Skip to content

Architecture · method selection

The right method
is rarely the
loudest one

Every computational task hides a scientific choice: which method fits this data. The literature makes one answer very loud, and for a given dataset that answer is often the wrong one.

  1. 01 CONTEXT
  2. 02 SPREAD
  3. 03 RECALL
  4. 04 SCORE
01 / 04premise 01

CONTEXT

one cohort · 92 lesional against 82 normal, as RPKM

corpus
35,897
RNA-seq DE articles
methods measured
16
named in that corpus
head share
92.5%
of named methods
tail evidence
5
papers, 4 one thread

Above: sixteen differential-expression methods, by how often PubMed names them. Bars and the curve are measured counts on a linear axis; the cohort is the real GSE54456 design; the search window is schematic, and what it contains is measured in fig. 04. Retrieved 2026-07-02.

The three premises

Why a scientific harness has to own method selection

What follows

None of the three is fixed by a larger model

One premise is a property of the biology, two are properties of the literature and of how a model is trained on it. A system that selects methods has to be built to handle all three.

  1. 01

    Read the context first

    Inspect the data before choosing anything. Sample size, the form it arrives in and the variation between samples are what make a method applicable or not.

  2. 02

    Retrieve around what the inspection found

    Search on the condition rather than on the generic task, and weight a few methodological sources above the volume of consensus usage.

  3. 03

    Score the candidates, and escalate the rest

    Grade every candidate on applicability to this context, keep a contested result contested, and hand back the cases no evidence settles.

Block 04 builds this, and names the ten capabilities it takes.Skip to the response

Block 01Biological context decides the method

Three cohorts, one question, three different right answers

Nothing about the question changes across these three. Only the biology does, and with it the method that can honestly answer it.

fig. 01 · three contexts, three answers

Read the context first, and the method follows from it

Each column is one real study design. Three properties of the cohort are shown and the rest are named, because three is enough to change the answer. Watch the last row: the same question answered from a model’s own weights returns one method in all three contexts, including the one where no single method is defensible.

fig. 01

The same question, asked of three cohorts: how do you find the genes that differ?

cycling

context 01

A controlled experiment

Three mice per group, one perturbation

samples per group
3
data as it arrives
raw counts
variation between samples
low, inbred and controlled

+ tissue, platform, read depth, batch, and the rest of the record

selected on this context

Count model · DESeq2 or edgeR

Three replicates carry too little data per gene to estimate variance, so the model borrows strength across thousands of genes. A rank test cannot reach significance at all.

context 02

A patient cohort

GSE54456 · 92 lesional against 82 normal skin

samples per group
92 against 82
data as it arrives
RPKM, not raw counts
variation between samples
high, patient to patient

+ lesion severity, treatment history, site, subtype, and the rest of the record

selected on this context

Rank test · Wilcoxon rank-sum

At population scale the count model’s distributional assumption is violated in ways that inflate false positives, and it cannot consume RPKM as provided. A rank test assumes no distribution and has ample power at this size.

context 03

A multi-site cohort

Forty patients per group, collected across three sites

samples per group
40
data as it arrives
raw counts
variation between samples
high, plus site and batch

+ protocol version, operator, storage time, subtype, and the rest of the record

selected on this context

No single answer · escalate

The count model’s assumption is the liability at this sample size, and the rank test cannot adjust for the site structure. Two credible options remain and no available evidence settles the trade-off, so the choice belongs to a scientist.

answered from a model’s own weights

DESeq2context 01

DESeq2context 02

DESeq2context 03

Recorded in the skill-validation log: an ungrounded model asked to build a DEG method for GSE54456 defaulted to a count model, and only grounding plus a context gate on the data flipped the constructed method.

what this block establishesA method is applicable or not applicable to a context. A system that answers before it has read the context is answering a different question from the one it was asked.

Block 02The applicable method is often sparse

Three tools hold most of the field. The rest is a long thin tail

Sixteen methods, measured inside one corpus of 35,897 articles. Move the cut in the figure and the concentration becomes a number.

fig. 02

Articles naming each method · within the 35,897-article RNA-seq differential-expression corpus

Click a method to move the cut between head and tail.

Article axis

Distribution-basedDistribution-freeSemiparametric

The cut, at rank 3

Head · ranks 1 to 3
72%

1,465 of 2,036 measured mentions

Tail · ranks 4 to 16
28%

13 methods share what is left

Mentions are non-exclusive. A paper may name several methods.

rank 3 · DE tool

edgeR

Statistical model
Negative-binomial GLM
Family
Distribution-based
Articles
385
Share of corpus
1.07%

Articles and share are measured. Statistical model and family are an editorial classification taken from the method’s primary reference.

Resolve 1.83% against 92.5%

The same counts answer two different questions

QuantityArticles% of 1,632
Full corpus35,897
Papers naming any recognized DE method1,632100%
Distribution-based (count / linear) tools, union1,50992.5%
of which the big three: DESeq2, edgeR, limma1,28278.6%
Distribution-free rank-test family, union1529.3%
Overlap, papers naming both291.8%

The method-labeled subset is the denominator for the sentence "X% of the field uses Y". B = 1,509 plus 152 minus 29 = 1,632 by inclusion and exclusion; shares exceed 100% slightly because 29 papers name both families.

Source · PubMed title/abstract counts, retrieved 2026-07-02, within one fixed corpus. Footprints are non-exclusive. The rank-frequency line spans more than two orders of magnitude; no power-law claim is made.

And the evidence for the tail method is thinner than the tail

A generous keyword search finds 23 papers that might recommend a rank test at large sample sizes. Reading all 23 leaves five that genuinely do, and four of those five are one paper and the debate it started.

fig. 03

From any mention of a rank test to genuine advocacy for it at large n

Select a stage to see how the number was obtained.

97 → 23 → 5 → 1

Bars are on a log scale. The final stage is a single article.

Consensus against this stage

256×

1,282 papers name one of the big three; 5 sit at this stage.

how obtained

All 23 abstracts read and classified A / B / C

Reading the 23 splits them into 5 genuine large-n advocacy, 10 distribution-free methods argued for a different reason, and 8 false-positive co-mentions. The keyword count overstates support by roughly 4.6×.

AGenuine large-n advocacy
5
BDistribution-free method, different rationale
10
CFalse-positive co-mention
8

The five advocacy papers

  1. PMID 352920872022anchorfor

    Anchor. Permutation analysis of human population RNA-seq; recommends the rank-sum test at large n.

    doi:10.1186/s13059-022-02648-4
  2. PMID 394785442024same threadfor

    Authors’ reply; reaffirms the rank-sum test as most robust.

    doi:10.1186/s13059-024-03232-8
  3. PMID 394786332024same threadagainst

    Rebuttal. Argues part of the effect is a normalization artifact.

    doi:10.1186/s13059-024-03231-9
  4. PMID 394786362024same threadagainst

    Winsorization comment on the anchor paper.

    doi:10.1186/s13059-024-03230-w
  5. PMID 308862782019independentfor

    The one independent source: a rank test on raw data outperforms limma-voom at large n.

    doi:10.1038/s41598-019-41315-w

Two of the five argue against the recommendation and are counted anyway. A contested result stays contested.

Source · PubMed, retrieved 2026-07-02. Stages 1 and 2 are Boolean counts; stages 3 and 4 are all 23 abstracts read and classified, so the last two numbers are manual work and are reported as such.

what this block establishesSparse is not the same as wrong, and popular is not the same as applicable. The evidence for the applicable method here is thin and actively contested, and both facts have to survive into the answer.

Block 03A language model returns the head, and so does a generic search

Reading the first page of results does not reach the tail

A tool-using system reads the top results, not the field. Probe those 50 papers for a rank test and nothing answers.

fig. 04

The 50 most relevant papers for the corpus query, under PubMed “Best Match”

Probe the window for a method and see how much of it responds.

n = 50 · read by hand

— / 50

Every square is one record in the window. 41 of them name no differential-expression tool or test at all.

Names DESeq2 and / or edgeR
8
Names another distribution-based tool only
1
Names no DE tool or test in title / abstract
41
Names any distribution-free test
0

A unit chart. The four classes partition the 50 records; square positions carry no meaning.

named methods in the window
DESeq2
8
edgeR
7
limma / voom
3
Cuffdiff / Cufflinks
1
Any other named tool
0
t-test / ANOVA
0
Wilcoxon / rank-sum / distribution-free
0
Papers advocating a rank test at large n
0

Mentions, non-exclusive, which is why they exceed the 9 papers that name any tool.

Paper types

22 method7 application21 other

Source · PubMed “Best Match” top 50 for the corpus query, retrieved 2026-07-02; all 50 titles and abstracts classified by hand. “Best Match” is a proprietary learned ranker, so a different engine would return a different window.

Two senses of the word “parametric”, kept apart

Parametric memory

a property of the language model

Knowledge stored implicitly in a model’s weights, as opposed to non-parametric memory, which is an external retrieval store (Lewis et al., 2020). Because the weights compress a training corpus, a model answering cold reproduces its most frequent answer. This is the sense used in the prose on this page.

Distribution-based method

a property of a statistical test

A method is distribution-based, or parametric in the statistical sense, if it assumes the data follow a specified distribution, such as the negative binomial. A distribution-free method assumes none and works on ranks. This page says distribution-based and distribution-free throughout, and never uses the bare word for a statistical method.

In this example the two senses happen to align. Parametric memory returns DESeq2, which is a distribution-based method, while the context-appropriate answer is a distribution-free one. The mechanism is training-corpus frequency, so the argument holds equally where the tail method is distribution-based.

what this block establishesReaching the tail is not a matter of searching harder. The head of a relevance ranking is the consensus, so the query itself has to be formulated around the condition the data showed.

Block 04The response

Score the method against the context, then escalate what it cannot settle

The three premises leave one way out. Inspect the data, build the retrieval around what the inspection found, and score every candidate on how applicable it is to this context rather than on how often the field names it.

This is the third requirement, built. Seven dimensions, both candidates, scored under a named context, with an honest third outcome when the dimensions disagree and no evidence settles them.

fig. 05

Seven dimensions of applicability, scored for both candidates

The same three contexts as fig. 01. Watch the scores move, not the popularity.

samples per group
82
Population scale · n ≥ 20 per group
data as it arrives
RPKM, as distributed
a count model cannot consume it
batch or covariates
None
no adjustment required
applicability, dimension by dimension
holdspartlyfailsunknown
dimensionCount model · DESeq2 or edgeRRank test · Wilcoxon rank-sum← selected

Distributional assumption

Experimental context

fails

Real cohorts carry outliers, heterogeneity and batch structure the model does not capture.

holds

Assumes no distribution. It asks only whether one group tends to outrank the other.

Input as provided

Data processing and QC

fails

Requires raw integer counts. RPKM cannot be consumed correctly.

holds

Indifferent to normalization, since ranks survive any monotone transform.

Error calibration at this n

Data-to-signal evaluation

fails

Under permuted labels, actual FDR sometimes exceeded 20% at a nominal 5%.

holds

Cannot be misspecified, so nominal error rates are honest at any n.

Power at this n

Domain-specific methods

holds

Strength borrowed across thousands of genes is what makes small designs work.

holds

Ample power at population scale, where the low-n weakness disappears.

Covariate and batch adjustment

Experimental context

holds

A GLM design matrix absorbs batch and covariate terms directly.

partly

No adjustment available; none required by this condition.

Precedent on the named dataset

Scientific retrieval

unknown

No published run on GSE54456 as distributed.

holds

The original GSE54456 study identified DEGs with the rank-sum test on RPKM.

Method evidence base for this condition

Contradiction detection

partly

1,282 naming papers, and a demonstrated failure mode at this scale.

unknown

5 verified advocacy papers, 4 of them one thread; actively contested.

corpus footprint

1,282

1,282 corpus papers name one of the big three.

5

5 verified advocacy papers, 4 of them one debate thread.

what the rubric selects

Rank test · Wilcoxon rank-sum

  • The data ship as RPKM, and a count model requires raw integer counts.
  • At 82 per group the rank test has ample power, so its low-n weakness does not apply.
  • This is what the original GSE54456 study did, verified from its Methods.

contestedThe large-n recommendation is well-supported and actively debated: a 2024 rebuttal attributes part of the effect to normalization, and the authors replied. Treat it as a strong contested result.

Rubric cells are an editorial reading of the cited primary literature evaluated against the context above. The sample-size bands read two published results, population scale of tens to hundreds per group and a benchmark tested only to n = 12, and are not a published threshold. Corpus footprints are measured PubMed counts.

Ten capabilities carry that, and they are architectural

Inspect the data, formulate retrieval around what the inspection found, score the candidates, keep a contested result contested, escalate what nothing settles, and record the choice so it can be challenged. Each is a capability in the architecture, in its own words.

Context and constraints

Experimental context

in method selectionReads the data before reasoning about it: input format, per-group n, batch structure, known confounders. This is what produces the triggering condition at all.

What it does
Captures study design, cohort definition, tissue and model system, platform, batch structure, sample quality and known confounders.
Why it is necessary
The same expression change means different things in lesional skin, a cell line, or a mouse model. Context is what licenses a mechanistic claim.
What a system does wrong without it
The system produces statistically valid, biologically meaningless results, and cannot detect that it has done so.
Open in the architecture

what this block establishesMethod selection is a function of the harness rather than a property of a larger model. Either it is grounded, context-gated and scored, or it is recalled.

05Pressure test

Where this analysis can be attacked

Naming frequency is a proxy, the family labels are hand-assigned, the keyword filter is arbitrary, and the recommendation at the centre of the example is contested. Open any objection.

Objections07

Raised against the analysis. Each response is a measurement, not a rebuttal.

  • True, these are naming-frequency proxies. The magnitude (92.5% against 9.3%) dwarfs any plausible measurement bias, the retrieval window points the same way with a starker 0 of 50, and the recorded model behavior matches the prediction.

Limitations05

What no response removes. These bound what the numbers can be used to claim.

Sources and queries14

Every count on this page was retrieved from PubMed (NLM) via E-utilities ESearch on 2026-07-02 against one fixed corpus: ("RNA-seq"[tiab] OR "RNA sequencing"[tiab]) AND ("differential expression"[tiab] OR "differentially expressed"[tiab]). Each method footprint is CORPUS AND <method token>[tiab]. Classifications of a method’s statistical family are editorial and labeled where they appear.

  1. 01Li Y, Ge X, Peng F, Li W, Li JJ. Exaggerated false positives by popular differential expression methods when analyzing human population samples. Genome Biol 2022;23:79.doi:10.1186/s13059-022-02648-4The anchor. Recommends the rank-sum test for large-n population RNA-seq.
  2. 02Hejblum BP, et al. Neglecting the impact of normalization. Genome Biol 2024;25:281.doi:10.1186/s13059-024-03231-9Rebuttal.
  3. 03Ge X, et al. Response to "Neglecting normalization" and "Winsorization". Genome Biol 2024;25:283.doi:10.1186/s13059-024-03232-8Reply.
  4. 04Winsorization comment. Genome Biol 2024.doi:10.1186/s13059-024-03230-w
  5. 05Independent large-n rank-test finding. Sci Rep 2019;9.doi:10.1038/s41598-019-41315-w
  6. 06Li B, Tsoi LC, Swindell WR, et al. Transcriptome analysis of psoriasis in a large case-control sample. J Invest Dermatol 2014;134:1828 to 1838.doi:10.1038/jid.2014.28The GSE54456 study. Used the rank-sum test on RPKM, verified from its Methods.
  7. 07Li D, et al. An evaluation of RNA-seq differential analysis methods. PLoS One 2022;17:e0264246.doi:10.1371/journal.pone.0264246Favours DESeq2 at n ≥ 6; tested only to n = 12.
  8. 08Warden CD, Wu X. Critical differential expression assessment for individual bulk RNA-seq projects. bioRxiv 2024.doi:10.1101/2024.02.10.579728At larger sample sizes standard tests gain utility; assess each project individually.
  9. 09Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 2014;15:550.doi:10.1186/s13059-014-0550-8
  10. 10Robinson MD, McCarthy DJ, Smyth GK. edgeR. Bioinformatics 2010;26:139 to 140.doi:10.1093/bioinformatics/btp616
  11. 11Ritchie ME, et al. limma powers differential expression analyses. Nucleic Acids Res 2015;43:e47.doi:10.1093/nar/gkv007
  12. 12Law CW, Chen Y, Shi W, Smyth GK. voom. Genome Biol 2014;15:R29.doi:10.1186/gb-2014-15-2-r29
  13. 13Lewis P, et al. Retrieval-augmented generation for knowledge-intensive NLP tasks. NeurIPS 2020.Origin of the parametric / non-parametric memory distinction.
  14. 14Clauset A, Shalizi CR, Newman MEJ. Power-law distributions in empirical data. SIAM Review 2009;51:661 to 703.

See these capabilities in a whole run

The work loop applies the same layers to one scientific problem, stage by stage, including the quality-control stage that names a confounder before anything is interpreted.