Scientific preflight for differential expression

Better biology starts with a defensible comparison.

AI is making biological discovery more accessible. Differential gene expression can reveal how disease, treatment, or environment changes gene activity—but even flawless analysis code cannot rescue messy metadata or a design that cannot support the question.

DEfensible closes that gap. GPT‑5.6 clarifies metadata and research intent; deterministic checks catch confounding, pairing, replication, and rank failures before they waste sequencing, laboratory, or translational effort.

Audit a study

No count matrix required
No uploaded data stored
Explicitly abstains

The missing quality gate

01
UnderstandGPT‑5.6 structures metadata and intent
02
ProveCode tests confounding, pairing, and rank
03
ActProceed, revise, or abstain with evidence

Interactive audit

Ask the question before you run the model.

Start with a curated study or upload a CSV/TSV sample sheet. The public demo accepts research metadata only—never patient-identifiable information.

Input

Sample metadata + research question

12 samples · confounded-treatment.csv

Use your own sample sheet

CSV or TSV · up to 150 rows · 250 KB

Choose file

What biological effect do you want to estimate?

Name target, reference, subset, and factors to account for.96/600

Processed in memory. API requests use store: false.

Human checkpoint

Confirm the interpretation

Curated example interpretation

Outcome factorTargetReference

Model factorsbatch + treatment

Inspect and correct normalized sample factors

Editable normalized factor values for each sample.
Sample	treatment	batch	cell_type	timepoint	replicate
vehicle_batch_1_r1	treatment for vehicle_batch_1_r1	batch for vehicle_batch_1_r1	cell_type for vehicle_batch_1_r1	timepoint for vehicle_batch_1_r1	replicate for vehicle_batch_1_r1
vehicle_batch_1_r2	treatment for vehicle_batch_1_r2	batch for vehicle_batch_1_r2	cell_type for vehicle_batch_1_r2	timepoint for vehicle_batch_1_r2	replicate for vehicle_batch_1_r2
vehicle_batch_1_r3	treatment for vehicle_batch_1_r3	batch for vehicle_batch_1_r3	cell_type for vehicle_batch_1_r3	timepoint for vehicle_batch_1_r3	replicate for vehicle_batch_1_r3
vehicle_batch_1_r4	treatment for vehicle_batch_1_r4	batch for vehicle_batch_1_r4	cell_type for vehicle_batch_1_r4	timepoint for vehicle_batch_1_r4	replicate for vehicle_batch_1_r4
vehicle_batch_1_r5	treatment for vehicle_batch_1_r5	batch for vehicle_batch_1_r5	cell_type for vehicle_batch_1_r5	timepoint for vehicle_batch_1_r5	replicate for vehicle_batch_1_r5
vehicle_batch_1_r6	treatment for vehicle_batch_1_r6	batch for vehicle_batch_1_r6	cell_type for vehicle_batch_1_r6	timepoint for vehicle_batch_1_r6	replicate for vehicle_batch_1_r6
drug_batch_2_r1	treatment for drug_batch_2_r1	batch for drug_batch_2_r1	cell_type for drug_batch_2_r1	timepoint for drug_batch_2_r1	replicate for drug_batch_2_r1
drug_batch_2_r2	treatment for drug_batch_2_r2	batch for drug_batch_2_r2	cell_type for drug_batch_2_r2	timepoint for drug_batch_2_r2	replicate for drug_batch_2_r2
drug_batch_2_r3	treatment for drug_batch_2_r3	batch for drug_batch_2_r3	cell_type for drug_batch_2_r3	timepoint for drug_batch_2_r3	replicate for drug_batch_2_r3
drug_batch_2_r4	treatment for drug_batch_2_r4	batch for drug_batch_2_r4	cell_type for drug_batch_2_r4	timepoint for drug_batch_2_r4	replicate for drug_batch_2_r4
drug_batch_2_r5	treatment for drug_batch_2_r5	batch for drug_batch_2_r5	cell_type for drug_batch_2_r5	timepoint for drug_batch_2_r5	replicate for drug_batch_2_r5
drug_batch_2_r6	treatment for drug_batch_2_r6	batch for drug_batch_2_r6	cell_type for drug_batch_2_r6	timepoint for drug_batch_2_r6	replicate for drug_batch_2_r6

See evidence behind the interpretation60 mappings

Sample	Source evidence	Normalized as
vehicle_batch_1_r1	`treatment=vehicle`	treatment → vehicle
vehicle_batch_1_r1	`batch=batch_1`	batch → batch_1
vehicle_batch_1_r1	`cell_type=monocyte`	cell_type → monocyte
vehicle_batch_1_r1	`timepoint=24h`	timepoint → 24h
vehicle_batch_1_r1	`replicate=1`	replicate → 1
vehicle_batch_1_r2	`treatment=vehicle`	treatment → vehicle
vehicle_batch_1_r2	`batch=batch_1`	batch → batch_1
vehicle_batch_1_r2	`cell_type=monocyte`	cell_type → monocyte
vehicle_batch_1_r2	`timepoint=24h`	timepoint → 24h
vehicle_batch_1_r2	`replicate=2`	replicate → 2
vehicle_batch_1_r3	`treatment=vehicle`	treatment → vehicle
vehicle_batch_1_r3	`batch=batch_1`	batch → batch_1
vehicle_batch_1_r3	`cell_type=monocyte`	cell_type → monocyte
vehicle_batch_1_r3	`timepoint=24h`	timepoint → 24h
vehicle_batch_1_r3	`replicate=3`	replicate → 3
vehicle_batch_1_r4	`treatment=vehicle`	treatment → vehicle
vehicle_batch_1_r4	`batch=batch_1`	batch → batch_1
vehicle_batch_1_r4	`cell_type=monocyte`	cell_type → monocyte

Showing 18 of 60 mappings. The complete provenance is included in the JSON export.

Not estimable

The data cannot separate the requested effect.

Running the analysis would produce coefficients, but they would not support the biological claim.

Target6drug

Reference6vehicle

Residual df10rank 2/3

Deterministic audit

3 checks with traceable evidence

2 unsafe0 review1 passed

PERFECT_CONFOUNDINGbatch is perfectly confounded with treatmentunsafe

Every batch level belongs to only one contrast arm. The model cannot distinguish a treatment effect from a batch effect.

Next step: Collect both contrast arms within the same batch levels or change the scientific claim.

batch_1: vehicle
batch_2: drug

SINGULAR_DESIGN_MATRIXThe design matrix is not full rankunsafe

Rank 2 is lower than 3 model columns. At least one coefficient is a linear combination of others.

Next step: Remove redundant covariates or redesign the comparison.

REPLICATION_PRESENTBoth arms have biological replicationPassed

drug: 6 samples; vehicle: 6 samples.

Observed overlap

Contrast arms by Batch

Every column should contain evidence from both arms when this factor must be adjusted.

Sample counts for each contrast arm and batch level.
Arm	batch_1	batch_2
vehicle	6	0
drug	0	6

Recommended modelFormula withheld

DEfensible will not generate analysis code for a design that cannot support the claim.

Reproducible handoff

Export the decision, not just the answer.

Built for scientific trust

Interpretation where needed. Proof where possible.

DEfensible separates semantic reasoning from statistical verification so every claim has the right kind of evidence.

01 · AI interprets

Messy labels become explicit intent.

GPT‑5.6 maps free text and metadata into factors, arms, subsets, and evidence—while surfacing uncertainty instead of hiding it.

02 · Code verifies

Math decides whether the claim is estimable.

Replication, overlap, pairing, confounding, rank, and residual degrees of freedom are checked by deterministic, tested code.

03 · Humans decide

Every assumption remains inspectable.

Researchers can correct normalized factors, review exact source cells, and export a transparent record for expert review.

The problem

Code can run successfully and still support the wrong conclusion.

“Treatment and batch encode the same partition of samples. No statistical model can distinguish their effects in this dataset.”

DEfensible catches that failure before researchers spend hours on an analysis—or carry an invalid comparison into a figure, manuscript, or decision.