ProVis - Webinterfaces für generative neuronale Netze

A modular, web-based platform for generating, analyzing, and visualizing synthetic protein sequences and structures using:

• the ZymCTRL model for sequence generation
• the Boltz-2 model for 3D structure prediction

Developed in collaboration with the Self-Organizing Systems (SOS) Lab, TU Darmstadt.

Overview

This project provides an intuitive graphical user interface (GUI) for:

• Generating novel protein sequences using generative language models
• Predicting 3D protein structures
• Visualizing molecular structures interactively in the browser
• Managing and exporting complete generation sessions
• Unlike existing workflows that rely on single-model pipelines and technical scripts, this platform integrates sequence generation, structure prediction, and visualization into one accessible web application.

The software is designed for research usage, particularly in synthetic biology and protein engineering, and is intended to be published as open source for the scientific community.

Screenshots

Sequence Generation Interface

Interface for generating synthetic protein sequences using the ZymCTRL model.

Structure Prediction & Visualization

Predicted 3D structures from the Boltz-2 model displayed in an interactive viewer.

Session Management

Export, Import and manage full protein design sessions directly from the UI.

Current Repository Layout

.
├── backend/        # Django API, worker runtime, model invocation
├── frontend/       # Next.js UI
├── scripts/        # runtime config + node/yarn setup helpers
├── config.json     # ports/bind host configuration
├── start.sh        # full bootstrap + production startup (screen)
└── stop.sh         # stop screen + kill Gunicorn fallback

Quick Start

For running the project on a server with sufficient GPU ressources use:

git clone https://github.com/nxlrth/DELIVERY_Webinterfaces-fuer-generative-neuronale-Netze.git
cd DELIVERY_Webinterfaces-fuer-generative-neuronale-Netze
./start.sh

Stop everything:

./stop.sh

Attach to the running screen session:

screen -r webinterface

Detach from screen: Ctrl+A, then D. Switch between screen windows: Ctrl+A, then Shift+2

What start.sh does

start.sh is not a lightweight launcher. It bootstraps and starts the full stack:

1. Sets up Node LTS + yarn (via nvm + corepack) if needed
2. Installs Miniconda (if missing) to ~/miniconda3
3. Creates/uses conda env backendProVis (Python 3.10)
4. Installs backend requirements
5. Runs python backend/worker/downloadModel.py
6. Runs Django migrations
7. Installs frontend dependencies
8. Rebuilds frontend (yarn build)
9. Starts backend + frontend in screen session webinterface

First run can take a long time (downloads + build + model setup).

Runtime Configuration

Runtime defaults are read from config.json:

{
  "backend": { "bind_host": "0.0.0.0", "bind_port": 8000 },
  "frontend": { "host": "0.0.0.0", "port": 3000 }
}

Additional Component Docs

• backend details: backend/README.md
• frontend details: frontend/README.md
• MUI theme guide: frontend/app/theme_README.md
• import/export internals: frontend/app/utils/ImportExportReadMe.md

Contributors:

Amin Al-Khatib
Oswin Bartels
Maxine Konz
Niklas Roth
David Schwab

License

This project is licensed under the MIT License. See the LICENSE file for details.

TU Darmstadt – Self-Organizing Systems Lab March 2026