What we are building

Software that makes the boundaries of sequencing evidence visible.

Our research asks where sequencing results hold and where they stop holding. These are the products and services we think that research could become, with an honest note on how far along each one is.

The path

Measurement → evidence → disagreement → uncertainty → reliability → decision support.

Each direction takes on one part of that path. None of this is clinical decision support; the decisions we mean are research and assay decisions.

  1. 01MeasurementWhat could the workflow observe?
  2. 02EvidenceWhat was recorded, under which conditions?
  3. 03DisagreementWhere do technologies differ?
  4. 04UncertaintyWhat stays unresolved?
  5. 05ReliabilityWhere do results hold, and where not?
  6. 06Decision supportResearch decisions, with boundaries shown
Where each direction sits along the path from raw measurement to decisions. The directions are independent research programmes: none uses another's data, methods or results. Decision support here means research and assay decisions, not clinical care.

01

RNA Measurement Intelligence

Research Research. One completed, locked study; manuscript submitted. No product exists yet.

Problem
RNA-sequencing workflows do not observe every transcript equally. Whether a transcript is reliably measured depends on abundance, sequence, architecture, identifiability, mappability, library preparation, platform and quantification method. Those limits are usually assumed, not modelled.
Idea
A measurement-intelligence layer that estimates, for a given workflow, which transcripts are likely to be reliably observable, and flags classes that are hard to measure.
How it could work
  1. Describe a transcript by properties that are measurable before sequencing: length, exon structure, sequence composition, similarity to other isoforms, mappability.
  2. Relate those properties to workflow-specific observation outcomes learned from paired, replicated datasets.
  3. Report expected observability per workflow, with explicit uncertainty, rather than a single quality number.
  4. Only add a property (such as measured RNA structure) when it has been shown to add information. Our first study tested exactly that.
Scientific foundation
Locked K562 study: measured in-vivo RNA structure added negligible predictive value beyond abundance, sequence, architecture, identifiability and callability, for the prespecified endpoint. Details
Potential users
  • Assay-development teams
  • Sequencing laboratories
  • Genomics researchers planning RNA-seq studies
  • Technology developers comparing workflows
Groups we think could benefit. Not current customers.

Transcript

Properties

IllustrativeHow transcript properties and sequencing workflow shape what can be observedTranscriptIllumina short-readNanopore direct-RNAObservationabundantE1E2E3E4E5long transcriptstrongstrongclassBOTHrareE1E2E3E4E5long transcriptno readsweakmissingclassILLUMINA_ONLYmoderateE1E2E3E4E5long transcriptisoform B shares these exonsuncertainstrongclassDIRECT_RNA_ONLYmoderateE1E2short transcriptlow mappability: sequence repeated elsewhereno readsuncertainmissingclassINDETERMINATE

Plenty of molecules and unique sequence: both workflows see it clearly.

Structure is drawn because it was tested. In the K562 study, transcript-wide measured in-vivo structure added negligible predictive value beyond the other properties shown.

Illustrative, not study data. Outcome classes use the study’s vocabulary (BOTH, ILLUMINA_ONLY, DIRECT_RNA_ONLY, NEITHER, INDETERMINATE). Real observability depends on the specific protocol, depth, quantifier and thresholds. Modelling that dependence explicitly is the research direction.
Illustrative outcomes by transcript type
TranscriptIllumina short-readNanopore direct-RNAClassWhy
Abundant, distinctstrongstrongBOTHPlenty of molecules and unique sequence: both workflows see it clearly.
RareweakmissingILLUMINA_ONLYFew molecules. Deep short-read sampling catches a little; a lower-throughput direct-RNA run may catch nothing.
Shares exons with an isoformuncertainstrongDIRECT_RNA_ONLYShort reads on shared exons cannot be assigned to one isoform. Full-length reads can tell the isoforms apart.
Short, repetitive sequenceuncertainmissingINDETERMINATESequence that reads cannot be placed uniquely, and too short to anchor long reads: neither workflow gives a clean answer.

02

Variant Evidence Adjudication

In development In development. Protocols and feature extraction are frozen; no model has been evaluated and no performance exists.

Problem
Illumina short reads and Oxford Nanopore long reads can disagree about the same candidate variant. Standard workflows reduce that to PASS or FAIL, which hides the reason and forces a confident answer where the evidence may not support one.
Idea
Evidence adjudication: given paired evidence from both platforms, return the interpretation the evidence supports (both, Illumina only, ONT only, or neither), or abstain when the evidence is insufficient.
How it could work
  1. Form candidate episodes where the two call sets disagree, before looking at truth.
  2. Extract 42 approved, deterministic features: call-level evidence from each platform and local reference context.
  3. Train against benchmark-derived support states, excluding cases the benchmark itself cannot judge (unassessable).
  4. Use selective prediction: emit ABSTAIN when the model’s own evidence is insufficient, and report coverage alongside accuracy.
Scientific foundation
Frozen modelling protocol, deterministic extraction and an explicit separation between benchmark uncertainty and model abstention. Details
Potential users
  • Laboratories running both short- and long-read sequencing
  • Bioinformatics teams reconciling call sets
  • Technology developers
  • Translational research groups
Groups we think could benefit. Not current customers.
Illustrative · not model outputPaired Illumina and Oxford Nanopore evidence converging on a candidate variant, weighed, and resolved to a support state or ABSTAINIllumina readsONT readscontext: non-difficultcontext: homopolymerreads cannot be placed uniquelycontext: low mappabilitycontext: segmental duplicationcontext: outside benchmark regionscandidatevariantEvidence weighedIllumina call evidenceONT call evidenceQuality consistencyContext difficultyInterpretationBOTH_SUPPORTEDILLUMINA_SUPPORTEDONT_SUPPORTEDNEITHER_SUPPORTEDABSTAINUNASSESSABLEbenchmark labelling only

Both platforms call the same allele with strong, consistent evidence in easy sequence. The supportable interpretation is that both calls hold.

Illustrative cases showing what each state means. They are not predictions: the adjudication model has not been evaluated and no performance figures exist. Evidence levels are drawn qualitatively on purpose; the eventual system reads 42 approved features.
  • Clear agreement: BOTH_SUPPORTED. Both platforms call the same allele with strong, consistent evidence in easy sequence. The supportable interpretation is that both calls hold.
  • Illumina-only indel in a homopolymer: ILLUMINA_SUPPORTED. Short reads call a small deletion in a run of one base; long reads do not. Context matters: homopolymers are a known difficulty for some long-read chemistries.
  • ONT-only call in low mappability: ONT_SUPPORTED. Long reads span a region where short reads cannot be placed uniquely. The absence of a short-read call is weak evidence here.
  • Conflicting evidence: ABSTAIN. Both platforms call something, with different genotypes, low quality and in duplicated sequence. The honest output is: not enough evidence for an automated conclusion.
  • Benchmark cannot judge: UNASSESSABLE. The truth set cannot establish whether either call is right. This is benchmark uncertainty, not model abstention: such cases are excluded from training labels.

03

OmicsEdge Evidence Atlas

Released v1.0.0 Released v1.0.0 (DOI 10.5281/zenodo.23085673). The Evidence Explorer searches it in the browser.

Problem
Benchmark evidence is conditional on sample, platform, pipeline, benchmark version, reference build and representation. Most people meet it as static files and aggregate accuracy, with those conditions stripped away.
Idea
Keep benchmark results as evidence: each run-scoped event and both of its observations, with the sample, platform, pipeline, comparator and artifacts that produced them. Not a universal trust score and not a predictive model: evidence you can inspect.
How it could work
  1. Regenerate benchmark comparisons under a frozen protocol and keep every run-scoped event and observation.
  2. Keep the comparator decision exactly as recorded, with the truth set, regions and versions it was recorded against.
  3. Link events to biological variants only by exact normalized-allele identity; leave everything else visibly unresolved or not evaluated.
  4. Deliver it as a static, checksummed bundle that the Evidence Explorer reads directly; later, more samples, platforms and benchmark versions.
Scientific foundation
Released v1.0.0: 11,553 events and 23,106 observations across HG002, HG003 and HG004, Illumina and Oxford Nanopore. Details
Potential users
  • Benchmark developers
  • Sequencing technology developers
  • Variant-calling method developers
  • Laboratories validating pipelines
Groups we think could benefit. Not current customers.
  1. Eventchr8:3,989,216 T>Crun-scoped record · 1-based GRCh38
  2. Sample and platformHG003 · Illumina NovaSeq 6000coverage 35x
  3. PipelineRN-IlluminaNeuSomatic, GATK 4.1.7.0, Picard 2.22.4
  4. Benchmark conditionshap.py 0.3.15 · vcfeval 3.12.1truth artifact-truth-hg003-grch38-v421
Observations (two sides)truth TP · query TPGT 0/1 · match gm · recorded comparator outcome
Provenanceartifact-query-60z59-hg003 → run-phase2c-hg003-60z59-grch38-v421obs-run-phase2c-hg003-60z59-grch38-v421-1226-query
One real record from the Evidence Atlas. Not a universal trust score: evidence you can inspect, condition by condition.

Open the Evidence Explorer

Supporting offerings

Available now, as applied work.

The methods behind the research, applied to a team’s own data and pipelines.

Sequencing Reliability Analysis

Available

Problem. Teams often know their aggregate accuracy but not where, or for which variant classes, results stop holding.

Idea. Context-stratified analysis of where a workflow’s results are and are not reliable, with the boundaries stated explicitly.

  • Stratify results by genomic context and variant class.
  • Compare platforms and pipelines without collapsing them into a ranking.
  • Report what remains unresolved.

Potential users: Sequencing laboratories, Biotechnology companies, Research cores.

Benchmark Design and Validation

Available

Problem. Benchmarks are easy to tune by accident: selecting data or regions after seeing results inflates confidence.

Idea. Benchmarks designed with outcome-blind selection, frozen protocols and prespecified pass/fail gates.

  • Freeze sources, panels and criteria before looking at performance.
  • Define validation gates up front.
  • Release with checksums and independent verification.

Potential users: Assay-development teams, Benchmark developers, Method developers.

Reproducible Genomics Engineering

Available

Problem. Analyses that cannot be rerun, or whose inputs cannot be traced, are hard to trust and harder to extend.

Idea. Pipelines and data products with pinned environments, deterministic outputs and provenance built in.

  • Pinned containers and environments.
  • Deterministic, content-derived identifiers.
  • Checksum manifests and reproducibility checks in CI.

Potential users: Bioinformatics teams, Research software groups, Translational research groups.

The long-term opportunity

An evidence layer between sequencing output and the decisions made from it.

Sequencing volume keeps growing, and so does the number of platforms, chemistries, pipelines and benchmark versions. Every new combination multiplies the conditions under which a result may or may not hold.

We think there is room for software that sits between raw output and downstream analysis and answers, condition by condition: what was measurable, what is supported, where the technologies disagree, and what remains unresolved.

This is a direction, not a claim of an existing platform. We would rather show a working demonstration and a negative result than a roadmap of promises.