A team of researchers from the Massachusetts Institute of Technology (MIT) Computer Science and Artificial Intelligence Lab (CSAIL) and Jameel Clinic, in collaboration with the clinical stage biotech company Recursion, has released a next-generation biomolecular foundation model, Boltz-2 (1). Designed to predict both molecular structure and binding affinity with high accuracy and speed, Boltz-2 is now available as open-source software under the MIT license (2).
The model builds upon its predecessor, Boltz-1, and represents a significant advancement over current tools, such as AlphaFold3— an advanced AI model that predicts the 3D structures of various biomolecular complexes, including proteins, nucleic acids (DNA and RNA), and small molecules—according to a June 6, 2025 news release from Recursion (1). Trained and validated using Recursion’s high-performance NVIDIA supercomputer, Boltz-2 is designed to jointly model complex molecular structures and predict their binding affinities—two critical factors in drug discovery and development workflows.
Joint Modeling Unlocks Faster, More Accurate Drug Candidate Screening
Key Takeaways
·Boltz-2, developed by MIT and Recursion, delivers 1000 times faster binding affinity predictions than free energy perturbation for virtual drug screening.
·Boltz-2 predicts both three-dimensional molecular structures and binding affinities, improving compound selection and accelerating pharmaceutical research and development.
·Released as open-source software, Boltz-2 allows drug developers to tailor advanced artificial intelligence tools for specific discovery and manufacturing needs.
Boltz-2 distinguishes itself through its capacity for joint modeling of 3D biomolecular complexes, enabling it to predict both binding affinity and structural dynamics in a single inference pass. According to Regina Barzilay, AI faculty lead at the Jameel Clinic, “Accurately predicting how strongly molecules bind has been a long-standing challenge in drug discovery—one that required novel machine learning and computer science techniques to address. Boltz-2 not only addresses this crucial problem but also helps scientists uncover new biological insights and ask questions they couldn't before with standard approaches that are more computationally intensive" (1).
One of Boltz-2’s most notable features is its ability to approach the accuracy of physics-based free energy perturbation methods—widely regarded as the gold standard in affinity prediction—while operating up to 1000 times faster. This efficiency enables large-scale virtual screening campaigns to be conducted more feasibly and at significantly reduced computational cost.
"Selecting the right molecules early is one of the most fundamental challenges in drug discovery, with implications for whether R&D programs succeed or fail," said Najat Khan, chief R&D officer and chief commercial officer at Recursion (1). “By predicting both molecular structure and binding affinity simultaneously with unprecedented speed and scale, Boltz-2 gives R&D teams a powerful tool to triage more effectively and focus resources on the most promising compounds."
Enhanced Control and Physical Realism in Molecular Predictions
Boltz-2 introduces several architectural innovations aimed at increasing model interpretability and control (1). These include the use of Boltz-steering for improved physical plausibility, and conditioning techniques that allow users to guide predictions using templates, methods, or specific molecular contacts. The model also integrates predictions of protein dynamics, such as B-factors, providing a richer understanding of biomolecular behavior.
Its training regimen leveraged a diverse and extensive dataset, combining molecular dynamics simulations, expanded distillation data, and approximately 5 million binding affinity assay measurements. On public benchmarks, Boltz-2 achieved leading performance, outperforming all participants in the CASP16 affinity challenge.
The model’s developers emphasize its openness and accessibility. All code, model weights, and the full training pipeline are freely available to both academic and commercial users, aligning with a broader push within the AI and pharmaceutical communities to democratize cutting-edge tools for scientific advancement.
References
Recursion. MIT and Recursion Release Boltz-2: Next Generation AI Model to Predict Binding Affinity at Unprecedented Speed, Scale, and Accuracy. Press Release. June 6, 2025.
MIT & Recursion. Boltz-2—Towards Accurate and Efficient Binding Affinity Prediction (accessed June 9, 2025).
