AI设计抗体盲测结果公布：29家机构511条序列参与，仅少数具备高亲和力
首个大规模AI抗体设计盲测表明，模型在真实未见过序列上预测能力有限，成功率高估严重。
AI设计抗体到底行不行？29家机构、511条序列被拉进同一场盲测。结果显示，在信息丰富的SARS-CoV-2 RBD靶标上，多数模型表现平平，最佳设计亲和力达95 pM，但整体成功率仅63%。这说明AI抗体设计仍远未成熟。
正文摘录
.jpg) Editor: & Over the past few years, papers on AI antibody design have multiplied. One model generates new antibody sequences, another predicts affinity, and still others leverage protein language models or structural information to search for better candidates. How much of this actually works? Hard to say—because antibody design and protein structure prediction are not the same thing. Antibody development is far more complex: a candidate must first bind its target with sufficient affinity; it should also express stably, resist aggregation, and avoid excessive non-specific interactions. Backtesting on existing data, therefore, easily masks the problem: a model can rationalize the answ…