How it works · stage by stage

One Run, five stages.

A researcher has a genomic question, not a bioinformatics team. Traceome runs the pipeline in five stages: Plan, Align and call, Investigate, Findings and Report. The researcher approves at three points, and every step lands in one record, the Journey. Each stage in plain words first, then the technical detail for those who want it. The example is development Run H on HG001, a public reference sample; the demo Run (#457) replays the same sample with a larger request R21. Screenshots are the GUI replay of development Run Q, whose answers equal Run H's item for item R5.

The input · Researcher

Ask

A researcher asks a question, cites a paper or names the variants (or both), and uploads the raw DNA reads.

Technical detail

Run H registered a question, a request of 24 rsIDs and the HG001 paired FASTQ (1.91 GB + 1.95 GB). The 24 rsIDs are a test list, not a paper's: it covers edge cases and two IDs planted so they should not resolve. With a paper, the Run finds and reads it itself: from a bare citation, the paper reader pulled rs429358, rs7412 and rs1801133 out of Mullins 2020, each with its quote R12.

The GUI early in a Run
Early in the Run: the Journey on the right already shows what has been recorded. Run Q (#141), Qwen comparison baseline.
Stage 1 · Nemotron

Plan

A planning agent identifies the reference genome and retrieves everything the alignment needs, ready for NVIDIA Parabricks.

Technical detail

The planning_agent (NVIDIA Nemotron 3.5 Super VL on Nebius Token Factory) inspects the inputs, identifies GRCh38 through NCBI and compiles one complete Parabricks proposal R1. Its output is a proposal, never a fact, until the researcher approves it.

Three times · Researcher

You approve

Nothing consequential happens without the researcher. They approve three exact things: the compute plan, the Work items, and what the report may say. Each approval can be more than a yes.

Technical detail

Each approval binds the exact proposal or basis by its digest, and records what it authorises and what it does not establish R2. At a gate the researcher can also approve a subset, revise, amend a target, reject one, or stop and still get an honest report R17.

A recorded approval
An approval as the Journey stores it. Run Q (#141), Qwen comparison baseline.
Stage 2 · NVIDIA Parabricks on a Nebius H100

Align and call

Once the researcher approves the plan, NVIDIA Parabricks aligns the reads and calls the variants on one GPU. The output files are verified before anything builds on them.

Technical detail

pbrun germline 4.7.0 runs as a Nebius Serverless AI job on one H100 R6. Inputs go to a fresh Nebius Object Storage bucket; the BAM and VCF (260,016 records) come back in parts and are verified byte for byte before Investigation starts. A health step then measures the Run (reads mapped, depth, bases covered) and each Work item's coverage, and the report shows it R19. Reads can also come from staged Object Storage files, copied server-side R20.

Screenshot: coming from the Nemotron Run replay
Stage 3 · Nemotron

Investigate

An agent looks up each variant with tools such as NCBI and EVA, and turns loose names into exact Work items for the researcher to approve. One name can hide two different letter changes, or have been merged into another.

Technical detail

The investigation_planner calls fetch_refsnp, fetch_assembly_report and compile_proposal. rs4149056 sits at NC_000012.12:21178615 with two alternate alleles, so it becomes two Work items, T>A and T>C. An older ID the researcher also listed, rs60037639, had been merged into it by dbSNP; the same two items serve it. 24 rsIDs became 29 Work items; the 2 planted bad IDs stayed unresolved R3 R4.

A question for the researcher
One typed question for the researcher while the other targets continue. Run Q (#141), Qwen comparison baseline.
Stage 4 · Code + Nemotron

Findings

Each Work item gets exactly one answer from the sample's own data: observed, not carried, no VCF record, or unresolved. "No record" is never turned into "not carried". When the agent runs into a conflict, it brings it to the researcher, and only that item waits. It's a collaboration.

Technical detail

The VCF has a T>C record at rs4149056, genotype 0/1, depth 26: T>C is observed, T>A is not carried. The table_reader flags two more records within 100 bases. Across Run H: 17 observed, 6 not carried, 6 no VCF record R3 R9.

The Findings screen
Every requested variant with its outcome. Run Q (#141), Qwen comparison baseline.
Stage 5 · Nemotron

Report

An agent presents the results and writes a report of the findings in plain words. It may cite only approved facts, and every citation opens that fact in the Journey.

"…the exact T>C allele at 21178615 is carried (heterozygous 0/1), while the T>A target at that same position is not carried. [3b84b6f1…, 331defe6…, …]"Run H's explanation for rs4149056.
Technical detail

The report_explainer may cite only facts in the approved basis. A malformed output gets a short repair note, a bounded number of times, and is then refused R9.

Evidence for rs4149056
The evidence behind the answer. Report view: coming to the GUI replay. Run Q (#141), Qwen comparison baseline.
Under every step

The Journey

The Journey sits under all five stages. Every model call, tool call, approval and file is a fact, stored once and never changed. The researcher reads it while the Run happens and afterwards; the agents read it too. A finished Run is not a dead end: fork starts a new branch from it with zero new external calls, leaving the original untouched.

Technical detail

The Journey file: append-only, immutable typed facts, plus write-once files stored under the SHA-256 of their bytes. Run H: 1,834 facts of 65 kinds and 337 files, all re-verified R14. Reading it makes no contact and changes no byte R15; agents query it through search_journey R16. A finished Run can be forked from a closed point with 0 contacts R10.

Want the architecture, the agent tools, the Journey's data model and the measurements? Read the technology page