VariPred is a novel and simple framework that leverages the power of pre-trained protein language models to predict variant pathogenicity. It was introduced in the 2023 preprint of the paper "VariPred: Enhancing Pathogenicity Prediction of Missense Variants Using Protein Language Models". Note: VariPred is not for the DMS data analysis

---

Recommand to implement this workflow on a Linux system.

For people fail to install openfold packages, please try pip install fair-esm.

---

Update on 2023.11.20:

• We uploaded a new requirements.txt file and also recommended you to create an environment with python=3.9. After install the packages with requirements.txt file, please install torch and fair-esm manually with the instruction given below.

Update on 2023.8.3:

• We fixed bugs in VariPred_test.txt and VariPred_train.txt and also in the embedding extraction.

Update on 2023.5.5:

• Embeddings can be generated in batches, meaning that you can obtain them much more quickly by adjusting the batch_size_for_embed_gen parameter in the config file. For example, if you set batch_size_for_embed_gen=16 and use a GPU with larger memory, you can generate embeddings for 16 data points simultaneously.

Step 1: Installation

VariPred was trained using the packages included in the file "requirements.txt". Please install requirements first.

$ git clone [email protected]:wlin16/VariPred.git

# if you failed to clone the repo to your machine, try:
# git clone https://github.com/wlin16/VariPred.git

$ cd VariPred
$ conda create -n varipred python=3.9 -y
$ conda activate varipred
$ pip install -r requirements.txt
$ pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu118
$ pip install fair-esm

---

Step 2: Prepare dataset

The input for VariPred must contains

1. mutation info (e.g. NP_001035957.1_L847P), a.k.a "target_id"

2. mutation position, a.k.a "aa_index"

3. wild-type amino acid, a.k.a "wt_aa"

4. mutant-type amino acid, a.k.a "mt_aa"

5. wild-type sequence, a.k.a "wt_seq"

6. mutant-type sequence, a.k.a "mt_seq"

   Label is optional for a prediction purpose but neccessary for preparing the training and test sets if re-train the model.

   Please name the column of the dataframe according to the example files under the "example" folder.

• For data with UniProt IDs, please fetch the wildtype with https://www.uniprot.org/id-mapping

• For data with RefSeq IDs (NP ids), please use the "prepare_dataset.py" script under the Dataset_preparation folder:

  • Inside Dataset_preparation folder, we prepared an example "target.txt". The first parameter is the name of the file.

    $ cd Dataset_preparation
    $ python3 prepare_dataset.py target

• Now, we have a dataframe named as "target.csv" under example/dataset/

---

Step 3: Train the model

Note: The weight of the trained model mentioned in the publication has been given under the directory of VariPred/model/model.ckpt. Running train_VariPred.sh script to re-train the model will replace the given weight.

If there is no need to customize the model for a specific task or evaluate the performance of VariPred, but only to use VariPred to make clinical impact predictions for variants, please skip Step 3 and Step 4 and proceed directly to Step 5.

We recommand you have an at least 12GB GPU, e.g. NVIDIA GeForce 1080Ti

"train.csv", "test.csv" are the example files as the training and test sets to re-train the model. "target.csv" is the example file for a simple prediction purpose.

• To prepare a training set and a test set, you can prepare the datasets with the following codes

  $ python3 prepare_dataset.py VariPred_train
  $ python3 prepare_dataset.py VariPred_test

• Now, we have two dataframe named as "VariPred_train.csv" and "VariPred_test.csv" under the directory of example/dataset

---

Setp 4: Fetch the embeddings and train the model

1. If you would like to re-train the VariPred, the embedding representations need to be generated for both the training and test sets.

   $ cd ../VariPred

2. Replace the variables based on the paths where you stored your datasets in "train_VariPred.sh" script. Then run the script. This will give you the performance of the model (MCC and AUC-ROC scores).

$ ./train_VariPred.sh

---

Setp 5: Fetch the embeddings and predict the effects of variants

1. To predict the effects of variants by VariPred, please replace the variables based on the paths where you stored your data in "predict.sh" script. Then run the script. This will give you the clinical impact of each variants

   $ cd ../VariPred
   $ ./predict.sh