Search the PGS Catalog for polygenic scores by free text.

Case-insensitive substring match across PGS ID, score name, reported trait, and EFO trait. Optionally filter to a genome build (GRCh37 / GRCh38, harmonized cross-build scores included). Returns up to limit matches with their key metadata.

Get cleaned metadata for a single PGS score by its ID (e.g. 'PGS000001').

Look up the best evaluation performance for a score (largest sample, EUR-preferred).

Returns parsed effect sizes (OR/HR/Beta) and classification metrics (AUROC/C-index) plus pre-formatted display strings.

Search the PGS Catalog REST API for traits by term.

Upstream matching is exact-substring over labels and synonyms, so this wrapper retries a few punctuation/order variants when the first query is empty. By default, results include counts of directly associated PGS IDs and child-trait PGS IDs; set include_pgs_ids for the full arrays.

Fetch a trait by ontology ID (EFO or MONDO) with its associated PGS IDs.

Poll a long scoring/download job started by this server.

compute_prs_by_trait, compute_prs_batch, normalize_vcf, and download_sample_genome return a PRSJob instead of blocking. Call this with that job_id. wait_seconds=25 waits inside one call (safe for Cursor) and returns as soon as the job finishes. When status is done, read html_link and result.

List recent scoring/download jobs on this server process.

Use after a client timeout to recover the job_id instead of starting the same compute again.

Normalize a VCF to a quality-filtered genotype Parquet.

This tool is optional preprocessing, not a prerequisite for scoring: compute_prs, compute_prs_batch, and compute_prs_by_trait accept raw VCF paths directly. Use normalization when you intentionally want a reusable Parquet for many later analyses or custom quality filters. Before normalizing, list_genomes can show whether a cached Parquet already exists.

Strips the chr prefix, renames id→rsid, computes genotype from GT, applies optional quality filters (FILTER allow-list, min DP, min QUAL), and writes zstd-compressed Parquet. The output is a drop-in genotype source for compute_prs / compute_prs_batch (genotypes_path) and for compute_prs_by_trait (a samples entry ending in .parquet), so each VCF is normalized once and reused. Normalize every genome you plan to compare; scoring still happens per genome.

Idempotent: if the target Parquet already exists it is reused and the (slow) normalization is skipped — reused_cache=True flags the hit. Custom filters (pass_filters / min_depth / min_qual / sex) always re-run, since the cached Parquet may not reflect them. Pass force=True to re-normalize unconditionally.

Normalization is the slow step (seconds to minutes depending on VCF size). Returns a PRSJob immediately; poll prs_job_status when status is running. When done, result is a NormalizeResult.

Download a public sample WGS VCF from Zenodo to try PRS without your own data.

Two whole-genome sequencing (WGS) datasets open-sourced by the just-dna-lite project are pre-configured — download both when you want a multi-person comparison:

• sample="anton" — Anton Kulaga's genome (~482 MB, CC0 public domain, Zenodo record 18370498, file: antonkulaga.vcf).

• sample="livia" — Livia Zaharia's genome (~349 MB, CC-BY-4.0, Zenodo record 19487816, file: SIMHIFQTILQ.hard-filtered.vcf.gz).

Pass record_url (e.g. 'https://zenodo.org/records/18370498') to fetch any other Zenodo record, and filename to pick a specific file when a record has several.

The downloaded VCF lands under <cache_dir>/samples/ (or output_dir) and is a drop-in path for normalize_vcf / compute_prs / compute_prs_by_trait (including samples=["Anton=<path>", ...]).

auto_normalize defaults to True: the download is normalized to a reusable Parquet in the same call, so data carries both path (the raw VCF) and normalized_path — a one-call, compute-ready genotype source with no separate normalize_vcf round-trip. (Normalization is idempotent, so a re-download of an already-staged sample is cheap.) Pass auto_normalize=False to fetch the raw VCF only.

Use list_genomes to see which genomes have already been downloaded and/or normalized.

Idempotent: if the target VCF already exists with the size Zenodo reports, the ~hundreds-of-MB download is skipped and the cached file is reused; likewise a present Parquet skips re-normalization. data echoes reused_cache (download skipped) and downloaded_bytes (bytes actually transferred, 0 on a cache hit) so the caller can tell a cache hit from a fresh fetch. Pass force=True to re-download/re-normalize regardless.

Returns a PRSJob immediately. When done, result is an OpResult whose data carries the local path. Poll prs_job_status if status is still running — the download is hundreds of MB.

List genomes available in the server's cache directory.

Scans <cache_dir>/samples/ for downloaded raw VCF files and <cache_dir>/normalized/ for normalized Parquet files. Also lists the pre-configured sample genomes that can be downloaded via download_sample_genome (even if not yet downloaded).

Use this to discover genomes you can score or compare (Anton + Livia are the built-in pair):

• Which genomes have already been downloaded (ready for direct scoring).

• Which genomes have already been normalized (ready for compute_prs / compute_prs_by_trait as an optional reusable fast path).

• Which pre-configured samples are available for download.

No network access required — reads the local filesystem only.

Detect a genome's build, input type (WGS/array/gVCF), and genetic ancestry.

One call that pairs WGS-vs-array/gVCF detection with population (super-population) inference, so before scoring you already know: the genome build (read from the VCF header), whether the input is whole-genome sequencing, a consumer genotyping array (and which chip), or a gVCF / all-sites callset, and the sample's inferred super-population.

Pass genotypes_path (a normalized Parquet from normalize_vcf / download_sample_genome) for the fast path, or a raw vcf_path. Genome-build detection reads the VCF header, so it is only available on the vcf_path branch.

The result carries ready-to-use recommendations: recommended_reference_restoration (feed to compute_prs / compute_prs_batch / compute_prs_by_trait as reference_restoration — though those tools default to "auto" and derive the same thing) and recommended_superpopulation (feed as superpopulation). Set infer_population=False to skip ancestry inference (it pulls a small model from HuggingFace on first use).

Compute a polygenic risk score for one genome against one PGS model.

Pass the raw VCF directly; a separate normalize_vcf call is not needed. If this VCF has a fresh normalized cache, it is reused automatically; genotypes_path is the explicit Parquet fast path. For a trait panel — one genome or several — use compute_prs_by_trait(samples=[...]), then compare_genomes / build_prs_prompt / plot_trait_panel.

Reference restoration (F15/F22). reference_restoration fills a scoring variant's missing reference allele from the precomputed reference-allele universe (pulled from HuggingFace on first use) so a locus the sample is confidently hom-ref at scores as dose-0 instead of being dropped as unscorable-absent — recovering the ~50% coverage gap on genome-wide WGS scores. Default "auto" detects the input: whole-universe restoration for WGS, the matching chip scope for a genotyping array, off otherwise. Pass "off", "wgs", or a chip id (e.g. "gsa_v3") to force a scope. The result's variants_ref_resolved_panel / variants_ref_resolved_fasta counters show how many loci were recovered.

Set infer_ancestry=True to also infer and attach the sample's genetic ancestry (sample_ancestry). genotype_input_mode ("auto" / "variant_only" / "all_sites") controls how absent loci are read.

Set attach_performance=True to embed the score's best published performance (effect sizes, AUROC/C-index, evaluation ancestry) on the result in the same call. The result also carries detected_genome_build / build_mismatch from the VCF (F4).

Recommended follow-up: percentile then absolute_risk (disease traits), or build_prs_prompt(kind='score') for a reusable LLM write-up.

Compute PRS for one genome against many PGS models.

Pass the raw VCF directly; a separate normalization call is not required. The tool transparently creates or reuses a collision-safe normalized cache, then uses the memory-safe DuckDB engine with spill-to-disk. It reuses the scoring caches across scores. This is still one genome × many PGS IDs. For a trait panel across people, use compute_prs_by_trait(samples=[...]) then compare_genomes / build_prs_prompt. Set attach_performance=True to embed each score's best published performance on its result in the same pass.

reference_restoration (default "auto") recovers absent-hom-ref coverage on WGS/array input — see compute_prs; the reference-allele universe is resolved once and reused across every score. genotype_input_mode ("auto" / "variant_only" / "all_sites") controls how absent loci are read.

Returns a PRSJob immediately. When done, result is the list of successful PRSResult objects. If status is queued or running, poll prs_job_status(job_id, wait_seconds=25).

Compute the PRS scores associated with a trait for one or more genomes.

REQUIRED REPORT DELIVERY: when html_link is set, your reply MUST include it as a markdown link (for example Open the interactive trait report) and explicitly tell the user to open it. This is the primary user-facing result, not an implementation detail. Do not skip the link, merely say the report was saved, fetch it, or paste the HTML. html_link is an HTTP URL when available and a file:// URI on local stdio.

Pass a raw .vcf/.vcf.gz path directly. Do NOT call normalize_vcf first as a prerequisite: multi-score tools transparently create or reuse a safe normalized cache, and single-score tools read the VCF directly. normalize_vcf is only an optional explicit step for custom filters. If list_genomes already shows a normalized Parquet, it can be passed directly.

trait_id may be an EFO or MONDO identifier. samples is always a list (one genome is fine): each entry is Label=path or a bare path; .parquet is a normalized genotype source. Every genome is auto-saved; the return carries reports, result_paths, and (when models have percentiles) the primary html_link plus transport-specific html_url / html_path — the just-prs trait HTML report (same page as prs plot trait -o *.html). Pass result_paths to compare_genomes (2+), plot_trait_panel, or build_prs_prompt.

Profile (curation): profile="all" (default) returns the raw associated panel. profile="curated" is the interpreted shortlist — it drops toy scores (<10 variants), scores with no performance evidence, scores below the C_wt coverage floor, and de-dups score families.

interpret=True (default) attaches percentiles/quality per model. Set interpret=False for raw scores only — the report then says so. reference_restoration defaults to "auto". superpopulation may be a 1000G code or "auto". Set include_prompt=True to also attach the just-prs LLM prompt covering every genome in this call.

This tool returns a PRSJob immediately. If status is queued or running, call prs_job_status(job_id, wait_seconds=25) until done or failed. Do not start a second identical compute — the server reuses the in-flight job. When done, html_link and result (TraitPRSReports) are on the job. Native multi-score work is isolated in a child process so a DuckDB/Polars crash fails the batch without killing the MCP transport.

Estimate the population percentile (0-100) for a computed PRS value.

Uses the 3-tier fallback: precomputed reference-panel distributions (best), then a theoretical distribution, then an AUROC approximation. superpopulation is a 1000G code (AFR/AMR/EAS/EUR/SAS) and should match that genome's ancestry — when comparing people, call this once per score×genome (or rely on compute_prs_by_trait(interpret=True)). Pass weight_mass_coverage (C_wt) from compute_prs so a deflated low-coverage percentile is flagged reliable=False with a caveat instead of presented as authoritative. Also returns the true z-score and reference mean/std used, so absolute risk can be computed without inverting the percentile.

Important next step: for disease traits, feed the returned z_score directly into absolute_risk to get the concrete lifetime probability and risk ratio vs the population average. This is more informative than the percentile alone.

Estimate absolute disease risk from a PRS z-score and population prevalence.

Joins the score's trait to prevalence + effect-size data. z_score is the PRS in standard deviations from the population mean. Raises if the required prevalence / effect-size data is unavailable for this score.

Classify and interpret a PRS result's quality (pure logic — no I/O).

match_rate is the fraction of scoring variants matched (0-1). Returns a quality label/color and a human-readable interpretation combining match rate, AUROC, and (optionally) the result percentile. Pass percentile_method / reliable / caveat from the percentile tool so the summary describes how the percentile was actually derived and echoes any low-coverage caveat.

Compare saved by-trait PRS reports across two or more genomes.

result_paths — the result_paths list from compute_prs_by_trait (or any saved report JSONs). At least two paths are required. Files are grouped by trait, so a family comparison across several traits is one call.

Rankings use just-prs summarize_trait_rows: the median in-scope percentile is the headline (not a single "best" model). High percentile is not automatically good or bad — interpret direction from the trait.

When every path is the same trait, include_prompt=True (default) attaches the same multi-sample LLM prompt as prs prompt / the UI Ask-AI buttons.

Build the reusable just-prs LLM prompt from saved by-trait results.

This is the MCP equivalent of prs prompt / the UI Ask-AI buttons. result_paths is always a list (one genome is fine; two or more becomes a comparison-first prompt). The returned prompt can be followed in this conversation or pasted into another assistant. Set include_url=True with assistant claude/chatgpt/perplexity/grok for a prefilled link.

kind: trait_results (default, full per-model panel), trait_summary (aggregated cards), or score (one PGS ID; pass pgs_id or the first scored row is used).

Write the just-prs trait HTML report and return a URL or file path.

REQUIRED REPORT DELIVERY: when html_link is set, your reply MUST include it as a markdown link (for example Open the interactive trait report) and explicitly tell the user to open it. This is the primary user-facing result, not an implementation detail. Do not skip the link, merely say the report was saved, fetch it, or paste the HTML. html_link is an HTTP URL when available and a file:// URI on local stdio.

This is the MCP equivalent of prs plot trait ... -o intel_o_family.html: just_prs.viz.plot_trait_scores + save_trait_report. The page has the Vega-Lite panel, median cards, model table, and Ask-AI buttons. result_paths is always a list (one genome is fine; two or more overlay people). On Streamable HTTP html_url is served from this same origin at /reports/<file> with no extra config.