Data
GDPa5: Predicting VHH-Fc Developability from Large-Scale IgG Data
GDPa5 is a public, 160-member VHH-Fc developability dataset profiled across 10 assays. Models trained on 559 IgGs predict VHH-Fc developability better than models trained on VHH-Fc data alone.
The VHH-Fc scaffold is an emerging therapeutic modality, but until now no large-scale, standardized public developability dataset existed for it. Format-specific training data are scarce for new scaffolds, and no one had systematically tested whether IgG-trained models carry over.
Introducing GDPa5
GDPa5 is a 160-member VHH-Fc library profiled across 10 biophysical assays on the PROPHET-Ab platform, released alongside a bioRxiv preprint.
Key findings
- IgG models generalize: cross-format models trained on 559 IgGs outperformed models trained on GDPa5 alone. The advantage comes from the scale of standardized IgG data, not from the format.
- Best-predicted properties: heparin binding (HAC, ρ = 0.82), hydrophobicity (HIC, ρ = 0.63), and self-association (AC-SINS, ρ = 0.62), all largely governed by antibody surface properties.
- Tabular neural networks (TabICLv2, TabPFN v2.5), applied to antibody developability for the first time, beat conventional approaches.
- Tiered assays help: adding measured HIC and HAC as model inputs improved polyreactivity prediction (PR-CHO, Δρ = +0.10).
The takeaway: IgG-trained models are a practical, data-efficient starting point for emerging antibody formats.