AI enables conditional protein design
- Nature Communications published ORI, a Tencent AI for Life Sciences framework that generates proteins from function prompts and updates designs with lab feedback. - In experiments, ORI produced a lysozyme with 100-fold higher activity, a chitinase stable at 85 degrees Celsius, and dual-function enzymes. - The work pushes protein design from trial-and-error toward controllable, multi-objective generation with experimental loops. (nature.com)
Proteins are chains of amino acids that fold into shapes, and those shapes decide what the molecules do inside cells. ORI is a new system that tries to design those chains from a requested function instead of by random mutation alone. (nature.com) (pmc.ncbi.nlm.nih.gov) Nature Communications published the ORI study on March 19, 2026, with authors led by Bing He, Chenchen Qin and Jianhua Yao from Tencent AI for Life Sciences. The paper describes “Ontology Reinforcement Iteration,” or ORI, as a framework that uses structured biology labels as prompts for protein generation. (nature.com) Those labels come from ontologies, which are organized vocabularies biologists use to describe what proteins do. ORI uses those labels to steer generation toward targets such as higher enzyme activity, better heat tolerance, or more than one function at once. (nature.com) The system does not stop at one round of computer output. The authors say ORI runs a closed loop of generation, experimental measurement and model updating, so lab results feed back into the next batch of designs. (nature.com) In the paper’s wet-lab tests, ORI produced a lysozyme variant with 100-fold higher activity than a natural baseline. It also generated a chitinase that remained stable at 85 degrees Celsius and enzymes that combined lysozyme and chitinase activity in one design. (nature.com) That is the practical problem protein engineers have been trying to solve for years. Traditional directed evolution usually starts from an existing protein, adds a few mutations, tests them, and repeats, which is slow and explores only a tiny slice of possible sequences. (pmc.ncbi.nlm.nih.gov) Conditional design means asking a model for a protein with a specified trait, the way an image model takes a prompt. Earlier work such as ProteoGAN in 2022 and ProCALM in 2024 showed that models could generate sequences conditioned on hierarchical functions, enzyme families or natural-language descriptions. (pmc.ncbi.nlm.nih.gov) (arxiv.org) ORI adds a stronger experimental loop to that idea. The paper says reinforcement learning from experimental feedback helps the model move from proteins that look plausible on a screen to proteins that perform better in measured assays. (nature.com) Tencent AI for Life Sciences has also released ORI code in a public GitHub repository. That makes the March 2026 paper part of a broader shift in protein design toward controllable generation systems that can be tuned for several properties at once. (github.com) (nature.com) The immediate result is not a finished drug or industrial enzyme. It is a design workflow that starts with a requested function, tests the output in the lab, and comes back with a better sequence on the next round. (nature.com)
