Skip to content
Knowledge PlanePreview · Included with every paid plan

Methods chosen from published rules.

The Knowledge Plane holds rules for when each method applies, drawn from the literature, with tested script templates. Inflexa matches your study design to those rules before it writes a plan, and checks the plan against them before you see it.

Knowledge Plane

Common practice

When a TPM, FPKM or log table has a known batch the design can hold, put the batch in the limma design. The test never takes adjusted values.

Consensus

When the values are TPM, FPKM or RPKM, use limma-trend on the log values. DESeq2, edgeR and voom need raw counts.

Back on your machine

  1. 1Log transform, log2(FPKM + 1)template
  2. 2PCA by site and strandednesstemplate
  3. 3limma-trend, sex, age, strandednesstemplate
  4. 4fgsea on hallmark setstemplate

Plan checked: no rule broken

  1. Field: Data (bulk RNA-seq, FPKM)
  2. Field: Groups (two-group contrast)
  3. Field: Replicates (14 vs 12)
  4. Field: Batch (batch: site, strandedness)
  5. Field: Covariates (covariates: sex, age)
  6. Field: Pairing (pairing: none)
  7. Rule: Log transform
  8. Rule: Sample structure
  9. Rule: Batch in the design
  10. Rule: Covariates
  11. Rule: FPKM input
  12. Rule: Ranked enrichment
  13. Method: PCA
  14. Method: limma-trend
  15. Method: fgsea
  16. Template: log-scale QC
  17. Template: limma-trend two-group
  18. Template: fgsea preranked
  19. Source: Zhao 2020
  20. Source: Conesa 2016
  21. Source: Nygaard 2016
  22. Source: Law 2014

  1. Your design arrives

    The typed situation crosses from your machine as counts and categories. Sample rows, file paths, column names and identifiers stay behind.

  2. Rules match

    Each field fires the rules whose conditions it meets, and the matching rules light in turn.

  3. The method path traces

    From each rule to the method it calls for, limma-trend and fgsea here, and on to the tested templates that run it.

  4. Sources light

    Every rule points at its evidence: Zhao 2020, Conesa 2016, Nygaard 2016 and Law 2014 light with their short citations.

  5. The procedure comes back

    Four steps return to your machine, each tied to its rule and filled in there from a template. The plan is checked against the rules before you see it.

  6. A design that cannot answer

    Now the batch is fully confounded with the condition. The flag rule lights red, the path refuses to trace, and Inflexa stops and says so.

    Knowledge Plane

    Consensus

    Batch is fully confounded with the condition. Stop and say so, or label every result as confounded. No method can separate batch from biology in this design.

    1. Field: Data (bulk RNA-seq, FPKM)
    2. Field: Groups (two-group contrast)
    3. Field: Replicates (14 vs 12)
    4. Field: Batch (batch: confounded with condition)
    5. Field: Covariates (covariates: sex, age)
    6. Field: Pairing (pairing: none)
    7. Rule: Confounded batch

What it sees

It sees the shape of your study.

Inflexa sends the plane the typed situation as counts and categories: the modality, the question, the groups, the batch, the covariates and the pairing. The plane never receives sample rows, file paths, column names or identifiers.

Your machine

The typed situation

modality
bulk RNA-seq, FPKM
question
two-group contrast
groups
14 vs 12
batch
site, strandedness
covariates
sex, age
pairing
none

Never sent

  • sample rows
  • file paths
  • column names
  • identifiers

Knowledge Plane

counts and categories

What comes back

A procedure, with the rule behind every step.

The plane returns the steps of a procedure, each tied to a rule and its sources, with tested script templates for the standard steps.

  1. 1Log transform, log2(FPKM + 1)template

    Rule: Log transformCommon practice

    • Zhao 2020. Misuse of RPKM or TPM normalization when comparing across samples and sequencing protocols
  2. 2PCA by site and strandednesstemplate

    Rule: Sample structureCommon practice

    • limma User's Guide
    • Conesa 2016. A survey of best practices for RNA-seq data analysis
  3. 3limma-trend, sex, age, strandednesstemplate

    Rule: FPKM inputConsensus

    • Love 2014. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2
    • limma User's Guide
  4. 4fgsea on hallmark setstemplate

    Rule: Ranked enrichmentConsensus

    • Subramanian 2005. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles
    • Korotkevich 2021. Fast gene set enrichment analysis

Plan checked: no rule broken

File paths and column names are filled in locally, after the template comes back.

Rules you can read

Each rule says when it applies, and why.

A rule carries its grade, the situation it applies to, what to do there and the sources it rests on. Some end in a stop.

ConsensusStop

Confounded batch

When

Batch is fully confounded with the condition

Then

Stop and say so, or label every result as confounded. No method can separate batch from biology in this design.

  • Leek 2010. Tackling the widespread and critical impact of batch effects in high-throughput data
  • Nygaard 2016. Methods that remove batch effects while retaining group differences may lead to exaggerated confidence in downstream analyses
Common practice

Batch in the design

When

A TPM, FPKM or log table has a known batch the design can hold

Then

Put the batch in the limma design. The test never takes adjusted values.

  • Nygaard 2016. Methods that remove batch effects while retaining group differences may lead to exaggerated confidence in downstream analyses
  • Johnson 2007. Adjusting batch effects in microarray expression data using empirical Bayes methods
Consensus

FPKM input

When

The values are TPM, FPKM or RPKM

Then

Use limma-trend on the log values. DESeq2, edgeR and voom need raw counts.

  • Love 2014. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2
  • limma User's Guide
Common practice

Interaction design

When

The question is an interaction of two factors with few samples per cell

Then

Use limma-voom. On shuffled labels DESeq2 called up to 93 genes where limma-voom called none.

  • Law 2014. voom: precision weights unlock linear model analysis tools for RNA-seq read counts
  • limma User's Guide
Consensus

Paired design

When

Samples are paired by donor

Then

Put the donor in the design before the condition, so the contrast is tested within each person.

  • Love 2015. RNA-Seq workflow: gene-level exploratory analysis and differential expression
  • Law 2020. A guide to creating design matrices for gene expression experiments
Consensus

Fractional counts

When

Counts are RSEM expected counts

Then

Import them with tximport and give the model the effective lengths as an offset. Do not round them by hand.

  • Li 2011. RSEM: accurate transcript quantification from RNA-Seq data with or without a reference genome
  • tximport vignette. Importing transcript abundance datasets with tximport
When the field disagrees

Disputes show both sides.

Where the field disagrees, the plane returns each position with its sources, and the plan states the side it takes.

Disputed

Ranking for a preranked enrichment

When

A preranked gene set enrichment needs its genes in order

Rank by a signed statistic

The Wald statistic, or the signed -log10 p-value when the p-values do not tie at zero, so each gene carries the direction of its change and the strength of the evidence for it.

  • Reimand 2019. Pathway enrichment analysis and visualization of omics data using g:Profiler, GSEA, Cytoscape and EnrichmentMap

Rank by the effect size

The log2 fold change shrunken with apeglm, which keeps the size of the change and removes the noise of low counts.

A survey of published enrichment analyses found the ranking often goes unreported, and that the choice changes the result.

  • Wijesooriya 2022. Urgent need for consistent standards in functional enrichment analysis

The plane returns both sides with their sources. The plan states the side it takes and reports it beside the gene set database and its release.

Benchmark

Smaller models with the plane out-planned a frontier model alone.

Grounding lifted every model we tested.

Plan quality, 0 to 100

Plan quality, 0 to 100. The frontier model alone, Claude Opus 5, scored 81.7.
ModelAloneWith the plane
Claude Opus 581.790.9
GLM 5.369.889.2
Qwen 3.8 27B62.983.4

0 invented references in any plan made with the plane

Inflexa planning benchmark, 54 bulk RNA-seq planning tasks, 14 Sep 2026. Single run, scored by a blind AI judge.

Which plans include it

Included with every paid plan.

On your machine

Inflexa Pro

A full workspace, served from your machine. Any model, including one you host.

About Pro

Inflexa Cloud

Biotech

For research groups and small biotechs who work together, with shared projects and Privacy Mode.

Claude Opus 5: 81.7 alone, 90.9 with the plane

Compare plans

Inflexa Cloud

Enterprise

For pharma and large institutions with compliance and deployment needs, in your own cloud.

Claude Opus 5: 81.7 alone, 90.9 with the plane

Compare plans

The open-source CLI and the free Cloud Personal plan do not include the plane.

See the plane on a question close to yours.

Tell us about your data and the question you are working on. We will walk you through Pro on a public dataset like it.

Talk to us about ProCompare plans