AI for science

AI for science in this catalog mostly means models trained on biological or chemical data rather than on general web text. Evo 2, from Arc Institute working with NVIDIA, is a DNA language model trained on genomes from all domains of life. ESM3 is a generative protein model that designs proteins from partial prompts of sequence, structure, and function. Boltz-2 jointly models biomolecular complex structures and binding affinities, and Chai-1 predicts the structures of proteins, small molecules, DNA, RNA, and other molecules together. ether0, from FutureHouse, is a 24B model fine-tuned from Mistral-Small-24B-Instruct-2501 that reasons in English and outputs molecules as SMILES.135678

The publishers range from nonprofits to companies. Arc Institute describes itself as an independent nonprofit research organization headquartered in Palo Alto, and FutureHouse as a 501(c)(3) AI-for-science research lab in San Francisco, while the Profluent-E1 license is offered by Profluent Bio Inc. FutureHouse's GitHub organization also hosts PaperQA2, an Apache-2.0 package for retrieval-augmented generation over scientific literature. By default it uses OpenAI models and embeddings, and any LiteLLM-compatible model can be configured instead, so the tool's own license does not cover the model service it calls.491110

Terms differ from release to release. Boltz says all of its code and weights are under the MIT license for academic and commercial use. Chai-1's code and weights are under Apache 2.0. The ESM3 open small model card points to an MIT license, and the Evo 2 checkpoints are tagged Apache 2.0, with Arc saying it shared the training data, training and inference code, and weights. Profluent-E1 uses a custom clickthrough license. Its attribution guidelines require commercial entities to credit the model, and require any drug, target, hit, or lead developed with it to be identified as Built with Profluent-E1, including in public disclosures to the U.S. Food and Drug Administration.675231112

Documented restrictions and safeguards take different forms. The MedGemma 27B model on Hugging Face can be downloaded only after signing in and agreeing to the Health AI Developer Foundations terms of use, and the MedGemma model card says its outputs are not intended to directly inform clinical diagnosis, patient management, or treatment. Other releases document measures in training data or model behavior instead. Arc says pathogens that infect humans and other complex organisms were excluded from Evo 2's base data set. The ESM3 open model card says virus-related data and some proteins from organisms on the USDA Select Agents and Toxins list were removed from training. ether0 was post-trained to refuse questions about compounds on OPCW Schedules 1 and 2.1314358