FLAb2 (Fitness Landscape for Antibodies 2)
Most extensive public benchmark for therapeutic antibody design. Developability assay data for >4 million antibodies from 32 studies covering 7 properties: thermostability, expression, aggregation, binding affinity, pharmacokinetics, polyreactivity, and immunogenicity.
Composite
91.9
Experimental validation
Wet-lab confirmed
Stages
IND-enablingLead ID / ADMET
Modalities
antibodyprotein_sequence
Task types
property_predictiondevelopabilityfitness_prediction
Size
antibodies: 4,000,000
studies: 32
properties: 7
studies: 32
properties: 7
License
MIT
First release
2025-12
Last updated
2025-12
Official site
Leaderboard
→ leaderboard
Dataset
Code / GitHub
HuggingFace
→ HF
Paper
Fitness Landscape for Antibodies 2: Benchmarking Reveals That Protein AI Models Cannot Yet Consistently Predict Developability Properties · · 2025 · paper · doi:10.64898/2025.12.27.696706 · 8 citations
Flags
biologicsmulti_property
Experts
—
Groups
—
Hosted by
—
Related benchmarks
Rubric (7-criterion)
rigor
5
coverage
5
maintenance
4
adoption
4
quality
4
accessibility
5
industry_relevance
5
Notes
Key finding: current protein AI models cannot consistently predict antibody developability. Critical for biologics pipeline. Covers therapeutically relevant properties beyond just binding affinity.