---
pipeline_tag: text-generation
---

<h1 align="center">
  TD3B: Transition-Directed Discrete Diffusion for Allosteric Binder Generation
</h1>

<div align="center Murphy">
 <a href="https://huggingface.co/papers/2605.09810"><img src="https://img.shields.io/badge/Paper-2605.09810-red?style=for-the-badge" alt="Paper"/></a>
 <a href="https://arxiv.org/abs/2605.09810"><img src="https://img.shields.io/badge/Arxiv-2605.09810-red?style=for-the-badge&logo=Arxiv" alt="arXiv"/></a>
</div>

![Screenshot 2026-05-10 at 6.52.29 PM](https://cdn-uploads.huggingface.co/production/uploads/64cd5b3f0494187a9e8b7c69/38-6PQ83pPraF7KAs3g3E.png)

TD3B is a sequence-based generative framework that designs peptide binders with specified agonist or antagonist behavior. It combines a target-aware Direction Oracle, a soft binding-affinity gate, and amortized fine-tuning of a pre-trained discrete diffusion model (MDLM).

The model was presented in [TD3B: Transition-Directed Discrete Diffusion for Allosteric Binder Generation](https://huggingface.co/papers/2605.09810).

## Installation

```bash
conda env create -f env.yml
conda activate td3b
pip install -e .
```

## Data and Checkpoints

Download the pretrained checkpoints and data from [Google Drive (TBA)](placeholder_link).

Place the files as follows:

```
TD3B/
├── checkpoints/
│   ├── pretrained.ckpt          # Pre-trained MDLM weights
│   ├── td3b.ckpt                # Fine-tuned TD3B model
│   └── direction_oracle.pt      # Direction Oracle weights
├── data/
│   ├── train.csv                # Training set (target-binder pairs)
│   └── test.csv                 # Test set
├── scoring/functions/classifiers/
│   ├── binding-affinity.pt
│   ├── hemolysis-xgboost.json
│   ├── nonfouling-xgboost.json
│   ├── permeability-xgboost.json
│   └── solubility-xgboost.json
└── tokenizer/
    ├── new_vocab.txt
    └── new_splits.txt
```

## Code Structure

```
TD3B/
├── inference.py                 # Generate binders (main inference entry point)
├── finetune_multi_target.py     # Multi-target TD3B training
├── finetune_utils.py            # Training utilities
├── launch_multi_target.sh       # Training launcher script
├── diffusion.py                 # MDLM backbone (TR2-D2)
├── roformer.py                  # RoFormer wrapper
├── noise_schedule.py            # Noise schedules
├── peptide_mcts.py              # MCTS tree search
├── td3b/
│   ├── direction_oracle.py      # Direction Oracle (f_φ)
│   ├── td3b_scoring.py          # Gated reward R = g_ψ · σ(d*·(f_φ−0.5)/τ)
│   ├── td3b_losses.py           # L_WDCE + λ·L_ctr + β·L_KL
│   ├── td3b_mcts.py             # TD3B-extended MCTS
│   ├── td3b_finetune.py         # Training loop
│   └── data_utils.py            # Data loading utilities
├── scoring/                     # Affinity predictor (g_ψ) and property classifiers
├── baselines/                   # CG, SMC, TDS, PepTune, Unguided baselines
├── tokenizer/                   # SMILES tokenizer (vocab + splits)
├── configs/                     # Model and training configs
└── utils/                       # Misc utilities
```

## Inference

Generate agonist/antagonist binders for target proteins:

```bash
python inference.py \
    --ckpt_path checkpoints/td3b.ckpt \
    --val_csv data/test.csv \
    --save_path results/ \
    --seed 42 \
    --num_pool 32 \
    --val_samples_per_target 8 \
    --resample_alpha 0.1
```

This generates 32 candidates per (target, direction), scores them with the Direction Oracle and affinity predictor, applies Algorithm 2 weighted resampling, and saves only valid peptide samples.

Output: `results/td3b_results_seed42.csv` with columns: target, sequence, direction, affinity, gated_reward, direction_oracle, direction_accuracy.

## Training

### Multi-target TD3B

1. Edit `launch_multi_target.sh` — set paths to checkpoints, data, and oracle:

```bash
BASE_PATH="/path/to/TD3B"
PRETRAINED_CHECKPOINT="${BASE_PATH}/checkpoints/pretrained.ckpt"
TRAIN_CSV="${BASE_PATH}/data/train.csv"
ORACLE_CKPT="${BASE_PATH}/checkpoints/direction_oracle.pt"
```

2. Launch training:

```bash
bash launch_multi_target.sh
```

Key hyperparameters (in `launch_multi_target.sh`):
- `CONTRASTIVE_WEIGHT=0.1` — λ for L_ctr
- `KL_BETA=0.1` — β for L_KL
- `SIGMOID_TEMPERATURE=0.1` — τ for gated reward
- `NUM_ITER=20` — MCTS iterations per round
- `NUM_CHILDREN=16` — Children per MCTS expansion

### Baselines

Run baseline methods (CG, SMC, TDS, PepTune, Unguided):

```bash
cd baselines/
bash run.sh --baseline cg --device cuda:0
bash run.sh --baseline smc --device cuda:0
bash run.sh --baseline tds --device cuda:0
```

## Citation

```bibtex
@inproceedings{
  cao2026td3b,
  title={TD3B: Transition-Directed Discrete Diffusion for Allosteric Binder Generation},
  author={Hanqun Cao and Aastha Pal and Sophia Tang and Yinuo Zhang and Jingjie Zhang and Pheng Ann Heng and Pranam Chatterjee},
  booktitle={Forty-third International Conference on Machine Learning},
  year={2026},
  url={https://huggingface.co/papers/2605.09810}
}
```