HZAU-China: Ab-Bind

This record contains the dataset used in our iGEM project on antibody mutation effect prediction. Our goal is to develop a transferable machine-learning pipeline to evaluate the impact of single mutations on the binding affinity and structural quality of VHH nanobodies targeting VEGF165.

The dataset uploaded here (“AbBind.csv”) is derived from the publicly available AB-Bind benchmark of antibody–antigen binding free-energy changes (Sirin S. et al., Protein Science 2016;25:393–409; doi:10.1002/pro.2829). It consists of curated single-mutation entries with measured ΔΔG values and associated structural information for conventional antibodies (heavy + light chains). This curated file was cleaned and reformatted for use in our pipeline but the underlying numerical values remain identical to the original AB-Bind dataset.

We used these data to train a low-rank bilinear energy model with ESM-1b sequence embeddings for ΔΔG prediction, and a parallel random-forest model with MolProbity Clashscore as labels for structural quality prediction. During inference, we adapted the models to our VEGF165 VHH nanobody, which contains two highly homologous heavy-chain segments and lacks a light chain, by zeroing the light-chain input and feeding only the first heavy-chain segment.

The full description of our project, methods, and results is available on our iGEM wiki page:
https://2025.igem.wiki/hzau-china/index.html