GYDE (Guide Your Design and Engineering)
A bio-design platform that integrates and runs multiple AI models, including AlphaFold2, Boltz, and Chai.
- Run AI models in an integrated environment: Directly execute over 10 cutting-edge protein AI models, including AlphaFold2, Boltz-1/2, Chai-1r, OpenFold3, ABodyBuilder2, Ibex, MOE Ab_workflow (structure prediction), ProteinMPNN, ThermoMPNN, LigandMPNN, BindCraft, RFDiffusion (protein design), MAFFT, Absolve, RaSP, Rosetta ฮฮG, TAP, and MolDesk (analysis) through the Slivka API in a web GUI. - Mol* based 3D structure visualization: Integrate interactive plotting, including automatic sequence-structure synchronization, MSA viewer (filtering, alignment, annotation), heatmap (saturation mutation data), scatter plot (predicted score vs. experimental value), sequence logo (residue conservation), and image viewer. - Antibody engineering workflow: A consistent pipeline of Absolve numbering โ ABodyBuilder2 structure prediction โ TAP/MolDesk developability analysis โ ProteinMPNN mutation design. - Flexible data model: Integrate sequence, structure, and experimental data using a column-based dataframe structure. Supports CSV, XLSX, FASTA, and PDB uploads, as well as connections to public databases like PDB, UniProt, and Pfam. - Session version control and collaboration: Supports collaboration among researchers with OAuth2 authentication, session access control, and version history. - Send-to-GYDE API: Programmatically import data from external systems via a curl endpoint (integrated pipeline automation). - HPC scheduler integration: Supports major HPC job schedulers such as LSF, Slurm, and SunGrid Engine through Slivka.
๐ป System Requirements
GYDE itself is a CPU-based web server. Running AI models such as AlphaFold2/RFDiffusion on the Slivka compute layer requires an NVIDIA GPU (requirements vary by model โ AlphaFold2 16GB+, ProteinMPNN 8GB+, RFDiffusion 16GB+).
GYDE core is approximately 500MB. With the Slivka compute tools and AI model weights included, the storage requirement is tens to hundreds of GB (AlphaFold2 database is separate, approximately 2.5TB, when deployed locally).
โก Installation
### 4-1. Quick Start (Docker, Recommended)
```bash
git clone https://github.com/proteinverse/gyde.git
cd gyde
export SLIVKA_DATA_DIR=/path/to/slivka/data
docker compose --profile setup run slivka-bio-installer
docker compose up gyde-server
# Access at http://localhost:3030
```
### 4-2. Local Development Environment
```bash
# Prerequisites: Node.js v14+, MongoDB v4.4+, Git
git clone https://github.com/proteinverse/gyde.git
# Frontend
cd gyde/gyde-frontend && npm install && npm start
# http://localhost:3000
# Backend (in a separate terminal)
cd gyde/gydesrv && npm install
export GYDE_SLIVKA_URL=http://your-slivka-server:4040
export GYDE_MONGO_CONNECTION=mongodb://localhost:27017/
npm start
```
### 4-3. Production Deployment
```bash
cd gyde-frontend && npm run build
cd ../gydesrv
export GYDE_HOST=0.0.0.0
export GYDE_PORT=3030
export GYDE_SLIVKA_URL=http://your-slivka-server:4040
export GYDE_MONGO_CONNECTION=mongodb://localhost:27017/
node index.js
```
### Key Environment Variables
| Variable | Description | Default Value |
|------|------|--------|
| `GYDE_HOST` | Binding address | 127.0.0.1 |
| `GYDE_PORT` | Port | 3030 |
| `GYDE_SLIVKA_URL` | Slivka compute API endpoint | **Required** |
| `GYDE_MONGO_CONNECTION` | MongoDB connection URI | mongodb://localhost/ |
| `GYDE_DB_NAME` | Database name | gydedb_prd |
| `GYDE_CA_BUNDLE` | Path to corporate SSL certificate | (None) |๐งฌ Bio Use Cases
Large-Scale Filtering of Protein-Protein Interactions (PPI)
From proteomics data, 1,381 interaction candidates were structurally predicted using AlphaFold2-multimer and filtered based on pTM/ipTM scores within GYDE. Interface residues were visually inspected in the Mol viewer to prioritize experimental validation.
Benchmarking of Structure Prediction Methodologies
Boltz-1 and Chai-1r co-folding results were directly compared within the same session. Prediction quality metrics, regions of structural disagreement, and confidence scores were visually inspected to select the optimal model for each project.
Antibody Engineering (Rational Design)
From B-cell repertoire sequences, antibody numbering, structure prediction, CDR variant heatmap visualization, developability (TAP/MolDesk) assessment, and experimental picklist generation were performed within a single platform. A case study demonstrating the design optimization of an Anti-PD-1 antibody using ProteinMPNN is presented.
Enzyme Optimization
A ProteinMPNN-based saturation mutagenesis workflow was applied to HyperTEV protease to computationally reproduce a previous study that achieved approximately a 20-fold improvement in catalytic efficiency compared to the wild type.
De Novo Binder Design
LRRC15 target AAV capsid retargeting was designed using BindCraft and RFDiffusion. Experimental binding/VLP yield data and predicted structures were integrated and analyzed in GYDE.
Collaborative Iterative Design
Through session version control and access management, computational biologists, experimental scientists, and project leaders can share the same design session and perform iterative design-experiment cycles.
๐ Update Notes
No update notes yet.
๐งช Related Code of Life
No related Code of Life posts yet.