Design-build-test-learn (DBTL) cycles are a fundamental part of protein or organism engineering in synthetic biology. In this article, we will focus on protein engineering, as the advances in machine learning have made it possible to speed up this process significantly. As we will discuss, cell-free approaches for DNA synthesis can also accelerate the build and test phases of DBTL cycles for protein engineering down to true one-shot workflows.1 This means that proteins with improved target properties can be designed in one DBTL cycle.
The DBTL cycle explained
Design: During the design phase, computational models generate hypotheses about which sequence changes may improve the function of a target protein. Inputs to the design phase include existing structural models, mutational data, or computational predictions. The result is a set of protein variants.
Build: The amino acid sequence of the protein variants is encoded by DNA molecules and then synthesised, whereby speed, accuracy, and representation of all protein variants is important for downstream data quality. Protein variants are expressed in cell-based or cell-free systems.
Test: In this phase, protein variants are experimentally characterized using protein type specific assays. Those assays can include measurement of the catalytic rate, binding constants, expression levels, or thermostability. The acquired data is quantitative and allows for comparison across protein variants.
Learn: Data is analyzed to identify which protein variant-function relationships drive performance. These insights are utilized for the next design round to iteratively improve each cycle.
Bottlenecks in the DBTL process
Building proteins and rapidly testing variants are the major bottlenecks today in protein engineering. Based on the complexity of the protein, having an accurate DNA molecule is a pre-requisite, and therefore being able to quickly build and iterate protein variants presents several challenges, as described below.
Time-consuming steps in a cell-dependent process
When using a living host system (e.g., propagating recombinant DNA in plasmids within bacterial cells), the DNA must undergo multiple processing steps. These typically include cloning into a vector, transformation into host cells, selection and colony screening, plasmid amplification through cell growth, and plasmid purification. Each one of these steps introduces room for error while delaying the transition from design to testing during the build phase.
Endless DBTL Cycles in Cell-Based Expression
Another constraint arises in the testing phase. When using a living host, protein expression is a laborious process where only a few protein variants can be tested per DBTL cycle. In cell-free protein engineering, protein expression is a biochemical reaction, allowing for multiple protein variants and expression conditions to be tested at once. In other words, protein variants move from digital, de novo, and rational design to functional testing without biological filtering, preserving the integrity of the original pool of variants. Moreover, cell-free expression systems accelerate DBTL cycles.
Watch our webinar to learn how we enable one-shot DBTL cycles.
Source: LinkedIn – Ribbon Bio
Cytotoxicity
Protein variants that affect cell viability or burden cellular metabolism get counter-selected, meaning that the molecules that are the hardest for the cells to express are least likely to get evaluated. As a result, experimental datasets are biased, reflecting what survives the expression system rather than what performs best biochemically. Cell-free approaches allow for the testing of cytotoxic protein variants.2
Lack of control over cloning and propagation
The dependency on cell biology can further complicate the building of protein variants. Clones that grow faster or pick up recombination events skew representation long before protein expression begins. By the time proteins enter an assay, the diversity of variants generated computationally will have narrowed.
The role of automation and machine learning
Automation links design, build, test, and learn steps together within biofoundries. Robotic liquid handling and standardized DNA assembly assisted by machine learning, allow for much higher throughput.
A successful DBTL workflow should involve a closed loop, allowing artificial intelligence to continuously propose, execute, and analyze DBTL cycles with minimal human intervention.
Conclusion
Cell-free protein engineering moves prediction closer to functional protein, by allowing for multiple variants of proteins to be tested within the same DBTL cycle. By absolving the host of its duties, variations in yield or activity can be attributed to sequence. This increases data quality for the learning phase, guiding following designs with cleaner signals. This effect compounds over multiple rounds and enables you to arrive at the optimal protein faster.
- Clark-ElSayed, A., Harrison, I.M., Olsen, M.L. et al. LDBT instead of DBTL: combining machine learning and rapid cell-free testing. Nat Commun 16, 9782 (2025). https://doi.org/10.1038/s41467-025-65281-2
- Pandi, A., Adam, D., Zare, A. et al. Cell-free biosynthesis combined with deep learning accelerates de novo-development of antimicrobial peptides. Nat Commun 14, 7197 (2023). https://doi.org/10.1038/s41467-023-42434-9
