How redomic/neuthera-drug-discovery-toolkit works
Project type and purpose
• NeuThera is a full-stack, modular AI-driven drug discovery toolkit. It blends data ingestion, multiple ML model services (TamGen, DeepPurpose, chemBERTa, a DTI predictor), external scientific tools (AutoDock, MMseqs2), a graph database (ArangoDB), and an interactive UI (Streamlit + Jupyter notebook).
• Its goal: end-to-end and interactive drug candidate generation, validation (docking + affinity prediction), embedding/indexing, retrieval, and lead optimization.
File structure
• Top level:
• app.py, start.ipynb, Streamlit config (.streamlit/) → interactive frontend & notebook
• docker-compose.yml → ArangoDB service
• data/ & database/ → raw molecular data, prebuilt caches
• DTI_model/, gpt_model/ → pretrained model artifacts
• TamGen/ → core generative model code & submodules (fairseq integration, molecule_utils, scripts, tests)
• preprocess.py, db.py, tools.py → orchestration & utility scripts
• scripts/ → shell/python helpers for data preparation, training, inference
• requirements.txt, setup.py → build & install
README for additional insights
• Core modules: TamGen (fairseq-based SMILES generator + docking), DeepPurpose (affinity prediction), chemBERTa (SMILES embeddings + vector indexing)
• Knowledge graph: ArangoDB holds entities & embeddings, vector index for similarity
• Workflow: preprocess → TamGen generate → docking → DeepPurpose score → chemBERTa embed → graph ingestion → retrieval/analytics → UI