Ginkgo Datapoints Antibody Developability Competition outcomes: limited model performance and a call for data...

Area of Science:

• Biotechnology
• Computational Biology
• Drug Discovery

Background:

• Antibody developability is crucial for successful drug development.
• Predicting antibody developability computationally is an active area of research.
• Existing datasets for model training are often small and heterogeneous.

Purpose of the Study:

• To benchmark computational models for predicting antibody developability.
• To assess the generalization capabilities of these models on unseen data.
• To identify limitations and future directions for antibody developability prediction.

Main Methods:

• Conducted the blinded Ginkgo Datapoints Antibody Developability (AbDev) Competition.
• Evaluated predictors on five biophysical properties: hydrophobicity, thermostability, self-association, expression titer, and polyreactivity.
• Used a public training set (246 antibodies) and a held-out test set (80 antibodies).

Main Results:

• Top performance varied by property, with hydrophobicity prediction being strongest (Spearman's ρ = 0.708).
• Cross-validation scores on training data consistently outperformed test set performance, indicating overfitting.
• Limited out-of-distribution generalization was observed across predictors.

Conclusions:

• Current antibody developability models show limited generalization due to small, heterogeneous datasets.
• Larger, standardized, and diverse experimental datasets with harmonized protocols are essential for robust model training and validation.
• Future antibody discovery campaigns require reliable, generalizable developability prediction models.