George Soto and team
MindNetQ Inc., Boynton Beach, FL, USA
---
We present a governed multi-omics computational pipeline that identified two previously unreported prostate cancer gene candidates — DNAH5 and PRR36 — from 19,010 genes in the TCGA-PRAD dataset. Both genes exhibit strong differential expression (DNAH5: 4.4-fold, adj_p = 1.22×10⁻¹⁷; PRR36: 3.7-fold, adj_p = 2.57×10⁻²¹) with zero prior prostate cancer publications, and both were independently replicated in the MSKCC GSE21034 cohort (179 samples, microarray platform). Root cause analysis revealed that both genes are overexpressed despite promoter hypermethylation, with H3K27ac ChIP-seq data confirming active enhancer marks at both loci — indicating enhancer-mediated overexpression overriding methylation silencing. DNAH5, a ciliary dynein motor protein localized to the cilium axoneme, is overexpressed without co-expression of any of its 10 high-confidence interaction partners, suggesting a moonlighting function in prostate tumors. PRR36, a brain-enriched gene (97 TPM in cerebellum) with no known function, represents ectopic activation in prostate tissue. Copy number analysis confirmed that overexpression is not amplification-driven (76% and 78% normal diploid, respectively). The pipeline employs a non-compensating four-gate statistical filter, cross-modal discordance detection with physics-based biological constraints, and authority-based compute allocation — all governed through a delegation lease system with hash-chained evidence enabling deterministic replay. Applied broadly, the pipeline identified 382 novel gene candidates from 1,207 screened, including 229 with druggable protein structures. These findings suggest that enhancer-mediated activation of normally silenced genes may represent an underexplored mechanism in prostate cancer, with potential implications for BET inhibitor patient selection.
Keywords: prostate cancer, multi-omics, gene discovery, enhancer, methylation, governed AI, DNAH5, PRR36
---
Prostate adenocarcinoma (PRAD) is the second leading cause of cancer death in men, with approximately 35,000 deaths annually in the United States. Despite extensive genomic characterization through projects such as The Cancer Genome Atlas (TCGA), novel gene candidates with zero prior disease association continue to be identified through computational reanalysis of existing datasets, suggesting that current analytical approaches may systematically overlook certain classes of candidates.
We hypothesized that a governed multi-omics pipeline integrating six data modalities — transcriptomics, genomics, proteomics, epigenomics, clinical/survival, and structural biology — with non-compensating statistical filters and cross-modal discordance detection could identify novel prostate cancer gene candidates that single-modality analyses miss.
The pipeline introduces several methodological innovations: (1) a four-gate non-compensating filter that requires independent satisfaction of statistical significance, effect size, cross-validation stability, and data completeness; (2) cross-modal discordance detection using physics-based biological constraints; (3) authority-based compute allocation that governs resource expenditure based on evidence strength; and (4) a delegation lease system that ensures every analytical decision is traceable and deterministically replayable.
---
Primary analysis was performed on TCGA-PRAD (n=500; 498 tumor, 52 normal) RNA-Seq data downloaded from the Genomic Data Commons (GDC). Somatic mutations, clinical/survival data, methylation (HM450), and CPTAC proteomics were obtained from the same source. Independent replication used MSKCC GSE21034 (n=179; 150 tumor, 29 normal) profiled on Affymetrix Human Exon 1.0 ST microarray.
RNA-Seq counts were filtered to protein-coding genes with minimum expression threshold, log2-CPM normalized. Differential expression was computed using Mann-Whitney U test with Benjamini-Hochberg correction for multiple testing.
Candidates were required to independently satisfy all four gates: - Gate 1: Adjusted p-value ≤ 0.05 (Benjamini-Hochberg) - Gate 2: Cross-validation selection rate ≥ 0.60 across 5-fold L1 logistic regression - Gate 3: Bootstrap sign consistency ≥ 0.80 across 100 resamples - Gate 4: Missing value rate ≤ 0.10
No cross-gate compensation was permitted: strong performance on three gates did not offset failure on one gate.
Six modalities were integrated per gene: transcriptomics (differential expression), genomics (mutation rate, damaging variants), clinical (survival association), epigenomics (differential methylation), proteomics (protein abundance, variance), and structural biology (AlphaFold pLDDT, binding pockets, druggability).
A physics rules engine evaluated biological consistency across modality pairs. Eight rules with quantitative thresholds detected contradictions including: expression without mutation (expression-driven dysregulation), methylation-expression discordance, and expression without survival association.
Each candidate underwent 25-step automated validation: PubMed literature search, COSMIC Cancer Gene Census check, Human Protein Atlas tissue expression, AlphaFold structural assessment, Reactome/KEGG pathway membership, STRING protein-protein interactions, UniProt function annotation, and cross-dataset replication.
For top candidates, regulatory mechanism was investigated through: (1) gene-level copy number variation from TCGA GISTIC2 data; (2) promoter methylation from HM450 beta values; (3) H3K27ac ChIP-seq from ENCODE prostate tissue datasets (ENCSR841AJO, ENCSR763IDK); (4) transcription factor binding analysis; and (5) regulatory region mutation analysis.
All analytical decisions were governed through the E-DICE-Edge SDK, which implements delegation leases, hash-chained evidence logging, and deterministic replay. Twenty molecular decision points were governed across the pipeline.
---
From 19,010 protein-coding genes, the four-gate non-compensating filter retained 1,434 candidates. Of 1,207 candidates screened through the automated validation pipeline, 676 received GO decisions (≥4/5 gates), 466 received REVIEW (3/5), and 65 received NO-GO (≤2/5). Among GO candidates, 382 had zero prostate cancer publications in PubMed.
DNAH5 (dynein axonemal heavy chain 5) was identified as 4.4-fold overexpressed in tumors (adj_p = 1.22×10⁻¹⁷) with zero prostate cancer publications and 35 publications in other cancer contexts. Cross-dataset replication in MSKCC confirmed overexpression (p = 3.55×10⁻¹⁵) with consistent direction of effect.
Interactome analysis revealed that none of DNAH5's 10 highest-confidence STRING interaction partners (DNAI2: 0.997, DNAI1: 0.995, DNALI1: 0.977, NME8: 0.976, DNAL1: 0.968) were present in the 1,434 candidate set, classifying DNAH5 as interactome-decoupled. Subcellular localization (UniProt) placed the protein at the cilium axoneme — a compartment not functionally relevant to prostate epithelium.
Copy number analysis showed 76% of tumor samples (38/50) had normal diploid CN=2, with 24% showing gain (CN=3-4). GTEx comparison showed low normal prostate expression (1.5 TPM), with highest expression in pituitary (3.38 TPM) and lung (2.23 TPM). Isoform analysis identified ENST00000265104 as the dominant transcript.
PRR36 (proline rich 36) exhibited the strongest statistical signal in the dataset (3.7-fold, adj_p = 2.57×10⁻²¹) with zero publications in any cancer context. Replication in MSKCC confirmed overexpression (p = 3.04×10⁻³).
GTEx analysis revealed PRR36 as brain-enriched (96.7 TPM in cerebellum, 83.7 TPM in cerebellar hemisphere) with only 4.9 TPM in normal prostate — indicating ectopic activation in prostate tumors. Copy number was normal in 78% of samples. The protein has no characterized function.
Root cause analysis revealed a paradoxical finding: both genes are overexpressed despite promoter hypermethylation.
DNAH5 promoter methylation (probe cg16126291): mean beta 0.918 in tumors, 0.873 in normals — both heavily methylated. PRR36 methylation (probe cg07107453): mean beta 0.549 in tumors, 0.416 in normals — increased methylation in tumors.
H3K27ac ChIP-seq from ENCODE prostate tissue (ENCSR841AJO) revealed active enhancer marks at both loci: - DNAH5: one peak at TSS (chr5:13944372-13945439, signal 2.99, p = 10⁻⁶·⁸) - PRR36: three peaks within ±50kb, including an extremely strong peak at chr19:35628264-35629759 (signal 13.26, p = 10⁻⁸⁶·⁴)
No androgen receptor binding was detected at either locus. DNAH5 had 7 scattered non-recurrent mutations; PRR36 had zero mutations.
These findings indicate enhancer-mediated overexpression overriding promoter methylation silencing — a mechanism where active enhancers drive transcription despite the promoter being in a methylated, nominally silenced state.
Beyond DNAH5 and PRR36, the pipeline identified 382 novel candidates (zero prostate publications), 229 with druggable protein structures (druggability ≥ 0.4), and an unstudied pathway — the axonemal dynein family (7 genes with zero prostate cancer publications collectively).
Physics-based discordance analysis across 19,010 genes identified 11,094 (58%) with at least one cross-modal discordance and 597 with physics rule violations. Both DNAH5 and PRR36 exhibited the "expression without mutation" pattern (physics rule PH06), consistent with expression-driven rather than genetically-driven dysregulation.
---
We report the identification of DNAH5 and PRR36 as novel prostate cancer gene candidates through governed multi-omics analysis, with independent replication and a computationally identified regulatory mechanism.
The enhancer override finding — overexpression despite promoter methylation, driven by active H3K27ac enhancer marks — has potential therapeutic implications. BET inhibitors, which block BRD4 binding to acetylated enhancers, are currently in clinical trials for prostate cancer (ZEN-3694, Phase II, in combination with enzalutamide). If enhancer-driven overexpression of DNAH5 and PRR36 is confirmed as a feature of a subset of prostate tumors, these genes could serve as biomarkers for BET inhibitor patient selection.
The interactome decoupling of DNAH5 — a ciliary motor protein overexpressed without any of its canonical interaction partners — suggests a moonlighting function unrelated to ciliary motility. The ectopic activation of brain-enriched PRR36 in prostate tumors raises questions about lineage-specific regulatory programs being aberrantly activated in cancer.
This study is computational. While cross-dataset replication strengthens the findings, wet-lab validation (RT-qPCR, Western blot, functional assays) is required to confirm protein-level expression and biological function. The methylation analysis used a single probe per gene; genome-wide bisulfite sequencing would provide higher resolution. The enhancer analysis used normal prostate H3K27ac data; tumor-specific ChIP-seq would be more informative.
Every analytical decision in this study was governed through the E-DICE-Edge SDK with hash-chained evidence. The complete analysis is deterministically replayable from raw TCGA data to final findings using the same pipeline version and governance policy.
---
TCGA-PRAD data is available from the Genomic Data Commons (https://portal.gdc.cancer.gov/). MSKCC data is available from GEO (GSE21034). ENCODE H3K27ac data: ENCSR841AJO, ENCSR763IDK. The MindNetQ Cortex pipeline code is available upon request.
---
G.S. conceived the study, designed and built the pipeline, performed all computational analyses, and wrote the manuscript. A.Ho. contributed to pipeline development and testing. A.Hu. provided biological interpretation and suggested additional validation steps. M.F. provided institutional support and wet-lab coordination.
---
G.S. is the inventor on patent applications related to the MindNetQ Cortex pipeline (US provisional 64/036,382; 64/036,705).
---
1. Cancer Genome Atlas Research Network. The Molecular Taxonomy of Primary Prostate Cancer. Cell. 2015;163(4):1011-1025. 2. Taylor BS, et al. Integrative genomic profiling of human prostate cancer. Cancer Cell. 2010;18(1):11-22. 3. Filippakopoulos P, et al. Selective inhibition of BET bromodomains. Nature. 2010;468(7327):1067-1073. 4. Welti J, et al. Targeting bromodomain and extra-terminal (BET) family proteins in castration-resistant prostate cancer (CRPC). Clin Cancer Res. 2018;24(13):3149-3162. 5. Jumper J, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583-589. 6. Szklarczyk D, et al. The STRING database in 2023. Nucleic Acids Res. 2023;51(D1):D483-D489. 7. ENCODE Project Consortium. An integrated encyclopedia of DNA elements in the human genome. Nature. 2012;489(7414):57-74. 8. GTEx Consortium. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science. 2020;369(6509):1318-1330. 9. Benjamini Y, Hochberg Y. Controlling the false discovery rate. J R Stat Soc B. 1995;57(1):289-300. 10. Whyte WA, et al. Master transcription factors and mediator establish super-enhancers at key cell identity genes. Cell. 2013;153(2):307-319.