Drug x Disease Interaction Prediction Package

A package for predicting some new drug x disease interactions

Create Conda Environment

To set up the environment:

conda env create -f bio-graph-env.yaml

conda activate bio-graph-env

Note: For production, a Docker image might be a better choice for consistency across systems, but the Conda environment should work reasonably well on most setups.

Running the Script

The script load_and_Represent_graph.py can be executed with:

python load_and_Represent_graph.py

Output Files

General Performance Metrics

• Random Forest performance metrics: results/evaluation_20250112_171908_rf.txt
• Neural Network performance metrics: results/evaluation_20250112_171909_nn.txt
• ROC curve overview: results/roc_curve_20250112_171909.png
• Computational performance metrics: results/performance_20250112_171910.txt

Quality Check

• Graph image overview (for input nodes and edges): results/graph_20250112_171602.png

Results

• Random Forest predictions: results/results_20250112_171910.tsv
• Neural Network predictions: results/results_nn_20250112_171910.txt

Testing

To run tests:

python -m unittest discover tests

Note: This is a simple testing framework designed as a marker rather than a comprehensive structure.

Main Script

The script load_and_Represent_graph.py performs the following functions:

1. Generate Embeddings:

   • Generate embeddings from the sample graph
   • Function: generate_node_embeddings(graph, dimensions=64, walk_length=30, num_walks=200, workers=4)

2. Load Embeddings:

   • Loading the pre-computed embeddings
   • Function: load_embeddings(embeddings_file)

3. Divide Data:

   • Split embeddings into training and testing sets with (70% training, 30% testing).
   • Function: create_dataset(graph, embeddings)

4. Train Classifier:

   • Methods used:
     • Random Forest (via sklearn for speed and effectiveness).
     • Neural Network.
     • Logistic regression
   • Function: train_classifier(X_train, y_train)

5. Graph Creation:

   • Creates graph from the provided test data
   • Function: create_graph_from_data(nodes, edges)

6. Model Evaluation:

   • Evaluates precision, recall, and F1-score.
   • Function: evaluate_model(classifier, X_test, y_test)

Package Structure

The package includes the following modules:

• Core Module:

  • drugXdisease/__init__.py

• Graph Generation:

  • Creates a simple simulated data graph
  • drugXdisease/dXd_graph.py

• Calculate Graph Metrics:

  • Calculates graph metrics in order to increase the number of learning features
  • drugXdisease/dXd_graph_metrics.py

• Logging:

  • Simple logging of computational resouces, monitoring computational costs for effective compute use (stub)
  • drugXdisease/dXd_logging.py

• Data Handling, input:

  • Read provided input files from the /data folder
  • drugXdisease/dXd_read_input.py

• Data Handling, output text and pictures:

  • Create output files and pictures
  • drugXdisease/dXd_results.py

• Learning Approaches:

  • Some different learning approaches tested
  • Logistic Regression: drugXdisease/dXd_logisticRegression.py
  • Neural Network: drugXdisease/dXd_neuralNet.py
  • Random Forest: drugXdisease/dXd_randomForest.py

Conclusions

It was fun and interesting to work with this rich dataset. Here are some reflections from the work.

Feature Engineering could improve results

This is a way to create more features for learning without adding more data, but just using the existing data.

• Added graph-based metrics to embeddings:
  • Shortest Path Length: Between drug and disease nodes.
  • Number of Paths: Connections between drug and disease nodes.
  • Shared Genes: Genes linking a drug to a disease.
  • Node Degree: Number of connections for high-degree diseases and genes.
  • Betweenness Centrality: Key regulators in the graph.
  • Clustering Coefficient: Local connectivity of nodes.

Embeddings needs to be carefully constructed

• Improved embeddings by increasing dimensionality and richness of the knowledge graph.

Data Choices are very important

• Gradually expanded the graph with additional nodes, edges, weights, and embeddings for complexity. Data choice seems to me to be crucial for predictions and performance.

Test Data is difficult to create

• It is really difficult to create test data, that has some similarity to real data. The real data existing is so complex and biased in different ways so it is almost impossible to reconstruct. In this case, it seems to make more sense just to use slices of real data for all training, testing and experiementation. For visualisation I did "dumb" filtering, of just taking out nodes that are not relevant, and "dumb" subsetting if the data for visualisation. But for real implementation it would make more sense to do slices of the real graph, during development.

Suggested Improvements

• The performance logging should be updated to include CPU, GPU etc which is suitable for the compute cluster we're at
• The graph image that is written is not very helpful, it should be updated to a nicer looking perhaps 3D graph in order to be effective
• The evaluation could be more skewed from not finding the overall optimal solution, towards finding more of what we really want; potential new drug-disease associations to explore further. For example a metrics like f1 is a bit helpful, but not optimal in that regard.
• Trying more different classification approaches (most likely to create stepwise change, along with really good data)
• Stability testing - it might be that new drug-disease associations that are found using more different classification approaches, or with the same approach re-run several times may be more reliable?
• Reverse engineer the classifications, so we also can understand which features/input data most contributes to the best solutions. There are elegant and more pragmatic ways of doing that, which I've tried.
• Any new drug-disease associations found needs additional orthogonal evidence to be pursued. For example scientific literature, animal model experiments, or similar to build a case.

Some Results

Example Predictions

1. Hyperpigmentation (MONDO:0019290):

   • Potential treatment: Hyperpigmentation of the skin (MONDO:0019290) could be treated by Azelaic acid (CHEMBL.COMPOUND:CHEMBL1238)
   • There is some evidence to support it, Azelaic acid is currently most used in treatment of acne, but effect on pigmentation has been noted: "Azelaic acid, a tyrosine inhibitor, also helps reduce pigmentation, therefore is particularly useful for darker skinned patients whose acne spots leave persistent brown marks (postinflammatory pigmentation) or who have melasma."

2. Cystic Fibrosis (MONDO:0009061):

   • Cystic fibrosis (MONDO:0009061) could be treated by CHLORHEXIDINE GLUCONATE (skin antiseptic) CHEMBL.COMPOUND:CHEMBL4297088
   • Obviously nonsense.

3. Coronary Aneurysm (MONDO:0006714):

   • Prediction: Coronary aneurysm (MONDO:0006714) could be treated by Urokinase (CHEMBL.COMPOUND:CHEMBL1201420).
   • Assessment: Limited utility, as Urokinase is effective for blood clots, but coronary aneurysm only occasionally involves blood clots. Those cases with Coronary aneurysm with blood clots would currently be treated with aspirin, P2Y12 inhibitors, or general anticoagulants like warfarin. This probably would not be a lead we'd follow up on, because it is a bit non-specific, and it seems good options exists already.

4. Lichen Planus (MONDO:0006572):

   • Prediction: lichen planus (a chronic, recurrent, pruritic inflammatory disorder of unknown etiology that affects the skin and mucus membranes, MONDO:0006572), could be treated by ALEFACEPT (CHEMBL.COMPOUND:CHEMBL1201571).
   • Supporting Evidence: Looking at the CHEMBL website, "Lichen Planus" is already listed as studied. https://www.ebi.ac.uk/chembl/explore/compound/CHEMBL1201571 . This is the study: https://clinicaltrials.gov/study/NCT00135733?term=NCT00135733&rank=1 The results were published https://pubmed.ncbi.nlm.nih.gov/18459520/ and showed that 2 out of 7 patients achieved significant improvement, and was well tolerated by all patients. Authors recommended further study.

Suggested Improvements

• Weigh results against broad-spectrum drugs (eg. anti-inflammatory) and diseases (e.g. headache, skin rash) to create a more targeted list.