The MIT team has open-sourced BoltzGen, which can design protein binders across molecular types, achieving nanomolar-level affinity for 66% of the targets.

To address the limitations of traditional protein design, which relies on physical calculations, incurs high computational costs, has a limited design space, and struggles to handle multi-modal targets simultaneously, the Massachusetts Institute of Technology (MIT) collaborated with multiple institutions to introduce BoltzGen. It replaces traditional discrete residue labels with geometric continuous representations, enables joint training for protein folding and binder design, and constructs a flexible design specification language to achieve controllable generation across different molecular types, thereby enhancing the model's design efficiency, versatility, and interpretability.

In the fields of drug development and biomolecular engineering, "De-novo protein design (De-novo Binder Design)" is one of the core methods for automated drug development. Researchers can leverage computational simulations and deep learning to generate peptide chains or protein structures with binding capabilities at specific targets, making it possible to develop new drug forms such as antibodies, nanobodies, and cyclic peptides.

However, traditional protein design strategies mostly rely on physical calculations such as molecular dynamics simulations and sequence optimization algorithms. Although they can achieve high precision in a single system, they have high computational costs, a limited design space, and find it difficult to handle multi-modal targets such as protein-small molecules and RNA simultaneously. While current deep generative models have improved the generation speed to some extent, they generally lack the ability to perform "atom-level" structural reasoning, are optimized for specific types of molecules, and have limited versatility. Additionally, their model evaluation often depends on similar complexes already present in the training set, making it difficult to verify their generalization ability for "unseen targets." They also lack a controllable generation mechanism and flexible structural constraint expression, resulting in limitations in design efficiency and interpretability.

To tackle this issue, MIT collaborated with multiple institutions including Boltz to propose the "All-atom Generative Model" BoltzGen, which unifies structure prediction and binder design. This model not only replaces traditional discrete residue labels with geometric continuous representations to enable joint training for protein folding and binder design in a single system but also constructs a flexible "design specification language" to achieve controllable generation across different molecular types.

Experimental results show that 66% of the targets in both nanobody and protein binder designs by BoltzGen achieved nanomolar-level affinity, demonstrating for the first time that a "single model system" can simultaneously optimize folding and binding performance in multi-modal biomolecular design.

Currently, the relevant research findings are published under the title "BoltzGen: Toward Universal Binder Design." 

GitHub link: https://github.com/HannesStark/boltzgen

Research highlights:

* Unifies structure prediction and binder design in a single all-atom generative model, enabling simultaneous protein folding, binding site modeling, and sequence generation at atomic-level precision, significantly enhancing the physical rationality and controllability of molecular design; 

* Proposes a universal "design specification language" that allows the model to flexibly switch between different systems such as proteins, nanobodies, cyclic peptides, and small molecules, achieving cross-modal structure generation and constraint control, and broadening the scope of application of generative AI in the field of biomolecular design. 

Paper link: https://go.hyper.ai/3sx2K

Mixed dataset: Multi-modal training strategy

The research team adopted a multi-level, cross-modal joint training framework when training BoltzGen. The core sources of the dataset they used include three types:

* High-quality experimentally resolved structures from the Protein Data Bank (PDB), covering various complex structures such as RNA, DNA, and protein-small molecules, providing the model with real chemical bond constraints and three-dimensional geometric distribution data; 

* Experimental data predicted and re-learned by AlphaFold2 from the AlphaFold Database (AFDB), covering reliable folding patterns of proteins in experiments; 

* Complex structure samples generated by the Boltz-1 model, covering multi-modal scenarios such as small molecule binding and RNA-DNA interactions, which can enhance the model's generalization ability across different biomolecular types. 

To prevent the model from being overly biased towards specific structure types, the research team removed the upsampled datasets of antibodies and TCRs to maintain the diversity of the generation space. Meanwhile, all structural samples were randomly cropped and multi-task processed during the training process, allowing the model to randomly undertake tasks such as folding prediction, binder design, and structure completion in each training iteration, realizing a unified multi-functional learning framework, and enabling the model to have cross-modal understanding ability while generating at the atomic level.

Model architecture: All-atom reasoning from noise to structure

This model retains the main components of the AlphaFold3 and Boltz-2 architectures and makes some improvements on this basis to introduce more conditional inputs.

As shown in the figure below, the entire model is divided into two main parts: a larger Trunk (backbone network) and a Diffusion Module. The Trunk is responsible for generating token representations and pairwise representations for conditional control, while the diffusion module generates three-dimensional structures based on these. The Trunk runs only once, while the diffusion module runs multiple times iteratively to gradually denoise the three-dimensional coordinates of all atoms.

BoltzGen model architecture diagram

During the Trunk stage, similar to the Trunk module of Boltz-2, it is responsible for parsing the input protein structure and target information. The Trunk module processes tokenized molecular structures. It mainly uses the PairFormer architecture to efficiently model the spatial relationships between atoms through Triangle Attention. At the same time, combined with Geometric Residue Encoding, it can simultaneously infer residue types and atomic coordinates in a continuous space without relying on discrete amino acid labels. This mechanism allows the model to truly understand the physical laws of the structure during generation rather than simply relying on data memory.

During the Diffusion Module stage, this module receives noisy three-dimensional atomic coordinates as input and predicts their denoised coordinates. Meanwhile, it uses a standard Transformer architecture and runs at both the atom level and the token level. BoltzGen uses a continuous space diffusion model to gradually "denoise" and generate atomic coordinates, realizes the transformation from a random initial state to a stable conformation by predicting the noise vector, and retains the constraints of the molecular energy surface during the generation process to avoid physical conflicts or structural collapse.

Experimental results: Universal design verification across 26 targets

In the experimental part, the performance verification of the BoltzGen model covered multiple dimensions from proteins to peptides, from new pathogens to small molecule targets, demonstrating excellent generalization and controllability.

The team tested 26 targets in eight independent wet-lab verification projects, involving various types of binders such as nanobodies, proteins, linear and cyclic peptides. The results showed that BoltzGen still maintained a high success rate on unseen complex targets: in nine experiments on new targets completely different from the training data, the designed proteins and nanobodies achieved nanomolar (nM) high-affinity binding on 66% of the targets, indicating the model's strong structural reasoning and cross-modal design capabilities.

Experimental results of protein design

In experiments on bioactive peptides with diverse structures, the proteins designed by BoltzGen can bind to different types of peptide molecules with nanomolar to micromolar (μM) affinity and effectively neutralize their antibacterial or hemolytic activities. For the disordered protein NPM1 related to acute myeloid leukemia, the peptides generated by the model showed nucleolar co-localization in living cells, providing the first in vivo evidence that AI-designed proteins can bind to natural disordered proteins. 

Design of peptides binding to the disordered region of NPM1

Significant results were also obtained in the design for the core enzymes of cell metabolism, RagC and the RagA:RagC dimer: 7 out of 29 candidate peptides successfully bound to RagC, with the highest affinity reaching 3.5 µM; 14 cyclic disulfide bond peptides showed stable binding. 

Design of peptides binding to specific sites of the RagC GTPase

BoltzGen also demonstrated cross-scale design capabilities on two biomedically significant small molecules. The generated protein binders showed detectable binding activity in the range of 50–150 µM, proving that the model can achieve small molecule recognition without expert chemical guidance. Additionally, in the design of antibacterial peptides against the bacterial DNA gyrase GyrA, more than 19% of the candidate sequences could reduce bacterial growth by more than four times, and some peptides could directly kill host cells. 

Design of proteins binding to small molecules

In tests on five benchmark targets with known binding structures (such as PD-L1, TNFα, PDGFR, etc.), BoltzGen also achieved a high hit rate—80% of the targets had nanomolar-level binders, verifying its accuracy comparable to the current best models. 

Test results of protein binders

Test results of protein binders

Overall, this series of experiments shows that BoltzGen can not only reproduce high-quality binding structures within the known data distribution but also achieve functional design in completely unfamiliar biological systems. Its unified all-atom generation architecture integrates the "design - prediction - verification" process, providing an open, controllable, and scalable AI infrastructure for future drug discovery and biomolecular engineering.

From prediction to generation, the Boltz series reshapes the landscape of AI-driven molecular design

In 2024, the research team from the Jameel Clinic at MIT introduced the Boltz-1 model. In the context of the global drug design field shifting from "structure prediction" to "function generation," although the AlphaFold series of models first revealed the computability of the protein folding problem, AlphaFold3 was not fully open-source, limiting the free iteration of the industry in real drug scenarios. Thus, Boltz-1 emerged in this context. It not only approaches AlphaFold3 in performance but is also fully open-source and commercially available, promoting molecular structure prediction into an open ecosystem in the industry.

Boltz-1 uses a generative system that combines a diffusion model and a Transformer architecture. It can predict the structures of proteins, RNA, DNA, and small molecule complexes at the atomic level. Its flexible conditional interface allows the model to accurately model specific binding sites or molecular conformations, greatly expanding its industrial application scope. From the design of new antibodies and the optimization of enzyme engineering to the screening of small molecule ligands, end-to-end prediction can be achieved within the Boltz-1 framework, significantly lowering the entry threshold for biological computing.

In 2025, the Jameel Clinic team at MIT introduced the Boltz-2 model based on Boltz-1. It pushed the accuracy of protein folding prediction to a new high and was hailed as the "GPT-4 of structural biology."

Compared with its predecessor, Boltz-2 achieved significant improvements in generation accuracy and computational efficiency. It also introduced multi-modal conditional inputs, enabling it to integrate sequence information, experimental data, and chemical properties for more refined molecular design. In the wave of global biological computing and drug discovery moving towards "full-scenario generation," the emergence of Boltz-2 further filled the demand of academia and the industry for highly available, scalable, and commercially available tools.

Boltz-2 inherits and optimizes the hybrid generative system of the diffusion model and the Transformer architecture. Its core Trunk module can extract multi-level representations of protein or nucleic acid complexes at once, while the Diffusion module generates and optimizes structures based on these. 

Boltz-2 structure diagram

Thanks to its flexible conditional interface, researchers can precisely control the output structure for specific binding sites, active pockets, or small molecule ligands, significantly expanding the application potential of the model in fields such as the design of new antibodies, the optimization of enzyme catalysis, and the screening of drug lead molecules. The open-source nature of Boltz-2 also ensures that academia and the industry can freely iterate, thereby accelerating the implementation of molecular generation computing in real drug development scenarios.

Now, BoltzGen proposes a universal "design specification language" that allows the model to flexibly switch between different systems such as proteins, nanobodies, cyclic peptides, and small molecules, achieving cross-modal structure generation and constraint control, and further broadening the scope of application of generative AI in the field of biomolecular design.

This article is from the WeChat public account "HyperAI Super Neural." Authors: YeYe