AI rushes into biology
- New AI models are being used to accelerate biological research and drug-discovery workflows. (x.com) - A protein-engineering effort generated about 10 million data points in three days to train those models. (x.com) - The coverage says major labs are pivoting toward biology and reports also noted surprising bacterial 'rewiring' of DNA machinery. ( )
Biology is turning into an AI data problem, and labs are racing to generate the measurements those models need to design drugs and proteins faster. (nature.com) (phys.org) Proteins are the cell’s working parts, built from strings of amino acids, and protein engineering means swapping those parts to change what a protein does. The hard part is that even a small protein has an enormous number of possible variants, so researchers need huge datasets linking sequence to function. (nature.com) (science.org) A Rice University team said on April 19 that its platform measured the activity of more than 10 million protein variants in a single experiment completed in three days. The group said those measurements can be used to train AI models to predict which mutations improve protein function. (phys.org) That push fits a broader shift from hand-tuned protein design toward model-guided cycles that mix machine learning, automation, and repeated lab testing. A 2025 Nature Communications paper described an autonomous enzyme-engineering platform that combined machine learning, large language models, and biofoundry automation to run protein optimization with minimal human intervention. (nature.com) Drugmakers are also feeding private data into shared AI systems. AbbVie and Johnson & Johnson said in March 2025 that they would contribute proprietary structural data to OpenFold3 through Apheris federation tools to improve models for small molecule-protein and antibody-antigen interactions. (statnews.com) The field is also moving beyond predicting static protein shapes toward models that try to capture how molecules interact inside cells. A Science review in 2025 described that goal as building a molecular “microscope” for the interactome, the network of contacts among proteins and other biomolecules. (science.org) At the same time, biologists are still finding basic mechanisms that no model could have assumed in advance. A Science paper published in April 2026 reported that multicellular cyanobacteria repurposed a DNA-segregation system, which normally helps split genetic material during cell division, into a filament system that helps control cell shape. (science.org) That result matters for AI-driven biology because the models are only as good as the biology they are trained on. Reviews in Nature Biotechnology and Nature Methods this year both argued that “generalist” biological AI and agent-like lab systems will depend on better experimental data and human oversight, not just larger models. (nature.com 1) (nature.com 2) The near-term pattern is clear: more robot labs, more proprietary datasets, and more claims that AI can shorten the path from sequence to candidate drug. Even companies and researchers pushing the technology say the wet lab still decides which predictions survive contact with real biology. (nature.com) (statnews.com)
