Precision Intelligence Engine

live   Domain: clinical   Type: engine

Variant annotation (ClinVar/AlphaMissense) + RAG interpretation over the shared variant foundation to identify and contextualize druggable targets. Decision support for a qualified clinician, not diagnosis. (An Ensembl VEP container ships as an optional, manually-run annotator; it is not in the automated interpretation path.)

A narrated, captioned explainer. Decision support for a qualified clinician.

Illustrative. Decision support for a qualified clinician — never autonomous diagnosis or prescribing.

In plain terms

If the Genomics Foundations Engine finds where a patient's variants are, the Precision Intelligence Engine explains what they mean. It is the factory's interpretation brain. It takes a raw variant file — millions of rows — and turns it into a short, cited, plain-language narrative a clinician can actually use: which variants matter, why, and which point to a druggable target. It is also the hub that coordinates the eight intelligence agents.

Why it matters

A variant list is not an answer. The clinically meaningful signal is usually a handful of variants buried in millions, and the value is in the context — is this variant known to be pathogenic, what does it do to the protein, is there a therapy that targets it. This engine is what turns data into decision support: grounded, sourced, and honest about uncertainty, never an autonomous diagnosis.

For a patient: the difference between a raw list of millions of variants and a short, plain-language explanation of the few that actually affect their care.

How it works

From a variant file to a cited clinical narrative — grounded in retrieved evidence. Illustrative.

1. Annotate — each variant is tagged against curated knowledge: ClinVar clinical significance and AlphaMissense pathogenicity predictions.
2. Retrieve — retrieval-augmented generation (RAG) pulls the relevant evidence from a vector database, so the narrative is built on real sources rather than a model's memory.
3. Reason — a language model writes a grounded interpretation with citations, surfaces druggable targets, and is built to refuse rather than fabricate when evidence is thin.
4. Coordinate — it routes questions to the eight specialist agents (pharmacogenomics, oncology, rare disease, and the rest) and pulls their answers together.

What goes in, what comes out

• In: a VCF of variants (from Engine 1) and a clinical query.
• Out: a cited clinical narrative — the interpreted, contextualized report.

Where it fits

The interpretation hub: it consumes the genomics substrate and coordinates the eight agents. Illustrative.

It sits directly downstream of genomics and upstream of the agents — the interpretation layer the rest of the clinical reasoning depends on.

Honest limits

• Decision support, not diagnosis. Every output supports a qualified clinician; it never diagnoses or prescribes on its own.
• Grounded, and it says when it can't be. The interpretation is retrieval-grounded and cited; when its vector database or model key are absent it returns an honest degraded response, never invented clinical content.
• VEP is optional and manual. An Ensembl VEP container ships as an optional, manually-run annotator — it is not in the automated interpretation path.

Interface

• Endpoint: localhost:5001 · Invoke path: /
• Serving: native · GPU: no · Cost class: medium

Inputs

Name	Shape	Semantic	Notes
vcf	file	vcf_variants
query	scalar	—

Outputs

Name	Shape	Semantic	Notes
report	map	clinical_narrative	targets + clinical context

Tags: rag · annotation · interpretation

---

↩ Back to the Engines index · the Capability Maturity Matrix · the Capability Brief.

Note

Status and interface are generated from the capability registry (lib/hcls_common/capabilities.json) — the site cannot claim ahead of the code. All clinical output is decision support for a qualified clinician, never autonomous diagnosis.