How hypnopump/minifold works
Core Components
a) Data Processing Pipeline
• Input Layer: Raw protein sequences and PSSM (Position Specific Scoring Matrix) data
• Preprocessing Module:
• Protein sequence filtering (< 200 amino acids)
• Angle calculation from coordinates
• Data preparation and formatting
b) Model Architecture
• Two parallel neural network branches:
1. Angle Prediction Branch:
• 1D ResNet
• Input: LxN tensor (L≤200, N=42 channels)
• Output: 4 values (sin/cos of Phi/Psi angles)
2. Distance Prediction Branch:
• 2D ResNet
• Input: LxLxN tensor (L≤200, N=41 channels)
• Output: LxLx7 tensor (6 distance classes + 1 padding class)
c) Post-processing Pipeline
• RR Format Conversion
• 3D Structure Modeling Interface
[Raw Input Data] → [Preprocessing Pipeline]
• ↓
• ┌────────────────┴───────────────┐
• ↓ ↓
[Angle Prediction Branch] [Distance Prediction Branch]
• ↓ ↓
• ┬───────────────┘
• ↓
• [Post-processing Pipeline]
• ↓
• [3D Structure Output]
Important Details
a) Input Specifications:
• Protein sequences (< 200 amino acids)
• PSSM matrices
• Van der Waals radius
• Surface accessibility data
b) Model Architecture Details:
• ResNet block structure
• Input/Output dimensions
• Key activation functions (ELU, Softmax)
c) Data Flow Annotations:
• Data transformations
• File formats (RR format)
Implementation Notes
• Parallel nature of angle and distance predictions
• Where model weights (*.h5 files) are used
• Evaluation points and metrics collection
• Interfaces with external tools (e.g., Rosetta)