ImmuneBuilder: Deep-Learning Models for Immune Protein Structure Prediction

Over 148,000 structural models for paired antibody sequences from the Observed Antibody Space (OAS) database, generated using the ABodyBuilder2 model. The dataset also includes the trained model weights for ABodyBuilder2, NanoBodyBuilder2, and TCRBuilder2, developed by Brennan Abanades at the University of Oxford. ImmuneBuilder models are reported to predict structures with state-of-the-art accuracy, such as a 2.86Å RMSD for antibody CDR-H3 loops, and are significantly faster than AlphaFold2.

Use Cases

• Training or benchmarking structure prediction models based on the provided model weights.
• Analyzing antibody structural diversity based on the over 148k predicted models from OAS sequences.
• Estimating prediction confidence for residues based on the ensemble prediction method described.
• Comparing prediction speed and accuracy against other tools like AlphaFold2 for immune receptors.

Strengths

• Includes over 148,000 predicted structural models for paired antibody sequences.
• Model weights are provided for three specific immune protein types: antibodies, nanobodies, and T-Cell receptors.
• Benchmark results are provided, such as a 2.86Å RMSD for antibody CDR-H3 loops, showing specific performance metrics.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the structural models dataset is known, but the specific file formats, size, and sample data are unavailable.
• Last update date is unknown; freshness unverified.

Provenance

Source

University of Oxford

Collection Method

Structural models generated by the ABodyBuilder2 deep-learning model applied to sequences from the Observed Antibody Space (OAS) database.