Bijux Proteomics

Bijux Proteomics is a composable Python platform for proteomics analysis, reproducible execution, evidence-aware interpretation, decision support, and laboratory follow-up. It is designed so a reviewer can trace a result from accepted and rejected scientific inputs through execution, grounding, recommendation, and observed consequence.

No layer receives authority over all the others. A completed run does not prove scientific acceptance. Grounded evidence does not authorize an action. A recommendation does not establish laboratory value.

Open Proteomics Documentation Follow The Scientist Journey Inspect Capability Limits View Source

Six Accountable Layers

flowchart TD
    request["Scientific question"] --> core["Core<br/>scientific calculation and acceptance"]
    core --> runtime["Runtime<br/>execution custody"]
    runtime --> knowledge["Knowledge<br/>support and contradiction"]
    knowledge --> intelligence["Intelligence<br/>challenge and recommendation"]
    intelligence --> lab["Lab<br/>readiness and consequence"]
    foundation["Foundation<br/>identity, schema, canonical bytes"] -. binds .-> core
    foundation -. binds .-> runtime
    foundation -. binds .-> knowledge
    foundation -. binds .-> intelligence
    foundation -. binds .-> lab
    core -. missing burden .-> narrow["Narrow or refuse"]
    runtime -. missing burden .-> narrow
    knowledge -. missing burden .-> narrow
    intelligence -. missing burden .-> narrow
    lab -. observed outcome .-> knowledge

Layer	Question it owns	Durable record
Foundation	how is a subject, schema, and serialized record identified?	canonical representation, digest, producer, compatibility, and typed disposition
Core	what scientific computation ran and what did it accept or reject?	inputs, assumptions, QC, result, rejection, benchmark lineage, and limitation
Runtime	what executed, under which state and environment?	request, selected capability, events, artifacts, terminal state, comparison, and replay evidence
Knowledge	which evidence supports or contradicts the claim?	source identity, context, support, contradiction, freshness, and unresolved gaps
Intelligence	why was an action ranked, downgraded, or refused?	candidate universe, policy, alternatives, sensitivity, confidence posture, and refusal conditions
Lab	what follow-up was feasible and what happened?	readiness, controls, custody, deviation, observation, consequence, and feedback

The stable join between layers is a typed identity or artifact reference. A filename, display label, dashboard color, or prose summary is not enough to join scientific records across packages or revisions.

Scientific Inputs Preserve Rejection

Proteomics input APIs return reports rather than silently filtering to a list of accepted values. A partially valid FASTA document, search result, or quantification table can therefore retain both the records used and the records excluded.

flowchart LR
    input["Scientific input"] --> contract["Typed scientific contract"]
    contract --> accepted["Accepted records"]
    contract --> rejected["Rejections and diagnostics"]
    contract --> policy["Assumptions and provenance"]
    accepted --> result["Scientific result"]
    rejected --> result
    policy --> result

The rejected portion is evidence about the population actually analyzed. It must travel with downstream QC and interpretation because hidden exclusions can change the conclusion.

Database Preparation Is Part Of The Result

Protein databases, spectral libraries, identifier maps, literature collections, and benchmark corpora are analytical inputs with their own scientific decisions. Their construction must remain visible beside downstream results.

flowchart LR
    sources["Versioned sources<br/>and licenses"] --> capture["Captured members<br/>and native identities"]
    capture --> normalize["Canonical identifiers<br/>and typed records"]
    normalize --> reconcile["Aliases, isoforms,<br/>decoys, conflicts"]
    reconcile --> qualify["Coverage, exclusions,<br/>context, and review"]
    qualify --> publish["Database or library<br/>manifest + digest"]
    publish --> workflow["Workflow-family input"]

Preparation decision	Evidence to retain	Scientific consequence if hidden
source and release selection	accession, release, retrieval context, license, and digest	silent database drift or biased coverage
identifier normalization	native ID, canonical ID, organism, isoform, and mapping rule	distinct entities collapse or one entity fragments
sequence and record admission	accepted and rejected members with diagnostics	analyzed population cannot be reconstructed
target/decoy construction	producer, rule, seed or determinism record, and manifest role	error-control assumptions become unverifiable
spectral or assay library transfer	source context, instrument and workflow compatibility, calibration and exclusions	library presence is mistaken for transfer validity
literature and ontology grounding	source version, claim relationship, context, contradiction, and freshness	citation count is mistaken for support
benchmark preparation	truth source, population, leakage controls, expected metrics, and limitations	performance is evaluated against self-confirming evidence

A database digest identifies bytes, not suitability. Suitability belongs to the workflow-family contract and intended use. The same database may be adequate for exploratory identification and inadequate for transferable quantification or an experimental recommendation.

Preserve Study Design And Error Control

Proteomics evidence depends on more than accepted peptide or protein records. The study design and its error-control policy determine which comparisons the data can support.

Scientific boundary	Evidence that must remain inspectable
sample structure	subject, condition, biological replicate, technical replicate, blocking, batch, and run order
identification error	target/decoy policy, search space, score direction, threshold scope, and retained population
protein inference	peptide-to-protein ambiguity, grouping rule, shared evidence, and unresolved alternatives
quantification	normalization, aggregation, censoring, imputation, and uncertainty propagation
contrasts and multiplicity	declared contrasts, tested family, correction method, denominator, and omitted tests
calibration and interference	calibrators, fitted range, residuals, carryover, selectivity, and out-of-range behavior

flowchart LR
    design["Study design and contrasts"] --> observations["PSMs, peptides, proteins,<br/>intensities, and missing values"]
    observations --> decisions["Identification, inference,<br/>normalization, and error control"]
    decisions --> result["Estimate, uncertainty,<br/>and complete decision population"]
    result --> transfer["Transfer test in a distinct context"]

False discovery rate and q-values are properties of a declared decision population under a declared procedure. They do not prove that an individual identification is correct, and a threshold computed for one family cannot be silently reused for a narrower favorable subset. Protein-level claims also need their own inference and error boundary; peptide-level control is not an automatic protein-level guarantee.

Missingness must preserve scientific meaning. Below-detection, absent from the search result, failed quality control, not measured, and structurally inapplicable are different states. An imputed value is a modeled value and must not replace the observed-state record from which it was derived.

Recompute Confidence At Every Inference Level

Proteomics decisions move through nested but non-equivalent populations. Confidence attached to a spectrum match cannot be copied to a peptide, protein group, abundance contrast, pathway, or experimental recommendation.

Inference level	Decision population that must remain visible
spectrum match	target and decoy candidates, score competition, search space, and accepted and rejected matches
peptide	contributing matches, modifications, mapping, collapse rule, and peptide-level error control
protein or group	unique and shared peptide evidence, grouping alternatives, parsimony or inference rule, and unresolved members
quantitative contrast	eligible samples and proteins, missingness, normalization, model, tested contrasts, and multiplicity
pathway or mechanism	selected entities, background universe, annotation revision, dependence, and enrichment family
recommendation	evidence tier, feasible action, counterfactual, uncertainty, and decision threshold

Filtering on a favorable intermediate score changes the later population and can invalidate a previously calibrated rate. Each aggregation should retain its parents, rejected alternatives, selection rule, denominator, and a confidence procedure defined at that level. A polished pathway result does not repair ambiguous protein identity or uncontrolled selection below it.

Keep Batch Correction From Becoming Evidence Erasure

Normalization and batch correction are models applied to observations. They can reduce nuisance variation, but they can also remove biological structure or manufacture comparability when batch and condition are confounded.

flowchart LR
    raw["Observed intensities and QC"] --> design["Design, batch, run order,<br/>controls, and missingness"]
    design --> correction["Declared normalization<br/>and correction model"]
    correction --> diagnostics["Before/after diagnostics<br/>and residual structure"]
    diagnostics --> estimate["Bounded estimate and limitation"]

Retain the uncorrected observations, transformation parameters, reference or control basis, excluded samples, before-and-after distributions, residual association with batch and condition, and sensitivity to reasonable alternate corrections. A visually compact embedding is not proof that batches are reconciled, and absence of a visible batch cluster is not proof that biological signal survived.

For longitudinal or repeated-measure designs, preserve subject and time dependence through normalization, missingness, and inference. Treating repeated samples as independent can create false precision even when protein-level error control and software execution are otherwise correct.

Test Transfer With A Separate Population

Internal repeatability and companion-package performance answer different questions. A transfer claim names both the development context and the target context, then tests the assumptions expected to survive between them.

For every transfer result, retain:

• the primary and companion or holdout package identities and their selection rules;
• cohort, specimen, preparation, acquisition, instrument, library, software, and processing differences;
• the invariant being tested, the allowed tolerance, and the complete eligible denominator;
• failures, exclusions, missing strata, and the context in which performance no longer supports the declared claim.

When one package determines the method, thresholds, or preferred model and also supplies the final score, the result is development evidence. Calling it external validation does not create independence. A successful transfer in one companion population remains evidence for that named context, not a universal claim across laboratories, instruments, or cohorts.

Evidence Ladder

A workflow family earns only the strongest language supported by every required layer.

flowchart LR
    asset["Source, license, benchmark manifest"] --> science["Scientific contract"]
    science --> execution["Runtime lane and run bundle"]
    execution --> grounding["Support, contradiction, context"]
    grounding --> judgment["Challenge, sensitivity, refusal"]
    judgment --> consequence["Feasibility and observed outcome"]
    consequence --> status["Bounded trust status"]
    asset -. missing .-> narrow["Narrow or refuse"]
    execution -. imported or unstable .-> narrow
    grounding -. contradicted .-> narrow
    consequence -. infeasible .-> narrow

A later success cannot promote an earlier weak record. A complete run bundle cannot compensate for an unclear scientific acceptance policy. A grounded claim cannot compensate for a decision that reverses under small policy changes. A recommendation cannot compensate for infeasible follow-up.

Workflow Families Have Independent Ceilings

Evidence is assessed separately for DDA, DIA, LFQ, multiplex, PTM, and targeted workflows. Scientific assumptions and execution modes differ, so strength in one family cannot be borrowed by another.

Family	Current documented posture	Essential limit
DDA	review_grade_bounded	primary evidence begins with external search-engine results; repository-owned raw search execution is not claimed
DIA	outsider_auditable_bounded	checked-report execution does not establish chromatogram-native or universal library transfer
LFQ	outsider_auditable_bounded	repeatability does not establish cross-cohort transfer or external quantitative truth
multiplex	internal_support_only	companion transfer remains fragile and outsider consequence closure is incomplete
PTM	outsider_auditable_bounded	localization does not establish occupancy, function, causality, or regulation
targeted	outsider_auditable_bounded	vendor parity, calibration transfer, interference, and assay burden remain bounded

These tokens are claim ceilings, not general maturity grades. The workflow-family guide is the authority for current evidence and blockers.

Read The Status Vocabulary Literally

Status	What it permits	What it does not permit
internal_support_only	useful implementation inside a restricted authority boundary	public recommendation or outsider consequence claims
review_grade_bounded	scientific review under named limits	raw execution parity, general transfer, or authority to act
outsider_auditable_bounded	external inspection and rerun of the declared bounded chain	clinical, universal, or decision-grade authority
release-ready	all required repository-wide categories pass for one source candidate	universal scientific validity

The exact status belongs in machine-readable evidence and public prose. Replacing it with a friendlier but stronger phrase creates an unreviewed claim.

Recommendation Records Preserve Counterfactuals

A defensible recommendation records why it would change. Comparator removal, literature removal, policy changes, laboratory burden, and observed outcomes are tested as separate drivers.

flowchart TD
    prior["Retained prior decision"] --> compare["Compare input revisions"]
    compare --> evidence{"Evidence changed?"}
    compare --> policy{"Policy changed?"}
    compare --> burden{"Burden changed?"}
    compare --> outcome{"Outcome observed?"}
    evidence --> attribution["Named driver set"]
    policy --> attribution
    burden --> attribution
    outcome --> attribution
    attribution --> revised["New posture and rationale"]
    revised --> audit["Old and new records remain inspectable"]

If removing one evidence axis or increasing downstream burden collapses the recommendation, that weaker posture is part of the truthful product surface. An observed outcome changes the next decision; it does not rewrite the prior record.

Operate Without Losing Scientific Meaning

Runtime scale and service behavior matter only if the evidence-bearing population remains equivalent. Optimizations that discard ambiguity, contradictions, provenance, rejections, or rare failure classes change the scientific operation even when throughput improves.

Operational pressure	Preserve	Unsafe shortcut
batched ingestion	source manifest, deterministic member identity, rejection population, and cross-batch integrity	treating batches as independent when identities or conflicts cross them
indexed identifier resolution	every candidate, evidence tier, organism and alias context	first-match collapse of ambiguity
partitioned evidence graph	stable partition key, cross-partition edges, conflicts, and final integrity audit	ignoring relationships outside the selected shard
parallel workflow execution	request and environment identity, event order where meaningful, artifacts, failures, and terminal state	aggregating only successful worker outputs
large recommendation universe	complete candidate population, pruning policy, alternatives, sensitivity, and burden	ranking only preselected favorable candidates
retained laboratory evidence	custody, controls, deviations, observations, and link to the prior decision	replacing analytical history with the eventual outcome

The repository exposes performance and behavioral evidence but does not turn that into a universal service-level objective. Capacity claims require the named workload, environment, dataset, evidence-completeness checks, and observation window. A faster result that changes ambiguity or conflict counts is not the same scientific result.

Release Authority Is Multi-Dimensional

A candidate may publish only when the required scientific, runtime, knowledge, recommendation, laboratory, compatibility, security, and channel owners agree for that revision. This is an intersection, not a majority vote.

flowchart TB
    candidate["Source candidate"] --> science["Scientific evidence"]
    candidate --> runtime["Execution evidence"]
    candidate --> knowledge["Grounding and freshness"]
    candidate --> decision["Recommendation stability"]
    candidate --> consequence["Lab consequence boundary"]
    candidate --> delivery["Compatibility, security,<br/>and channel evidence"]
    science --> verdict{"Release disposition"}
    runtime --> verdict
    knowledge --> verdict
    decision --> verdict
    consequence --> verdict
    delivery --> verdict

One blocking owner narrows or refuses the affected claim even when every other lane is green. The disposition must retain blocker codes, affected workflow families and claims, evidence identities, and the condition that would permit reconsideration.

Verify A Proteomics Claim

Claim	Evidence route
a record entered the calculation	canonical identity, accepted/rejected report, schema, and provenance
a scientific workflow supports a result	family contract, benchmark lineage, QC, comparison, and claim ceiling
the result can be rerun	runtime request, environment, state, artifact ledger, and comparison record
a claim is grounded	cited context, supporting and contradicting evidence, freshness, and gaps
a recommendation is proportionate	alternatives, ranking policy, sensitivity, burden, confidence, and refusal behavior
a follow-up changed the evidence	readiness, custody, controls, deviation, observation, and linked consequence record
a repository candidate may publish	revision-specific readiness matrix, blocker codes, governed outputs, and channel decision

Readiness does not use majority voting. A green runtime result cannot erase a benchmark blocker, and a strong benchmark cannot erase an unreviewable laboratory consequence path.

Scope And Non-Claims

The repository implements sequence, peptide, spectrum, confidence, protein inference, quantification, DIA, LFQ, PTM, targeted, evidence-grounding, recommendation, and laboratory-follow-up surfaces. Coverage is not a blanket accuracy claim.

The platform does not claim universal transfer across cohorts, instruments, search engines, acquisition modes, or experimental designs. It does not convert execution success into biological truth, or advisory output into clinical or autonomous authority.

Continue with Applied Domains for the wider scientific evidence model or Operational Assurance to compare run evidence with delivery and recovery qualification.