连破两局，技术又开始降维打击了

本周科技领域涌现两项重要AI进展。由艾伯维、Astex、BMS、强生、武田五家全球知名药企组成的AISB联盟1，通过联邦学习方法联合训练了蛋白结构预测模型AISB-1-Fed1。该模型在1056个未参训蛋白-配体测试结构上的高质量预测占比达到52.1%1，较公开版本OpenFold3的35.6%和开源模型Boltz-2的40.9%均有明显提升1。

联邦学习技术的应用是这一突破的关键。这一方法将模型训练过程搬到各企业数据所在地，确保原始蛋白结构始终留在公司内部，只有模型参数在联盟间流动1。AISB联盟于2025年3月成立1，最初仅由艾伯维与强生参与1，后来扩展至现有规模。

同时，阿里巴巴达摩院与浙江大学医学院附属第一医院等机构发布了通用医疗影像AI模型DAMORADAR1，采用视觉-语言学习方法，可识别146种病症1。相关成果已发表于《科学》期刊1。

A consortium of five major pharmaceutical companies has achieved a significant advancement in protein structure prediction through federated learning. The AISB alliance—comprising Abbvie, Astex, Bristol Myers Squibb, Johnson & Johnson, and Takeda—unveiled the AISB-1-Fed model, which demonstrated high-quality predictions on 52.1% of 1,056 untrained protein-ligand test structures, substantially outperforming the publicly available OpenFold3 model's 35.6% accuracy and the open-source Boltz-2 model's 40.9% performance 1. The AISB alliance, originally formed in March 2025 with Abbvie and Johnson & Johnson as founding members, has since expanded to include these five organizations 1.

The federated learning approach employed in developing AISB-1-Fed represents a novel collaborative methodology in which training occurs at the location of the data itself, ensuring that original protein structures remain confidential within each company while only model parameters are exchanged across the network 1. In a separate development, Alibaba's DAMO Academy, in collaboration with the First Affiliated Hospital of Zhejiang University School of Medicine and other institutions, released DAMORADAR, a general-purpose medical imaging AI model capable of identifying 146 different disease conditions through a vision-language learning approach 1. The research findings were published in Science on September 17, 2026 1.