D3AI-Spike

D3AI-Spike predicts binding affinity changes between multiple amino acid mutations in the SARS-CoV-2 spike receptor binding domain (RBD) and human angiotensin-converting enzyme 2 (hACE2) to assess effects of variants on receptor binding.

Key Features:

• Deep learning models: Implements convolutional neural networks (CNNs) and a CNN-RNN hybrid to predict RBD–hACE2 binding affinity.
• Multi-mutation modeling: Predicts effects of multiple amino acid substitutions and deletions in the SARS-CoV-2 spike RBD on hACE2 binding.
• Combinatorial capacity: Handles the large combinatorial space of potential RBD mutations arising from random genetic variation.
• Performance: Demonstrated a concordance index of approximately 0.8 against experimental data.
• Validation example: Predicted average affinity scores of 0.483 for the wild-type RBD and 0.438 for the B.1.640.2 (IHU) variant, corresponding to experimental dissociation constants (K_aff) of 5.39 ± 0.38 × 10^7 L/mol (wild type) and 1.02 ± 0.47 × 10^7 L/mol (IHU).
• Output characteristics: Provides rapid and quantitative affinity predictions for mutated spike RBD sequences.

Scientific Applications:

• Affinity prediction: Predicts how specific RBD amino acid changes alter binding affinity to hACE2.
• Variant assessment: Assesses variants such as B.1.640.2 (IHU) including substitutions N501Y and E484K and multiple deletions in the spike protein.
• Transmission and pathogenicity inference: Informs interpretation of potential impacts of RBD mutations on viral transmission dynamics and pathogenicity by quantifying receptor binding changes.
• Experimental comparison: Enables comparison of predicted affinity scores with experimental dissociation constants (K_aff) for validation of predictions.

Methodology:

Predictions are generated using convolutional neural networks (CNNs) and CNN-RNN hybrid deep learning models, with evaluation reported by concordance index and comparison to experimental dissociation constants (K_aff).