PheMART: Phenotypic Prediction of Missense Variants

This repository hosts the official implementation of PheMART, a method that predicts the phenotypic effects of missense variants (MV) via deep contrastive learning.
All the source data will be published on: https://doi.org/10.6084/m9.figshare.26036227 and https://doi.org/10.5281/zenodo.13138603.
We offer visualizations of high-confidence phenotypic predictions at: https://shiny.parse-health.org/PheMART/. These visualizations are categorized by phenotypes, genes and protein domains.

---

Table of Contents

• Introduction
• System Requirements
• Installation
• Input Data
• Usage
  • Running Inference with Pre-trained Model
  • Fine-Tuning or Training a New Model
• Expected Outputs
• Computation Steps

---

Introduction

PheMART is a computational framework designed for predict the clinical phenotypic effects of missense variants. Users can either:

• Use a pre-trained model for quick inference.
• Fine-tune or train the model using their own dataset.

This README provides step-by-step instructions for setting up, running, and interpreting the results.

---

System Requirements

To ensure smooth execution, we recommend the following system specifications:

• Operating System: Ubuntu 20.04
• Python Version: 3.7+
• GPU (for training): NVIDIA GPU with at least 16GB VRAM
• CUDA: 11.2+

---

Installation

1. Clone the Repository

git clone https://github.com/celehs/PheMART.git
cd PheMART

2. Create a Virtual Environment (Optional but Recommended)

conda create -n phemart_env python=3.7.4
conda activate phemart_env

3. Install Dependencies

pip install -r requirements.txt

---

Input Data

PheMART requires different input datasets depending on the use case.

1. Using the Pre-trained Model

• Required Input: A list of missense variants in CSV format and a file containing the variant embeddings with *Numpy array format. Each row represents the embedding vector of a variant.
• Example format (CSV):

  variants
  NM_002074.5(GNB1):c.230G>A (p.Gly77Asp)
  NM_022787.4(NMNAT1):c.205A>G (p.Met69Val)
  

2. Fine-Tuning or Training a New Model

• Variant Embeddings: Precomputed embeddings for missense variants.
• Phenotype Embeddings: Precomputed phenotype representation. The LLM phenotype embeddings have been provided and the EHR embedding will be provided upon request.
• Training Labels: Variant-phenotype annotations in CSV format.
  • Example format (CSV):

  variant_id,gene,phenotype_CUI
  NM_002074.5(GNB1):c.388G>A (p.Glu130Lys),GNB1,C3276355
  NM_003036.4(SKI):c.68A>C (p.Gln23Pro),SKI,C1321551
  

Note: We also provide scripts to preprocess user-provided patient-level data for generating EHR embeddings.

---

Usage

Running Inference with Pre-trained Model

To run inference using our pre-trained model:

python predict.py --file_snp_prediction variants.csv  --dirr_results_main  result/ --dirr_pretrained_model data/model_pretrained/

Arguments

• --file_snp_prediction: File containing the list of variants to predict (CSV).
• --dirr_results_main: Path to save the predictions.
• --dirr_pretrained_model: Path to the pre-trained model.

---

Fine-Tuning or Training a New Model

To fine-tune or train a model using your own dataset:

bash submit.sh --train --file_annotations /path/to/annotations --file_snps_labeled  /path/to/list of labeled variants  --file_snps_labeled_embedding  /path/to/embeddings of labeled variants  --dirr_results_main /path/to/results  --dirr_save_model  /path/to/saved model

Arguments

• --train: Flag to indicate training mode.
• --file_annotations: File containing the annotated variant-phenotype pairs.
• --file_snps_labeled: File containing the list of annotated variants.
• --file_snps_labeled_embedding: File containing the embedding vectors of the annotated variants.
• --dirr_results_main: Path to save result files.
• --dirr_save_model: Path to save the trained model.

---

Expected Outputs

PheMART generates different outputs depending on the mode of operation.

1. Inference Mode (Using Pre-trained Model)

• variant_ID.csv: For each variant, the result file contains the scores to the 4,179 phenotypes;
• For example, in rs5453.csv:

  C1321551, 0.532
  C1535926, 0.125
  C1321557, 0.021
  

  • Higher scores indicate a higher likelihood of pathogenicity to the phenotype.

2. Training Mode

• results_validations.txt: Training loss,hyperparameters, model performance on the validation dataset.

---

Computation Steps

The submit.sh script automates the following computational steps:

1. Data Preprocessing

   • Get variant embeddings.
   • Get phenotype embeddings.
   • Split data into training and validation sets.

2. Model Training / Fine-Tuning

   • Loads the dataset and initializes the neural network.
   • Runs training with mini-batch gradient descent.
   • Performs validation and saves the models.

3. Prediction and Evaluation

   • If in inference mode, loads the trained model and predicts the variant's relevance to all the phenotypes investigated.
   • Saves predictions.

---