MIT Researchers Unveil Boltz-2: AI Model Predicts Protein Structure, Binding Affinity in Seconds

June 6, 2025

By Allison Proffitt

June 6, 2025 | Researchers at MIT and Recursion today announced the release of Boltz-2, an updated artificial intelligence model that can predict protein structure (building on Boltz-1) and also binding affinity—how strongly potential drugs bind to their target proteins—in just 20 seconds with a single GPU.

Boltz-2 was developed in Regina Barzilay’s Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT in conjunction with researchers at Recursion. The model is available today under an MIT license for commercial and non-commercial use.

The model builds upon the success of Boltz-1, which became the most widely adopted co-folding model in the industry after its release last fall. “We're really far outperforming methods… pretty much doubling the performance of previous methods,” said Saro Passaro, an AI research scientist in Barzilay’s group at CSAIL.

“Boltz-2 [structure predictions] outperformed Boltz-1 pretty much everywhere across the board and has largely bridged the gap with AlphaFold3,” added Jeremy Wohlwend, a member of the research team.

Beyond Structure: A Unified Approach to Binding Affinity

Unlike previous AI models that focus solely on predicting protein structures, Boltz-2 is a “biomedical foundation model” that produces both the structure and internal representations that unify structure prediction with binding affinity calculations.

Determining binding affinity—or how tightly a drug molecule attaches to its intended target—is one of the most persistent bottlenecks in drug discovery, explained Najat Kahn, Chief R&D Officer and Chief Commercial Officer at Recursion. This fundamental property determines both a drug's effectiveness and its potential for harmful side effects.

“Binding affinity is core to developing a therapeutic start to finish,” said Kahn in a briefing about the new model. “It's been the fundamental issue that a lot of us have been trying to grapple with for various reasons—you want to bind to the right areas and you don't want to bind to the wrong proteins to limit off-target effects.”

Current computational methods for predicting binding affinity rely on physics-based simulations called Free Energy Perturbation (FEP) calculations, which can cost hundreds of dollars per prediction and require 6-12 hours to complete. Boltz-2 achieves comparable accuracy in mere seconds at a fraction of the cost—dropping from approximately $100 per prediction to just a few cents.

In benchmark tests, Boltz-2 achieved a 0.6 correlation with experimental results on the gold-standard FEP+ benchmark, matching the performance of computationally intensive simulations that run for hours or days. The model also demonstrated unprecedented performance in drug discovery scenarios, significantly outperforming existing methods in identifying potential drug compounds.

“[Predicting binding] affinity was an open problem for decades,” said Barzilay. Groups have been actively working on it but have not make progress until now. “It really required very novel machine learning to develop this technique,” she said, calling Boltz-2 not only an advancement in life sciences research, but also in computer science.

Collaborative Effort

The development of Boltz-2 was a collaborative effort, according to a statement from the team, combining the MIT team’s academic expertise with computational resources and practical drug discovery insights from Recursion.

Recursion had first incorporated Boltz-1 into drug discovery workflows after it was released late last year. “We're now getting even better results by replacing Boltz-1 with Boltz-2. This pipeline is being used by our team internally and has applications in screening, target deconvolution, and lead optimization,” said Khan.

The MIT team led the foundational work on the structural prediction model of Boltz-2 and initiated the affinity prediction model by building its core code, selecting the initial data, training it, and gathering public benchmarks.

Recursion validated the affinity model using its NVIDIA-based supercomputer, BioHive-2 (approximately 400K-500K A100 hours). The company also tested Boltz-2 Affinity against its own internal drug discovery data and developing additional machine learning and physics-based benchmarks to ensure its accuracy.

Affinity Advantages: Molecular Dynamics, Physical Steering

The model incorporates several innovative features, including the ability to use molecular dynamics simulations in training, template-based predictions using known structures, and “physical steering” mechanisms that protect against structural hallucinations and ensure predictions remain grounded in chemical reality.

The implications for drug discovery are substantial. Currently, developing a single drug can require testing thousands of compounds, with each experimental validation costing hundreds or thousands of dollars and taking weeks to complete. Boltz-2's speed could enable researchers to screen millions of potential compounds computationally before synthesizing the most promising candidates.

Companies are already seeing dramatic improvements. At Recursion, a biotechnology company using the technology, drug discovery programs were completed in 18 months, instead of the industry average of 42 months, while the number of compounds requiring synthesis has dropped to just a few hundred from an industry average of 5,000-10,000, Kahn reported.

“The ability to get FEP-level accuracy that early on helps you select the right candidates to synthesize much, much earlier,” she said. “That improves both the efficiency… and the efficacy of how you're actually developing the right small molecule therapeutics.”

Open Science Approach

Released today under an open MIT license, Boltz-2 will be freely available for both academic research and commercial applications, said Gabriele Corso, from the CSAIL lab. “It basically tells you that you can take the code and do whatever you want with it. Use it, you know, modify it, directly use it, deploy it, and so on, whether you are academic or a company.”

“The reason we believe that it’s really important for these models to be out there in the open is, at the end of the day, 99.9% of drug developers or biologists are outside of companies like Isomorphic [Labs, spinout from DeepMind and developers of AlphaFold],” noted Corso. “Part of the reason why we are releasing this fully open source is because we want all these biologists to have access to it.”

The open-source approach has already proven successful with Boltz-1, which has attracted nearly 40 contributors who have improved the model's performance and efficiency. The community has implemented optimizations ranging from specialized algorithms for cyclic proteins to GPU acceleration kernels.

Looking Forward

While Boltz-2 focuses on small molecule drugs binding to proteins, the researchers see potential for expansion to other therapeutic modalities, including protein-protein interactions and biological drugs. The model's foundation could enable prediction of additional properties beyond binding affinity, potentially allowing researchers to optimize multiple drug characteristics simultaneously.

“We believe this is going to be a game changer in preclinical development,” Passaro concluded. “The accuracy and speed of Boltz-2 enables large and diverse virtual screening, significantly improving the enrichment factors of previously developed methods.”