mims-harvard/skills/tooluniverse-phewas

Cross-Biobank PheWAS & Replication

A PheWAS is the inverse of a GWAS: a GWAS fixes a phenotype and scans variants; a PheWAS fixes a variant (or gene) and scans the entire phenome. The scientific payoff of running it across several biobanks is replication and ancestry resolution — an association seen in one population that reappears in another (with the same effect direction) is far more credible, while one that appears only in East-Asian cohorts may reflect ancestry-specific LD, allele frequency, or biology.

This skill orchestrates one variant/gene across four ancestries plus a gene-burden layer. Look up, don't guess — never assert a variant's phenotype associations or effect sizes from memory; the whole point is the live cross-biobank numbers.

The biobank panel

| Tool | Biobank | Ancestry | Build | Evidence type | |---|---|---|---|---| | UKBTOPMed_phewas_by_variant | UKB-TOPMed | European (UK) | GRCh38 | per-phenotype assoc (phecodes, ~1,400) | | FinnGen_get_variant_finemapping | FinnGen | Finnish | GRCh38 | credible-set membership (fine-mapping) | | BioBankJapan_phewas_by_variant | BioBank Japan | Japanese | GRCh37 | per-phenotype assoc | | TPMI_phewas_by_variant | TPMI | Taiwanese (Han) | GRCh38 | per-phenotype assoc (ICD-based) | | Genebass_gene_burden_phewas | Genebass (UKB exomes) | European | GRCh38 | gene-level rare-variant burden |

The four *_phewas_by_variant tools (UKB-TOPMed, BBJ, TPMI) and Genebass_gene_burden_phewas accept an rsID (Genebass also accepts a gene symbol) and resolve coordinates themselves via Ensembl, picking the correct allele for multi-allelic SNPs; the build difference between BBJ-GRCh37 and the others is handled internally — just pass the rsID. You normally do not need to hand-convert coordinates for these.

FinnGen_get_variant_finemapping is the exception: it does NOT accept an rsID. It requires an explicit GRCh38 variant string in chr:pos:ref:alt format (e.g. "10:112998590:C:T"). Resolve the rsID to GRCh38 coordinates first — the simplest way is to read the variant/rsids field returned by the UKB-TOPMed (GRCh38) call, then pass that coordinate string to FinnGen.

Workflow

Step 1 — Decide variant-level vs gene-level

• Have an rsID / specific variant → variant-level panel (the four *_phewas_by_variant + FinnGen). This is the common case. Pass the rsID directly to the four *_phewas_by_variant tools; for FinnGen_get_variant_finemapping you must supply an explicit GRCh38 chr:pos:ref:alt string (resolve the rsID first — see "The biobank panel" note above).
• Have a gene and care about rare coding burden (e.g. "what does loss-of-function of PCSK9 do?") → Genebass_gene_burden_phewas with burden_set: "pLoF". Add the variant panel too if a specific common variant is also of interest.

Step 2 — Query the panel in parallel

Call the relevant tools for the same rsID. Always pass max_pval: 5e-8 (genome-wide significance) for the variant tools when you want only robust hits, or omit it to see suggestive associations too. Example anchor call:

UKBTOPMed_phewas_by_variant(rsid="rs7903146", max_pval=5e-8, limit=25)
BioBankJapan_phewas_by_variant(rsid="rs7903146", max_pval=5e-8, limit=25)
TPMI_phewas_by_variant(rsid="rs7903146", max_pval=5e-8, limit=25)
FinnGen_get_variant_finemapping(variant="10:112998590:C:T")   # FinnGen needs a variant string (GRCh38) or use an rsID-resolved coord

For gene-burden:

Genebass_gene_burden_phewas(gene="PCSK9", burden_set="pLoF", max_pval=2.5e-6, limit=25)

Step 3 — Align phenotypes across biobanks

Phenotype coding differs per biobank (UKB/TPMI use phecodes, BBJ uses curated endpoints, FinnGen uses its own endpoint codes, Genebass uses UKB field codes). Match on the human-readable name / disease concept, not the code. Group associations into shared concepts (e.g. "Type 2 diabetes" across all of them) before comparing.

Step 4 — Judge replication (see interpretation table)

For each shared phenotype concept, compare significance, effect-size direction (sign of beta), and allele frequency across biobanks.

Step 5 — Report

Use the report template below. Lead with what replicates, then ancestry-specific signals, then caveats.

Interpretation

| Pattern across biobanks | Meaning | |---|---| | Significant + same beta sign in ≥2 ancestries | Replicated, robust association — highest confidence | | Significant in European but not East-Asian | May be European-specific, OR underpowered / lower allele frequency in East-Asian cohort — check af and num_cases before concluding biology | | Significant only in BBJ/TPMI | Candidate East-Asian-specific effect (ancestry-specific LD or biology) — flag for follow-up | | Opposite beta signs between biobanks | Caution: possible strand/allele-coding mismatch, or genuine flip — verify the ref/alt allele each biobank reported before interpreting | | Variant-level null but Genebass burden significant | Phenotype driven by rare coding variation in the gene, not the common variant | | FinnGen credible-set hit but weak elsewhere | Variant is a fine-mapped candidate in Finns; credible-set membership ≠ p-value, so treat as corroborating, not quantitative |

Allele-frequency caveat is central. A "missing" association in one population is only interesting after you confirm the variant is actually common enough there to be powered. Each *_phewas_by_variant result carries af, num_cases, num_controls — use them. A variant at AF 0.30 in Europeans but 0.02 in Japanese will look "European-specific" purely from power, not biology.

Report structure

# PheWAS: <rsID / gene> (<nearest gene>)

## Replicated associations (≥2 ancestries, concordant direction)
- <phenotype> — UKB-TOPMed p=<>, β=<>; BBJ p=<>, β=<>; TPMI p=<>, β=<>  → concordant

## Ancestry-specific / unreplicated signals
- <phenotype> — significant in <biobank> only; AF=<> elsewhere (powered? yes/no)

## Gene-burden layer (if run)
- <phenotype> — Genebass <burden_set> p=<>

## Caveats
- allele-frequency / power notes, build notes, coding-direction checks

Worked example (rs7903146, TCF7L2)

Running the panel on rs7903146 returns Type 2 diabetes as the top hit with the T (risk) allele increasing risk (β>0) in every ancestry — UKB-TOPMed (p~1e-134), BioBank Japan (native Japanese T2D endpoint p~2e-47; its external Suzuki-2024 meta entries underflow to pval: 0.0), and TPMI (p~6e-11) — concordant positive effect across European, Japanese, and Taiwanese ancestries → textbook replicated association. (Exact p-values drift as biobanks update; treat these as illustrative.) The weaker TPMI p-value tracks the lower T-allele frequency in East Asians (AF~0.29 in Europeans vs ~0.02–0.05 in East Asians) and smaller case count, not a weaker biological effect — exactly the power caveat above. A nice secondary signal: in BBJ the T2D-risk allele associates with lower BMI/body weight, the known TCF7L2 feature that it acts through insulin secretion rather than adiposity. This is the canonical robustly-replicated common-variant association and a good sanity check that the panel is working.

Reading pval: 0.0: a returned p-value of exactly 0.0 is floating-point underflow on an astronomically significant association (the true value is below ~1e-308), not "no association." Treat it as the strongest possible hit, not a null.

BioBank Japan catalogue note: the BBJ PheWeb surfaces both native BBJ endpoints and external multi-ancestry GWAS meta-analyses (e.g. Suzuki 2024, with separate EUR/EAS/AFA/SAS/HIS rows). This is a bonus — you get per-ancestry betas from one call — but do not double-count the meta-analysis rows as independent "Japanese" evidence; the native BBJ endpoint is the Japanese-specific one.

Limitations

• Variant-level only reports what each biobank pre-computed. PheWeb instances return summary associations, not raw genotypes; you cannot run a custom phenotype not in their catalogue.
• TPMI / BBJ phenotype catalogues are ICD/endpoint-based and less granular than UKB phecodes; absence of a phenotype may mean "not catalogued", not "no association".
• Genebass is European (UKB exomes) only — it adds a rare-variant burden layer but not an ancestry dimension.
• No automatic meta-analysis. This skill compares biobanks qualitatively (replication, direction, power). For a formal cross-ancestry meta-analysis or fine-mapping, hand off to dedicated tooling (tooluniverse-mendelian-randomization, tooluniverse-gwas-finemapping).
• FinnGen evidence is credible-set membership, a different currency than the PheWeb p-values; do not put its "hits" on the same numeric axis.