Traceback

One blood draw. Three computed cfDNA signals. Every claim traceable.

The public evidence demo turns Oxford Nanopore reads into fragment-length, methylation cell-origin, and chromosome-dosage evidence, then uses a bounded AI reviewer to check whether the written interpretation is supported.

Open the live demo · View the repository · Demo walkthrough · Algorithms

Research prototype only. It does not diagnose cancer or replace a validated clinical assay.

Choose a local workflow

Workflow	Use it for	Record and qualification
Signed local runner	A BAM-to-record journey for fragment length; optional cell-origin and ichorCNA copy-number analyses are described in the operator guide.	Development-signed, local, unqualified records.
Traceback Lab	Exploratory measurements on stored ONT or supported prealigned Illumina BAMs, with local sample and cohort views.	Unsigned results with execution provenance; research use only.
Public evidence demo	Inspect three fixed aggregate readouts and a recorded or live bounded AI evidence review.	A demonstration, not a new analysis of your BAM.

Lab includes read quality and depth checks; fragment, methylation, end-pattern, and microbial measurements; exploratory copy-number and cohort comparisons; and input-scoped prenatal or transplant calculations where an operator supplies matching controls. Each method has its own platform, depth, asset, and tool requirements. Run traceback lab algorithms for the registered catalog and see the plain-language algorithm explainer for what each method may teach you and the Lab algorithm guide for requirements and limits.

What the public demo shows

• Fragmentomics: a deterministic BAM-derived read-length distribution with explicit filters, denominator, mode, median, and long-fragment fraction.
• Cell origin: CpG methylation calls classified against Loyfer markers, followed by count-weighted NNLS deconvolution and seeded bootstrap intervals.
• Chromosome dosage: an experimental whole-chromosome screen using high-confidence read starts in fixed 5 Mb bins.
• AI evidence review: the model can select only registered checks and must cite the exact result or source passage behind its conclusion.
• A real catch: the reviewer identifies that the report's updated aligned-span method conflicts with its fixed 45 bp subtraction instructions.

The third readout is deliberately narrow: it is not ichorCNA, does not estimate tumor fraction, and cannot resolve focal or subclonal events.

The diagram below describes the public evidence demo, not the Lab run path.

flowchart LR
    A[MinION modBAM] --> B[Fragment-length summary]
    A --> C[CpG methylation extraction]
    A --> J[Whole-chromosome dosage screen]
    C --> D[Loyfer UXM classification]
    D --> E[NNLS cell-mixture estimate]
    B --> F[Registered evidence]
    E --> F
    J --> F
    G[Source passages] --> F
    F --> H[Bounded AI review]
    H --> I[Cited assessment + limitations]

Run locally

Requires Python 3.11 and uv.

uv sync
uv run pytest
uv run streamlit run app.py

To take your own Nanopore BAMs through the signed CLI (align, reference register, preflight, run, catalog import, serve), start with the short scientist quick guide. The full reference is Real local BAM (unqualified) in the operator guide. It needs samtools (brew install samtools) to index the inputs. These local records are unqualified and not for clinical use. For the separate, unsigned multi-algorithm workflow, follow the Lab guide.

The checked-in public demo bundles contain aggregate measurements and provenance only. They do not contain BAMs, read IDs, local paths, source documents, or credentials.

Review modes

• Live: the default local mode calls the Bedrock OpenAI-compatible Responses endpoint after the user presses Run check.
• Recorded: set TRACEBACK_DEMO_REPLAY=1 to replay validated assessments without AWS credentials. The UI labels the result as a fixture and does not claim a live provider call.

For live review, configure the standard AWS credential chain and optionally:

Variable	Default
BEDROCK_MODEL_ID	openai.gpt-5.5
BEDROCK_REGION	us-east-2
BEDROCK_MAX_TOKENS	1024

Regenerate cell-origin results

Private inputs belong under ignored data/local/ paths:

• hg38 FASTA and minimap2 index
• Loyfer U250 regions, markers, atlas, and healthy-plasma reference workbook
• an aligned modBAM with valid MM, ML, and MN tags, or a normalized modkit extract calls TSV

TRACEBACK_FRAGMENT_HASH_SALT='local-private-value' \
  uv run python scripts/regenerate_cell_origin.py \
  --extract-tsv data/local/cell-origin/calls.normalized.tsv

The validated result is written to data/local/cell-origin/result.json and takes precedence over the public aggregate bundle.

Regenerate the experimental chromosome-dosage result:

uv run python -m scripts.regenerate_copy_number \
  --bam data/local/cell-origin/full.hg38.sorted.bam

Documentation

• Browse all documentation
• Scientist quick guide: your BAMs to a browsable record
• Operator guide: every command, error code and recovery step
• Lab guide: every cfDNA algorithm on a kit BAM, unsigned, research only
• Lab algorithm explainer: what each measurement may reveal
• Canaries: the nightly real-data and CI checks
• Three-minute demo
• 75-second recording script
• Algorithm and evidence design
• Railway deployment

Privacy boundary

Never commit .env files, credentials, sequence files, BAMs, read IDs, source reports, local paths, or patient/sample identifiers. Tests are fully offline. Only bounded excerpts and aggregate check results may enter model prompts.