Overview

This project explores whether graph-based machine learning methods can be used to identify novel genes involved in the host response to tuberculosis and uncover potential drug targets.

Methods

• Data: Disease–gene associations (DisGeNET), ontology data (MONDO), protein–protein interaction networks (BioGRID)
• Approach:
  • Constructed a bipartite disease–gene graph
  • Applied a Graph Autoencoder (GAE) using a Graph Convolutional Network (GCN) encoder
  • Performed link prediction to identify potential new disease–gene associations
  • Ranked predicted associations using embedding similarity scores
• Evaluation:
  • Assessed model performance using AUC, average precision (AP), and ROC curves
• Tools: Python, PyTorch, PyTorch Geometric, NetworkX

Key Insight

The model generated a large set of candidate gene associations for tuberculosis, highlighting the potential of network-based approaches to expand gene sets for underrepresented diseases. However, extending this approach to drug–gene prediction revealed limitations in model performance, emphasizing the importance of data quality and representation.

Notes

Completed as part of coursework in molecular informatics and inspired by prior research experience. This project reflects early experience building graph-based machine learning pipelines for biological data.

---

Execute Scripts in this order

1. get_disgenes.Rmd

Input: Disease gene association data fromm DisGeNET and MONDO obo file. Output: Propagated genes associations for all collected DisGeNET data.

Files needed:

1. mondo_2024_03.obo
2. all_gene_disease_associations.tsv

Functions

1. map_mondo_to_ontology (DisGeNET df, obo_file): Maps MONDO disease identifiers to the disease identifiers contained in DisGeNET database(UMLS).
2. get_ances (list of disease identifiers, obo file): gets the ancestor term of all the terms found in the DisGeNET data set.
3. propagate_genes (list of disease identifiers from DisGeNET df, ancester df, DisGeNET data with mapped MONDO identifiers): function that propagates genes from children terms up to parent terms.

Output:

1. not_propagated_disgenet_genes.tsv
2. tb_propagated_disgenet_genes.tsv

**2. get_drug_data.RMD **

Input: Raw data downloaded from DGIdb website. Output: Processed drug data containing CHEMBL drug identifiers.

Files needed:

1. interactions.tsv

Output:

1. dgidb_final_drugdata.tsv

**3. karla_vela_gae_code.ipynb **

Input: BioGRID data, propagated genes obtained from DisGeNET, and processed DGIdb drug data Output: performance metric plots for each GAE and checkpoint data files to provide for class presentation.

Files needed:

1. propagated_disgenet_genes.tsv
2. biogrid_network.txt
3. dgidb_final_drugdata.tsv

Functions

1. def plot_roc_curve(title, model, data): plots ROC curve for GAE and VGAE models.

2. plot_training_stats(title, losses, test_auc, test_ap, train_auc, train_ap): plots the AP and AUC after each epoch.

3. get_edge_dot_products( data, model, num_dz_nodes): computes the dot products between the encoded disease and gene nodes for the VGAE and GAE models.

4. get_ranked_edges(data_object, model, num_dz_ndoes): ranks the edges based on the dot products of the encoded nodes and returns the ranked edge list and dot products.

5. Class GCNEncoder(torch.nn.Module): defines and initializes the GAE model.

6. gae_train(train_data, gae_model, optimizer): trains the GAE model on the provided training data.

7. gae_test(test_data, gae_model): evaluates the GAE model on the provided test data.

3. karla_vela_class_version.ipynb

• This is a shortened version with added background and images for the class presentation.

Input: BioGRID data, propagated genes obtained from DisGeNET, and processed DGIdb drug data Output: performance metric plots for each GAE and checkpoint data files to provide for class presentation.

Files needed:

1. disgene_final_dataset.tsv
2. new_filtered_disgenet.tsv
3. filtered_biogrid_df.tsv