Sdrf:metascreen

SDRF Meta-Analysis Screening Protocol

You are screening proteomics datasets before annotation or autoresearch. Your job is to resolve a candidate study set, read repository metadata and the associated publication for each candidate, apply the user's inclusion criteria, extract requested study-level metadata, and write one TSV row per accession.

Do not guess. Use the MCP tools to verify every field.

This skill is complementary to sdrf:autoresearch: run it first when the user needs a more precise, publication-aware study screen before deciding which datasets should enter annotation. This skill produces a curation TSV; it does not create, validate, fix, or improve SDRF files.

Step 1: Parse the Request

Normalize into these fields:

target

What dataset set to screen. Accepted forms:

• accessions:PXD001234,MSV000078958 — use the comma- or whitespace-separated

accession list directly. Do not put file paths after accessions:.

• Any file path (.txt, .tsv, .csv) — resolve automatically by extension:
• .txt — one accession per line, strip whitespace, skip blank lines
• .tsv / .csv — read the first available accession column from:

id, accession, project_accession, project, or the first column

• all PRIDE datasets — use PRIDE MCP search to discover matching

datasets. Examples:

• all PRIDE human gut metaproteomics datasets
• all PRIDE crosslinking datasets
• all PRIDE human plasma proteomics datasets

If a manifest contains previous labels, scores, or notes, use only the accession column unless the user explicitly asks to reuse those fields.

When target is a free-text PRIDE category, resolve it into a concrete accession list using PRIDE MCP search before screening begins (see Step 2).

criteria

Inclusion/exclusion rules. Accepted forms:

• Path to a .md or .txt file — read the file and use its contents as the rules
• Inline text — use directly as the rules
• Omitted — no filtering; label all records include and only extract columns

extract

What fields to extract per record. Accepted forms:

• Comma-separated names: "sex,age,region"
• Path to a .txt file — one field name per line
• Path to a .md file — parse field names and any per-field extraction

instructions from the file. If the file contains instructions for how to fill a field (format, source, yes/no semantics), follow those instructions exactly.

• Omitted — no extra fields beyond the required output columns

When extract and criteria point to the same .md file, parse both the inclusion rules and the extraction instructions from it.

Always include id, label, reason, and evidence in the output, even if they are not listed in extract. Append requested extraction columns after those required columns.

output

Path for the output TSV. Default: metascreen_results.tsv

---

State the resolved config before processing begins:

Target     : 
Criteria   : 
Extract    : 
Output     : 

Step 2: Resolve the Target Set

Fixed lists and manifests

For accessions:... and file paths, resolve directly into a working list. Keep the original order. Preserve duplicate accessions only if the user explicitly asks; otherwise deduplicate and report the number removed.

PRIDE category discovery

For all PRIDE datasets:

1. Translate the category into repository search terms. Use repository metadata

from a few known examples only if needed to calibrate terminology.

1. Search PRIDE for matching projects. Use relevant keywords from the category

description (sample type, organism, experiment type, etc.).

1. Fast pre-filter using PRIDE metadata only — before fetching any publications,

discard obvious mismatches using structured PRIDE fields:

• organisms — filter by species if the criteria specify one
• experimentTypes — filter by acquisition mode (DDA, DIA) if specified
• quantificationMethods — filter by quantification strategy if specified
• instruments — filter by instrument family if specified

This avoids spending publication-fetch calls on clear non-candidates.

1. Report the discovered and pre-filtered candidate list before publication

screening: `` Discovered : N datasets matching "" Pre-filtered: N remaining after PRIDE metadata check ``

Step 3: Screen Each Candidate

Process each accession in the working list independently.

3.1 Fetch repository metadata

Tool: get_project_details(project_accession="PXD######")

Read all fields. Pay close attention to free-text protocol fields — they are often the highest-signal source for instrument, acquisition, sample type, quantification, and study design information.

3.2 Fetch the publication

Use publication identifiers in the repository record first. If no identifier is available, search by project title and authors.

Tool: get_article_metadata(ids=["PMID:XXXXXXX"])

If the paper is open access, fetch the full text:

Tool: get_pdf_by_unpaywall(identifiers=["doi:10.xxxx/xxxxx"])

Read the Methods section for any information relevant to the criteria and the requested extract fields. Use repository protocols, the abstract, and title only as fallback evidence when full text is unavailable.

If no open-access full text is available, do not infer missing eligibility details from vague title or keyword matches. Use uncertain when the evidence is not sufficient.

3.3 Apply the inclusion criteria

If criteria was provided:

• Apply each rule in order
• include — all required inclusion rules pass and no exclusion rule applies
• exclude — at least one exclusion rule clearly applies; record the decisive

failed rule as reason

• uncertain — the study may be relevant, but repository metadata and available

publication evidence are insufficient to decide; state what is missing

Be conservative: prefer uncertain over guessing when evidence is absent.

If criteria was omitted, label all records include.

Use only these lowercase labels: include, exclude, uncertain.

3.4 Extract the requested fields

For each field in the resolved extract list:

• If the extract source provided instructions for this field, follow them exactly
• Otherwise, infer the extraction approach from the field name and available evidence
• Use unclear if the value cannot be determined from PRIDE metadata or the publication
• Keep values short and analysis-ready. Do not paste long abstracts or methods text

into metadata fields.

3.5 Print progress

After each accession:

[N/total] PXD###### → LABEL

Step 4: Write the TSV

• First row: column headers, with id always first
• Required columns: id, label, reason, evidence
• One row per accession in the order processed
• Tab-separated, UTF-8
• Use unclear for any field that could not be determined
• In evidence, cite the evidence source briefly, e.g. PRIDE protocol,

paper methods, abstract, or title/keywords only

Print a summary:

Saved N records to 
include:   N
exclude:   N
uncertain: N

Step 5: Report unresolved records

List uncertain accessions and state specifically what evidence was missing.

Example invocations

# Discover from PRIDE and screen
/sdrf:metascreen target="all PRIDE human gut metaproteomics datasets" criteria="criteria/human_gut_metaproteomics.md" extract="criteria/human_gut_metaproteomics.md" output="results/human_gut_screen.tsv"

# Fixed accession list from a txt file
/sdrf:metascreen target="data/accessions.txt" criteria="criteria/eligibility.md" extract="instrument,acquisition,sex,age,region" output="results/screen.tsv"

# Inline accessions
/sdrf:metascreen target="accessions:PXD001234,MSV000078958" criteria="human fecal metaproteomics only; exclude animal-only studies" extract="organism,sample_type,instrument" output="results/screen.tsv"

# From a manifest TSV
/sdrf:metascreen target="data/candidates.tsv" criteria="criteria/eligibility.md" extract="criteria/extract_fields.txt" output="results/screen.tsv"

Notes

• MSV (MassIVE) accessions: PRIDE MCP may return limited metadata; use the

publication as the primary evidence source.

• Never fill a field with a value not supported by evidence. Use unclear instead.
• Process accessions sequentially — do not skip any.
• Do not run sdrf:annotate, sdrf:validate, sdrf:fix, or sdrf:improve here.

Those belong to sdrf:autoresearch.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: bigbio
• Source: bigbio/sdrf-skills
• License: MIT
• Homepage: https://sdrf.quantms.org

Install and usage instructions live in the source repository linked above.