PhysChemDiff: Physicochemical Property-Guided Protein Generation Using Latent Diffusion Models

This repository is the official implementation of PhysChemDiff: Physicochemical Property-Guided Protein Generation Using Latent Diffusion Models.

Note: The training code and protein sequences will be released after the paper is accepted.

---

Requirements

To install the required dependencies, use the following command:

pip install -r requirements.txt

---

Data Preparation

We provide relevant datasets with protein sequence length of 128 .

1. swiss_onehot_and_feature_128.pt: Contains the one-hot encoded protein sequences and the corresponding physicochemical properties of the protein sequences.
2. swiss_text_128.csv: Contains the protein sequences along with their textual descriptions.

You can download these datasets from the Hugging Face Model Page.And then place them in the data folder of this repository.

---

Pre-trained Models

The pre-trained models for each module of the PhysChemDiff model can be downloaded from the Hugging Face Model Page and you can place them in the pretrained_models folder of this repository.

---

Inference

The instructions for running inference are available in the inference.py file. You can run the following command to perform inference:

python inference.py --use_dataset_for_features <True|False> --use_dataset_for_text <True|False> --num_samples <number> --seq_length <length> --custom_text <text> --num_timesteps <number>

---

Notes:

• Ensure you have the correct environment setup as described in the requirements section.
• Refer to the inference.py file for additional arguments and detailed usage.

---