Atlas Data Consortium launches with five founding members, each designing a million protein-interaction affinity measurements per quarter to share data rather than models

The Atlas Data Consortium shifts the shared object from a trained model to affinity measurements newly generated under uniform experimental conditions, so members receive raw measurements rather than a consensus model. Earlier sharing attempts centered on federated learning, such as the MELLODDY project, which trained shared models across 10 major pharma companies without moving data, followed by efforts such as TuneLab and the FAITE consortium. The consortium publicly launched in July with five founding members including GSK, Boltz, Cradle and Dyno Therapeutics; A-Alpha says all data are generated on the AlphaSeq platform under standard conditions and then pooled and released quarterly.

Each member designs one Data Block per quarter, a unit of one million protein-protein affinity measurements built around a machine-learning objective, and A-Alpha designs additional blocks alongside them, scaling as the consortium grows. Publicly available data were generated to advance a specific program, answer a specific question or support a specific filing, not to train a generalizable protein design model. The first consortium focuses on VHHs, the variable domains of heavy-chain-only antibodies; members share a scaffold, so each member's Data Block is highly relevant to every other member's models, and consortia for other scaffolds such as scFv/IgG are expected from 2027.

The demonstrated scope is the consortium's structure, Data Block specification and uniform measurement pipeline on AlphaSeq, with the first consortium limited to the VHH scaffold and aimed at model developers and pharma R&D teams working on antibodies and protein design. The material states that consortia focused on other scaffolds such as scFv/IgG are expected to launch beginning in 2027, and that the team is preparing a companion analysis of affinity landscapes for publication later this year.

The material says its extrapolations of the required data volume may be wrong in detail but that it is confident about the order of magnitude, which it says is far beyond what most of the field is planning for; the authors also note that harder epitopes and objectives such as pH sensitivity or dual specificity will require more data still. The material further notes that affinity is context dependent, so data aggregated across laboratories can let a model learn to recognize the laboratory instead of the biology.