AI tools reshape protein design
- MIT said on April 17 that OpenProtein.AI is expanding no-code protein-engineering software, giving biotech and academic labs access to design models. - New 2026 studies showed the pace shift: one Rice-led method made 10 million protein-activity data points in three days for training. - The field is moving from trial-and-error toward prediction, with wet-lab checks still deciding what works. (nature.com)
Proteins are the tiny machines inside cells, built from amino-acid chains that fold into precise shapes. Artificial intelligence is now being used to guess which chains will fold and work before scientists build them in the lab. (nature.com) That matters because the number of possible protein sequences is too large to test one by one. A 50-amino-acid protein has about 1.13 × 10^65 possible combinations, according to Rice University researchers. (phys.org) Older protein engineering relied heavily on trial and error: make a change, test it, then repeat. The 2025 Nature Reviews Bioengineering survey said AI tools now generate, evaluate, and optimize proteins much faster across the design cycle. (nature.com) One recent sign of that shift came from MIT’s April 17 profile of OpenProtein.AI, founded by Tristan Bepler and former MIT professor Tim Lu. The company said its platform gives scientists no-code access to models for protein design, structure prediction, function prediction, and model training. (news.mit.edu) Another came from Dyno Therapeutics, which announced Dyno Psi-Phi at NVIDIA GTC on March 18. Dyno said the suite combines generative protein-design models with filters trained on experimental results so users can screen candidates for wet-lab success. (dynotx.com) Academic labs are also pushing the workflow forward by generating better training data. A Rice University team, with collaborators from Johns Hopkins and Microsoft, reported in April that its Sequence Display method can produce more than 10 million protein-activity data points in a single experiment and build useful models in three days. (phys.org) Other groups are teaching models to handle interactions, not just single proteins. A National University of Singapore team reported on April 20 that its paired protein language model trained on more than 3 million protein pairs and improved interaction-prediction accuracy by up to about 17% on benchmarks. (phys.org) Researchers are also moving from prediction into design. A Nature Biotechnology paper published April 15 described artificial protein switches built by combining reporter proteins with machine-learning-designed receptor domains. (nature.com) A Nature Computational Science paper published April 15 said its DualGPT-AB system generated antibody candidates with multiple desired properties at once. The authors reported that 8 of 100 randomly selected candidates showed strong HER2 binding, and lab tests found designs with stronger tumor-cell killing than Herceptin. (nature.com) The promise is speed, but the limit is still biology. Dyno says its filters are calibrated with experimental data, and the broader literature still treats synthesis and lab validation as the step that decides whether an AI-designed protein is useful. (dynotx.com) (nature.com) The same tools are also raising security questions as they spread. A Frontiers review published March 31 said generative protein design could aid therapeutics and pandemic preparedness, while also lowering barriers to creating harmful proteins that evade sequence-based screening. (frontiersin.org) So the story in 2026 is not one lab or one product. It is a broader shift toward software that proposes proteins first and experiments that verify them second. (news.mit.edu) (nature.com)
