Source data and model checkpoints for scFvCAR-GPT, a scaffold-aware HER2 scFv-CAR design framework

This record contains the processed data, model checkpoints, evaluation outputs and figure source data supporting the scFvCAR-GPT study. scFvCAR-GPT is a conditional generative framework developed for HER2-targeted scFv-CAR design, integrating CDRH3 sequence generation with HER2 enrichment, scFv scaffold compatibility, Fv developability constraints and CAR-context sequence evaluation. The archive includes the HER2 feedback classifier checkpoint, the final scFvCAR-GPT generator checkpoint, tokenization metadata, training logs, formal 10,000-sequence evaluation outputs, benchmark summary tables, final top-three scFv-CAR amino-acid and CDS sequences, CAR domain annotations, source data for the main figures and structure-prediction outputs for the selected HER2 scFv-CAR candidates. Raw public database downloads are not redistributed in this record. Natural antibody sequence data were obtained from the Observed Antibody Space database, and public antibody structural records were obtained from SAbDab and the RCSB Protein Data Bank. The archive provides source information for these reused public datasets and includes only the processed outputs generated in this study. Associated code repository: https://github.com/tangaode/scFvCAR-GPT