Energy metric prediction for double insertion mutants via the RoseNet deep learning framework

Area of Science:

• Computational biology
• Bioinformatics
• Protein engineering

Background:

• Understanding protein mutation impacts is crucial.
• Previous work introduced the RoseNet neural network architecture.
• RoseNet predicts energy metrics for protein mutations.

Purpose of the Study:

• Evaluate RoseNet's prediction accuracy for double amino acid insertions/deletions (InDels).
• Analyze how protein domain features influence RoseNet's performance.
• Expand RoseNet's applicability to new proteins and mutation types.

Main Methods:

• Trained RoseNet models on benchmark datasets with exhaustive and random double InDel mutations.
• Evaluated model performance across six proteins.
• Analyzed the impact of secondary structures (alpha-helix, beta-sheet) and solvent accessible surface area (SASA) on predictions.

Main Results:

• RoseNet demonstrates superior generalization to unseen residue combinations compared to unseen insertion positions.
• Predictions are more accurate when InDels occur in beta-sheets versus alpha-helices.
• Higher performance is observed when InDels are in regions with high SASA.

Conclusions:

• RoseNet is a proficient tool for predicting energy metrics of proteins with double InDels.
• Protein structural features significantly impact RoseNet's prediction accuracy.
• Further research can refine RoseNet for specific structural contexts.