DDI_single (Domain Distance Inference, single-sequence) is a domain assembly method for multi-domain proteins that relies solely on the amino acid sequence. It uses the protein language model ESM-1b to extract sequence features and an improved Gated Cross-Attention module to predict inter-domain residue pair distances and orientations, enabling correct spatial assembly of structural domains.
Pipeline
DDI_single Pipeline
The pipeline takes a protein sequence and domain definitions as input. ESM-1b extracts per-residue embeddings, which are split by domain and refined through 4 Gated Cross-Attention layers with 4 recycle iterations. The model predicts inter-domain distance and orientation distributions (d, ω, θ, φ), which are combined with intra-domain constraints and fed into Rosetta for 3D structure assembly.
Key Features
Feature
Description
Single-sequence input
No MSA or homologous templates required; uses ESM-1b for sequence embeddings
Gated cross-attention
Bidirectional cross-attention with gating to capture inter-domain residue interactions
Recycle refinement
4 rounds of iterative feature extraction to strengthen inter-domain relationship modeling
Multi-task prediction
Outputs distance (d) and dihedral angle (ω, θ, φ) distributions for inter-domain residue pairs
Rosetta integration
Predictions serve as spatial constraints for domain assembly and energy minimization
Performance
Inter-domain residue pair distance prediction accuracy (given known domain definitions):
Inter-domain constraints predicted by DDI_single can be combined with Rosetta for energy-minimization-based domain assembly:
Known domain conformations — Extract intra-domain constraints from PDB, predict inter-domain constraints with DDI_single, concatenate, and fold with Rosetta
Unknown domain conformations — Predict intra-domain constraints with trRosettaX_single first, then replace/enhance inter-domain constraints with DDI_single, and assemble with Rosetta
Model Architecture
Module
Description
Single_DDI_sESM
Wraps ESM-1b and the domain assembly network
GAU
Gated cross-attention unit with Pre-LN and GELU feed-forward layers
feature_extractor
Bidirectional cross-attention updating both domain representations
inter_head
Predicts d / ω / θ / φ from attention maps
Hyperparameters (config.py)
Parameter
Value
GAU layers
4
Attention heads
16
Head channels (GAU)
256
Prediction heads
40
Prediction head channels
128
Recycle iterations
4
Project Structure
DDI_single/
├── config.py # Model and training hyperparameters
├── model.py # Network architecture (GAU, Single_DDI, Single_DDI_sESM)
├── predict.py # Inference script
├── pipeline.png # Full pipeline diagram
├── DDI_single.pth # Model weights
├── README.md # English documentation
└── README_ch.md # Chinese documentation
Citation
If you use DDI_single in your research, please cite:
bibtex
1@article{ddi_single,
2 title = {DDI_single: Single-sequence-based Domain Assembly via Gated Cross-Attention},
3 author = {Shengyi Zong},
4 journal = {},
5 year = {}
6}
