| # PockLigGPT |
|
|
| PockLigGPT is a pocket-conditioned molecular generation framework based on GPT architectures and reinforcement learning (RL) for structure-based drug design. |
|
|
| --- |
|
|
| ## 🌐 Access & Resources |
|
|
| ### 🔬 Web Interface |
| 👉 https://pockliggpt.streamlit.app |
|
|
| Researchers and companies can: |
| - submit PDB structures |
| - request molecule generation |
| - request full computational studies |
|
|
| ### 💻 Code |
| 👉 https://github.com/pvaras8/PockLigGPT_official |
| |
| --- |
| |
| ## 🧬 Key Features |
| |
| - 🧠 **Protein-conditioned generation** using sequence + embeddings |
| - 🧪 **Docking-aware optimization** via reinforcement learning |
| - 🔬 **Structure-based drug design pipeline** end-to-end |
| - ⚡ **Scalable GPT-based architecture** for molecular generation |
| |
| --- |
| |
| ## 📦 Available Models |
| |
| ### Pretraining |
| - `ckpt_zink_M.pt` → ZINC pretraining |
| - `ckpt_zink_chembl.pt` → ZINC + ChEMBL pretraining |
| |
| ### Fine-tuned (CrossDocked) |
| - `ckpt_zink_chembl_cross_sequence_add.pt` → enhanced conditioning |
|
|
| --- |
|
|
| ## 🚀 What Can It Do? |
|
|
| PockLigGPT enables: |
|
|
| - Generation of novel molecules conditioned on protein pockets |
| - Optimization of binding affinity using RL + docking |
| - Exploration of chemical space under structural constraints |
| - End-to-end workflows for early-stage drug discovery |
|
|
| --- |
|
|
| ## 🧪 Example Workflow |
|
|
| 1. Load a pretrained checkpoint |
| 2. Provide protein pocket: |
| - amino acid sequence |
| - ProtT5 embeddings |
| - docking grid |
| 3. Run RL optimization |
| 4. Generate candidate molecules |
|
|
| --- |
|
|
| ## ⚙️ Technical Stack |
|
|
| - SELFIES molecular representation |
| - GPT-based autoregressive models |
| - ProtT5 protein embeddings |
| - AutoDock Vina (reward signal) |
| - Reinforcement learning (PPO-style) |
|
|
| --- |
|
|
| ## 💼 Commercial Use |
|
|
| PockLigGPT is available for research use. |
|
|
| For: |
| - industrial applications |
| - large-scale screening |
| - full structure-based optimization workflows |
|
|
| 👉 please contact via the web interface. |
|
|
| --- |
|
|
| ## ⚠️ Limitations |
|
|
| - RL requires docking setup (external tools) |
| - Computationally intensive for large-scale runs |
| - Requires protein preprocessing |
|
|
| --- |
|
|
| ## 📜 License |
|
|
| MIT License |
|
|
| --- |
|
|
| ## 📄 Citation |
|
|
| Coming soon |
