Geometric potentials from deep learning improve prediction of CDR H3 loop structures.

Area of Science:

• Structural biology
• Computational biology
• Bioinformatics

Background:

• Antibody structure is conserved, with six variable loops in the complementarity-determining region (CDR).
• Five CDR loops have predictable folds, but the CDR H3 loop exhibits high conformational diversity, posing a prediction challenge.
• Deep neural networks excel at learning complex protein structure patterns.

Purpose of the Study:

• To develop DeepH3, a deep residual neural network for predicting inter-residue distances and orientations in antibody heavy and light chains.
• To utilize DeepH3's output to refine antibody structures and predict novel CDR H3 loop conformations de novo.

Main Methods:

• A deep residual neural network (DeepH3) was trained to predict residue-pair distances and orientations from antibody sequences.
• The network's output was converted into geometric potentials.
• These potentials were used with RosettaAntibody to discriminate between decoy structures and predict CDR H3 loop structures.

Main Results:

• DeepH3-predicted potentials outperformed the standard Rosetta energy function in identifying near-native CDR H3 loop structures on a benchmark dataset.
• The average root-mean-squared distance (RMSD) of predictions was improved by 32.1% (1.4 Å).
• DeepH3 achieved an average de novo CDR H3 loop prediction RMSD of 2.2 ± 1.1 Å.

Conclusions:

• DeepH3 effectively predicts CDR H3 loop structures and refines antibody models.
• Inter-residue orientations proved more critical than distances for discriminating near-native CDR H3 loops.
• The DeepH3 tool is publicly available for further research.