Boltz

1. Introduction

Boltz^1,2 is a fundamental model in structural biology jointly launched by institutions including MIT, focusing on the modeling of biomolecular interactions.

Currently, AbSeek has deployed the latest version, Boltz-2¹, which covers structure prediction and binding affinity prediction. Its features are reflected in multiple aspects such as multimodal capabilities, performance breakthroughs, methodological innovations, data breadth, and open-source ecology:

• Comprehensive functions: It can accurately predict static complexes and dynamic assembly structures of various biomolecules such as proteins, DNA, RNA, and small molecules, and is particularly good at analyzing complex interactions like antibody-antigen interactions. Meanwhile, it can predict the binding affinity between small molecules and proteins, providing key support for evaluating molecular functions and therapeutic efficacy in drug design.
• Performance breakthroughs: In terms of structure prediction, it significantly improves the accuracy of crystal structure prediction compared to Boltz-1². It performs outstandingly in modalities such as RNA chains and DNA-protein complexes, approaching the dynamic property prediction capabilities of specialized models like AlphaFlow and BioEmu. In terms of affinity prediction, it is the first AI model to approach the accuracy of Free Energy Perturbation (FEP) methods, with computational efficiency improved by at least 1000-fold, surpassing mainstream deep learning models in benchmark tests such as FEP+ and CASP16.
• Methodological innovations: It introduces controllable features such as experimental method conditions, distance constraints, and multi-chain template integration, supporting users to incorporate prior knowledge. It optimizes computational efficiency through mixed-precision training and trifast kernel optimization, and enhances the physical rationality of structures with Boltz-steering¹. It optimizes affinity prediction based on the latent representation of structural modeling, achieving a balance between performance and efficiency.
• Rich data: The training data includes experimental structures from PDB, molecular dynamics ensembles (MISATO, ATLAS, mdCATH), and distilled data (AlphaFold2 monomer predictions, Boltz-1 complex predictions). The binding affinity data is from databases such as PubChem and ChEMBL, which have been standardized to extract effective signals, supporting hit discovery and optimization scenarios.

The architecture and performance of Boltz-2 are shown in the following figures:

Fig. 1. Architecture diagram of Boltz-2.

Fig. 2. Performance of Boltz-2 in structure prediction.

Fig. 3. Performance of Boltz-2 in small molecule affinity prediction.

2. Parameter Description

This tool supports multiple molecule types, including proteins, DNA, RNA, and small molecules. You can select the type via the Molecule Type parameter. Currently, only standard amino acids, standard bases, and small molecules in SMILES format are supported.

Click the Add Sequence button to add a new sequence. Chains will be automatically labeled using uppercase letters A–Z based on the order of input.

Click the Add Constraints button to incorporate different types of constraints, such as:

• Residue/Atom Distance Constraints (contact)
• Pocket Distance Constraints (pocket)
  • Requires input of two different chains for constraint
  • The first object supports any type of chain, while the second object only supports protein, DNA, or RNA chains
• Atom Covalent Bond Constraints (bond)

Among them, the input parameter position for Ligand corresponds to the order of atoms in the SMILES expression; the atom names in standard residues need to be correctly filled in by the user.

Note that constraints are not strictly enforced—the final output structure is determined by the prediction algorithm.

For more detailed information on constraints, please refer to the official documentation:https://github.com/jwohlwend/boltz/blob/main/docs/prediction.md

3. Result Explanation

Five predicted structure files in .cif format will be returned, along with a .csv file of prediction metrics corresponding to each structure. The explanations of core metrics are as follows:

4. References

[1] Passaro, Saro and Corso, Gabriele and Wohlwend, Jeremy and Reveiz, Mateo and Thaler, Stephan and Somnath, Vignesh Ram and Getz, Noah and Portnoi, Tally and Roy, Julien and Stark, Hannes and Kwabi-Addo, David and Beaini, Dominique and Jaakkola, Tommi and Barzilay, Regina. Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. bioRxiv 2025. https://doi.org/10.1101/2025.06.14.659707

[2] Wohlwend, Jeremy and Corso, Gabriele and Passaro, Saro and Getz, Noah and Reveiz, Mateo and Leidal, Ken and Swiderski, Wojtek and Atkinson, Liam and Portnoi, Tally and Chinn, Itamar and Silterra, Jacob and Jaakkola, Tommi and Barzilay, Regina. Boltz-1: Democratizing Biomolecular Interaction Modeling. bioRxiv 2024. https://doi.org/10.1101/2024.11.19.624167

[3] Mirdita, M., Schütze, K., Moriwaki, Y. et al. ColabFold: making protein folding accessible to all. Nat Methods 19, 679–682 (2022). https://doi.org/10.1038/s41592-022-01488-1