Biohub Pushes Protein AI Onto an Open Stage, Moving From Structure Prediction to Molecular Design Validation

When biomedical AI talks about a "world model," it can easily slide into abstract slogans. What makes Biohub's newly released ESM system notable is that it brings the question back to proteins, a level that can be tested experimentally: the model does not only predict sequences or structures, but also attempts to help design protein interfaces that bind to specific targets, and sends some results into the laboratory for testing.

According to Axios reporting and Biohub's public materials, the release includes three main components: the protein language model ESMC, the structure prediction and design model ESMFold2, and the expanded ESM Atlas. On its GitHub page, Biohub calls it a "world model for protein biology" and says ESM Atlas covers 6.8 billion protein sequences, with more than 1 billion structures predicted and generated by ESMFold2.

ESMC plays the role of an underlying model for reading protein sequences. In its Hugging Face model card, Biohub describes ESMC as being trained on billions of protein sequences and usable for protein representation and therapeutic protein engineering. The ESMC-6B version has 6 billion parameters and 80 layers; its training data sources include UniRef, MGnify, and sequence data from the U.S. Department of Energy Joint Genome Institute, clustered at 70% sequence similarity. The public page also lists model families including ESMC-300M, 600M, and 6B, meaning researchers can choose different scales according to their computing resources.

ESMFold2 pushes sequence information toward three-dimensional structures and hypotheses about molecular interactions. Biohub's model card says ESMFold2 can predict full-atom three-dimensional structures from amino acid sequences, and can optionally incorporate multiple sequence alignment data. It is also described as being able to handle biomolecular contexts beyond proteins, including small molecules, DNA, RNA, and modified amino acids. However, these capabilities remain model outputs, and Biohub also explicitly notes that predictions are machine-generated hypotheses and cannot replace experimental structure determination.

One aspect of this release with greater biomedical significance is Biohub's claim that some protein interfaces designed by the model have passed experimental testing. Its GitHub materials state that ESMFold2 experimentally validated de novo designed minibinders and antibody-derived scFvs against five therapeutically relevant targets, and reported a high hit rate, nanomolar affinity, target specificity, and functional activity. That distinguishes it from AI tools that simply display attractive structural images, but the public summary is still not enough to judge how far these molecules remain from drug candidates, animal studies, or human trials.

The expanded ESM Atlas may also change how researchers enter unknown protein space. In the past, many proteins from environmental microbes or uncultured organisms had only sequences and lacked structural clues. If billion-scale predicted structures can be queried quickly, researchers may be able to propose hypotheses about enzyme function, viral protein interactions, or new binding molecules more efficiently. The value of this kind of atlas is not that every prediction is correct, but that it transforms a vast sequence black box into candidate questions that can be prioritized and revalidated.

But from research tool to medical product, there is still a long chain of evidence. If protein design is to enter therapeutic use, it must confront immunogenicity, stability, manufacturing, in vivo distribution, safety, and regulatory review. Model performance in benchmarks or in vitro experiments also does not mean it can predict complex responses inside the human body. Biohub's choice to open its code, model weights, datasets, and design system helps more laboratories reproduce, challenge, and expand these results. The real test will be whether the open community can turn this toolchain into reliable and reproducible biological discoveries.