Graph representation learning using molecular interaction networks for modelling omics data in Parkinson's disease

This repository contains an implementation to perform graph representation learning modelling using molecular interaction networks of transcriptomics (protein-protein interactions) and metabolomics (metabolite-metabolite interactions), which is able to learn PD-specific fingerprints from the spatial distribution of molecular relationships in an end-to-end fashion. The scripts apply the graph representation learning modelling pipeline on networks of molecular interactions, where transcriptomics and metabolomics data from the PPMI and the LuxPARK cohort, respectively, are projected.

If something is not clear or you have any question, please open an Issue.

Repository structure

The analyses on both PPMI ans LuxPARK cohorts include some pre-processing steps prior to the modelling pipeline. In addition, the analyses on PPMI include an R script and R functions to compute functional enrichment analysis on MeSH, KEGG and GO databases from the most relevant genes identified by the modeling pipeline. Corresponding scripts do not exist for LuxPARK because the functional analysis was done with MetaboAnalyst software.

The main script to run the modelling pipeline, including model building, training, hyperparameter tunning and cross-validation, is the file executed by the wandb agent: cv_wandb.py (or cv_wandb_DENOVO.py). This file includes all the code necessary to extract the vector containing the omics profile of each patient, project it in the molecular interaction graph (PPI or MMI), read the hyperparameters defined from the wandb agent, train, and evaluate a GCN, GAT or ChebyNet model accordingly. The files it requires are in the same directory:

• utils.py: Many utility functions, including those for network construction, training and evaluation, feature relevance, etc.
• plot_utils.py: Functions to create plots about the training and validation, as well as to project the node (sample) embeddings in 2D and 3D. They were used for debugging and experimentation.
• wandb_config_*.yaml: Config file for the hyperparameter search of each model.

In addition, cv_wandb.py requires as input information from the topology of the molecular interaction networks, given by the following lines of code (to be replaced if pipeline implemented on a different dataset):

rna_file = "../data/data_rnaseq_ppi_4ML.csv"
labels_file = "../data/labels_4ML.csv"
ppi_score_file = "../data/ppi_score_4ML.csv"
adj_file = "../data/A_ppi_weighted_4ML.csv" 

metab_file = "../data/data_metab_mmi_STITCH_700.csv"
labels_file = "../data/labels_4ML.csv"
mmi_score_file = "../data/mmi_score_STITCH_700.csv"
adj_file = "../data/A_mmi_weighted_STITCH_700.csv"

• rna_file, metab_file: Matrix of gene expression and metabolomics abundance (n x m). Rows being samples and columns representing molecules.
• labels_file: Patient/control binary labels (n x 2), columns being sample id, diagnosis.
• ppi_score_file, mmi_score_file: Edge list (e x 3), columns being molecule1, molecule2, weight. Each row defines an edge (interaction) in the network.
• adj_file: Weighted adjacency matrix (n x n).

Other scripts used for the modelling pipeline: * `cv_results.py`: Extracts the cross-validated results by looking at the minimum validation loss and generates figures and results tables based on node and edge importance. * `features_plot.py`: Generates barplot with most relevant nodes and their functional annotation. * `pw_embeddings.py`: Adds a "pathway embeddings layer" using a masked linear layer with sparse mask (defined by molecules' pathway membership) and weight matrix, mimicking the representation of a pathway functional embedding.

Data pre-processing

In the PPMI cohort:

• Building_network_data_4ML.ipynb: A jupyter notebok that shows how to build a fully connected PPI network from STRING database files matching a matrix of pre-processed transcriptomics data.
• ppmi_data4ML_class.R: This script performs unsupervised filters to generate data for ML modelling of snapshot data (BL) from RNAseq data. *ppmi_data4ML_class.R employs as input transcriptomics and phenotypical data resulting from previous pre-processing scripts described in repository statistical_analyses_cross_long_PD for parsing data and Baseline (T0) PD/HC (ppmi_filter_gene_expression.R, ppmi_norm_gene_expression.R, ppmi_generate_pathway_level.R).

In the LuxPARK cohort:

• STITCH_cleansing.ipynb, KEGG_cleansing.ipynb: Perform prunning in the original network of interactions to get only edges with medium confidence score in STITCH (>700). They also filter molecules mapping both to the metabolomics matrix and the MMI network, merging annotations based on multiple identifiers: PubChem, KEGG, ChEBI, inchikeys, CAS, SMILES.
• Building_network_data.ipynb: A jupyter notebok that shows how to build a fully connected MMI network from pre-processed STITCH or KEGG database files matching a matrix of (log-transformed) metabolomics data.
• func_kegg_reactions.R: Functions that download compounds and reactions data from kegg API, and subset relevant fields of information.

*Unsupervised filters were not applied to metabolomics data. Building_network_data.ipynb employs as input metabolomics and phenotypical data resulting from previous pre-processing scripts described in repository statistical_analyses_cross_long_PD for parsing data and Baseline (T0) PD/HC (lx_extract_visit.R, lx_denovo_filter.R, lx_generate_pathway_level.R).

Running the experiments

The code in this repository relies on Weights & Biases (W&B) to keep track and organise the results of experiments. W&B software was responsible to conduct the hyperparameter search, and all the sweeps (needed for hyperparameter search) used are defined in the wandb_config_*.yaml files. All the runs and sweep definitions are publicly available at the project's W&B page. Each of the sub-projects displays a different experiment (e.g., using a different model, or a different dataset).

In particular, W&B sub-projects starting with ppi_* or mmi_* reflect the implementation of this modelling pipeline (while those starting with psn_* reflect the implementation of the pipeline described in repository GRL_sample_similarity_PD).

We recommend that a user wanting to run and extend our code first gets familiar with the online documentation. As an example, one would create a sweep by running the following command in a terminal:

$ wandb sweep --project mmi-LUXPARK-gcn wandb_config_gcn.yaml

Which yielded an identifier (in this case 8phdgg09), thus allowing us to run 130 random sweeps of our code by executing:

$ wandb agent psn-metabolomics/mmi-LUXPARK-gcn/8phdgg09 --count=130

The wandb agent will execute cv_wandb.py with its set of hyperparameters (as defined in wandb_config_gcn.yaml or corresponding yaml file, depending on the experiment). Note that for a given experiment, the same sweep file with random hyperparameter search is utilized.

Data

The public transcriptomics data used in this project was derived from the Parkinson’s Progression Markers Initiative (https://www.ppmi-info.org/, RNAseq - IR3). The metabolomics data from LuxPARK is not publicly available as it is linked to the Luxembourg Parkinson’s Study and its internal regulations. Any requests for accessing the dataset can be directed to [email protected].

License

The code is available under the MIT License (see LICENSE).