Thyroid Cancer Multi-Omics: Clinical Covariates, Transcriptomic Programs, Somatic Drivers, and Precision-Oncology Integration

Expanded technical portfolio case study of non-medullary thyroid cancer integrating clinical phenotype engineering, covariate-adjusted RNA-seq, pathway/signaling analysis, inflammatory programs, somatic driver architecture, demographic associations, and staged DNA–RNA integration.

Why this repository exists

This project is not a single differential-expression analysis. It was designed around four linked precision-oncology questions:

Sociodemographic genomic disparities — do somatic drivers, molecular groups, pathway alterations, and mutation burden differ by race, age, sex, or disease status?
Immune and inflammasome architecture — do inflammatory/immune programs differ by AWD/NED, autoimmune thyroiditis, or demographics?
Prognostic pathway stratification — are MAPK, PI3K/AKT/mTOR, stress-response, and other tumor programs associated with AWD versus NED?
Multi-omics integration / validation — do DNA driver states map to RNA phenotypes, and can major findings be compared with TCGA-THCA after appropriate harmonization?

Real project scale

Layer	Aggregate real project value
Final non-medullary RNA AWD/NED cohort	130
NED	91
AWD	39
RNA samples with reported race	107
Genes tested	24,364
DNA MAF samples	139
DNA samples with clinical metadata	121
DNA AWD/NED subset	101
Low-depth DNA samples flagged	20

End-to-end architecture

Clinical metadata
AWD/NED + race + age + sex + histology + thyroiditis evidence
                           |
             cohort / endpoint harmonization
                           |
             +-------------+-------------+
             |                           |
          RNA-seq                        DNA
     Salmon gene counts            Funcotator MAFs
             |                           |
    count QC / filtering          variant QC / reconciliation
             |                           |
      DESeq2 adjusted model        curated driver layer
 age + sex + disease status             |
             |                   patient alteration matrix
        VST / PCA                       |
             |                    MAPK / PI3K / repair
      gene-level effects                |
             |                   clinical exact models
  Hallmark / GO / KEGG / Reactome       |
             |                           |
 thyroid + stress + immune programs     |
             +-------------+-------------+
                           |
                   matched DNA–RNA
                     integration

1. Clinical phenotype engineering and covariates

The RNA analysis does not treat the clinical table as passive metadata.

Disease endpoint

NED is the reference and AWD is the disease group of interest. Endpoint mapping is kept auditable because biochemical/structural incomplete response, recurrence, persistence, or similar labels can materially change the phenotype.

Histology

Histology is refined using structured fields and pathology text. Medullary thyroid cancer is excluded. A study-level covariate collapses histology to:

Papillary / Follicular / FVPTC
All other non-medullary

Demographics

The analysis retains:

age as a continuous covariate;
age groups <45, 45–54, 55–64, 65+;
a <55 versus 55+ indicator;
sex;
harmonized race group.

Adjusted clinical model

A separate clinical model evaluates AWD in relation to available demographics and histology:

glm(
  AWD ~ age_at_biopsy + sex + race_group + study_histology_group,
  family = binomial()
)

See docs/COVARIATE_MODELING.md.

2. RNA-seq: QC, modeling, and global structure

Salmon-derived counts are validated, aligned to clinical metadata, and filtered before DESeq2.

The primary source model is:

design = ~ age_at_biopsy + sex + disease_group

Samples missing required model covariates are excluded explicitly rather than silently retained.

DESeq2 outputs include:

all genes;
BH-FDR significant genes;
strong-effect genes (FDR < 0.05, |log2FC| >= 1);
genes higher in AWD;
genes higher in NED;
normalized counts;
VST expression;
PCA coordinates.

Real aggregate DE result

Result	N
Genes tested	24,364
FDR < 0.05	354
Higher in AWD, FDR < 0.05 and \|log2FC\| ≥ 1	110
Higher in NED, FDR < 0.05 and \|log2FC\| ≥ 1	41
Race-adjusted FDR < 0.05 sensitivity result	230

The race-adjusted result is kept conceptually separate from the primary age/sex-adjusted source model.

3. Candidate expression biomarkers

The project evaluated effect size together with descriptive discrimination rather than ranking genes on P value alone.

Gene	log2FC	AUC	95% CI
IL1A	4.59	0.67	0.570–0.777
SLC2A1	2.30	0.64	0.528–0.749
GJB6	3.18	0.61	0.505–0.721
GJB2	3.37	0.60	0.495–0.714
ZPLD1	2.55	0.59	0.483–0.697

ROC analyses are discovery-level and are not presented as clinically validated classifiers.

4. Pathway enrichment and signaling biology

DESeq2 statistics are used for preranked FGSEA across:

Hallmark
GO Biological Process
KEGG
Reactome

The analysis tracks normalized enrichment score, FDR, and whether a pathway is enriched toward AWD or NED.

Prespecified tumor/thyroid programs

The project also evaluates coordinated sample-level programs:

Program	Genes detected	NED median	AWD median	FDR
Mitochondrial metabolism	10	-0.0009	0.148	0.590
Oxidative stress and ferroptosis	10	-0.0343	0.112	0.590
MAPK signaling	12	-0.212	-0.0415	0.749
PI3K/AKT signaling	10	-0.350	-0.322	0.749
Thyroid differentiation / iodine metabolism	11	0.052	-0.0481	0.846
EMT / invasion	10	0.0387	0.0033	0.896
Inflammation / interferon response	10	0.0052	-0.130	0.915

No prespecified program remained FDR-significant in the unadjusted AWD/NED program-level comparison. Mitochondrial and oxidative-stress distributions nevertheless motivated covariate-adjusted follow-up.

See docs/PATHWAY_AND_SIGNALING.md.

5. Inflammasome and immune architecture

The prespecified inflammasome panel contains:

NLRP3, NLRP1, NLRC4, AIM2, PYCARD, CASP1, CASP4, CASP5, IL18, GSDMD, NFKB1, RELA, TXNIP.

Individual genes

Gene	NED median	AWD median	Difference	P	FDR
CASP4	7.95	8.58	+0.627	0.0154	0.200
GSDMD	7.72	7.86	+0.132	0.121	0.514
NLRP3	7.29	7.02	-0.270	0.206	0.535
CASP1	7.04	7.17	+0.126	0.382	0.828

CASP4 is the strongest individual-gene signal, but it does not remain FDR-significant.

Composite program and covariates

The composite inflammasome score is constructed from gene-wise standardized VST expression.

Aggregate project results include:

race association: Kruskal P=0.0155, FDR 0.031;
Black vs White pairwise BH P=0.019;
adjusted Black vs White coefficient β=-0.439, P=0.0285;
adjusted AWD coefficient β=0.367, P=0.0317.

These results illustrate why covariate-aware modeling changes the interpretation of pathway-level signals.

Thyroiditis

Autoimmune-thyroiditis evidence is defined conservatively using pathology and/or clearly positive thyroid-autoantibody evidence. When immune-deconvolution outputs are available, cell-type estimates can be evaluated by thyroiditis, disease status, race, age group, and sex.

See docs/INFLAMMASOME_AND_IMMUNE.md.

6. Somatic driver architecture

The DNA workflow reconciles provisional and curated calls into an analysis-ready driver layer and collapses them to patient-level alteration matrices.

Real aggregate driver prevalence

Gene	Mutated	Prevalence
BRAF	68	48.9%
HRAS	10	7.2%
NRAS	9	6.5%
PIK3CA	5	3.6%
PTEN	5	3.6%
KRAS	4	2.9%
RET	4	2.9%
TP53	4	2.9%

Rare alterations are analyzed preferentially at the pathway level rather than overinterpreted individually.

7. Canonical MAPK molecular groups

The project defines three patient-level states:

Group	N	%
BRAF V600E	64	46.0%
Other / no canonical MAPK hotspot	56	40.3%
RAS hotspot	19	13.7%

This grouping is used for clinical interpretation and future DNA–RNA pathway integration.

8. DNA signaling pathways

The public configuration explicitly records pathway membership.

MAPK

BRAF, NRAS, HRAS, KRAS, NF1, RET

PI3K / AKT / mTOR

PIK3CA, AKT1, PTEN, MTOR, TSC1, TSC2

Cell cycle / tumor suppression

TP53, RB1, APC

DNA repair

ATM, ATR, CHEK2, BRCA1, BRCA2

Thyroid-associated

DICER1, MEN1

This makes the biological decision logic reviewable instead of burying it in code.

9. Race-stratified molecular architecture

In the current mutation-only analysis, canonical MAPK molecular-group distribution differed between Black and White patients:

Fisher exact P = 0.000583.

The observed pattern suggested relatively greater RAS-hotspot and lower BRAF V600E representation among Black/African American patients in this cohort.

This is a discovery-level association, not a causal ancestry claim. The project explicitly calls for adjustment/sensitivity analysis incorporating histology, age, sex, sequencing depth, and sample size.

10. AWD/NED driver comparisons

Current mutation-only observations indicate that:

BRAF prevalence is similar in AWD and NED;
NRAS is concentrated more heavily in AWD in the plotted subset;
RET appears more frequent in NED in the plotted subset;
sparse drivers require exact/pathway-level modeling.

The prognosis question cannot be fully resolved from mutation-only data because CNV and fusion events are not yet fully integrated.

11. DNA–RNA integration

The matched multi-omics layer is designed to compare DNA states such as BRAF/RAS/PI3K with RNA phenotypes including:

MAPK signaling
mTORC1 / PI3K activity
E2F / G2M / MYC
glycolysis
oxidative phosphorylation
EMT
interferon / inflammation
oxidative stress
thyroid differentiation

See docs/MULTIOMICS_INTEGRATION.md.

12. What is completed versus staged

A major principle of this public repository is not to present planned work as completed.

Supported by current project materials

cohort / metadata harmonization
histology refinement and medullary exclusion
clinical Table 1 / missingness summaries
adjusted AWD demographic model
DESeq2 + VST/PCA
race-adjusted RNA sensitivity result
FGSEA
thyroid/stress program scoring
inflammasome gene / composite analyses
MAF/driver reconciliation
primary driver prevalence
canonical MAPK molecular groups
discovery-level race-stratified MAPK comparison

Conditional / partially implemented

thyroiditis annotation
immune deconvolution
expanded robust demographic sensitivity modeling

Staged

full CNV integration
full fusion/rearrangement integration
final mutational-signature layer
fully integrated mutation+CNV+fusion outcome models
TCGA-THCA validation

See docs/COMPLETION_STATUS.md.

Repository structure

config/
  analysis_parameters.yml
  gene_sets.json
data/synthetic/
docs/
scripts/R/
results/
  aggregate_real_results/
  example_tables/
scripts/python/
.github/workflows/

Reproducibility

The synthetic public layer is designed to exercise the workflow without redistributing restricted data.

python scripts/python/validate_inputs.py
python scripts/python/privacy_check.py

Rscript scripts/R/01_build_clinical_cohort.R
Rscript scripts/R/02_table1_and_missingness.R
Rscript scripts/R/03_adjusted_awd_clinical_model.R
Rscript scripts/R/04_prepare_counts.R
Rscript scripts/R/05_deseq2_primary_awd_ned.R
Rscript scripts/R/06_vst_pca.R
Rscript scripts/R/09_thyroid_program_scores.R
Rscript scripts/R/13_driver_reconciliation.R
Rscript scripts/R/14_mapk_molecular_groups.R
Rscript scripts/R/15_driver_prevalence.R

Skills demonstrated

Clinical bioinformatics: endpoint harmonization, histology curation, missing-data review, covariate construction, logistic regression.

RNA-seq: Salmon counts, DESeq2, LFC shrinkage, VST/PCA, adjusted differential expression, effect-size prioritization.

Pathway biology: Hallmark, GO, KEGG, Reactome, FGSEA, sample-level pathway scoring, MAPK, PI3K/AKT, DNA repair, oxidative stress, hypoxia, EMT, interferon, thyroid differentiation.

Tumor immunology: thyroiditis phenotype engineering, immune deconvolution scaffolding, inflammasome gene and composite-score analysis.

Cancer genomics: MAF processing, curated driver reconciliation, hotspot grouping, mutation prevalence, pathway-level alteration matrices, exact testing.

Multi-omics: DNA–RNA matched analysis design, molecular-group to transcriptional-program integration, external-cohort harmonization strategy.

Statistical rigor: covariate adjustment, sparse-data exact tests, BH-FDR, sensitivity analyses, discovery-versus-validation distinction.

Data governance

No real patient-level identifiers, sequencing matrices, institutional paths, protected clinical records, or restricted CARIS data are distributed. Real results are limited to aggregate non-patient-level summaries; synthetic data are used for public reproducibility.

Name		Name	Last commit message	Last commit date
Latest commit History 6 Commits
.github/workflows		.github/workflows
config		config
data/synthetic		data/synthetic
docs		docs
results		results
scripts		scripts
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
UPGRADE_INSTRUCTIONS.md		UPGRADE_INSTRUCTIONS.md

Folders and files

Latest commit

History

Repository files navigation