Boltz-2: Open-Source AI for Biomolecular Structure Prediction and Drug Discovery

Drug discovery often requires 10 to 15 years and can cost billions of dollars, while failure rates remain high at every development stage. Artificial intelligence aims to shorten these timelines and reduce costs by quickly predicting molecular structures and molecular interactions.

Over roughly the past six years, many models have appeared in the field of biomolecular structure prediction. Boltz-2 marks an important step forward and is notable for several reasons. Created by MIT researchers as an open-source AI model, it shows strong performance in predicting binding affinity and provides model weights, inference code, and training code under a permissive open license. Based on the foundations of AlphaFold-3 and Boltz-1, Boltz-2 goes beyond protein structure prediction and supports a wider range of molecular systems that are important for drug design.

Unlike earlier tools that mainly focused on proteins, Boltz-2 can predict the 3D structures of:

• Proteins and protein complexes
• Protein-ligand interactions, which help explain how drugs bind to their targets, also known as binding affinity
• Nucleic acids such as DNA and RNA
• Small molecules and their interactions with biological macromolecules

Key Takeaways

• MIT’s Boltz-2 is fully open-source, including weights, inference code, and training code under a permissive license.
• Boltz-2 performs especially well when predicting how strongly drugs bind to their targets, which is essential for therapeutic effectiveness.
• The model goes beyond AlphaFold-3 and Boltz-1 by predicting structures for proteins, protein-ligand interactions, nucleic acids such as DNA and RNA, and small molecule interactions.
• Boltz-2 comes close to gold-standard Free-Energy Perturbation accuracy while requiring far less computational effort.
• Its training approach uses distillation from high-confidence AlphaFold-2 and Boltz-1 predictions, supported by Boltz-steering, an inference-time technique that integrates physics-based potentials for better accuracy.

Binding Affinity

Any improvement in binding affinity prediction is highly significant. Binding affinity describes how strongly small molecules attach to proteins. This is a central factor in determining whether a drug can engage its target and produce a therapeutic effect. At present, atomistic simulations such as Free-Energy Perturbation provide the highest accuracy for measuring binding affinity. However, these methods are computationally intensive and expensive, making them difficult to use at scale. Faster approaches such as docking are available, but they do not provide the precision required for dependable predictions. So far, no AI model has matched the accuracy of FEP methods or laboratory assays for binding affinity prediction.

Data

Boltz-2 is trained on diverse biomolecular data and improves on Boltz-1 by including ensembles from experimental and computational methods. Its training dataset includes structures from the Protein Data Bank and molecular dynamics trajectories, helping expose the model to both local fluctuations and global ensembles. Distillation methods are also used to expand the training data and strengthen the supervision signal by incorporating high-confidence predictions from AlphaFold2 and Boltz-1.

Architecture

Boltz-2 improves on the architecture of Boltz-1 and Boltz-1x by increasing controllability and strengthening the affinity module. It introduces method conditioning, template conditioning, and contact or pocket conditioning for more accurate predictions. The affinity module predicts binding likelihood and affinity values with a PairFormer model, using protein-ligand and intra-ligand interactions.

How Was Boltz-2 Trained?

Boltz-2 training is split into three phases: structure training, confidence training, and affinity training. Affinity training uses pre-computation, custom sampling strategies, and robust loss functions to improve scalability and generalization. Boltz-2 is also used to train a molecular generator called SynFlowNet, which creates small molecules with high binding scores. AlphaFold-2 was distilled to expand the training set by using its high-confidence predictions on single-chain monomers.

Performance

Boltz-2 outperforms its predecessor Boltz-1 in crystal structure prediction, especially for RNA chains and DNA-protein complexes. It also delivers competitive results compared with models such as Chai-1, ProteinX, and AlphaFold3, particularly in antibody-antigen structure prediction and the Polaris-ASAP challenge. In addition, Boltz-2 shows a stronger ability to capture local protein dynamics and approaches the accuracy of free-energy simulations on public benchmarks for affinity prediction.

Boltz-2 outperforms existing methods on the CASP16 affinity challenge and internal assays from Recursion. It also performs strongly in virtual screening, reaching high average precision and enrichment factors on the MF-PCBA dataset. Its combination of scalability and accuracy makes Boltz-2 a promising option for large-scale virtual screening in drug discovery.

Boltz-Steering

Boltz-steering, introduced as part of the Boltz-1x release, is an inference-time method that uses physics-based potentials, also known as interatomic potentials or empirical force fields. This improves physical plausibility without reducing accuracy. The same approach was integrated into Boltz-2 to create Boltz-2x.

Implementation

Set Up a GPU Server

Boltz-2 is designed to run on a GPU by default, although it can be switched to CPU execution with the --accelerator option. However, inference without a GPU is much slower.

Start by preparing a suitable GPU-based server environment.

Clone the Repository

Copy and run the following commands in your terminal one after another. These commands download Boltz-2 and install the required dependencies.

git clone https://github.com/jwohlwend/boltz.git
cd boltz; pip install -e .[cuda]

Prepare Your Input File

Boltz-2 needs information about what it should predict. This is provided through a YAML file, which is a simple text file. Create a file named my_protein.yaml. Inside this file, list the sequences of the molecules you want to study. If you are unsure about the required format, review the examples/ folder inside the Boltz directory.

By default, input_path should point to a YAML file or to a directory containing YAML files for batch processing. These files describe the biomolecules to model and the properties to predict, such as affinity.

For additional command-line options, refer to the documentation.

boltz predict input_path --use_msa_server

After the model has run, the generated results are stored in the output directory using the following structure:

out_dir/
├── lightning_logs/                                            # Logs generated during training or evaluation
├── predictions/                                               # Contains the model's predictions
    ├── [input_file1]/
        ├── [input_file1]_model_0.cif                          # The predicted structure in CIF format, with the inclusion of per token pLDDT scores
        ├── confidence_[input_file1]_model_0.json              # The confidence scores (confidence_score, ptm, iptm, ligand_iptm, protein_iptm, complex_plddt, complex_iplddt, chains_ptm, pair_chains_iptm)
        ├── affinity_[input_file1].json                        # The affinity scores (affinity_pred_value, affinity_probability_binary, affinity_pred_value1, affinity_probability_binary1, affinity_pred_value2, affinity_probability_binary2)

        ├── pae_[input_file1]_model_0.npz                      # The predicted PAE score for every pair of tokens
        ├── pde_[input_file1]_model_0.npz                      # The predicted PDE score for every pair of tokens
        ├── plddt_[input_file1]_model_0.npz                    # The predicted pLDDT score for every token
        ...
        └── [input_file1]_model_[diffusion_samples-1].cif      # The predicted structure in CIF format
        ...
    └── [input_file2]/
        ...
└── processed/                                                 # Processed data used during execution

FAQ

What Is Boltz-2?

Boltz-2 is an open-source AI model developed by MIT researchers. It is designed to predict the 3D structures of different biomolecular systems and, most importantly, to predict binding affinity, meaning the strength with which a small molecule such as a drug binds to a protein target.

How Is Boltz-2 Different from Predecessors Like AlphaFold-3 or Boltz-1?

Boltz-2 goes beyond protein structure prediction and supports a much broader set of systems, including protein-ligand interactions, nucleic acids such as DNA and RNA, and small molecule interactions. It also uses advanced training and inference methods, including Boltz-steering, to reach near gold-standard accuracy in binding affinity prediction while remaining computationally efficient.

Is Boltz-2 Truly Open-Source?

Yes. Boltz-2 is fully open-source, including model weights, inference code, and training code, all released under a permissive open license.

What Is Binding Affinity and Why Is Its Prediction Important?

Binding affinity describes the strength of the molecular interaction between a drug candidate, usually a small molecule, and its biological target, usually a protein. Accurate prediction is essential for evaluating therapeutic effectiveness and guiding lead optimization in drug discovery.

What Is FEP Accuracy, and How Does Boltz-2 Compare?

FEP (Free-Energy Perturbation) is a class of atomistic simulation techniques widely regarded as the gold standard for predicting molecular binding affinities due to their exceptional accuracy. However, these methods are computationally intensive and often require significant time and resources to produce results.

Boltz-2 aims to deliver binding affinity predictions that approach the accuracy of FEP-based methods while dramatically reducing computational costs. This improved efficiency makes it a practical solution for large-scale virtual screening workflows, where evaluating thousands or even millions of candidate molecules would be impractical using traditional FEP simulations alone.

What Is Boltz-Steering?

Boltz-steering is an inference-time method integrated into Boltz-2 and Boltz-1x. It uses physics-based potentials, also known as empirical force fields, to improve the physical plausibility of predicted structures and interactions without reducing accuracy.

What Hardware Is Required to Run Boltz-2?

Boltz-2 is intended to run on a GPU for best performance. Inference can also run on a CPU using the --accelerator option, but this is significantly slower.

Where Can I Find the Source Code and Documentation?

The repository is available on GitHub. The documentation in the repository provides detailed information about command-line options and usage.

Final Thoughts

The drug discovery field is moving quickly. Boltz-2, released as an open-source model by MIT researchers, enables broader experimentation with biomolecular simulation. By approaching gold-standard Free-Energy Perturbation accuracy in binding affinity prediction, a major bottleneck in drug development, while remaining computationally efficient, Boltz-2 expands what is possible in AI-assisted drug discovery.