DeepH3

DeepH3 predicts probability distributions of inter-residue distances and orientation angles to model and evaluate CDR H3 loop structures in antibodies.

Key Features:

• Deep Residual Network Architecture: Employs a deep residual neural network tailored to model CDR H3 loop structure and to predict probability distributions over inter-residue distances and orientation angles.
• Input Data: Uses heavy- and light-chain antibody sequences as input to capture residue dependencies that influence CDR H3 conformations.
• Output and Geometric Potentials: Produces probability distributions over pairwise distances and orientation angles that are transformed into geometric potentials for structural evaluation.
• Decoy Discrimination with RosettaAntibody: Applies the derived geometric potentials to discriminate and rank decoy structures generated by RosettaAntibody.
• Benchmark Performance: On the Rosetta antibody benchmark dataset of 49 targets, identified better structures in 33 cases, equivalent structures in 6 cases, and worse structures in 10 cases, improving average RMSD by 32.1% (1.4 Å reduction) versus the Rosetta energy function.
• Orientation versus Distance: Found inter-residue orientations to be more effective than distances for distinguishing near-native CDR H3 loops.
• De novo Prediction Capability: Achieved an average de novo CDR H3 RMSD of 2.2 ± 1.1 Å on the same benchmark dataset.

Scientific Applications:

• Antibody Design and Engineering: Supports modeling and optimization of CDR H3 loops in therapeutic antibody development.
• Structural Evaluation and Decoy Selection: Improves discrimination and selection of near-native CDR H3 conformations from RosettaAntibody decoys.
• Benchmarking and Method Comparison: Serves as a comparative method to the Rosetta energy function for evaluating CDR H3 prediction approaches.
• De Novo Loop Prediction: Enables de novo prediction of novel CDR H3 loop structures with reported RMSD performance.

Methodology:

Uses a deep residual neural network that takes heavy- and light-chain sequences as input, predicts probability distributions over inter-residue distances and orientation angles, converts those distributions into geometric potentials, and uses the potentials to discriminate and rank RosettaAntibody-generated decoy structures; evaluation was performed by RMSD on the Rosetta antibody benchmark dataset (49 targets) against the Rosetta energy function.