AbAdapt

AbAdapt predicts antibody-antigen complex structures and scores docked poses derived from unbound homology models from antibody and antigen amino acid sequences.

Key Features:

• Input Handling: Accepts antibody and antigen amino acid sequences as inputs.
• 3D Structure Modeling: Models three-dimensional structures of antibodies and antigens from sequences, producing unbound homology models for docking.
• Epitope and Paratope Prediction: Predicts antigen epitopes and antibody paratopes to inform interaction analysis.
• Docking Simulations: Generates docking poses using the Piper and Hex docking engines.
• Machine-learning Optimization: Scores and ranks docked poses using machine-learning models trained on a diverse dataset.
• Validation: Validated by leave-one-out cross-validation on 622 antibody-antigen pairs and a holdout set of 100 unrelated pairs, identifying at least one 'Adequate' pose for 550/622 queries (88.4%).
• Performance Metrics: Reported median rank of 'Adequate' poses was 22 (IQR 5–77); repredicting epitopes with docking-derived features increased median ROC AUC from 0.679 to 0.720 in cross-validation and from 0.694 to 0.730 on the holdout set.

Scientific Applications:

• Antibody Design: Predicts interaction sites to support design of antibodies with desired binding properties.
• Vaccine Development: Identifies potential antigenic epitopes relevant to vaccine target selection.
• Therapeutic Discovery: Provides structural and scoring information to inform development of therapeutic antibodies.

Methodology:

From antibody and antigen sequences, AbAdapt builds 3D unbound homology models, predicts epitopes and paratopes, performs docking with Piper and Hex, scores and ranks poses using machine-learning models trained on a diverse dataset, and validates performance by leave-one-out cross-validation on 622 pairs and a 100-pair holdout set, with epitopes optionally repredicted using docking-derived features.