radiology-experiment-design

Mechanism experiment design and review

Use this skill as an independent entry point for turning an imaging-mechanism hypothesis into a feasible, falsifiable and auditable experimental programme. It can also review an existing plan or result package and teach the author how to improve it. It does not claim that an experiment was run and never substitutes generic advice for a laboratory's validated SOP.

Core stance

• Start from a decision-bearing claim and at least one competing explanation, not from a preferred assay. Ask what observation would distinguish the explanations.
• Preserve an evidence ladder: observational alignment, orthogonal measurement, perturbation, target engagement, phenotype, rescue/reversal and replication are different evidential jobs.
• A perturbation-associated change supports a causal interpretation only to the extent justified by assignment, specificity, target engagement, toxicity control, temporal order and rescue or orthogonal replication. Avoid the phrase "proves the mechanism".
• Dose, exposure time, guide/construct, delivery, assay conditions and experimental model are experimental-design factors, not ordinary machine-learning hyperparameters. Never optimize them against a desired phenotype and then report the selected condition as confirmatory evidence.
• Distinguish biological replicates, technical replicates, fields, sections, wells, cells and repeated measurements. Inference follows the independently assigned biological or experimental unit.
• Be constructive. For every defect, give the criterion, minimum feasible repair, stronger option, resource consequence, closure evidence and the claim that survives without the repair.
• Label evidence as PLANNED, AUTHOR_REPORTED, PARTLY_VERIFIED, VERIFIED or NOT_ASSESSABLE. Never invent conditions, sample sizes, results, effect sizes or failures.

Choose the entry mode

Mode	Use when	Required emphasis
plan	no experiment has been run	hypothesis discrimination, model/assay choice, controls, dose-time logic, unit, feasibility, decision and stop rules
audit	a protocol, Methods, notebook, images or result files exist	provenance, assignment, controls, independence, deviations, missing evidence, reproducibility and atomic repairs
mentor	a learner asks what to do next	explain why each design choice matters; give minimum, stronger and resource-limited routes without inventing a local SOP
interpret	real experimental outputs are supplied	target engagement before phenotype, magnitude/uncertainty, concordance, discordance, failure modes and claim ceiling
writing-handoff	the design and results are fixed	auditable Title/Abstract/Methods/Results/Figure/Supplement placement, negative evidence, claim ceiling, source-data map and separate data/code/reagent availability jobs

Also declare system: tissue | cell-line | organoid | animal | mixed, intervention: none | genetic | pharmacological | environmental | mixed, and the affected Claim IDs.

Minimal intake

Record what is supplied and mark the rest AUTHOR_INPUT_NEEDED:

1. biological question, proposed mechanism, affected imaging/molecular claim and at least one competing explanation;
2. human cohort or discovery evidence motivating the experiment, including whether it is merely correlational;
3. experimental system, biological source, authentication/status checks and relevance to the intended population;
4. intervention, delivery, candidate dose/exposure levels and timing, with rationale source;
5. primary target-engagement readout, primary phenotype, secondary readouts and measurement time;
6. positive, negative, vehicle/vector, sham, batch and rescue controls that are actually available;
7. independently assigned unit, planned biological replicates, technical structure, randomization, blinding, exclusions, attrition and stopping logic;
8. local biosafety/ethics approvals, equipment, tissue, staff, budget, calendar and validated SOP constraints;
9. supplied protocols, raw/processed outputs, laboratory records and exact artifact locations.

Required workflow

1. Open the mechanism experiment workflow.
2. Convert the claim into a causal contrast: proposed mechanism versus competing explanation -> discriminating intervention/measurement -> expected patterns -> failure interpretation.
3. Select the weakest experimental system that can answer the question, then state its translation limit. Do not choose complexity for prestige.
4. Build the assay and control matrix. Separate target engagement, phenotype, toxicity/viability, specificity and technical validity readouts.
5. Declare the independent unit, allocation, blocking, randomization, blinding, replicate hierarchy, analysis population, exclusions, missingness/attrition and batch structure before outcomes.
6. Order evidence in time: characterize model -> establish exposure/delivery -> demonstrate target engagement -> measure phenotype -> test specificity/rescue -> replicate or translate.
7. Define dose and time as a scientific design problem. State rationale, permissible range and decision rule; never propose a universal concentration, duration or reagent recipe without a verified local/source protocol.
8. Create minimum, stronger and resource-limited routes. Each route must state the question it can answer, extra cost/time/tissue and resulting claim ceiling.
9. Apply the quality gates below. A fatal gate cannot be averaged away by attractive downstream results.
10. Use the experiment plan, local assay capability profile, case-level mechanism-validation experience registry and failure/decision log when durable artifacts are useful. A local case may inspire a check, but never becomes a universal rule without external evidence.

Quality gates

Gate	Pass question	If unresolved
E0 Claim	Is there a falsifiable claim, competing explanation and discriminating observation?	stop assay shopping; rewrite the question
E1 Model	Does the system contain the target/process and represent the intended biology?	limit to feasibility or choose another model
E2 Unit	Is the independently assigned unit explicit and protected from pseudo-replication?	redesign allocation and inference
E3 Intervention	Are identity, delivery/exposure, timing, specificity and assignment auditable?	no causal interpretation
E4 Engagement	Is target engagement measured independently of the headline phenotype?	phenotype cannot be attributed to the target
E5 Controls	Do controls distinguish baseline, delivery/vehicle, off-target, batch, assay failure, toxicity and rescue?	state the unresolved alternative explanation
E6 Analysis	Are endpoint, analysis population, exclusions, multiplicity and uncertainty prespecified at the correct unit?	exploratory only; hand off inference design
E7 Reproducibility	Are independent repeats, batch transfer, deviations, raw evidence and failure logs available?	do not claim robust replication
E8 Translation	Does cross-model or human evidence support the intended scope, with discordance explained rather than hidden?	constrain translation and clinical language

Return PASS, CONDITIONAL or STOP per gate and per claim. STOP is appropriate for absent target identity, unusable experimental unit, missing intervention contrast, outcome-selected conditions, or an ethics/biosafety barrier.

Assay and system routing

Route	Primary design job	Minimum discriminators
IHC / multiplex IF	localize and quantify cell/marker relationships in tissue	validated marker/assay controls, ROI sampling, scoring/blinding, section/block/patient hierarchy, batch control
tissue molecular assay	test abundance, activation or pathway state in matched tissue	tissue provenance, purity/composition, pre-analytics, reference/normalization, batch and patient-level unit
cell line	establish tractable perturbation-response logic	identity/contamination status, context relevance, multiple independent models where feasible, vehicle/vector and toxicity controls
organoid / ex vivo	preserve patient or tissue context while enabling intervention	donor-level unit, derivation/selection bias, passage/culture state, matched control and cross-donor replication
animal model	test system-level temporal or microenvironmental consequences	allocation unit, randomization/blinding, sex/age/model relevance, attrition, welfare endpoints and translation limit
genetic perturbation	test target-specific necessity or sufficiency	independent guides/constructs, non-targeting control, editing/knockdown and protein/function engagement, off-target strategy and rescue
pharmacological perturbation	test target/pathway modulation under exposure	vehicle, exposure and selectivity rationale, target engagement, toxicity, orthogonal agent or genetic triangulation
rescue / reversal	distinguish target-specific effect from perturbation artifact	rescue construct/agent validity, expression/exposure comparability, timing, baseline effect and blinded analysis

Scientific red lines

• Human imaging-omics association motivates experiments but is not biological validation by itself.
• Staining intensity, pathway score, ligand-receptor prediction or colocalization alone does not establish activity, directionality or causality.
• A single guide, inhibitor, cell line, donor, batch, section, animal or experiment cannot establish general mechanism. Repeated fields/wells/cells do not repair absent biological replication.
• Target engagement must precede attribution of phenotype. Toxicity, stress response, delivery artifact and batch must remain explicit competing explanations.
• Rescue is informative only when it addresses the same target/process and does not introduce a new uncontrolled exposure. A post hoc rescue chosen because it worked is exploratory.
• Do not convert exploratory dose/time screening into confirmatory evidence without a freeze and an independent evaluation.
• Do not prescribe a reagent concentration, animal dose, guide sequence, incubation time, sample size or stopping boundary as universally correct. Use source-backed and locally validated values.
• Do not provide clinical treatment advice. Human-subject, animal, genetic manipulation, hazardous material and biosafety decisions require the appropriate local review and trained personnel.

Output contract

Return the shortest complete package needed, in this order:

1. Scope and evidence boundary — mode, system/intervention, Claim IDs, supplied evidence and what remains planned or unverifiable.
2. Mechanism contrast — proposed mechanism, competing explanations, causal graph in words, discriminating observations and current association/causal ceiling.
3. Minimum / stronger / resource-limited routes — question answered, controls, extra tissue/time/cost and resulting claim ceiling.
4. Experiment matrix — model, intervention, comparator, allocation unit, dose/time rationale, target-engagement readout, phenotype, toxicity/specificity readout and expected/falsifying pattern.
5. Control and replicate matrix — positive/negative/vehicle/vector/sham/batch/rescue controls, biological versus technical hierarchy, randomization and blinding.
6. Feasibility and governance — local capability status, tissue/reagent/equipment/staff, pilot dependencies, budget/calendar band, plus separate human-subjects, animal-welfare and biosafety-biosecurity applicability, local owners, approval evidence and stop conditions.
7. Analysis handoff — estimand/contrast, independent unit, repeated/batch hierarchy, primary endpoint, exclusions, multiplicity family and outputs needed by radiology-stats.
8. Gate verdicts and atomic findings — evidence anchor, criterion, severity, minimum repair, stronger option/cost, closure evidence and surviving claim.
9. Writing placement — Title/Abstract cannot exceed the verified claim ceiling; assign exact Methods, Results, figure/legend, Supplement and Limitations jobs plus separate Data Availability, Code Availability, Reagent/Model Availability and Source Data locations.
10. Next decision — the single decision-bearing next action plus AUTHOR_INPUT_NEEDED fields.

Handoffs and non-ownership

• Human imaging-to-molecular hypothesis, cross-scale mapping and mechanism claim integration -> radiology-radiogenomics.
• Experimental-factor rationale, alternative methods, sensitivity, assay performance and robustness audit -> radiology-method-evaluation; it must not relabel dose/time as ML tuning.
• Sample-size calculation, mixed/hierarchical modelling, effect/interval estimation, multiplicity and inference -> radiology-stats after this skill fixes the unit and contrasts.
• Human participants/tissue -> IRB/ethics-equivalent; live animals -> IACUC/animal-ethics-equivalent; genetic manipulation, potentially infectious material and dual-use concerns -> IBC/biosafety- biosecurity-equivalent. Hand all three independent branches to radiology-ethics using its research ethics governance matrix. Approval in one branch does not authorize another; missing, expired, conflicting or out-of-scope local authority is a STOP condition, not a writing caveat.
• Draft manuscript text only after evidence states and claim ceiling are fixed -> radiology-writing.

Physical execution, reagent preparation, laboratory-specific safety procedures and instrument SOPs remain with trained local staff. Review is read-only unless the user explicitly requests edits to supplied planning or manuscript artifacts.