building, and testing protein variants. The team says the approach can reduce a workflow that traditionally takes days or weeks into a process completed within 24 hours.
Proteins are essential biological molecules with applications spanning medicine, biotechnology, and industrial manufacturing. Scientists routinely engineer proteins to improve drug therapies, treat disease, develop biosensors, and create industrial enzymes used in products such as detergents and food ingredients. However, conventional protein engineering remains highly labor-intensive because every new protein variant must be physically constructed and tested in living systems.
Although AI systems can recommend potential protein improvements, researchers still need to experimentally validate each candidate in the laboratory. Traditional workflows require assembling DNA instructions for proteins inside microbes such as bacteria or yeast, growing those microbial clones, extracting DNA, and then transferring the genetic material into mammalian cells for testing. The process can take many days for a single protein and even longer for proteins requiring mammalian-cell validation.
In a study published in Molecular Systems Biology, Michael Z. Lin, a professor of neurobiology and bioengineering at Stanford, along with graduate students Yan Wu and Pengli Wang, described how MIDAS bypasses those traditional microbial cloning steps entirely.
Instead of relying on circular DNA structures known as plasmids, the researchers used polymerase chain reaction (PCR) technology to rapidly amplify linear DNA fragments directly. This enables scientists to create entire genes for mammalian cell expression without microbial cloning and then transfer the resulting gene variants directly into mammalian cells for testing.
The only essential inputs for the system are short DNA fragments known as primers, which can be ordered for next-day delivery. According to the researchers, this allows a highly compressed workflow where scientists can receive primers in the morning, assemble genes by midday, and begin testing protein function in mammalian cells later the same day.
The researchers said MIDAS can also scale efficiently. Hundreds or even thousands of protein variants can be assembled and tested in parallel, significantly expanding the number of proteins researchers can evaluate during a single experiment.
Traditional protein engineering workflows become particularly slow after researchers identify promising variants because each gene must then be cloned into plasmids, transferred into bacteria or yeast for amplification, purified, and subsequently moved into mammalian cells for functional validation. MIDAS eliminates those bottlenecks by treating DNA as linear information compatible with PCR amplification.
The Stanford team said the insight came from recognizing that the circular structure of plasmids is unnecessary for PCR-based workflows. By removing plasmids from the process, researchers were able to dramatically simplify and accelerate protein engineering.
In one practical demonstration, the researchers used MIDAS to test 384 protein variants with approximately four hours of hands-on lab work and roughly $2,000 in reagent costs. By comparison, existing cloning-based methods would require an estimated 192 hours of work and around $20,000 in reagents to evaluate just 24 variants.
Based on those results, the researchers estimate that MIDAS is nearly 50-times faster and roughly one-tenth the cost of traditional cloning-based approaches.
The team believes MIDAS could have significant implications across biological research. Potential applications include accelerating enzyme engineering, improving biosensor development, enhancing robotic laboratory automation workflows, and generating much larger datasets for AI-driven molecular biology systems.
The researchers also said MIDAS could improve machine-learning models by rapidly generating extensive sequence-fitness datasets that show how protein variants perform under experimental conditions. Those datasets could help AI systems better predict which molecular designs are most likely to succeed.
Looking ahead, the researchers envision tighter integration between MIDAS, robotics, and AI-driven protein design systems. They believe the technology could accelerate the design-build-test cycle for protein engineering and enable faster advances in computational biology and therapeutic development.
The study included contributors from Fudan University and Promega Corporation. Funding support came from the NIH, Stanford Bio-X, and the Stanford Wu Tsai Neurosciences Institute.
The paper also noted that Pengli Wang passed away in May 2026 after completion of the research while serving as a fourth-year PhD student in chemical engineering.
KEY QUOTES:
“The fundamental questions of molecular biology remain: how do we make better proteins and how do we understand what makes a protein work? Doing that work takes valuable time and resources, but we’ve found a way to dramatically reduce those demands.”
“We decided there’s nothing magical about the circular structure of plasmids. For PCR, you just need the genetic data. That was the moment of inspiration.”
“MIDAS is at least an order-of-magnitude faster at real-world validation. It compresses the engineering design-build-test cycle for proteins to just a couple of days, and we think it could drive rapid advances in AI-inspired molecular biology.”
Michael Z. Lin, Professor Of Neurobiology And Bioengineering, Stanford University
“With MIDAS, we can receive PCR primers in the morning, assemble the necessary genes by mid-day, and by late afternoon transfer the genes into cells to observe how the proteins function. And we can do this all for hundreds or thousands of protein variants in parallel at a time.”
Yan Wu, Graduate Student, Stanford University
“We used MIDAS not only to find the best-performing version of a protein but also to understand how well closely related variants work, which is information we can use to train AI models. MIDAS is so easy that we can use it to create large data sets very quickly.”
Pengli Wang, Graduate Student, Stanford University
