Skip to content

GoliathOmics Gaps

Wang’s program is four different method classes. GoliathOmics covers one of them well and only touches the others at the edges.

Snapshot

Wang / Moffitt technique Completeness Fit today
cfDNA methylation (targeted EM-Seq, MHB/MHL, mCRPC OS) High (~75%) First-class procedure
Clinico-genomic scores (labs + ctDNA fraction + methylation risk → nomogram) Medium (~50%) Cox sidecar exists; clinical calculator and CNA scores do not
Exosomal miRNAs / proteins Low (~10%) Generic RNA/proteomics packs only; no EV / small-RNA path
Long-read methylation of noncoding risk variants Near zero Explicitly deferred; not a product path

1. Cell-free DNA methylation — mostly supported

This is the Wong / Wang 2026 plasma EM-Seq + MHB/MHL work.

Shipped: cfdna analyte; cfdna_emseq_targeted and cfdna_emseq_mhl_survival; linear EM-Seq SamplePrep; panel BED extract; {chrom}-CG.mhap.h5; native MHB/MHL with their defaults (window 3, (r^2>0.3), (p<0.05), ≥3 CpGs, median reads >50, MHL lengths 1–10); Cox / KM / time-AUC / optional nomogram.

Still missing for a true Wang-lab match:

  • Their Twist panel BED, control-region BEDs, and locked 15-MHB list (operator assets, not code)
  • GREAT on DMR coordinates
  • MBD-seq ultra-low-input (Huang & Wang, Epigenetics 2022) — no MBD-seq SamplePrep
  • 5hmC in plasma cfDNA (Li et al. 2023, ADT-resistant PCa) — no 5hmC chemistry or calling
  • Plasma copy-number risk scores (Huang et al. 2022 Cancers) — fragmentomics mentions CNV; there is no ichorCNA / cfDNA CNA gene-score action
  • Bismark/mHapSuite bit-for-bit parity (intentionally native, not wrapped)

Plasma WGBS + fragmentomics (cfdna_wgbs_plasma) is a different cfDNA methylation geometry (discovery / MCED-adjacent), not their capture-MHL assay.

2. Exosomal miRNAs / proteins — not supported as their assay

Wang-adjacent work includes plasma exosomal miR-423-3p / miR-375 / miR-1290 (CRPC prediction), EV isolation-method effects on lipidome/metabolome, and EV cargo more generally.

What we have: RNA-Seq pack = STAR/kallisto mRNA/lncRNA gene counts; proteomics pack = DIA-NN / Sage / Olink-style protein abundance. Those are bulk (or panel) modalities, not EV isolation + small-RNA.

To add if this is a partnership pillar:

  1. Analyte or matrix: plasma_ev / exosome (isolation method as config, not a hardcoded kit)
  2. Small-RNA SamplePrep: adapter-aware miRNA-seq (or qPCR panel ingest) → mirna.h5 (mature_id, count/RPM), not STAR gene counts
  3. Optional EV-protein ingest into the existing proteomics abundance.h5 seam (ingest_mode: panel)
  4. A procedure that joins EV-miRNA (± EV-protein) with cfDNA MHL on matched sample_id for Cox or tabular fusion
    Isolation chemistry stays wet-lab; we only need the digital contract.

Do not treat rna_parabricks_star_de as “we support exosomal miRNA.”

3. Long-read methylation of noncoding risk variants — not supported

This is Tian, Wang et al.: Nanopore native 5mC (and 5hmC in later nanoASM), adaptive sampling on GWAS/mQTL intervals, haplotype-specific DMRs, allele-specific methylation vs ATAC/ChIP, large MHBs that short reads cannot span (HGG Adv 2025; nanoASM 2026; earlier rs7247241 / PPP1R14A work).

What we have: Illumina short-read WGBS/EM-Seq. Long-read is deferred in the SamplePrep plan. mojo-align still has experimental methylGrapher longread.py / MM:Z graph align — not a worker action, not Dorado, not adaptive sampling, not nanoASM.

To add (large):

  1. ONT SamplePrep: pod5/BAM ingress → Dorado (5mCG ± 5hmCG) → GRCh38 (linear first; graph later)
  2. Adaptive-sampling BED (GWAS + mQTL) as target_panel_bed analogue
  3. Phasing + allele-specific methylation (nanoASM-class): SNP-split reads, haplotype DMRs, optional overlap with GWAS/mQTL catalogues
  4. Long-read MHB/entropy on native calls (their point vs short-read MHL)
  5. Optional ATAC/ChIP overlays — not required for v1

This is tissue/cell-line functional genomics (22Rv1, prostate tissue), not the plasma EM-Seq product path. Keep it a separate procedure (tissue_ont_asm or similar).

4. Clinico-genomic scores in mCRPC — half supported

Wong nomogram = MHL risk + PSA, ALP, LDH + predicted ctDNA fraction (ctdna.org: yield, labs, ECOG, mets) + nested test vs ctDNA-only.

Shipped: survival_path (sample_id, time, event, optional psa/alp/ldh/predicted_ctdna_fraction); Cox; time-AUC; write_nomogram; nested_lrt.

Missing:

  • In-process ctdna.org (or a documented clone of that calculator)
  • First-class EMR fields (ECOG, viscera mets, cfDNA yield) vs a free-form CSV
  • Plasma CNA / AR-locus scores (Huang 2022; Du multiplex dPCR AR amp)
  • CTCs, treatment-line covariates
  • Their exact LOOCV (\sum\beta_i M_i) recipe and 0.5/1/2-year timeROC defaults as a locked Moffitt score object (we have the engine, not their published score)

HiTIMED tumor_fraction is not their ctDNA predictor.

What to add, in partnership order

If the goal is “we can work with Wang’s lab,” not “we reimplement the entire CV”:

  1. Close the MHL+OS loop (small): operator BED + survival dictionary; optional GREAT; document nested LRT + time-AUC horizons 0.5/1/2 years; keep ctdna.org as a sidecar column unless they insist on the API.
  2. Clinico-genomic v2 (medium): typed clinical schema (PSA/ALP/LDH/ECOG/mets/yield) + optional cfDNA CNA/ichor action; do not invent AR-dPCR in Python.
  3. EV cargo (medium, only if they want it): small-RNA + EV-protein panel ingest and fusion with MHL — reuse RNA/proteomics seams, new analyte/QC.
  4. Long-read ASM (large, separate product): Dorado + adaptive sampling + haplotype ASM. Do not block plasma MHL on this.

A partnership can start on (1)+(2) with assets they already have. (3) and (4) are new process packs, not overlays on cfdna_emseq_mhl_survival.