Reading the Genome

HELENA BIOINFORMATICS

RUO · MMXXVI

READING THE GENOME
The genome is an inherited story.
Folklore reads it.

SOFIA · BULGARIA

A

CLINICAL GENOMICS PLATFORM FOR VARIANT INTERPRETATION

VOL. I

FOLKLORE

AN

INHERITED

STORY

Folklore is knowledge passed from
generation to generation. Never
formally written, yet carrying the
identity of a whole nation.
The genome is exactly that kind of inherited memory, a story recopied at every
generation, with an occasional new addition. Every variant is a line added along
the line of descent, a heritage carried through time.

The folklorist does not invent the tale. Only reads it.
Folklorists read, interpret and systematise them, gathering scattered oral
knowledge and drawing out meaning, structure and significance.
That is exactly what Folklore does: reading the genome and interpreting the
results. The platform takes called variant data, runs it through a deterministic
interpreter backed by dozens of curated evidence sources, and returns every
variant sorted and classified: by pathogenicity across the five ACMG classes, and
by closeness to the patient’s phenotype.

We do not invent the story. We read it.

FOLKLORE

002

FOLKLORE

THE

PLATFORM

One story.
Eight ways to read it.
Each module is one reading of the genome, and together they turn a raw VCF into
a prioritised, review-ready result.
Variant classification

V 3.39.1

Millions of variants → five ACMG classes, ranked by priority.

Mitochondrial DNA
Rule-based mtDNA under the MMDWG 2020 framework.

Structural variants · SV / CNV
Deletions and duplications, Riggs 2020, five tiers.

Phenotype matching
HPO terms → clinical priority, five tiers.

Clinical screening
Review order, not pathogenicity, Tier 1–4.

Cohort analysis
Groups of samples → burden, enrichment, candidates.

Family & trio analysis
Child–mother–father → the origin of a variant.

Literature evidence
Variant → relevant, anchored publications.

C LINGEN MT-VCEP

R IGGS 2020

H PO · 0–100

7 COMP · 9 BOOSTS

F ISHER · SKAT-O

T RIO · DE NOVO

P UBMED-ANCHORED

The modules sit within a navigable genome browser, an on-premise AI assistant, a review
board, and a tiered clinical report. Pathogenicity is decided by the deterministic
classifier, not by AI. Folklore is a decision-support tool for the qualified geneticist.
(Research Use Only.)
I

· THE PLATFORM

003

FOLKLORE

INTERPRETATION

·

THE

ENGINE

Nuclear variants, under ACMG/AMP.
Every variant is annotated through Ensembl VEP, then classified against the 2015
ACMG/AMP guidelines as deterministic, rule-based logic.
Folklore implements the ACMG/AMP 2015 framework (Richards et al.) through the
Tavtigian Bayesian point system, calibrated to the ClinGen Sequence Variant
Interpretation recommendations.
Of the 28 evidence criteria, 19 are computed automatically against a curated
reference layer of 45 databases: gnomAD, ClinVar, dbNSFP, SpliceAI,
AlphaMissense, HPO, UniProt, and more, billions of records held locally on EU
infrastructure. The remaining 9, which need segregation, functional, or de novo
evidence, stay with the reviewing geneticist. Each applicable criterion carries a
weight and the combined score maps onto a single verdict.
Every classification is recorded with its evidence and version, so the result is
reproducible, auditable and resolves to one of five classes:
PATHOGENIC

LIKELY PATHOGENIC

VUS

LIKELY BENIGN

BENIGN

Every classification traces back to the rule and the source behind it.

II · INTERPRETATION

004

FOLKLORE

WHAT

YOU

SEE

Every result carries its evidence.
The variants arrange by gene and by ACMG priority, Pathogenic through Benign.
Along the top, five class cards hold the live totals for each class and filter the list
beneath them. A gene search narrows it to one.

3

7

214

1.1k

18k

PATHOGENIC

LIKELY PATH.

VUS

LIKELY BENIGN

BENIGN

#1

BRCA1

c.5266dupC · p.Gln1756fs

#2

MYH7

c.1988G>A · p.Arg663His

#3

TTN

c.21A>G · p.Ser7=

PATH

DE NOVO?

CURATED

LIK. PATH

VUS

Each variant carries its whole case. The ACMG criteria that fired, shown as colourcoded evidence. ClinVar significance and review stars, with any ClinGen expertpanel assertion. In-silico predictions, gnomAD frequency, conservation and
constraint. The sequencing quality behind the call. Every value traces to the
record that produced it.

The record is always one step beneath the result.

III

· WHAT YOU SEE

005

FOLKLORE

MITOCHONDRIAL

DNA

Mitochondrial variants, under MMDWG.
Mitochondrial variants are classified under MMDWG 2020 (McCormick), an
independent module running alongside nuclear ACMG/AMP. Every variant in the
output carries an explicit framework-provenance label, so the geneticist always
knows which framework produced the class. Of the 28 nuclear criteria, seven are
excluded for the mitochondrial genome with verbatim rationale. The rest are
respecified with mtDNA-specific thresholds: 11 automated, 10 curated with
strength tiers. Combining rules from Richards 2015 are preserved unchanged.
A curation panel exposes the ClinGen mtDNA VCEP criteria for the reviewer,
alongside heteroplasmy assessment and haplogroup-aware interpretation
(APOGEE2, MitoTIP, MITOMAP). The classification follows the genome’s biology:
PP1 cannot contribute when a variant is homoplasmic across all maternal
members (VCEP §5.2.7). The panel then shows the pipeline class and the curated
class side by side:
P IPELINE CLASS
VUS

W ITH CURATION

→

PM2 · PP3

LIKELY PATH.

+ PS3 functional · heteroplasmy 78%

Two frameworks, each named on every variant.

IV · MITOCHONDRIAL DNA

006

FOLKLORE

STRUCTURAL

VARIANTS

Structural variants, under ClinGen/ACMG.
Copy-number variants are classified under the ClinGen/ACMG 2020
semiquantitative point system (Riggs 2020): losses on Table 1, gains on Table 2,
as two distinct point systems. Dosage sensitivity comes from pHaplo / pTriplo
(Collins 2022) and overlap with established haploinsufficient (HI) and
triplosensitive (TS) regions from the ClinGen dosage map. Population frequency
comes from gnomAD-SV. Points sum through a single five-tier ladder into the
same P / LP / VUS / LB / B labels.
chr16:29,580,000–30,180,000

del · 600 kb

S V SPAN

GRCh38

DELETION

G ENES

C LINGEN · HI
A structural variant renders as a span (deletion red, duplication blue) over the genes it crosses. A
haploinsufficient region overlap and the count of genes and exons feed the Riggs score.

16p13.11 deletion · 600 kb

Pathogenic

The CNV classification (Riggs) panel shows the total score and tier with each criterion met or not met:
established-HI overlap, gene count, dosage, population frequency, alongside copy state, breakpoint precision
and paired/split-read support.

Its own tab, its own metric, landing on the same five-class scale.

V

· STRUCTURAL VARIANTS

007

FOLKLORE

PHENOTYPE

MATCHING

Variants, ranked by phenotype.
ACMG classification tells you how pathogenic a variant is. It does not tell you
which variant explains your patient. Phenotype matching answers that second
question.
The patient’s HPO terms are correlated with gene–disease associations by
semantic similarity, scored 0–100 for each term, and folded into a clinical priority
score. Genes are ranked and placed into five tiers (Tier 1, Tier 2, Incidental
Findings, Tier 3, Tier 4). Each variant shows its strongest HPO-term matches.
THE CORRECT BEHAVIOUR

A VUS with strong phenotypic relevance ranks above a Pathogenic variant for an
unrelated condition. Relevance to this patient, not pathogenicity in the abstract.
The HPO terms entered at intake become filters over the gene list. A selected
phenotype collapses the list to the genes that match it.
#1

SCN1A

HPO match 0.92 · seizures, DEE

TIER 1

#2

KCNQ2

HPO match 0.71

TIER 2

#3

TTN

HPO match 0.08 · unrelated

TIER 4

The story is read for the patient in front of you, not for a textbook.

VI · PHENOTYPE

008

FOLKLORE

CLINICAL

SCREENING

Clinical screening, tier by tier.
Screening sets the review order from the patient’s full context: age, sex, ethnicity,
family history, sample structure, and the chosen panels.
A seven-component score (constraint, deleteriousness, phenotype, dosage,
consequence, compound-het, age relevance) is adjusted by up to nine clinical
boosts (ACMG strength, ethnicity, family history, de novo, and more), then placed
into four tiers. Each gene carries a clinical-actionability signal: immediate,
monitoring, or future. Screening modes tune the run for neonatal, paediatric, adult
diagnostic, or carrier testing.
T1

RYR1

score 0.86 · malignant hyperthermia

IMMEDIATE

T1

BRCA2

score 0.79

MONITORING

T2

MYBPC3

score 0.55

FUTURE

Screening sets the order of reading. Pathogenicity stays with ACMG.

VII

· SCREENING

009

FOLKLORE

COHORT

ANALYSIS

Cohort analysis, across samples.
A study gathers classified samples into one matrix and runs population-level
statistics, from two samples upward.
Gene-burden testing reports Fisher exact p, FDR q and odds ratios (with CMC and
SKAT-O), read off a volcano plot. A loss-of-function pass adds pLI and LOEUF.
Pathway enrichment and cross-sample compound-heterozygote detection round
out the signal. A candidate-gene nomination fuses seven independent lines of
evidence (burden, pLoF, disease, constraint, pathway, GWAS, compound-het) into
a ranked list and an evidence-matrix heatmap. Where samples disagree on a
class, the discordance is surfaced, not hidden.
#1

LDLR

Fisher p 3.1e-5 · OR 6.4 · 5/7 evidence axes

SIGNIFICANT

#2

APOB

Fisher p 8.0e-4 · OR 3.2

FDR Q<0.05

#3

PCSK9

pathway + GWAS + constraint

CANDIDATE

The same reading, now across a population, with every disagreement between samples
shown.

VIII

· COHORT

010

FOLKLORE

FAMILY

&

TRIO

Inheritance, read from the trio.
A trio (proband, mother, father) adds what a single sample can never carry: the
parental genotypes. Three inheritance workflows run on them: de novo,
compound heterozygous, and segregation.
De novo. The proband carries the alternate allele; both parents are reference. A
per-member table shows the genotype, depth, quality and class behind each call.
A clinical-grade filter scopes the drill.
M EMBER

GENOTYPE

DP · QUAL

Proband

het (0/1)

42x · 99

Father

hom_ref (0/0)

38x · 99

Mother

hom_ref (0/0)

45x · 99

Compound het. Two variants in one gene, one inherited from each parent, are
phased from the trio genotypes: paternal when father is het and mother reference,
maternal for the reverse. Only a trio-phased pair is called in trans.
PATERNAL

GENE · variant A
father het · mother 0/0

IN TRANS

↔

MATERNAL

GENE · variant B
father 0/0 · mother het

Cis-ambiguous pairs are shown as such. The trio phases only what the genotypes support.

IX · FAMILY & TRIO

011

FOLKLORE

FAMILY

&

TRIO

Segregation, and its ceiling.
Segregation testing computes a LOD score for each variant. But a trio (one
proband and two parents) has too few informative meioses. The maximum LOD
attainable is about 0.30, which sits exactly at the ClinGen SVI 2016 supporting
evidence threshold.
THE CEILING

A trio reaches PP1_Supporting at most, never PP1 Moderate or Strong. The platform
states the ceiling and recommends extending the pedigree for stronger evidence,
rather than borrowing a strength the data cannot support.
Before any of this, cross-sample QC checks that the family is the family: PLINK
identity-by-descent relatedness catches sample swaps, unexpected
consanguinity, and duplicates. A critical alert holds the line: inheritance findings
should not drive clinical interpretation until the pedigree checks out.
MAX LOD 0.30

Trio ceiling = ClinGen SVI 2016 supporting threshold. Extend the pedigree for more.

CROSS-SAMPLE QC

PLINK IBD (PI_HAT): sample-swap, consanguinity and duplicate detection.

The limit is named, not worked around. A stated ceiling holds where a borrowed strength
would not.

X

· FAMILY & TRIO

012

FOLKLORE

LITERATURE

EVIDENCE

Literature, anchored to the variant.
A local, genetics-filtered PubMed mirror with pre-extracted gene, variant and
phenotype mentions turns literature search into a sub-second, variant-anchored
query.
Publications rank by a six-component model aligned to ACMG evidence
categories, each badged by strength: strong, moderate, supporting, weak. An
exact-match badge flags a paper that names this variant. A functional badge flags
experimental data. The gene ranking fuses clinical priority and literature relevance
(60/40), and every hit keeps full PMID / PMC / DOI traceability.
EXACT MATCH

The paper reports this exact variant, not just the gene.

SIX-COMPONENT SCORE

Phenotype, publication type, gene focus, functional data, variant, recency.

TRACEABLE

PMID / PMC / DOI on every citation, evidence you can open.

Every citation opens to its source. The folklorist names where the claim came from.

XI · LITERATURE

013

FOLKLORE

INTERPRETATION

&

REPORT

AI assistant, under the deterministic gate.
An on-premise AI clinical assistant, a self-hosted open-weight model on EU
infrastructure where no external AI API ever sees patient data, integrates
classification, phenotype and literature into a structured clinical narrative. You can
ask it questions in plain language: it runs queries against the session database and
charts them, searches the literature, and explains which ACMG criteria support a
class.
The division of labour is the whole point: the AI drafts interpretation text, the
deterministic classifier decides pathogenicity. The assistant carries its own
caution. The finished output is a single tiered PDF: Tier 1 (actionable) and Tier 2
(potentially actionable) variants, each with a complete evidence chain, ready for
the geneticist’s review and sign-off.
THE DIVISION OF LABOUR

AI writes the sentence; rules assign the class. The geneticist makes the call. Research
Use Only. Not for diagnostic use without professional review.

CLINICAL REPORT · TIERED · EVIDENCE CHAIN
TIER 1

BRCA1

c.5266dupC · Pathogenic · PVS1 PS4 PM2 · ClinVar ★★★

TIER 2

MYH7

c.1988G>A · Likely path. · PM1 PM2 PP3

RUO

Reviewed & signed: ________________ · qualified clinical geneticist

A draft is a draft until a human signs it. Folklore never forgets which is which.

XII

· INTERPRETATION & REPORT

014

FOLKLORE

GENOME

BROWSER

The genome browser.
A navigable per-session window over the patient’s GRCh38: variants as ACMGcoloured lollipops sized by clinical salience, structural variants as spans, genes as
MANE models with strand direction, ClinGen dosage regions, alongside evidence
lanes for BayesDel, SpliceAI, pext, conservation, ClinVar and constraint. Zoom,
pan, jump to a locus, an HGVS change, or a gene. The browser adds no reading of
its own. It shows what was already computed.
chr17:43,044,295–43,170,245

MANE Select

GRCh38

V ARIANTS

S TRUCTURAL

DELETION

G ENES · MANE
C LINGEN ·
D OSAGE
Variants as lollipops by ACMG class, height tracks salience. Structural variants as spans, genes with
direction along MANE. A graph mode reads insertions and deletions as divergence. Where phase is not
measured, it says “inferred”.

BRCA1 · c.5266dupC (p.Gln1756fs)
PVS1

PS4

Pathogenic

PM2

Click a feature for the full readout: fired ACMG criteria as badges, BayesDel and SpliceAI, the pext level,
MANE, and the ClinVar stars. Only what has a value is shown.

One coordinate frame holds every finding, each shown where the analysis placed it.

XIII

· GENOME BROWSER

015

FOLKLORE

THE

REVIEW

BOARD

The review board.
A case carries the exact classifier version it was run with, so every result stays
tied to the rules that produced it. The review board is where a qualified geneticist
takes over from the pipeline.
A reviewer can reclassify any variant, with a written justification that is attributed
and revertable, or leave threaded notes for a colleague. The reviewer’s call always
overrides the pipeline’s. A case-level outcome (solved, likely solved, unsolved) is
tracked with its full change history.

Every reclassification is signed, sourced, and reversible. The audit trail is the product.

XIV

· THE REVIEW BOARD

016

FOLKLORE

PROCESS

From VCF to clinical report.
1

Ingest & build detection

2

Quality control

3

Annotation

4

Classification

5

Prioritisation

6

Report

G RCH37 / 38

VCF in; automatic GRCh37/38 detection and liftover to GRCh38.
C LINVAR-SAFE

Quality filters; ClinVar-pathogenic variants protected from being dropped.
4 5 DATABASES

Against the reference layer: 45 databases, billions of rows.
A CMG · MMDWG · RIGGS

ACMG/AMP for nuclear variants; MMDWG for mtDNA; Riggs for CNVs.
3 –20

Phenotype matching and screening; down to 3–20 candidates.
R UO

A clinical-interpretation draft for review by a geneticist.

Every stage traceable, every stage reproducible.

XV · PROCESS

017

FOLKLORE

SPECIFICATION

The specification.
Genome builds

GRCh38 native; GRCh37 auto-lifted

Variant classes

SNV, indel, mtDNA, SV / CNV (Riggs 2020), all active

Reference layer

45 databases; SpliceAI ~3.4B rows, gnomAD ~909M, dbNSFP ~80.7M,
ClinVar ~4.1M

Classification

ACMG/AMP 2015, deterministic; MMDWG (ClinGen mt-VCEP); Riggs
2020; 5 classes

Genome browser

Navigable GRCh38 view; variants, SVs, genes, dosage; evidence
drawer; graph mode

Speed

Analysis pipeline ~7–12 min (WGS)

EU data residency

Helsinki, Finland; data never leaves the EU; GDPR Art. 9

Audit & security

SHA-256 append-only chain; encryption in transit and at rest;
RBAC

Artificial intelligence

Self-hosted open-weight model (Qwen) on EU infra; no external
AI API

Status

Research Use Only. A decision-support tool, not a diagnostic
device

EU DATA RESIDENCY

GDPR ART. 9

SHA-256 AUDIT

RBAC

RUO

One reading, traceable down to the last rule.

XVI

· SPECIFICATION

018

FOLKLORE

FOR

WHOM

The tale.
Every capability carries both a clinical proof and a signal of scale.
THE READING

THE MOAT

Clinical-grade, traceable, reproducible.

Category. A European clinical-genomics

Five classes and an audit trail on every

software pure-player, no wet lab, no

action. A decision-support tool that leaves

capex. Moat. Curation, not code: a

the last word to the qualified geneticist.

reference layer (45 databases, billions of

R are disease

N ewborn screening

C arrier screening

rows) and a discordance-gated classifier,
built over years. Defensibility. Live rulebased mtDNA and live Riggs CNV,
capabilities others treat as afterthoughts.
Sovereignty. Self-hosted EU AI where no
external API sees patient data; GDPR by
design.
E uropean software
R eference layer

S elf-hosted AI
E U data

The tool reads the story. The last word is the human’s.

XVII

· FOR WHOM

019

FOLKLORE

VALIDATION

Validation, against an external lab.

0

68.4%

3

BENIGN ↔ PATHOGENIC
INVERSIONS (OPPOSITE
POLES), ACROSS ALL THREE
COHORTS

EXACT FIVE-CLASS
AGREEMENT · COHORT 2,
CLASSIFIER V3.28.0, 19
EVALUABLE

REAL-WORLD COHORTS ·
20 / 20 / 57 CASES,
ACROSS CLASSIFIER
GENERATIONS

Five-class agreement sits in the 60–75% range that the literature reports between
laboratories (Amendola 2016 ~66%, Harrison 2017 72–76%). Where Folklore
differs, it differs by one step and in one direction: it returns VUS rather than
asserting a pathogenicity it cannot fully evidence. No benign becomes
pathogenic, nor the reverse.
WHAT WE HAVE VALIDATED

WHAT WE HAVE NOT

Concordance against an external clinical

Large-scale prospective outcomes, multi-

laboratory across three real-world

laboratory benchmarking, SV/CNV

cohorts. Zero opposite-pole inversions. A

phenotype correlation. The claim does not

conservative skew toward VUS.

exceed the evidence.

External clinical laboratory as a peer comparator, not a reference standard. An ongoing,
iterative real-world audit.
XVIII

· VALIDATION

020

FOLKLORE

REGULATORY

STATUS

Regulatory status.
Folklore is a Research Use Only tool. It is a clinical decision-support system, not a
diagnostic device, and it is not a CE-marked IVD.
Every classification Folklore produces requires review and confirmation by a
qualified clinical geneticist. The platform drafts and orders evidence. The clinical
decision belongs to the reviewing professional.
Folklore does not replace clinical judgement, established laboratory procedure, or
the reporting standards of the laboratory that uses it. Results are intended to
support qualified interpretation, not to stand alone.
Research Use Only

Not a diagnostic device

Not a CE-marked IVD

Geneticist sign-off

The tool reads the evidence. The clinical decision is the geneticist’s.

XIX

· REGULATORY STATUS

021

HELENA BIOINFORMATICS EOOD

RUO · MMXXVI

The genome is an inherited story. Folklore reads it.

A

CONTACT@HELENA.BIO

HELENA BIOINFORMATICS PLATFORM · SOFIA, BULGARIA

HELENA.BIO