Radiology Transcriptomics Analysis

Transcriptomics analysis execution

Use this skill when the requested object is an executable or auditable bulk RNA, sc/snRNA, or spatial-transcriptomics workflow. It can be entered directly with raw matrices, processed objects, code, logs, or only a proposed analysis. It does not require an imaging dataset.

Core contract

• Separate PLAN_ONLY, CODE_READY, RUNNING, RUN_COMPLETE, RUN_FAILED, AUDIT_ONLY, and REPRODUCED. Never turn a plan, code draft, screenshot, reported result, or partial log into a completed run.
• Preserve the chain source files -> sample/donor map -> QC decisions -> transformed object -> analysis object -> result table/figure -> claim handoff with exact paths and versions.
• Use the biological sampling unit for inference. Cells, nuclei, spots, genes, tiles, fields, technical replicates, bootstrap draws, seeds and clusters are not new patients or donors.
• Build the shortest workflow that answers the locked question. Optional trajectory, cell-cell communication, GRN, deconvolution or integration is activated only when its assumptions and decision value are explicit.
• Never invent data availability, sample annotations, reference labels, package versions, QC thresholds, commands, run success, results, cell identities, spatial domains or mechanisms.
• Review is read-only by default. Build, edit, or execute code only when the user requests it and the target inputs, working directory, output root, environment and protected-data boundary are known.

Choose the entry mode

Declare exactly one primary modality: bulk-rna, scrna, snrna, or spatial. Declare secondary modalities only when there is a real matched or bridge analysis. If the user asks whether a parameter or method choice is defensible rather than to configure or run it, route that decision to radiology-method-evaluation.

Minimal intake

Record what is available and mark the rest AUTHOR_INPUT_NEEDED:

1. question, endpoint/contrast, intended use, affected Claim IDs and whether imaging is involved;
2. source type and exact locations: FASTQ, counts, features/barcodes, matrices, object files, histology images, coordinates, annotations, code, logs and prior outputs;
3. donor/patient/specimen/region/section/batch/library hierarchy and matched-sample map;
4. platform/chemistry/reference build or assay version when known; strandedness and feature definition for bulk; spot/bin/cell resolution for spatial;
5. clinical covariates, group/contrast, repeated measures, confounders and protected test state;
6. environment, package lock, compute limits, output root, seed policy and execution authority;
7. pre-existing QC thresholds, exclusions, reference atlases, labels, checkpoints and deviations.

If donor identity, sample-to-donor mapping, contrast, output root or execution authority is missing, planning may continue but execution returns RUN_BLOCKED.

Required workflow

1. Open the modality workflows and the runtime/reproducibility contract.
2. Freeze an intake manifest and choose the narrowest modality branch.
3. Define the true independent unit, analysis unit, repeated-measure structure and aggregation rule before filtering, integration or testing.
4. Create a decision ledger for every irreversible or conclusion-bearing choice: inclusion, filtering, normalization, reference/annotation, integration, contrast, spatial scale and exclusion.
5. Write the analysis plan and expected-output contract. Preserve a protected-test boundary where applicable.
6. In build, create code/configuration with explicit inputs, outputs, versions, seeds, assertions, checkpoints and failure behavior. Run only safe dry checks until real execution is authorized.
7. In run-audit, record the exact command, working directory, environment, start/end state, log, exit status, warnings, QC decisions, deviations and artifact checksums where feasible.
8. Stop on mapping ambiguity, corrupt/incompatible objects, sample-count mismatch, output overwrite risk, unresolved privacy boundary, or a requested analysis that changes the locked estimand without approval.
9. Package real outputs for statistical inference, biological interpretation, method evaluation, writing and review without silently upgrading evidence state.

Use the analysis plan, run manifest, QC decision log, local runtime profile, and local analysis experience registry when persistent artifacts are useful. Local experience can guide failure diagnosis only inside its recorded platform/tissue/data-state boundary; it never becomes an unreported default.

Modality routes

Scientific red lines

• Do not treat cell/spot-level P values as patient-level evidence. For case-control or clinical claims, use donor/patient-aware models, pseudobulk or another justified hierarchical strategy.
• Integration is not automatic quality improvement. It can erase real biological or clinical differences; preserve unintegrated evidence and test whether the intended contrast remains.
• Cluster labels are hypotheses supported by declared markers/references and discordance checks, not ground truth supplied by an algorithm.
• Post hoc QC tuned until the desired biology appears is outcome leakage. Record timing and inspect conclusion stability under defensible alternatives.
• Gene-set enrichment depends on the tested universe, ranking/statistic, database version and multiplicity family. A pathway label is not direct pathway activity or causal mechanism.
• Deconvolution, imputation, virtual cells/spots, spatial completion and inferred interactions are estimates. They are not measured cells, molecules, physical contact, perturbation or validation.
• Trajectory ordering is an inferred geometry; it does not establish real time, lineage or causal direction without external evidence.
• A spatial neighborhood result must state coordinate system, radius/kernel/k, edge handling, null model, section/patient aggregation and sensitivity to scale.
• Cross-modal agreement in the same cohort is orthogonal concordance, not independent replication.
• Never overwrite raw inputs or the only copy of a processed object. Write to an explicit output root and preserve provenance.

Output contract

Return only the sections needed for the current mode, but preserve this order:

1. Execution scope — mode, modality, question/contrast, hierarchy, independent unit, evidence state, protected-test state and supplied artifacts.
2. Input and identity audit — source paths, dimensions/checksums when available, donor/sample mapping, missing/duplicate identities and blockers.
3. Workflow and decision ledger — each step's input, output, rationale, alternatives, assumption, decision owner, freeze state and failure condition.
4. Run record — exact command/environment/seed, start/end/exit status, log path, deviations and monitored output paths; omit only in plan.
5. QC and exclusion ledger — metric, level, rule, timing, affected units, retained/excluded counts, reason, sensitivity need and author approval.
6. Artifact inventory — object/table/figure/report path, schema/dimensions, checksum or other integrity evidence, producer and verification state.
7. Findings and failure boundary — real observed results only, alternative explanations, unresolved confounding and what failed or was not run.
8. Handoffs — exact packet for inference, method judgement, mechanism interpretation, wet-lab validation, writing or prereview.
9. State and next action — one of the declared execution states, blockers, surviving claim ceiling and the single next decision-bearing action.

In mentor, every major recommendation also reports: why it matters, minimum repair, stronger alternative, resource/time burden, expected failure signal, fallback, closure evidence and the claim that survives if it cannot be repaired. In writing-handoff, do not draft around missing evidence: map frozen workflow/version/unit/QC to Methods; verified prespecified or labeled exploratory outputs to Results; source-data locators to figures/tables; full exclusions/sensitivities/environment/ deviations to Supplement; alternatives and limits to Discussion; and enforce the weakest-link claim ceiling in Abstract/title.

Handoffs and non-ownership

• Biological interpretation, alternative mechanism hypotheses, cross-scale imaging bridge and claim ceiling -> radiology-radiogenomics, using only verified result artifacts.
• Parameter provenance, method fit, integration/normalization alternatives, sensitivity and fair comparison -> radiology-method-evaluation.
• Donor-aware effect estimation, confidence intervals, multiplicity, hierarchical modelling or survival inference -> radiology-stats after this skill freezes units and contrasts.
• Functional/perturbational wet-lab validation design -> radiology-experiment-design.
• Data governance, de-identification and sharing -> radiology-data and radiology-ethics.
• Frozen replay packaging, prespecified tolerance, operator independence and cumulative evidence level -> radiology-reproducibility; this skill retains assay/workflow-specific rerun and QC.
• Fixed-evidence manuscript construction -> radiology-writing; whole-manuscript review -> radiology-prereview; upload-file audit -> radiology-submission.

This skill may explain what an output means operationally, but a computational association, estimated cell state, inferred spatial relation or computational perturbation is not causal proof.