mirror of
https://github.com/NousResearch/atropos.git
synced 2026-04-19 12:57:58 +00:00
64 lines
2.7 KiB
Markdown
64 lines
2.7 KiB
Markdown
# 🧬 LLM-Guided De Novo Protein Design Environment
|
||
|
||
**De novo protein binder design** is one of the hardest problems in bioengineering: you're tasked with inventing an amino acid sequence that folds into a 3D structure that binds to a given target protein. This environment lets **Large Language Models (LLMs)** tackle that problem using reinforcement learning (RL) — not by predicting sequences blindly, but by *learning to use the right tools in the right order* to produce functioning binders.
|
||
|
||
---
|
||
|
||
## 🤖 Why LLM-based RL Instead of Classic RL?
|
||
|
||
Classic RL works well for Atari, but it could never work for de novo protein binder design. Why?
|
||
|
||
- **Simulation is slow.** Each step—AlphaFold, RFdiffusion, ProteinMPNN—can take minutes. You don’t get to run millions of episodes like in classic RL.
|
||
- **State/action spaces are vast and weird.** Proteins are not 2D boards or pixel arrays. Designing them involves sequences, structures, config files, hotspots, and domain hacks.
|
||
- **Heuristics and intuition matter.** LLMs are pretrained on a *world model*—language, code, protein sequences, scientific papers. They come in with baked-in priors that help them reason, even under sparse rewards.
|
||
|
||
**Classic RL policy networks?** They’d need to learn everything from scratch, which is impossible!
|
||
|
||
---
|
||
|
||
## 🧪 The Protein Design Pipeline
|
||
|
||
Each episode consists of an LLM navigating a 4-step design pipeline, using state-of-the-art tools as function calls:
|
||
|
||
### Step 1: Target Sequence → Structure (`AlphaFold`)
|
||
- **Input:** Target protein sequence
|
||
- **Output:** 3D `.pdb` file (structure)
|
||
- **Reward:** Format validity
|
||
|
||
### Step 2: Target Structure → Binder Backbone (`RFdiffusion`)
|
||
- **Input:** `.pdb` file of target
|
||
- **Output:** `.pdb` backbone of potential binder
|
||
- **Reward:** Format validity
|
||
|
||
### Step 3: Backbone → Full Binder Sequence (`ProteinMPNN`)
|
||
- **Input:** Binder backbone
|
||
- **Output:** `.fasta` with side chains
|
||
- **Reward:** Format validity
|
||
|
||
### Step 4: Evaluate Binding (`AlphaFold-Multimer`)
|
||
- **Input:** Target + binder sequences
|
||
- **Output:** Complex structure prediction
|
||
- **Reward:**
|
||
- Format OK
|
||
- No steric clashes
|
||
- **Bonus:** Contact interface, binding affinity metrics (Not yet implemented)
|
||
|
||
---
|
||
|
||
## 🏆 Reward Function
|
||
|
||
The reward is cumulative:
|
||
- **+0.2**: Successfully generate output in correct format at each step
|
||
- **+0.0 to +1.0:** Structural reward based on complex validity smoothly interpolated on AlphaFold2 multimere confidence
|
||
- **+1**: High predicted binding affinity (Not yet implemented)
|
||
|
||
Sparse, but real. LLMs must *plan* tool use, not just spam actions.
|
||
|
||
---
|
||
|
||
## 🔧 Setup
|
||
|
||
Access to hosted NVIDIA APIs:
|
||
```env
|
||
NVIDIA_NIM_API_KEY="YOUR_API_KEY"
|
||
```
|