BioChirp Help & Documentation
Ask a plain-language question about a drug, gene, disease, pathway or biomarker and get back a complete, structured table from 11 curated biomedical databases — not a chatbot's best recollection. AI only interprets what you asked; rule-based code does the actual data fetching, so the same question returns the same rows every time.
Who it's for: bench scientists, bioinformaticians and pharmacologists who need an exhaustive, reproducible view of known associations. Safety notice: research use only, not a certified medical device — never a substitute for clinical judgement.
2. How BioChirp works
Four stages. The first two involve AI; the last two don't — that boundary is why the same question always returns the same table.
- A. Gateway — your question arrives over a WebSocket connection, which streams live progress back as each stage runs.
- B. Interpretation (AI-assisted) — multiple AI models parse your question in parallel into a structured record — e.g. drug: (return all), disease: tuberculosis — with abbreviations expanded. This is the last step AI touches.
- C. Data retrieval (rule-based) — Open Targets is queried live via its GraphQL API. The other 10 databases are pre-processed local snapshots — a schema-graph planner walks the shortest path between the tables your question needs. Same input, same join order, every time.
- D. Summary & delivery — an AI model writes a short narrative from a preview of the results. The narrative is a convenience; the downloadable CSV is the authoritative output.
3. The 11 databases
Open Targets is queried live over the internet; the other 10 ship as pre-processed local snapshots refreshed periodically. If a question falls outside all 11, BioChirp falls back to a clearly-labelled biomedical web search.
| Database | Covers |
|---|---|
| Open Targets Platform (live) | Drug–disease–target evidence: approved drugs, clinical candidates, target associations across 20+ evidence sources. |
| Therapeutic Target Database (TTD) | Drug-target-disease relationships: 3,000+ drugs, 2,600+ curated protein targets. |
| Comparative Toxicogenomics Database (CTD) | Chemical–gene–disease interactions including environmental exposures and toxicology. |
| Highly Confident Drug-Target Database (HCDT) | Validated high-confidence drug-target interactions in human cancer. |
| ClinVar | Pathogenicity classifications of germline variants with submitter evidence. |
| Human Phenotype Ontology (HPO) | Standardized phenotype terms linked to diseases and genes. |
| Orphanet | Rare disease database with phenotypes, gene causes, prevalence, and epidemiology. |
| Reactome | Curated biological pathways (2,870+) with ChEBI/UniProt/Ensembl cross-links. |
| STRING | Protein-protein interaction network with functional and physical associations. |
| Universal Protein Resource (UniProt) | Protein sequences, functions, domains, and annotations from Swiss-Prot. |
| Molecular Signatures Database (MSigDB) | Curated gene sets for enrichment analysis (GSEA): human Hallmark/C1–C9 and mouse collections. |
4. Getting started
From a blank chat window to a downloadable results table in under a minute.
- Pick a database from the home page — each has its own chat.
- Ask a question naming at least one entity (drug, gene, disease, pathway, variant, biomarker). Abbreviations, brand names and gene aliases resolve automatically —
TB,CML,ERBB1all work. - Send it and watch progress through interpretation, name matching, retrieval and summarisation.
- Read the answer — a short narrative above a preview table; expand the tool panels to see exactly what happened at each step.
- Download the CSV for the full, untruncated table.
Good questions
Which drugs are prescribed for tuberculosis?What diseases does aspirin treat?Show me pathways involving TP53
Won't work well
- Clinical scenarios with no named drug/gene/disease
- Molecular structures, 3-D structures, raw sequence data
- Trial logistics, regulatory filings, insurance coverage
5. Reading BioChirp outputs
Every answer has two layers: a short AI-written narrative on top, and a structured table below — the table is the definitive result and the summariser can never add, remove or reorder its rows. The narrative explains how your question was understood and flags any gaps; the table has fixed column headers, a scrollable in-chat preview, and a full CSV download. Each internal step (interpretation, name matching, database query) is also a collapsible panel showing elapsed time and its actual input/output, useful for auditing a surprising result.
6. Limits & troubleshooting
What it can't (yet) do
- Only as comprehensive as its 11 sources — very new indications may be missing
- The 10 offline databases refresh periodically, not in real time
- Highly ambiguous terms can still trip up the resolver
If two sources disagree, both appear — BioChirp doesn't editorially pick a winner. For high-stakes decisions, verify against the original source.
Fixing a bad result
- No matches? Add a concrete entity, e.g.
drugs targeting JAK2 in myelofibrosis, or try a broader term - Timeouts usually clear up on retry after a minute
- Table changes between runs → name-matching issue; stays the same → check the source database directly
- Still stuck? Open a GitHub issue with the query and what you expected
7. Frequently asked questions
How is this different from ChatGPT or Claude?
Chatbots recall biomedical facts from training data, which can miss records or vary between runs. BioChirp uses AI only to understand your question — every row in the output is read directly from a curated database via rule-based joins, so it can't hallucinate entries or silently drop rows.
Which databases are included?
11: Open Targets (queried live), plus TTD, CTD, HCDT, ClinVar, HPO, Orphanet, Reactome, STRING, UniProt, and MSigDB (local, refreshed snapshots). See section 3. Questions outside all 11 fall back to a clearly-labelled web search.
Can I use brand names, abbreviations or gene aliases?
Yes — string matching, embedding similarity, and curated alias lookups run in parallel, then a dual-AI filter removes false positives. "Vazalore" resolves to aspirin, "ERBB1" to EGFR, "CML" expands to chronic myeloid/myelogenous leukemia.
Why does the summary wording change between runs but the table doesn't?
The table comes from rule-based code and is identical on every run against the same snapshot. The narrative is AI-written and may be phrased differently each time — for any downstream use, rely on the CSV, not the narrative.
Does BioChirp store my queries?
No login is required and no personally identifiable data is collected. Queries are processed in real time, not stored for profiling or tracking.
Is it open source?
Yes — code, prompts, schema definitions, and pipelines are on GitHub at github.com/abhi1238/biochirp.
8. Contact & support
Reasons to get in touch
- Repeated errors or outages that don't clear on retry
- Output that contradicts the original database
- Feature ideas, UI feedback, or data accuracy concerns
How to reach us
Email abhishekh@iiitd.ac.in, debarka@iiitd.ac.in, or open a GitHub issue.
Please include the exact query, approximate time, and whether it reproduces on a second attempt.
Reminder: BioChirp is a research aid, not a clinical decision-support system. For medical emergencies, contact local emergency services immediately.