CycDiff-DPO

CycDiff-DPO uses preference-aligned diffusion to design target-specific macrocyclic peptides with enhanced membrane permeability, balancing binding competence and cell permeability in one framework.

---

Table of Contents

• Setup
• Quick Start
• Training from Scratch
• Preference Pairs Construction
• Permeability Predictor Training

---

Setup

Environment

conda env create -f env.yaml
conda activate CycDiff_DPO

Note: This environment uses Python 3.9.19.

Data Download

Download datasets and pre-trained models from Zenodo:

# Download checkpoints
wget https://zenodo.org/records/19429073/files/ckpts.tar.gz?download=1 -O ./ckpts.tar.gz
tar -xzf ./ckpts.tar.gz && rm ./ckpts.tar.gz

# Download train_valid dataset (for training DPO)
wget https://zenodo.org/records/19429955/files/train_valid.tar.gz?download=1 -O ./datasets/train_valid.tar.gz
tar -xzf ./datasets/train_valid.tar.gz -C ./datasets/ && rm ./datasets/train_valid.tar.gz

# Download LNR_CPSea dataset
wget https://zenodo.org/records/19429073/files/LNR_CPSea.tar.gz?download=1 -O ./datasets/LNR_CPSea.tar.gz
tar -xzf ./datasets/LNR_CPSea.tar.gz -C ./datasets/ && rm ./datasets/LNR_CPSea.tar.gz

# Download SciBERT model
wget https://zenodo.org/records/19429073/files/scibert_model.tar.gz?download=1 -O ./scibert_model.tar.gz
tar -xzf ./scibert_model.tar.gz && rm ./scibert_model.tar.gz

Pre-trained Weights

The following weights are included in this repository:

File	Description
./ckpts/base_model.ckpt	Base model from CP-Composer
./ckpts/autoencoder.pth	Pre-trained full-atom autoencoder (from PepGLAD)
./ckpts/dpo/epoch44_step513090.ckpt	DPO-fine-tuned model (final checkpoint)
./ckpts/xgboost_ensemble/	XGBoost ensemble for membrane permeability prediction
./datasets/train_valid/generated_pairs.pkl	Pre-generated DPO preference pairs

---

Quick Start

The full pipeline consists of three steps. Default: 5 samples per target on the LNR_CPSea test set.

Step 1: Generation

conda activate CycDiff_DPO
GPU=0 bash scripts/inference_forw.sh

Step 2: Filter

bash scripts/filter_success.sh ./results/LNR_CPSea/condition2_w5_5samples/results.jsonl

Step 3: Postprocessing

INPUT_DIR=./results/LNR_CPSea/condition2_w5_5samples/candidates
OUTPUT_DIR=./results/LNR_CPSea/condition2_w5_5samples/relaxed
NUM_CORES=10 bash scripts/batch_relax_good_results.sh

---

Training from Scratch

DPO training fine-tunes the pre-trained LDM to align with membrane permeability preferences. We provide the pre-generated preference pairs and the trained DPO model. To train from scratch:

conda activate CycDiff_DPO
GPU=0 bash scripts/train.sh

---

Preference Pairs Construction

Preference pairs are used to train the DPO model. We provide pre-generated pairs at ./datasets/train_valid/generated_pairs.pkl.

To regenerate pairs with a custom permeability predictor, run:

bash scripts/run_build_pairs_xgboost.sh

This requires the training PDB structures. Download and place them in:

./datasets/train_valid/pdbs/   # Reference PDB structures

---

Permeability Predictor Training

We provide the trained XGBoost ensemble at ./ckpts/xgboost_ensemble/, which includes:

• model_*.pkl — 10 individual XGBoost models
• scaler.pkl — feature scaler
• extractor.pkl — ECFP + descriptor feature extractor
• config.json — ensemble configuration

To retrain from scratch using Caco-2 permeability data at ./datasets/caco2/caco2_dedup.csv:

bash scripts/train_xgb.sh