AWS and Johns Hopkins announce groundbreaking database for AI/ML antibody design
New database with 50 seed antibodies and 42 antigens aims to accelerate therapeutic antibody discovery.
Amazon Web Services (AWS) and the Gray Lab at Johns Hopkins Whiting School of Engineering have announced the launch of the Antibody Developability Benchmark. This new public database is designed to evaluate AI and machine learning models for therapeutic antibody design. It addresses a critical bottleneck in the field: the lack of large, diverse, and standardized experimental data against which to test predictive algorithms. The benchmark features 50 seed antibodies in four structural formats targeting 42 different antigens, and it measures six key developability properties that determine if an antibody can be manufactured and used safely as a drug.
The dataset is engineered to include variants with both favorable and unfavorable outcomes, all validated through wet lab experiments. A key feature is its support for zero-shot learning, meaning AI models can be evaluated on it without having been trained on the data first, which enhances confidence in their real-world performance. The collaboration aims to overcome the limitations of existing public datasets, which are often biased toward single formats or clinically advanced antibodies, thereby hindering the development of trustworthy in-silico tools. The team plans to expand the benchmark with additional models and properties to foster continuous improvement in AI-driven drug discovery.
- Database includes 50 seed antibodies targeting 42 antigens across four structural formats.
- Measures six critical developability traits (e.g., solubility, specificity) with wet-lab validated data.
- Supports zero-shot evaluation to reliably benchmark AI models like protein language models (pLMs).
Why It Matters
Provides a standardized foundation to accelerate AI-driven drug discovery, potentially reducing development timelines and costs for new therapeutics.