How to Use Boltz-2 Online
Try Boltz-2
Commercially Available Online Web Server
Boltz-2: A Unified Model for Structure and Affinity Prediction
Boltz-2 is a revolutionary structural biology folding model that sets a new standard for computational drug development. It is the first AI model of its kind to accurately and efficiently perform binding affinity prediction while also predicting the 3D structure of proteins and other biomolecular interaction models.
Developed by researchers at MIT and Recursion, Boltz-2 achieves the accuracy of traditional Free Energy Perturbation (FEP) methods—the gold standard for affinity prediction—but with a 1000x computational efficiency gain. This leap in performance makes large-scale virtual screening and optimization of small molecule drugs practical for the first time.
Key Innovations: Jointly Modeling Complex Structures and Binding
Boltz-2 is a powerful open source model released under the MIT license, offering capabilities crucial for accelerating life science research:
Unmatched Affinity Prediction: Boltz-2 is the first learning model to approach the accuracy of FEP methods on public benchmarks like FEP+, making its predicted affinity values highly reliable for lead optimization in drug development.
Unified Prediction: Unlike models that only predict the 3D structure of proteins, Boltz-2's architecture is a single, cohesive framework designed for jointly modeling complex structures and binding affinities and likelihoods. This enables seamless use in high-throughput screening.
Comprehensive Structural Modeling: It accurately predicts the complex structures and binding poses of proteins, DNA, RNA, and small molecules, building upon predecessors like Boltz-1 and AlphaFold3. It demonstrates improved accuracy on challenging modalities like DNA-protein and antibody-antigen complexes.
Capturing Dynamics: Boltz-2 is trained on expanded training data that includes structural ensembles from molecular dynamics (MD) simulations. This, combined with B-factor supervision, allows it to better capture local protein dynamics and predict conformational changes.
Enhanced Controllability: The model allows researchers to guide predictions with "controllability features," including:
Conditioning on experimental methods (e.g., X-ray crystallography, molecular dynamics).
Applying user-defined distance constraints for contacts or binding pockets.
Integrating multi-chain templates to inform structure prediction.
Performance Benchmarks
Task | Metric | Boltz-2 Result | Key Finding |
Binding Affinity Prediction | Pearson R on FEP+ (4 targets) | 0.66 | Approaching FEP accuracy at 1000x speed |
Hit Discovery | Average Precision (MF-PCBA) | Doubled over ML/Docking | Enables accurate large-scale virtual screening |
Structure Prediction | Performance vs. Boltz-1 | Modest, consistent gains | Strong structural foundation for all modalities |
Blind Challenge | Pearson R (CASP16 Affinity) | 0.65 (Out-of-the-box) | Tops all submitted entrants for affinity ranking |
Virtual Screening with Boltz-2: High-Throughput Hit Discovery
Boltz-2 offers a powerful solution for large-scale virtual screening (VS) in early-stage drug discovery by combining high speed with high precision, addressing the critical performance/compute time trade-off.
Key Advantages for Hit Discovery
Speed and Scale: Boltz-2 is approximately 1000x more computationally efficient than Free-Energy Perturbation (FEP) methods. This speed enables the screening of hundreds of thousands of compounds per day, making it practical for exploring vast chemical libraries.
Accuracy for Classification: The model is trained on a hybrid dataset, including large-scale binary labeled data (binder/non-binder) specifically for hit discovery.
Superior Enrichment: On the retrospective MF-PCBA benchmark for high-throughput screens, Boltz-2 substantially outperforms prior machine learning and docking approaches. It nearly doubles the average precision and achieves significant enrichment gains (e.g., an enrichment factor of 18.4 at a 0.5% threshold), demonstrating its reliability in retrieving actives from large, imbalanced datasets.
Generative Virtual Screening: When coupled with a molecular generator (like SynFlowNet), Boltz-2 acts as a fast and robust scoring function to enable de novo generation of new binders from ultra-large, synthesizable chemical spaces (e.g., Enamine 76B REAL space).
This pipeline was demonstrated prospectively on the TYK2 target, where Boltz-2 successfully steered the generation of diverse, synthesizable, high-affinity candidates, as estimated by absolute FEP simulations (ABFE).
Lead Optimization with Boltz-2: Refining Chemical Series
In the later stages of drug development, such as hit-to-lead and lead optimization, the challenge shifts from finding binders to accurately ranking subtle differences in binding affinity among closely related analogues. Boltz-2 excels at this precise task.
Performance for Ranking and Affinity
Approaching FEP Accuracy: Boltz-2 is the first AI model to approach the performance of FEP methods in estimating protein-small molecule binding affinity.
On the competitive FEP+ benchmark, Boltz-2 achieves an average Pearson correlation of 0.66, outperforming all available inexpensive physical methods and deep learning baselines.
It also approaches the performance of OpenFE, a widely adopted open-source relative FEP method.
Outperforming Competition: In the rigorous CASP16 affinity challenge (a blind benchmark), Boltz-2, run entirely out-of-the-box, outperformed all top-ranking participants.
Predicting Continuous Affinity: The model's affinity module is specifically designed to regress continuous affinity values log10 scale from µM) for quantities like Ki, Kd, and IC-50. This quantitative output is critical for guiding molecular refinement and prioritizing candidates.
Robustness via Pairwise Loss: During training, affinity training is supervised primarily by minimizing pairwise intra-assay differences in affinity values. This technique helps cancel out assay-specific confounding factors (like those accounted for by the Cheng-Prusoff equation), resulting in predictions that are highly effective for ranking molecules within the same experiment or chemical series.
Structural Grounding: The affinity module operates on Boltz-2's core structural predictions, leveraging the refined pair representation and predicted coordinates, focusing exclusively on protein-ligand and intra-ligand interactions. Reliable affinity prediction fundamentally relies on the accuracy of the predicted 3D protein-ligand structures and binding interfaces.
Try Boltz-2 now for virtual screenings, lead optimization, etc, by following the below guide on the Tamarind platform, and start predicting the 3D structures of proteins and their structures and binding affinities to revolutionize your drug development projects.
What is Tamarind Bio?
Tamarind Bio is a pioneering no-code bioinformatics platform built to democratize access to powerful computational tools for life scientists and researchers. Recognizing that many cutting-edge machine learning models are often difficult to deploy and use, Tamarind provides an intuitive, web-based environment that completely abstracts away the complexities of high-performance computing, software dependencies, and command-line interfaces.
The platform is designed to provide easy access to biologists, chemists, and other researchers who may not have a background in programming or cloud infrastructure but want to run experimental models with their data. Key features include a user-friendly graphical interface for setting up and launching experiments, a robust API for integration into existing research pipelines, and an automated system for managing and scaling computational resources. By handling the technical heavy lifting, Tamarind empowers researchers to concentrate on their scientific questions and accelerate the pace of discovery.
Boltz-2 on Tamarind Bio: Accelerate Your Discovery
The integration of Boltz-2's advanced capabilities with Tamarind's user-centric platform creates a powerful synergy, significantly accelerating the drug discovery and research process. Tamarind democratizes access to this state-of-the-art AI model. Our no-code platform abstracts away the complexities of high-performance computing, GPU orchestration, and dependencies.
Speed and Accuracy: Boltz-2 provides FEP-level accuracy at an unmatched speed, and Tamarind Bio provides the scalable cloud infrastructure to run these predictions on a massive scale. This enables researchers to perform large-scale virtual screens that were previously infeasible, identifying promising drug candidates much faster.
Streamlined Workflow: By abstracting away the computational and technical complexities of running Boltz-2 locally, Tamarind Bio allows researchers to focus on the science. The user-friendly interface and automated resource management drastically reduce the time spent on setup and troubleshooting, allowing for rapid iteration and a faster discovery pipeline.
Enhanced Controllability: The ability to incorporate experimental data and user-defined constraints in Boltz-2 is amplified by Tamarind Bio's platform. This allows researchers to guide the model with their expertise and rapidly explore specific design hypotheses, making the entire process more efficient and effective.
How to Use Boltz-2 on Tamarind Bio
Tamarind makes using Boltz-2 straightforward and efficient, regardless of your technical expertise. The no-code platform streamlines the entire workflow for molecular structure and binding affinity prediction.
Here is a simple, step-by-step guide for researchers to get started:
Access our Platform: Begin by logging in to the tamarind.bio website.
Select Boltz-2: From the list of tiles, either search or select Boltz-2.
Specify Inputs: Once you are in the Boltz-2 tool, you will next want to specify your protein by providing its sequence, and your ligand (SMILES string or a file). Any necessary file conversions and preprocessing are carried out automatically.
Configure Parameters: In a simple, graphical user interface, you are able to specify all of your prediction parameters. You can choose to run a structural prediction or a binding affinity prediction, as well as set custom constraints or use a template to guide the model.
Submit and Monitor: In order to submit your job, click "Submit" and your job will start running. The Tamarind Bio platform handles the allocation of powerful GPU resources and will perform the Boltz-2 simulation, which can take as little as 20 seconds on a single GPU. You can monitor the progress of your job directly from the dashboard.
Analyze the Results: After completion of the job, you will receive a comprehensive report within the Tamarind application. The results include a ranked list of predicted structures, binding affinity values (log(IC50)), and confidence scores. From this data, you can explore interactive 3D visualizations of the structures, enabling you to inspect the predicted binding poses and evaluate the affinity predictions.