FLAb2 基准:30 个蛋白 AI 模型难以稳定预测抗体可开发性
FLAb2: Benchmarking reveals that protein AI models cannot yet consistently predict developability properties

Chungyoun 等构建了包含 400 多万条抗体的公开基准 FLAb2,并测试了 30 个模型,发现多数模型未能通过,也没有模型能同时准确预测各项可开发性指标。作者指出,不少模型依赖与天然 germline 抗体的相似性等捷径,而非底层生物物理特征;提升抗体 AI 设计更依赖更多样化的实验数据,而非更大的模型。
原文摘要
出版方登记 · OpenAlexWe introduce the Fitness Landscape for Antibodies (FLAb2), the largest public benchmark for therapeutic antibody design, comprising developability data for over 4 million antibodies across 32 studies and covering 7 properties: thermostability, expression, aggregation, binding affinity, pharmacokinetics, polyreactivity, and immunogenicity. Evaluating 30 AI and biophysical models, we find that most (80%) fail to produce statistically significant correlations. No model performs well across all properties. Some pretrained models (IgLM, ProGen2, Chai-1, ESM2, ISM, and IgFold) show significant zero-shot correlations for select properties. Fine-tuning with ≥10 2 points helps for thermostability, aggregation, and binding though polyreactivity and pharmacokinetics lack sufficient data. With enough data (10 3 points), a fine-tuned one-hot encoding model matches billion-parameter models. Training data matters more than architecture, and intrinsic properties (thermostability) are predicted more readily than extrinsic ones (immunogenicity and pharmacokinetics). Partial correlation analysis shows that germline distance accounts for ∼37% of language models' apparent predictive power.
页面未能读取全文,以上为出版方登记的摘要;方法与图表见原文
公开基准测试30个蛋白AI模型对400万抗体可开发性的预测,结果显示多数模型依赖与天然germline抗体的相似性等捷径而非生物物理信号。
来源:Patterns · cell.com