MIT团队开源BoltzGen，可跨分子类型设计蛋白结合物，66%靶标获纳摩尔级亲和力

Core Insights - The article discusses the introduction of BoltzGen, a new model developed by MIT and several institutions to address the limitations of traditional protein design methods, which rely heavily on physical calculations and have high computational costs [1][2][3] Group 1: Model Overview - BoltzGen utilizes a unified all-atom generative model that replaces traditional discrete residue labels with geometric continuous representations, allowing for joint training of protein folding and complex design [2][3] - The model incorporates a flexible design specification language that enables controllable generation across different molecular types, enhancing design efficiency and interpretability [1][3] Group 2: Research Highlights - The model has demonstrated a 66% success rate in achieving nanomolar affinity for designed nanobodies and protein complexes, showcasing its ability to optimize folding and binding performance simultaneously [2][12] - BoltzGen's architecture integrates a trunk network for token representation and a diffusion module for generating three-dimensional structures, allowing for effective modeling of atomic relationships [10][11] Group 3: Experimental Validation - In experiments involving 26 targets, BoltzGen maintained a high success rate, achieving nanomolar affinity in 66% of cases for previously unseen complex targets [12][25] - The model has shown versatility in designing peptides that bind to various structures, including those related to acute myeloid leukemia and specific enzymes, with binding affinities ranging from nanomolar to micromolar levels [15][17][19] Group 4: Data Utilization - The training of BoltzGen involved a multi-modal dataset sourced from high-quality experimental structures, AlphaFold predictions, and generated complex structures, enhancing the model's generalization capabilities [7][9] - The research team ensured diversity in the training data by excluding over-sampled datasets, maintaining a broad generation space [9]