Trust_DTI

Imported from Itsabhirajurs/Trust_DTI (AGENTS.md). Install upstream with npx skills add Itsabhirajurs/Trust_DTI. Copyright stays with the author.

TrustDTI: instructions for any AI assistant (read this first)

Project. TrustDTI is a 12-week MSc research project. It predicts whether an existing drug interacts with a human protein, attaches a calibrated confidence to each prediction, looks up (never predicts) the protein's disease link, and uses both signals to decide what literature an AI agent searches for. Confident → look for supporting evidence. Uncertain → look for counter-evidence. Details: docs/PROJECT_GUIDE.md.

The owner is learning this project and is not yet fluent in it. Explain simply, give the why before the how, go one pipeline stage at a time, and do not jump ahead until they say so. When you explain something, save it under docs/stages/.

Start of every session (do this before anything else)

1. Read docs/STATUS.md: where we are, what is next, open decisions.
2. Skim the top 3 entries of docs/PROJECT_LOG.md (newest first).
3. Need background? docs/PROJECT_GUIDE.md. Need one stage in depth? docs/stages/.
4. Tell the owner in ~5 lines where we are and what you propose next. Then wait.
5. The repo and data are ground truth. If a doc disagrees with the code or data, trust the repo and flag the conflict.

Rules that never change silently

• Research question (exact wording in the guide §2) and the blinding protocol: never open cases/sealed.json before the Week 9 step that allows it.
• UNTESTED is never INACTIVE. Keep censored rows (>, <, >=, <=, >>, <<). Never filter on pchembl_value IS NOT NULL. "inconclusive" is not inactive.
• The router is deterministic, unit-tested code, never an LLM decision. UNKNOWN is never turned into CONTRADICT.
• Compound–protein edges (predicted) and protein–disease edges (curated) live in separate tables or carry a provenance column.
• Every methodological choice goes into DECISIONS.md (D-NNN, dated, with why and evidence) when it is made.
• No new scope (knowledge graph, docking, LLM fine-tuning, etc.) without a flagged discussion first.
• Raw and interim data are never committed. Small derived results in reports/ are.
• Do not invent numbers. Every figure must come from a file in reports/ or a query on the data.

End of every session

1. Add a dated entry at the top of docs/PROJECT_LOG.md (template inside that file).
2. Update docs/STATUS.md (current state, next actions, open decisions).
3. Record new decisions in DECISIONS.md.
4. Commit with a message that says what changed and why, then push to main.

Working environment

• Open the folder C:\Users\91701\code\Trust_DTI. The copy in OneDrive is an old backup with no data. Do not work there.
• Python 3.12 virtual environment at .venv (Windows: .venv\Scripts\activate). Install: pip install -r requirements.txt.
• Tests: pytest (fast, about 2 s). pytest -m slow is a ~1 hour full-database check, run only before a data-version change.
• Raw data are in data/raw/ (gitignored). How they were obtained and verified: data/MANIFEST.md. No GPU on this laptop; heavy embedding jobs go to Colab/Kaggle/HPC (DECISIONS.md D-003).
• Run scripts as modules from the repo root, e.g. python -m src.ingest.profile_chembl.