Structure Prediction

for antibody-antigen complex prediction (still a hard task).

predict the bound complex from sequence + ligand SMILES. Headline improvement: no need for a crystal structure of the apo target.

for antibody-antigen complex prediction (still a hard task).

weights publicly released.

predict the bound complex from sequence + ligand SMILES. Headline improvement: no need for a crystal structure of the apo target.

predict the bound complex from sequence + ligand SMILES. Headline improvement: no need for a crystal structure of the apo target.

protein-ligand cofolding.

retraining structure predictors as flow-matching generative models that produce conformational ensembles instead of single structures.

Microsoft Research, generates Boltzmann-weighted ensembles.

retraining structure predictors as flow-matching generative models that produce conformational ensembles instead of single structures.

combining AF outputs with classical MD or coarse-grained simulation; ML-accelerated force fields (CGSchNet, MACE, Allegro, AIMNet2) are dramatically extending accessible simulation timescales.

predict the bound complex from sequence + ligand SMILES. Headline improvement: no need for a crystal structure of the apo target.

the active toolkit. Accuracy degrades sharply on novel folds without homologs.

the active toolkit. Accuracy degrades sharply on novel folds without homologs.

the active toolkit. Accuracy degrades sharply on novel folds without homologs.

the active toolkit. Accuracy degrades sharply on novel folds without homologs.

the active toolkit. Accuracy degrades sharply on novel folds without homologs.

the watershed; ~1 Å median accuracy on monomers with good MSAs.

predict the bound complex from sequence + ligand SMILES. Headline improvement: no need for a crystal structure of the apo target.

open-source AlphaFold 3 reimplementation; competitive performance.

open-source AlphaFold 3 reimplementation; competitive performance.

weights publicly released.

another open AF3-class model.

language-model-only structure prediction, no MSA required, fast but less accurate than AF2.

weights publicly released.

three-track architecture; the open-source AlphaFold counterpart.

protein-ligand cofolding.

three-track architecture; the open-source AlphaFold counterpart.

open-source AF2 reproductions and extensions.