graph-molecular-phenotype-prediction

Code accompanying the paper "A graph-based learning approach to predict the effects of gene perturbations on molecular phenotypes" by Jin et al.

---

📁 Folder Structure

Legend: `/` = directory, no suffix = file
├──Datasets/
   ├──Influenza/
       ├──split_feature_group/
       ├──split_subgraph/
       ├──other data files
   ├──SREBP2/
       ├──split_feature_group/
       ├──split_subgraph/
       ├──other data files
   ├──LDLR/
       ├──split_feature_group/
       ├──split_subgraph/
       ├──other data files
   ├──Mitomics/
       ├──split_feature_group/
       ├──split_subgraph/
       ├──other data files
   ├──targets/ (target list for each phenotype)
   ├──Raw/
       ├──cellular-localization/ (celltype-specific protein abundance)
       ├──{phenotype}/ (instance list for each phenotype)
       ├──go_embedding_64.csv
       ├──subcellular-localization.csv 
       ├──uniprot_reactome_hpa_merged_stringid.csv (raw file for subcellular localization)
   ├──config/ (tuned hyperparameters for each model for each phenotype for each fold for each subset)
├──Code/
   ├──Dataset/
   ├──Figures/
├──Plot/
   ├──{phenotype}/ (Files necessary for feature importance replication)
   ├──Transfer_tuned/ (Files for transfer learning replication)

🚀 Getting Started

Prerequisites

• necessary packages are listed in file environment.yml

Instructions

• Here is the workflow of our proposed method

• In this paper we collect four phenotypes as well as four machine learning methods, each phenotype and method has its own abbreviation in the script name.
  Cholesterol homeostasis ── SREBP2
  Cholesterol uptake ── LDLR
  Influenza A virus replication ── Influenza
  Mitochondrial protein abundance ── Mitomics
  Logistic Regression ── LR
  Random Forest ── RF
  XGBoost ── XGB
  Neural Network ── NN

• To replicate Figure 2 and 3 for a phenotype, run base_{phenotype}. e.g. For phenotype cholesterol uptake, run base_LDLR.py

• To replicate Figure 4, first run feature_subgraph_{phenotype}.py for each phenotype, then run Feature_importance.ipynb.

• To replicate Figure 5, first run feature_feature_group_{phenotype}.py for each phenotype, then run Feature_importance.ipynb.

• For phenotype Mitochondrial protein abundance, feature_Mitomics.py would generate files for both Figure 4 and 5.

• To replicate Figure 6, first run Transfer_{ML-method}.py, then run Transfer_learning.ipynb.