Common practiceNygaard 2016 · Johnson 2007
Batch in the design
WhenA TPM, FPKM or log table has a known batch the design can hold
ThenPut the batch in the limma design. The test never takes adjusted values.
- Inflexa Pro
- Available now
- Runs on your machine
- Per seat, per year
Your co-scientist, on your own machine.
A workspace for computational biology that runs in your browser and keeps your data on your computer. Talk the question through, check the literature, approve the plan, then read results with the record beside them.
localhost · Inflexa Pro
Melanoma anti-PD-1›What is this dataset about?Profile · 2 files · 2 groups
You: What is this dataset about?
Inflexa: Read the data profile. GSE78220, the Hugo et al. anti-PD-1 melanoma cohort: bulk RNA-seq of metastatic melanoma biopsies from patients treated with pembrolizumab, with a clinical table that joins to the matrix completely. Four things will shape any analysis here:
- The deposit is an FPKM table, which a count model cannot take. The test is limma-trend on log2(FPKM + 1).
- Effective n is 26 patients: Pt27 contributes two runs, and Pt16 was biopsied on treatment.
- Two technical variables are on record: study site (UCLA 23, VIC 5) and library strandedness (7 stranded), and they nearly coincide. The plan checks PCA by both and keeps strandedness in the design.
- 14 responders against 12 non-responders: few genes will survive FDR.
Want me to draft a plan for the responder contrast, or start with the survival angle?
You: Draft the responder plan. Keep survival for later.
Plan: Anti-PD-1 responder DE and GSEA, bulk RNA-seq · 10 steps · 5 agents.
Ask a follow-up…GLM 5.3
Inputs 2
Add a path- GSE78220_PatientFPKM.xlsxfile
- GSE78220_metadata.csvfile
Inputs are references to paths on this computer. Inflexa never copies them.
Data profile
CompletedAn FPKM table of 25,268 genes by 28 libraries, keyed by gene symbol, and a 28-row annotation table that joins to it with no orphans. The outcome is anti-PD-1 response in three levels; overall survival and vital status are usable too.
- Effective patients
- 26
- Responders
- 14 vs 12
- Batch variables
- Site, strandedness
- Values
- FPKM
Anti-PD-1 responder DE and GSEA
Proposed10 steps
- T0S1Cohort assembly, depends on no step
- T0S2Log transform, depends on no step
- T1S1Responder DE, depends on T0S1, T0S2
- T2S1Immune composition, depends on T0S1
- T3S1GSEA, hallmark, depends on T1S1
- T4S1Covariate check, depends on T1S1
- T3S2Leading-edge genes, depends on T3S1
- T1S2Sensitivity fit, depends on T4S1
- T2S2Immune score contrast, depends on T2S1, T4S1
- T5S1Drug repurposing, depends on T2S1, T3S2, T1S2
- bulk transcriptomics
- enrichment
- immune profiling
- statistical modeling
T0S1Cohort assembly and the pre-treatment filter
- Agent
- bulk-transcriptomics-agent
- Depends on
- No step
- Acceptance
- One library per patient, all pre-treatment, with response as a two-level factor.
- Grounding
- Knowledge Planerule matched · consensus
Ask for changesApprove
Anti-PD-1 responder DE and GSEA
c98cfe · 10 steps · started 15:42
Running 6/10
4 completed2 running4 pending14m08s
- Cohort assemblycompleted, 2m11s
- Log transformcompleted, 0m52s
- Responder DEcompleted, 4m40s
- Immune compositioncompleted, 3m02s
- GSEA, hallmarkrunning, 1m19s
- Covariate checkrunning, 0m48s
Site and strandedness nearly coincide
no referenceAll five VIC libraries are stranded, and so are both Pt27 runs. PCA is checked by each, and strandedness stays in the design.
High confidenceT0S1
Files
- out/
- de_results.tsv1.9 MB
- gsea_hallmark.tsv48 KB
- immune_scores.tsv12 KB
- figures/
- volcano_responders.png
- hallmark_dotplot.png
- provenance/
- record.prov.json
- record.sig.json
- report.html
Signed record · intact
Each output here is listed in the signed record, with the step and run that wrote it.
Sessions built from public datasets: GSE78220, the Hugo et al. anti-PD-1 cohort.
Before anything runs
Talk the question through first.
Pro reads your data, the literature and the databases with you, and says what will shape the analysis before it proposes a plan. The hypotheses get argued while they are still cheap.
It reads your data first
Every input gets a profile: shape, types, design, and the traps in it.
- Read the data profileGSE78220ok
Effective n is 26 patients, 14 against 12. The values are FPKM, and site and strandedness are both on record.
It reads the literature with you
PubMed, Semantic Scholar and arXiv, plus a tool that checks references against the DOI registry, Crossref and PubMed.
ECM remodeling genes dominate the CCl4 response
confirmedHigh confidencePMID 18471545It brings the databases
Targets, drugs, variants, pathways and safety, queried live while you talk.
- Queried Open TargetsGLP1Rok
- Queried ChEMBLGLP1R agonistsok
- Searched FAERSsemaglutideok
- Queried ClinVarGLP1R variantsok
Knowledge PlanePreviewIncluded with every paid plan
Methods chosen from published rules.
The Knowledge Plane holds rules for when each method applies, drawn from the literature, along with tested script templates. Pro describes your study design to it, gets back a procedure with the rule behind every step, and checks the plan before you see it.
- 01
On your machine
What Pro sends
- 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
- 02
Knowledge Plane
The rule it matched
ConsensusLove 2014 · limma User's GuideFPKM input
WhenThe values are TPM, FPKM or RPKM
ThenUse limma-trend on the log values. DESeq2, edgeR and voom need raw counts.
- 03
Back on your machine
A procedure, checked
- 1Log transform, log2(FPKM + 1)template
- 2PCA by site and strandednesstemplate
- 3limma-trend, sex, age, strandednesstemplate
- 4fgsea on hallmark setstemplate
Plan checked: no rule broken
File paths and column names are filled in locally, after the template comes back.
Your machine
Never sent
sample rowsfile pathscolumn namesidentifiersKnowledge 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
- 1Log transform, log2(FPKM + 1)template
- 2PCA by site and strandednesstemplate
- 3limma-trend, sex, age, strandednesstemplate
- 4fgsea on hallmark setstemplate
Plan checked: no rule broken
- Field: Data (bulk RNA-seq, FPKM)
- Field: Groups (two-group contrast)
- Field: Replicates (14 vs 12)
- Field: Batch (batch: site, strandedness)
- Field: Covariates (covariates: sex, age)
- Field: Pairing (pairing: none)
- Rule: Log transform
- Rule: Sample structure
- Rule: Batch in the design
- Rule: Covariates
- Rule: FPKM input
- Rule: Ranked enrichment
- Method: PCA
- Method: limma-trend
- Method: fgsea
- Template: log-scale QC
- Template: limma-trend two-group
- Template: fgsea preranked
- Source: Zhao 2020
- Source: Conesa 2016
- Source: Nygaard 2016
- Source: Law 2014
Disputes show both sides
Where the field disagrees, the plane returns each position with its sources, and you choose.
Bad designs get a stop
A design that cannot support the inference returns a flag or a stop, and Pro tells you why.
Rules tested on real data
Rules are checked against public GEO series with shuffled labels, so a method that finds signal in noise loses its place.
Benchmark
Smaller models plan far better with it.
Smaller models with the plane out-planned a frontier model alone. The lower-cost models gained the most, including Qwen 3.8 27B, which a group can host on its own GPU. That makes a self-hosted model a serious option for data that cannot leave.
Grounding lifted every model we tested.
Plan quality, 0 to 100
| Model | Alone | With the plane | Cost per plan |
|---|---|---|---|
| Claude Opus 5 | 81.7 | 90.9 | $0.72 alone |
| GLM 5.3 | 69.8 | 89.2 | $0.14 with the plane |
| Qwen 3.8 27B | 62.9 | 83.4 | $0.09 with the plane |
Cost per plan
$0.72
Claude Opus 5 alone
$0.14
GLM 5.3 with the plane
up to 80% lower cost per plan
with a model you choose on your machine
Qwen 3.8 27B with the plane: $0.09 per plan
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.
Models
Any model, including one you host.
Pick the model per analysis: a frontier model from Anthropic or OpenAI, or any OpenAI-compatible endpoint, including one running on your own GPUs. The model behind every step goes into the record by name.
Ask a follow-up…GLM 5.3
In control
Every step approved. Every step recorded.
Pro asks before it installs a package, runs a command or starts a plan. Every step runs in a sandbox where outbound network access is blocked, and the runtime signs a record of what it read, ran and wrote.
Approved
inflexa store link limma fgsea --lang rAnti-PD-1 responder DE and GSEA
bulk RNA-seq · 10 steps · 5 agents
- T0S1Cohort assembly, depends on no step
- T0S2Log transform, depends on no step
- T1S1Responder DE, depends on T0S1, T0S2
- T2S1Immune composition, depends on T0S1
- T3S1GSEA, hallmark, depends on T1S1
- T4S1Covariate check, depends on T1S1
- T3S2Leading-edge genes, depends on T3S1
- T1S2Sensitivity fit, depends on T4S1
- T2S2Immune score contrast, depends on T2S1, T4S1
- T5S1Drug repurposing, depends on T2S1, T3S2, T1S2
Approved by youRun c98cfe
inflexa · prov
$ inflexa prov lineage out/de_results.tsv
out/de_results.tsv sha256:8e07…13cd
limma_trend_responders.R exit 0
GSE78220_PatientFPKM.xlsx sha256:3b7a…e41c
GSE78220 source
model glm-5.3 · your endpoint · step T1S1
rule Knowledge Plane · consensus
signed ed25519:9f3c…a71b
ran inflexa store link limma fgsea --lang r exit 0
$ inflexa prov verify
ok · chain intact · signature valid
Local by design
Your files stay in the folders they are in.
A local server, a browser tab
Pro runs a server on your own computer and you work in your browser. Your files never touch our infrastructure.
Inputs are references
Point Pro at the folders where your data already lives. Inflexa reads the paths and never copies the files.
The open engine underneath
Pro runs on the same Apache 2.0 engine as the CLI, so the methods, the sandbox and the record are the same.
Runs on macOS, Linux and Windows with Docker or Podman. Bring your own model and key.
Where Pro sits in the family.
| Inflexa CLI | Inflexa Pro | Inflexa Cloud | |
|---|---|---|---|
| Where you work | Your terminal | A browser workspace | A browser workspace |
| Where it runs | Your machine | Your machine | Inflexa Cloud |
| Knowledge Plane | Not included | Included | Biotech and Enterprise |
| Models | Any, including one you host | Any, including one you host | Claude |
| Signed record | Yes | Yes | Yes |
| Support | GitHub Issues and Discussions | Community on Personal, email on Biotech, a dedicated CSM on Enterprise | |
| Price | Free, Apache 2.0 | Per seat, per year · pricing on request | Free Personal plan, then Biotech and Enterprise |
Questions about Pro.
See Pro 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. Pricing on request.