How csubiogroup/dtiam works
Project type and purpose
• DTIAM is a research‐oriented, full‐stack Python framework for three ML tasks:
• Drug-Target Interaction (DTI) prediction
• Drug-Target Affinity (DTA) prediction
• Mechanism of Action (MoA) classification (activation/inhibition)
• It combines data preparation, self-supervised molecular representation learning (BerMol), and downstream model training/validation.
File structure
• Top‐level directories/files:
• code/ — contains all Python modules, scripts, and package code
– BerMol/bermol/ — reusable Python package: data loading, tokenizer/vocab, model, trainer, utils
– scripts/ — standalone helpers, e.g., featurize.py
– data_process/ — dataset splitting and feature extraction pipelines
– training_validation.py — main entrypoint orchestrating experiments
– run.sh, test.sh — convenience wrappers
• data/ — raw benchmark datasets grouped by task and source
• img/ — architecture/flowchart reference
• requirements.txt, README.md, setup files, environment scripts
• Architectural patterns:
• Modular Python package (BerMol) for core representations
• Layered pipeline: Data → Preprocessing → Feature Extraction → Model → Training/Validation
• CLI entry point driving experiments with parameterized tasks/datasets/settings
README for additional insights
• High‐level flowchart (img/flowchart.png) outlines:
Raw data → Preprocessing → Self-supervised pretraining (BerMol) → Downstream fine-tuning and evaluation
• Instructions for conda environment and dependencies
• How to invoke training: python training_validation.py <task> <dataset> <setting>