FILE HL–2026–05

ENGINE explicit rules · no ML

CLASS. TECHNICAL · PUBLIC

DOMAIN GERMLINE MENDELIAN

DATABASES HPO · ClinVar · ClinGen

11 Clinical · genetic diagnosis

Helion

It turns the patient's phenotype into a prioritised genetic testing plan — panel, exome or whole genome — with every recommendation traceable to a specific rule.

Explicit code, zero machine learning, zero black boxes. This is not an aesthetic stance: it is what documenting clinical reasoning for an accreditation demands.

How to read this page

Every recommendation arrives with its evidence level and the version of the database behind it. Without that, a testing plan is an opinion with formatting.

19.000 phenotype terms · directed graph
2,8 M clinical variants · 3,500 associations

opaque models in the decision chain: every recommendation is traceable to a specific rule

19.000

phenotype terms over a directed graph, with semantic similarity by information content

2,8 M

clinical variants over the GRCh38 reference assembly

3.500

gene-disease associations with their curated validity level

What makes it unique

Every recommendation is traceable to an explicit rule. Exomiser and Amelie — the field's references — base their prioritisation on models that are hard to audit: they give an ordered list, but not a chain of reasoning a laboratory can defend before an auditor.

Here the criterion is readable code: which phenotype term weighed, which disease profile it was compared against, which variants support it and at what review level. The same input gives the same output, and that output can be read line by line.

01 The state of the art

The decision is made by hand, and the bias leaves no trace.

Faced with a complex presentation, someone has to decide which test to order. That means crossing phenotypes with variant databases, gene-disease validity and published profiles: a manual process, exposed to omissions and hard to reconstruct afterwards.

Tool

How far it goes

What Helion adds

Exomiser, Amelie

They prioritise genes and variants by combining phenotype and evidence, with good published performance.

Delivers the reasoning chain alongside the ranking: every decision traceable to an explicit rule with its code, documentable before an accreditation body.

Manual search of the clinical databases

It is what the geneticist does today, and with expert judgement it works.

Leaves the judgement to the geneticist and takes the searching away: it goes through the databases exhaustively and records, reconstructably, why what was ruled out was ruled out.

Closed commercial panels

They cover common presentations with manufacturer validation and a fast turnaround.

Answers also when the presentation fits no panel, and quantifies what escalating to exome or genome adds, so the clinician's decision has support.

Asking a language model

It gives an immediate answer with the appearance of reasoning.

Is reproducible and auditable: explicit rules, no ML, with the version of every database declared — the condition for a result to enter a regulatory file.

02 The evidence layers

Not everything published weighs the same.

The review level of a variant in the reference clinical database decides which layer it enters. Mixing the three layers into a single list is what produces recommendations that cannot be defended.

Layer 1 · clinical

Variants with solid review

Those the reference clinical database marks at the highest review level. This is the evidence on which a recommendation can be made without qualification.

Layer 2 · supportable

Partial evidence, declared

Variants with intermediate review: they enter the reasoning, but flagged, and are never presented with the weight of the previous layer.

Layer 3 · exploratory

No review, with its validity level

Unreviewed variants accompanied by the gene-disease validity level. They keep a hypothesis in view; they do not justify a test.

The phenotype specificity meter

Before recommending anything, the system measures whether the terms entered discriminate enough. A presentation described with three generic terms does not allow prioritisation, and saying so is more useful than returning a ranking that looks informative and is not.

The four domains, in order

Germline Mendelian is the built domain. Somatic, infectious — where it will connect with metagenomic analysis — and complex multifactorial are declared evolution, not current functionality.

03 How it works

From phenotype to plan, with the why behind each step up.

Three plans with their cost and their scope: targeted panel, exome or whole genome. Escalation is justified by gene coverage of the differential, not by preference.

Phenotype capture

The terms of the clinical presentation are read over the official ontology, which is a directed graph: a term inherits the meaning of its ancestors, and that matters when comparing.

Specificity measurement

Before prioritising, the system quantifies whether the terms entered discriminate. If they do not, it says so instead of producing a worthless ranking.

Comparison with disease profiles

Semantic similarity by information content against the published phenotype-disease profiles, to build the differential.

Gene prioritisation

Crossing with the clinical variants and the gene-disease validity curations, split across the three evidence layers.

Testing plan

Panel, exome or whole genome according to the gene coverage of the differential and the available evidence, with the justification and estimated cost of each option.

Plan A

Targeted panel

The genes of the most likely diseases. The cheapest and fastest when the differential is concentrated.

Plan B

Exome

All coding regions. The option when the differential is broad but still Mendelian.

Plan C

Whole genome

Maximum recall and maximum cost. Proposed when the coverage of the previous options leaves part of the differential out.

04 The contract

What goes in, what comes out, what it chains to.

Terms from the human phenotype ontology describing the patient's presentation, with the case reference. No direct identification: the case travels pseudonymised.

Out

Ranked differential diagnoses, candidate genes with their evidence level, a test recommendation with justification and cost estimate, and a complete, exportable audit timeline for the case.

Chains to

It is an independent subsystem by design: its knowledge is the official clinical databases, not the platform's microbiological corpus. The infectious domain is where it will connect with metagenomic analysis.

05 Technical sheet

Input

Terms from the human phenotype ontology with the pseudonymised case reference.

Output

A ranked differential, candidate genes by evidence layer, the recommended testing plan with justification and cost, and an exportable audit timeline.

Auditability

Explicit code end to end. Neither machine learning nor generative models in the decision chain: every step is a readable rule and the same input gives the same output.

Database versioning

Every snapshot of the clinical databases is recorded with its source, its version and its record count, so an old case can be reinterpreted knowing which knowledge resolved it.

Domain built

Germline Mendelian. Somatic, infectious — with a connection to metagenomic analysis — and complex multifactorial are declared as evolution, not as present functionality.

Limit · updating

Updating the clinical databases is manual and not yet scheduled. Since the version travels with the result, an analysis always states which snapshot resolved it.

Limit · cost

The cost estimate for each plan comes from a fixed catalogue that does not track the market. It serves to compare plans against one another, not as a quotation.

Limit · personal data

The handling of identifying data is not formalised in the module: pseudonymisation and data protection compliance are an open, priority requirement before any real clinical use.

Knowledge bases: the Human Phenotype Ontology, with its disease annotations; NCBI ClinVar over the GRCh38 assembly; and ClinGen's gene-disease validity curations. Semantic similarity by information content after Resnik. Every database snapshot is versioned with its date and its record count.

A laboratory cannot defend a black box.

Request access See the twelve modules

BIOTECNO.org · Vitoria-Gasteiz EU + Codex