How this works
A cited answer, or none at all
How siRNA Atlas is built: a curated corpus of FDA labels and review packages, EMA assessment reports, pivotal trials and siRNA ADME literature, searched per question so every answer carries a citation back to a primary document.
What it is
siRNA Atlas answers questions about the 8 approved small interfering RNA (siRNA) therapeutics — the mechanism, the approvals, the trial evidence, and the clinical pharmacology and ADME behaviour of the class. It is built on retrieval-augmented generation: every question triggers a search across a curated corpus, and the assistant answers only from what it retrieves.
The reason for building it is narrow and practical. The answers to the interesting questions about this class — how GalNAc conjugation changes distribution, what the nonclinical packages required, how renal and hepatic impairment were assessed at review — sit inside several-hundred-page regulatory PDFs that are rarely read end to end. A search over the primary record is a better instrument for that than a model’s memory.
The corpus
- FDA labels — Current prescribing information for all 8 drugs
- FDA reviews — Integrated, clinical, clinical-pharmacology and pharmacology/toxicology reviews
- EMA EPARs — European Public Assessment Reports + product information (6 EU-approved drugs)
- Pivotal trials — APOLLO, ENVISION, ILLUMINATE, ORION, HELIOS, PHYOX, ATLAS, PALISADE and others
- ADME / PK-PD — GalNAc conjugation, hepatocyte targeting, metabolism, delivery chemistry
- Regulatory science — FDA / EMA / ICH / OSWG oligonucleotide guidance
- Drug development history — Founding-CEO account of building Alnylam and taking RNAi from concept to approved medicine
Per-drug source lists live on the drug pages, each linking to the primary FDA and EMA documents.
How an answer is assembled
- RetrieveThe question is expanded with the spelled-out forms of regulatory acronyms, scoped to the drugs it names, and searched against the vector index. Multi-drug questions get a per-drug quota so one drug cannot take every slot.
- GroundThe retrieved passages, and nothing else, are given to the model as the material it may use. Where the corpus does not cover a named drug, the prompt says so explicitly rather than letting a neighbouring drug's data fill the gap.
- CiteEach claim carries a numbered marker resolved to a real document — a regulatory filename, a PubMed record, or a DOI. Markers are positional and never reordered, so the number in the text is always the source that was used.
- VerifyOpen the citation. Every one resolves to the label, assessment report or paper it came from.
For agents
The same corpus is available to your own AI agent over a retrieval-only MCP server at https://sirna-atlas.pkpdbuilder.com/api/mcp — four tools (search_corpus, corpus_info, get_citations, fetch_document_section), no authentication, no model call server-side. Connect instructions are behind the “For agents” button on the home page. A plain-text brief for machine readers is at /llms.txt.
Questions and answers
What is siRNA Atlas?
siRNA Atlas is a retrieval-grounded assistant covering the 8 approved small interfering RNA (siRNA) therapeutics. Each question runs a vector search over a curated corpus of regulatory documents and peer-reviewed literature, and the answer is written only from the passages retrieved, with a numbered citation on each claim.
Which drugs does it cover?
Patisiran (Onpattro), Givosiran (Givlaari), Lumasiran (Oxlumo), Inclisiran (Leqvio), Vutrisiran (Amvuttra), Nedosiran (Rivfloza), Fitusiran (Qfitlia), Plozasiran (Redemplo, EU) — every siRNA therapeutic approved to date, covering FDA and, where applicable, EMA records.
How is this different from asking a general chatbot?
A general model answers from training data it cannot show you, and can invent both facts and citations. Here the facts come from a vetted document store assembled for this class of drugs; the model supplies reasoning and wording only. Each answer links to the document and page it drew from, so a claim can be checked rather than trusted.
What is in the corpus?
20 FDA documents (current labels plus integrated, clinical, clinical-pharmacology and pharmacology/toxicology reviews), 12 EMA assessment reports and product information documents, and 79 publications of which 66 are indexed as full text — 10,807 searchable passages in total.
How are tables in regulatory PDFs handled?
Tables are the part that ordinary PDF extraction destroys, and PK and toxicology data live in tables. Documents are converted with a table-preserving pipeline and each table is kept whole in a single passage, so numeric dosing, exposure-response and NOAEL values remain searchable and are read in their original rows and columns.
Can an AI agent query the corpus directly?
Yes. siRNA Atlas exposes a retrieval-only Model Context Protocol (MCP) server at https://sirna-atlas.pkpdbuilder.com/api/mcp. It returns source passages with citations and runs no model server-side, so your own agent does the reasoning over verified sources. No authentication is required; connect instructions are behind the "For agents" button on the home page.
Can the answers be wrong?
Yes. Retrieval can miss a relevant passage, and a model can still misread one it retrieved. The citation is the safeguard: open the linked label, assessment report or paper and confirm the claim at source. This site is a research and reference tool, not medical advice.
Who built it
siRNA Atlas was built by Husain Attarwala, who works in clinical pharmacology and PK/PD. It is an independent project and is not affiliated with, endorsed by, or speaking for any sponsor of the drugs it covers.