Boltz-1：让生物分子交互建模更普及 Democratizing Biomolecular Interaction Modeling.docx

1

Boltz-1

DemocratizingBiomolecularInteractionModeling

JeremyWohlwend*12,GabrieleCorso*12,SaroPassaro*12,

MateoReveiz12,KenLeidal3,WojtekSwiderski3,TallyPortnoi12,

ItamarChinn12,JacobSilterra12,TommiJaakkola12,ReginaBarzilay12Correspondenceto{jwohlwend,gcorso,saro00}@csail.mit.edu

Abstract

Understandingbiomolecularinteractionsisfundamentaltoadvancingfieldslikedrugdiscoveryandproteindesign.Inthispaper,weintroduceBoltz-1,anopen-sourcedeeplearningmodelincorporatinginnovationsinmodelarchitecture,speedoptimization,anddataprocessingachievingAlphaFold3-levelaccuracyinpredictingthe3Dstructuresofbiomolecularcomplexes.Boltz-1demonstratesaperformanceon-parwithstate-of-the-artcommercialmodelsonarangeofdiversebenchmarks,settinganewbenchmarkforcommerciallyaccessibletoolsinstructuralbiology.Byreleasingthetrainingandinferencecode,modelweights,datasets,andbenchmarksundertheMITopenlicense,weaimtofosterglobalcollaboration,acceleratediscoveries,andprovidearobustplatformforadvancingbiomolecularmodeling.

Contents

1Overview2

2Datapipeline3

2.1Datasourceandprocessing 3

2.2Validationandtestsetscuration 3

2.3DenseMSApairingalgorithm 4

2.4Unifiedcroppingalgorithm 4

2.5Robustpocket-conditioning 5

3Modeling5

3.1Architecturalmodifications 6

3.2Trainingandinferenceprocedures 6

3.3Confid