How kivancbayraktar/enhancing-ddi-classification-via-textual-arguments works
Project type and purpose
• This is a machine-learning research pipeline (full-stack data-processing + modeling) for Drug-Drug Interaction (DDI) classification that leverages textual arguments from DrugBank, UMLS, and NER outputs. It orchestrates data ingestion, feature extraction (embeddings, similarity matrices), dataset preparation, and runs a sequence of deep‐neural‐network experiments, logging results via MLflow.
File structure
• colab_notebooks/: Jupyter notebooks that orchestrate each major step (data prep, embedding gen, experiments).
• drugbank/: Parser + processor modules for ingesting the DrugBank XML and building a local SQLite DB.
• ner/: Named-entity-recognition pipeline using cTAKES, producing .pkl outputs.
• embedding_configurations/: JSON defining embedding sources/parameters.
• experiment_configurations/: JSON configs for each experimental scenario, plus a models.py helper.
• research_datasets/: Modules and raw/index files for constructing train/validation/test splits for two datasets (DDI-MDL, MDF-SA-DDI).
• results/: Excel outputs for each experiment.
• name_rxcui.json: UMLS ID mapping for drugs.
• Configuration & dependency files: .gitignore, LICENSE, README.md.
README for additional insights
• Four major steps: build drugbank.db, run NER, generate embeddings, run experiments.
• Dependencies: Python, TensorFlow, Apache cTAKES, Transformers, SentenceTransformers, MLflow.
• Experiments driven by changing experiment_configuration_file in the notebook.
Detail
• Experimental scenario JSON appears as an input to the Experiment Runner.
• Optional paths (e.g., precomputed vs. regenerate embeddings, use existing NER output vs. rerun).