huang-sir1 Radiogenomics and Imaging-Multi-Omics
Use this skill to plan, analyse, report, submit, and revise studies that connect imaging phenotypes (radiomics, deep features, or spatial habitats) to molecular data: bulk genomics or transcriptomics, single-cell RNA-seq, deconvolution, and spatial omics. This is one of the highest-difficulty corners of imaging research because the analysable cohort is the matched intersection of imaging and omics, both data spaces are high-dimensional, and scanner/site and sequencing batch effects can masquerade as biology.
Core stance
- Match first, then mine. State the patients with both usable imaging and usable omics first; that matched n drives design, power, claims, and journal tier.
- Map tissue to image. A molecular sample is not automatically the whole tumour. Record timing, lesion, region, treatment interval, and whether the analysis is patient-, lesion-, habitat-, or section-level.
- Separate confirmation from discovery. Pre-specify the primary hypothesis and analysis plan; FDR-control discovery scans and validate independently whenever possible.
- Batch can look like biology. Scanner/site/protocol and sequencing batch/platform/center must be recorded, adjusted or harmonised appropriately, and tested in sensitivity analyses.
- Reproducible imaging and omics. Radiomics must be IBSI/CLEAR-aligned; omics QC, filtering, normalization, batch correction, accessions, and software versions must be explicit.
- Interpret as association unless proven otherwise. Pathways, cell types, and spatial evidence strengthen biological interpretation but usually do not prove mechanism.
- Submission-ready integrity. Never invent cohort counts, accessions, p values, effect sizes, approvals, validation results, or reviewer-response locations.
When to use
- Designing a TCIA-TCGA, GEO, dbGaP/EGA, cBioPortal, in-house, or multi-center radiogenomics study.
- Linking radiomic/deep features with mutations, gene expression, methylation, CNV, proteomics, molecular subtypes, pathway activity, immune/cell-type composition, or prognosis.
- Integrating imaging with multi-omics using MOFA/MOFA+, iCluster, SNF, DIABLO/mixOmics, sparse CCA, multi-block PLS, NMF, or related methods.
- Connecting imaging habitats to scRNA-seq deconvolution or spatial transcriptomics.
- Drafting a protocol, statistical analysis plan, Methods, Results, Discussion, supplement, or submission package for a radiogenomics manuscript.
- Auditing a manuscript or reviewer comments for leakage, batch confounding, small-n optimism, tissue-image mismatch, overclaiming, and incomplete data/code availability.
When to open extra files
| File | Open when |
|---|---|
| references/cohort-design.md | Choosing data sources, matching imaging to omics, estimating matched n, planning validation and ethics |
| references/sample-to-image-mapping.md | Determining whether tissue, biopsy, histology, spatial omics, or lesion sampling actually matches the imaging ROI/habitat |
| references/radiomics-pipeline.md | Building the imaging side: segmentation, IBSI features, deep features, habitats, registration, harmonisation |
| references/omics-qc-preprocessing.md | Preparing RNA-seq, mutation, methylation, CNV, proteomics, or other molecular matrices with QC, normalization, batch handling |
| references/analysis-plan-sap.md | Writing a protocol/SAP, defining primary versus discovery analyses, covariates, FDR, validation, sensitivity analyses |
| references/multi-omics-integration.md | Choosing integration strategy and methods: MOFA+, iCluster, SNF, DIABLO/mixOmics, sparse CCA, fusion strategies |
| references/deep-radiogenomics-fusion-strategies.md | Deep radiomics, foundation-model embeddings, radiopathomics, cross-attention/joint embeddings, pathway-informed fusion, disease-endpoint prioritisation, or modern hybrid fusion strategy |
| references/single-cell-spatial.md | scRNA-seq deconvolution, pseudobulk, spatial transcriptomics, and habitat linkage |
| references/association-validation.md | Feature-gene/pathway association, GSEA/ssGSEA, multiple testing, radiogenomic signatures, validation |
| references/biological-validation.md | Calibrating biological claims and adding pathway, IHC, spatial, single-cell, or orthogonal validation support |
| references/pitfalls.md | Auditing leakage, batch effects, double-dipping, small-n optimism, spatial mismatch, overclaiming |
| references/radiogenomics-submission-package.md | Preparing the manuscript, supplement, checklists, data/code availability, cover-letter angle, reviewer suggestions |
| references/reviewer-playbook.md | Simulating radiogenomics reviewer concerns or drafting rebuttal logic for batch, validation, mapping, and mechanism critiques |
Workflow
- Define the linkage question: imaging phenotype, molecular layer, endpoint, disease context, discovery versus prediction, and whether the intended claim is association, prediction, or biology.
- Assemble the matched cohort: sources, eligibility, the imaging-omics intersection, exclusions, and a validation cohort before modelling.
- Map sample to image: tissue source, lesion/region/habitat, time interval, treatment exposure, and mapping level. Bound claims when mapping is weak.
- Write the protocol/SAP: primary hypothesis, covariates, multiplicity, validation criterion, missing-data handling, sensitivity analyses, and leakage controls.
- Build the imaging side: segmentation/annotation, IBSI radiomics or deep features, registration, habitat definitions, stability filtering, and scanner/site harmonisation.
- Prepare omics: assay-specific QC, filtering, normalization, batch correction, feature definitions, accessions, software versions, and controlled-access constraints.
- For deep/hybrid radiogenomics, open
deep-radiogenomics-fusion-strategies.mdand decide whether early fusion, late fusion, joint embedding, pathway graph, radiopathomics, or foundation-model adapter is justified by matched n and validation. - Integrate or associate: choose association, pathway analysis, supervised prediction, or formal multi-omics integration. Keep all data-dependent operations inside training or discovery only.
- Validate and interpret: replicate direction/effect size, add biological corroboration where available, and keep mechanistic language bounded.
- Report and submit: map to CLEAR/IBSI and the appropriate prediction/diagnostic/observational guidelines, prepare supplement and data/code statements, then run pre-review and journal selection.
- Revise with traceability: classify reviewer comments, perform feasible analyses, soften unsupported claims, and cite exact manuscript/supplement locations.
Output contract
For design or audit tasks, return as many of these as the task requires:
Linkage design: phenotype, omics layer, endpoint, claim type, integration strategy.Matched cohort: sources, intersection n, exclusion logic, validation cohort, limiting count.Sample-to-image map: timing, lesion/region/habitat, tissue source, mapping level, uncertainty.Protocol/SAP: primary hypothesis, covariates, FDR/test family, validation rule, sensitivity analyses.Pipeline: imaging pipeline and omics QC/preprocessing with leakage-prone steps marked train-only.Analysis: association/integration/prediction method, multiplicity control, validation, code/tool route.Fusion strategy: baseline ladder, early/late/joint/pathway/radiopathomics route, matched-n justification, missing-modality and overfitting controls.Biological interpretation: claim level, pathway/cell/spatial evidence, alternative explanations.Reporting map: CLEAR/IBSI plus TRIPOD+AI/STARD/STROBE/REMARK/omics standards as applicable.Submission package: main-manuscript requirements, supplement tables, checklists, data/code/accessions.Reviewer risk list: likely radiogenomics critiques and concrete fixes or response strategy.Author input needed: any missing counts, accessions, approvals, line numbers, software versions, or results.
Handoffs
- Study feasibility, validation strategy, and cohort architecture ->
radiology-design. - Dataset/literature search and accession verification ->
radiology-search/radiology-citation. - Segmentation SOP and reproducibility ->
radiology-annotation. - IBSI/CLEAR/METRICS/RQS, TRIPOD+AI, STARD, STROBE routing ->
radiology-reporting. - ROC, calibration, DCA, survival, DeLong, MRMC, sample size, multiplicity ->
radiology-stats. - Figures: habitats, MOFA factors, heatmaps, deconvolution bars, KM, flow diagrams ->
radiology-figure. - Data/code availability, DICOM de-identification, GEO/dbGaP/EGA/Zenodo/GitHub wording ->
radiology-data. - Ethics, consent, DUA, HIPAA/GDPR/PIPL, genomic re-identification risk ->
radiology-ethics. - Manuscript drafting and claim calibration ->
radiology-writing/radiology-polishing. - Pre-submission mock review ->
radiology-prereview; journal ladder ->radiology-journal; rebuttal ->radiology-response. - Emerging linkage themes (liquid biopsy/ctDNA, pathology-foundation-model fusion) and whether the
data can carry them ->
radiology-frontier. - Reframing this research as a funding proposal ->
radiology-grant.
This skill guides design, analysis logic, reporting, and submission readiness. It does not replace a genomics/bioinformatics collaborator for production pipelines or institutional legal/ethics review.