PRS Pipeline — Stage 0 Findings

Date: 2026-05-07 Status: Stage 0 complete; falling back to zero-PRS for Track A.

Question being answered

Can we obtain a UKB-shaped sample patient (CSV row + biomarker values + comorbidity PRS scalars) to use as the Track A pipeline burn-in input for build_pdr1_input_csv.py and build_pdr2_input_pickle.py, before Maureen’s genome-wide genotype data arrives?

Sources checked

Local snapshots

Path	What’s there	Sample patient?
docs/References/PDR1/	LICENSE, README, code/, figures/, patient_merge.py, preprocessing_utils.py, split_GPKG.py, requirements.txt, Supplementary_Material/	No
docs/References/PDR1/Supplementary_Material/	Baseline_Patient_Characteristics.csv, _Read_Me.txt, supp_clusters/ (PDFs + interactive HTML cluster plots)	No — Baseline_Patient_Characteristics.csv is a 127-line TriNetX cohort aggregate table (ketamine vs antidepressants in the AD validation cohort), not a per-patient feature row.
docs/References/PDR2/	LICENSE, README, supplementary.pdf, the paper PDF, reasoning/ (config + script + ultra), split_KG.py, requirements.txt	No
docs/References/PDR2/.gitignore	Standard Python boilerplate plus *.txt, *.tsv, *.pkl, *.pt	The data formats are explicitly excluded from the repo.

Live GitHub

• github.com/ceragoguztuzun/PrecisionDrugRepurposing — same structure as our local snapshot. README references {patients_data_filepath}.csv as a placeholder; no actual example file. References data/ and test_data/ folders but they aren’t shipped.
• The same is true for the PDR2 paper’s repo (it’s the ULTRA fork — code only, no patient pickles).

UK Biobank

• Resource 5202 readme PDF confirms 28 binary disease traits + 25 quantitative traits in the Standard PRS Release, Field IDs in 26200-26285. Pre-computed scoring is not publicly downloadable — the PRS values themselves are access-controlled UKB data. Only trait definitions are public.
• Genomicsplc/ukb-pret/src/ukb_pret/data/traits.yaml — full trait code list. No Field-ID mapping in this YAML; it’s just (code, full_name, sex, scale, follow_up_months). The exact code→Field-ID mapping is in UKB Showcase per-trait pages and is not trivially extractable in batch.
• The 28 binary disease trait codes (relevant subset for PDR’s PRS slot): AD, AF, AMD, AST, BC, BD, CAD, CD, CED, CRC, CVD, EOC, HT, ISS, MEL, MS, OP, PC, PD, POAG, PSO, RA, SCZ, SLE, T1D, T2D, UC, VTE. (CRC = “Bowel cancer”; relevant for Maureen.)

Email path (deferred)

The PDR paper authors were already contacted previously, per an internal memory note, re: the original split files. We could append a request for a sample patient pickle to that thread, but per user policy emails require explicit approval — flagged as a follow-up the user can choose to send.

Conclusion

No usable sample patient data is obtainable through automated channels. Falling back to the planned zero-PRS placeholder for Track A:

• prs_values = numpy.zeros(len(disease_names)) — after PDR2’s z-score normalization, this becomes 0 everywhere; PersonalizedRepurposingLoss’s tanh(weight · prs_values) term becomes tanh(0) = 0, contributing nothing to the loss.
• expr_values is real (from Caris WTS via mods_gene_expression).
• This makes Track A a controlled baseline: the zero-PRS run is the foundation-model-with-expression-only baseline, against which Track B’s real-PRS run can be A/B’d to measure PRS contribution.

Implications for downstream stages

1. Stage 1 (PGS Catalog connector): still needed — once Maureen’s genotypes arrive, we map UKB trait codes (AD, CRC, T2D, ...) → UMLS CUIs → PGS Catalog scoring files. The trait code list above is our default disease_names set. PGS Catalog has scoring files for all 28 UKB diseases.
2. Stage 4 (PDR2 pickle builder): --source synthetic_zeros becomes the default invocation for Track A. Generates a pickle with real expr_values + zeros prs_values.
3. Stage 5 (PDR1 CSV builder): emit the CSV with Standard PRS for alzheimer's disease (AD) = 0 and PRS_{CUI} = 0 for each comorbidity column.
4. Stage 6 (Track B): when genotype data arrives, swap the PRS source by re-running the builders with --source maureen. No code change needed beyond Stages 1-3.

CUI mapping (preliminary, will be finalized in Stage 1)

These are the trait-code → UMLS CUI mappings we’ll use to populate disease_names. UMLS CUIs taken from standard mappings; final list will be cross-checked against GP-KG entity names at Stage 1.

UKB code	Trait	Likely CUI	In GP-KG?
CRC	Bowel cancer (colorectal)	C1527249	TBD
AD	Alzheimer’s disease	C0002395	Yes (paper-confirmed)
T2D	Type 2 diabetes	C0011860	TBD
CAD	Coronary artery disease	C1956346	TBD
HT	Hypertension	C0020538	TBD
AF	Atrial fibrillation	C0004238	TBD
ISS	Ischaemic stroke	C0948008	TBD
BC	Breast cancer	C0006142	TBD
PC	Prostate cancer	C0376358	TBD
PD	Parkinson’s disease	C0030567	TBD
RA	Rheumatoid arthritis	C0003873	TBD
AST	Asthma	C0004096	TBD
MS	Multiple sclerosis	C0026769	TBD

13 entries here = CRC + 12 comorbidities — matches the user’s choice in the planning question (“CRC + ~12 PDR-paper comorbidities”).