boltz-2
Datasets
All datasets matching “boltz-2”mgtbind-boltz2
MGTBind Boltz-2 Structure Predictions
MGTBind
Bindbridge
Tenstorrent
Boltz-2 ternary complex structure predictions for all 3,607 molecular glue complexes in the MGTBind database. Predicted in partnership with Bindbridge and computed on Tenstorrent AI accelerators using TT-Boltz.
Results
Comparison against the existing AlphaFold 3 predictions in MGTBind (all 3,607 complexes):
Metric
AF3
Boltz-2
Avg MG Chain ipTM
0.7443
0.7896
Delta MG Chain ipTM… See the full description on the dataset page: https://huggingface.co/datasets/Tenstorrent/mgtbind-boltz2.boltz2_accuracy_architecture_20260919_032813
Boltz2 accuracy and architecture study
Correction: The original Cα scoring selected the wrong atom. Use this revision’s corrected measurements and see the correction notice before using any prediction archive. Raw archived ca arrays are superseded; recover Cα from full coordinates and token_to_center_atom.
Measured small-protein ablations, an evaluation RNG/precision repair for the local Boltz quotient adapter, and a cyclic Fourier triangle-contraction prototype. No new weights… See the full description on the dataset page: https://huggingface.co/datasets/raftbioworks/boltz2_accuracy_architecture_20260919_032813.boltz2_accuracy_recovery_20260919_064306
Boltz2 accuracy recovery — ongoing research campaign
Work in progress. This snapshot does not establish accurate 100k-residue folding.
The campaign investigates numerical fidelity, symmetry compression, native-supervised adaptation and independent AF2 controls. It uses public small PDB assemblies and the unchanged Boltz2 checkpoint, plus explicitly identified learned adapters. Native coordinates are training labels and evaluation references; they are not supplied as inference… See the full description on the dataset page: https://huggingface.co/datasets/raftbioworks/boltz2_accuracy_recovery_20260919_064306.boltz2_quotient_validation_20260918_060803
Boltz 2 quotient-validation study
Inference-only experiments with group-indexed pair features and tiled attention, compared with the supplied Anthropic optimization kit's actual big mode on H100 GPUs.
Memory savings were measured; improved or preserved folding accuracy was not established. This is not a newly trained model. The checkpoint is the unchanged upstream Boltz 2 structure checkpoint, SHA-256 090e82ac8c92f5e943fa1b39e7410a44027bea7243c0bbb3caa67a77fc1428e1 (2,286,561… See the full description on the dataset page: https://huggingface.co/datasets/raftbioworks/boltz2_quotient_validation_20260918_060803.