Methods and limits

How it works, and what it will not tell you.

Public data enter the pipeline below. Its limits live here.

71,378qualifying ClinVar nonsense variants
70,376placed on a transcript and scored
20260715ClinVar release
01

The pipeline

Two lanes. One touches raw reads, one never does.

Two chains that share no inputs. Keeping them apart is what lets a figure say which kind of number it shows.

The measurement line

~18 GB of raw sequencing

Public ribosome-footprint and matched RNA sequencing, taken from raw reads through to positions on the genome resolved to a single codon.

  1. SRA toolkitstages the public sequencing runs, verified against their checksums
  2. FastQCreads the raw quality: base scores, duplication, adapter contamination
  3. cutadapttrims sequencing adapters and drops reads outside the usable length range
  4. STARaligns short footprints to the human genome across splice junctions
  5. samtoolssorts, indexes and filters the alignments
  6. riboWaltzfinds each footprint's reading position and checks the three-nucleotide rhythm
  7. featureCountscounts transcripts in the matched RNA data, for baseline abundance
  8. MultiQCcollects every quality metric above into one report
  9. pigzdecompresses the read archive in parallel

The prediction line

70,376 scored variants

Variant databases and genomic sequence context. This lane never touches a raw read, and runs on a laptop.

  1. cyvcf2parses the ClinVar variant file
  2. Pydantic + Panderaenforces the data contracts that stop coordinates drifting between modules
  3. Biopythoncodon tables, translation and the sequence vocabulary
  4. bioframeoverlaps genomic intervals — chosen over the alternative after it needed 13 GiB
  5. statsmodels + SciPyfits the readthrough model and runs the resampling behind every interval
  6. aenmda published rule engine for whether the transcript survives
  7. NMDetective-AIa deep model on the same question, run on a GPU as an independent third opinion
  8. BLOSUM62scores how chemically similar the inserted amino acid is to the original
  9. Editor panelBE4max and ABE7.10 geometry against SpCas9 targeting rules

Underneath both

one lockfile

The parts that make a number reproducible rather than merely produced.

  1. Nextfloworders the steps and resumes them; the scientific decisions stay in Python
  2. Pixione lockfile across Python, R, the JVM and the command-line tools
  3. R + caretthe independent implementation the Python model is checked against
  4. matplotlibthe report figures, drawn from the same tables and palette as this site
02

Reading a number

Four kinds of number, never blended.

Every figure carries a marker for where its numbers came from. A measurement and an estimate are never combined.

Predictions are not resultsModel estimate from reporter data—not a patient, cell line, or tissue measurement.
Counts are variants, not peopleCounts describe ClinVar variants, never people or population burden.
Absence is not evidenceMissing records and unreachable placements are unknowns, not proof of absence.
03

Denominators

Every row accounted for, including the useless ones.

A count is only meaningful against the set it came from. This is where the variant-condition rows went before any coverage figure was drawn.

A quarter of the rows name no condition at all

Variant-condition rows by how completely they map to a real condition

rule
Every variant-condition row accounted for · 120,478 totalORANGE WOULD HAVE BECOME THE COMMONEST CONDITIONmapped to a real condition: 80,558 rows (66.9%)USABLE80,55866.9%mapped to a real conditioncondition not provided: 27,176 rows (22.6%)PLACEHOLDER27,17622.6%condition not providedidentifier but no name: 9,241 rows (7.7%)PARTIAL9,2417.7%identifier but no nameno cross-reference at all: 3,503 rows (2.9%)
Details & limits
The orange slice is a placeholder the database uses for “condition not provided”. It looks like an ordinary identifier, and counting it would have made a label meaning nothing into the most common disease in the set.
04

Stated limits

The caveats that belong to the data.

Carried in the exported data itself, so a figure and its limit cannot drift apart.

entities

A MedGen concept is a ClinVar condition — a disease, but possibly a finding, susceptibility, or broad umbrella label. These are condition entities, not verified diseases.

coverage

Model coverage is exact restoration — a suppressor design decodes the stop and reinserts the native residue. Not a therapeutic or clinical claim.

three metrics

reach = at least one eligible variant covered; covered_fraction = covered over all eligible variants for the entity; complete = every eligible variant covered.

cross references

OMIM and Orphanet are ClinVar-provided cross-reference identifiers only — they add no prevalence, treatment availability, or unmet-need evidence.

unmet need

No unmet-need claim is made; that needs a treatment-status source not used here.

nmd

NMD escape is two hand-rolled rule predictors — the 50-nt guideline and the fuller Lindeboom rule set — checked against aenmd (the published rule tool) and NMDetective-AI (a deep model whose continuous efficiency is not thresholded into a verdict here); predNMD is not in. A rule disagreement is where the fuller rules turn the classification, not evidence of which rule is correct.

confirmation scope

confirmatory, one dataset, one laboratory and protocol family

05

Words

The vocabulary, in plain terms.

Defined once here rather than glossed differently on each page.

Nonsense variant

A single-letter change that turns an amino-acid instruction into a stop instruction, so the protein ends early.

Readthrough

Persuading the ribosome to ignore that premature stop and keep going, without changing the DNA.

Nonsense-mediated decay

A quality-control system that destroys messages carrying a premature stop. If the message is gone, there is nothing left to read through.

Suppressor tRNA

An engineered adapter molecule built to recognise one stop codon and deliver one specific amino acid.

Base editing

Rewriting a single DNA letter in place. Whether an editor can be positioned on a given stop is decidable from sequence.

Ribosome profiling

Freezing cells, digesting the RNA that is not protected by a ribosome, and sequencing what survives — a snapshot of where ribosomes were.

Interval

The range a number could plausibly take. Two numbers whose intervals overlap have not been shown to differ.

Scoreable

A variant that could be placed on a reference transcript and given a prediction. Variants that could not are counted, not dropped.