Ryan-Hu-Hu-Hu/De_novo_generation_platform

De Novo Protein Generation Platform

An automated pipeline for de novo enzyme design, integrated with a LINE chatbot frontend. Users send a PDB template code and a target reaction temperature via LINE; the platform generates novel protein sequences, evaluates them through a multi-stage screening pipeline, and returns the best candidate.

---

Architecture Overview

LINE user
   │ (LINE Messaging API)
   ▼
Wix Velo webhook  (https://graceng-ncku.com/_functions/lineWebhook)
   │ validates signature + forwards
   ▼
Flask server  (local HPC, exposed via cloudflared tunnel)
   │
   ▼
Pipeline Orchestrator
   ├─ PDB download + SMILES extraction
   ├─ P2Rank      → active-site pocket prediction
   ├─ RFdiffusion → backbone generation
   ├─ ProteinMPNN → sequence design
   ├─ CLEAN       → EC number prediction  (filter)
   ├─ SoDoPE      → solubility prediction (filter)
   ├─ UniKP       → kinetics prediction — kcat, Km  (filter)
   └─ Seq2Topt    → optimal temperature prediction  (selection)

---

Requirements

System

• Linux (tested on Ubuntu 22.04)
• NVIDIA GPU with CUDA 12+ (RTX 5090 recommended; ≥16 GB VRAM)
• Miniconda / Anaconda
• cloudflared — for exposing the local server via tunnel

Conda environments

Environment	Python	Used for	Environment file
lin	3.9	Flask chatbot, ProteinMPNN, SoDoPE, Seq2Topt, RFdiffusion	envs/lin.yml
Uni_test	3.10	UniKP kinetics prediction	envs/Uni_test.yml

Recommended — restore from the provided environment files (exact reproducibility):

conda env create -f envs/lin.yml
conda env create -f envs/Uni_test.yml

Alternative — create from scratch:

conda create -n lin python=3.9 -y
conda activate lin
pip install -r requirements_chatbot.txt

For the UniKP environment see UniKP/README.md.

Note: UniKP models (~1.2 GB) are pre-trained separately and stored in UniKP/models/ (gitignored). Run the training step below before starting the platform.

---

Installation

1. Clone the repository

git clone https://github.com/Ryan-Hu-Hu-Hu/De_novo_generation_platform.git
cd De_novo_generation_platform
git submodule update --init --recursive

2. Download external tools

P2Rank (pocket prediction)

mkdir -p tools/p2rank
# Download the latest release from https://github.com/rdk/p2rank/releases
# Extract and place the binary at:  tools/p2rank/prank
chmod +x tools/p2rank/prank

Seq2Topt (temperature prediction)

git clone https://github.com/SizheQiu/Seq2Topt tools/Seq2Topt
mkdir -p tools/large_model_pth
# Download model weights from the Seq2Topt GitHub releases and place at:
#   tools/large_model_pth/model_topt_window=3_r2=0.57.pth

CLEAN pretrained weights

Download the following files from Google Drive and place them into CLEAN/app/:

CLEAN/app/weights/split100.pth
CLEAN/app/weights/split70.pth
CLEAN/app/data/split100.csv
CLEAN/app/data/split70.csv

Also clone the ESM scripts required by CLEAN:

git clone https://github.com/facebookresearch/esm CLEAN/app/esm

3. Configure credentials

Create a .env file in the project root (never commit this file — it is gitignored):

# .env
LINE_CHANNEL_SECRET=your_line_channel_secret_here
LINE_CHANNEL_ACCESS_TOKEN=your_line_channel_access_token_here

Get these values from the LINE Developers Console.

4. Start the cloudflared tunnel

./cloudflared tunnel --url http://localhost:5000
# Note the generated HTTPS URL, e.g.: https://xxx.trycloudflare.com

Update FLASK_BACKEND_URL in your Wix http-functions.js to:

https://xxx.trycloudflare.com/callback

5. Pre-train UniKP ensemble models

UniKP requires 10 pre-trained models (5 for kcat, 5 for kcat/Km). This only needs to be run once (~5–10 minutes):

conda activate Uni_test
python UniKP/train_models.py

Models are saved to UniKP/models/ (≈1.2 GB total, gitignored). At inference time the platform loads them directly — no retraining on each query.

6. Run the platform

conda activate lin
python main.py

---

LINE Chatbot Usage

1. Add the bot as a LINE friend (QR code from LINE Developers Console)
2. Send any message to start
3. Enter a 4-letter PDB code (e.g. 1HEB)
4. Select a target reaction temperature (10–100 °C)
5. Wait for the pipeline — progress updates are sent automatically
6. Receive the best candidate sequence with EC number, solubility, and predicted Topt

Commands available at any time:

Command	Action
help / ?	Show help message
cancel / reset / stop	Abort current job and return to start

---

Wix Webhook Setup

The LINE Messaging API webhook URL should point to:

https://graceng-ncku.com/_functions/lineWebhook

The Wix http-functions.js validates the LINE HMAC-SHA256 signature and forwards valid requests to the cloudflared tunnel. See docs/LINE_setup_guide.md for the full setup walkthrough.

---

Project Structure

De_novo_generation_platform/
├── main.py                    # Entry point — Flask server + DB init
├── requirements_chatbot.txt   # Python dependencies
├── .env                       # Credentials (local only, gitignored)
│
├── pipeline/                  # Core pipeline package
│   ├── config.py              # All paths and tunable parameters
│   ├── orchestrator.py        # Full pipeline loop
│   ├── pdb_utils.py           # PDB download + SMILES extraction (3-strategy)
│   ├── active_site.py         # P2Rank pocket prediction
│   ├── rfdiffusion_runner.py  # RFdiffusion backbone generation
│   ├── proteinmpnn_runner.py  # ProteinMPNN sequence design
│   ├── clean_runner.py        # CLEAN EC number prediction
│   ├── sodope_runner.py       # SoDoPE solubility (SWI) prediction
│   ├── unikp_runner.py        # UniKP kinetics prediction
│   └── seq2topt_runner.py     # Seq2Topt optimal temperature prediction
│
├── chatbot/                   # LINE chatbot
│   ├── app.py                 # Flask /callback webhook endpoint
│   ├── line_handler.py        # State machine (IDLE→PDB→temp→processing)
│   └── job_manager.py         # SQLite job queue + background threads
│
├── tools/
│   ├── p2rank/                # P2Rank binary   (download separately)
│   ├── Seq2Topt/              # Seq2Topt source  (git clone separately)
│   └── large_model_pth/       # Model weights    (download separately)
│
├── RFdiffusion/               # (git submodule)
├── ProteinMPNN/               # (git submodule)
├── CLEAN/                     # (git submodule)
├── SoDoPE_paper_2020/         # (git submodule)
├── UniKP/                     # (git submodule)
│
├── envs/
│   ├── lin.yml                # Exact conda env for lin (main environment)
│   └── Uni_test.yml           # Exact conda env for Uni_test (UniKP)
│
├── docs/
│   └── LINE_setup_guide.md
└── tests/
    ├── test_seq2topt.py
    └── test_smiles_from_pdb.py

---

Configuration (pipeline/config.py)

Key parameters:

Parameter	Default	Description
NUM_DESIGNS	1	RFdiffusion backbone designs per iteration
NUM_SEQ_PER_TARGET	2	ProteinMPNN sequences per backbone
MAX_ITERATIONS	2	Max generation loops before returning best-so-far
MAX_GENERATED_LENGTH	None	Cap backbone length (e.g. 100 to save VRAM)
RFDIFFUSION_ENV	"lin"	Conda env for RFdiffusion
CLEAN_ENV	"lin"	Conda env for CLEAN
UNIKP_ENV	"Uni_test"	Conda env for UniKP
SEQ2TOPT_ENV	"lin"	Conda env for Seq2Topt

---

Security Notes

• .env is gitignored — LINE credentials are never committed to the repository
• jobs.db (SQLite with user job history) is gitignored
• data/ (all generated PDBs, sequences, prediction results) is gitignored
• Large model weights (*.pth, *.pt) are gitignored — download them separately per the instructions above

---

Acknowledgements

This platform integrates the following open-source tools:

• RFdiffusion — Watson et al., Nature 2023
• ProteinMPNN — Dauparas et al., Science 2022
• CLEAN — Yu et al., Science 2023
• SoDoPE — van Kempen et al., PLOS CB 2021
• UniKP — Yu et al., Nature Comm. 2023
• Seq2Topt — Qiu et al., ACS Synth. Biol. 2023
• P2Rank — Krivák & Hoksza, J. Cheminformatics 2018