DenseCPD

DenseCPD predicts per-residue probabilities of the 20 natural amino acids from protein backbone structures using a deep neural network that models the three-dimensional density distribution of backbone atoms for applications in computational protein design.

Key Features:

• Per-residue amino-acid probabilities: Predicts the probability distribution over the 20 natural amino acids for each residue in a given protein structure.
• Deep neural network: Leverages a deep-learning neural network that incorporates the three-dimensional density distribution of protein backbone atoms.
• Cross-validation performance: Achieved 51.56% ± 0.20% accuracy in a 5-fold cross-validation on the training set.
• Independent test performance: Reported accuracies of 54.45% and 50.06% on two independent test sets.
• Improvement over prior methods: Demonstrates more than a 10% improvement over previous state-of-the-art methods on the independent test sets.
• Output strategies: Provides accumulative probability cutoff to narrow sequence search space and top-k predictions as an alternative search strategy.
• Compatibility with Rosetta redesign: Using the accumulative probability cutoff narrows the search space and yields higher sequence identity when redesigning proteins with Rosetta.

Scientific Applications:

• Computational protein design: Guides sequence selection for designing protein sequences for specified structural backbones.
• Search-space reduction: Uses accumulative probability cutoff to restrict candidate residues and reduce combinatorial sequence search space.
• Protein redesign with Rosetta: Improves sequence identity in Rosetta-based redesign workflows when using accumulative probability cutoff outputs.
• Sequence–structure modeling: Models the relationship between protein backbone geometry and residue identity probabilities for structure-informed sequence prediction.

Methodology:

Uses a deep neural network trained to predict per-residue probabilities of the 20 natural amino acids from the three-dimensional density distribution of protein backbone atoms, validated by 5-fold cross-validation and evaluated on two independent test sets, and exposes outputs usable via accumulative probability cutoff or top-k prediction strategies.