Qtl candidate gene

Candidate-Gene Interpretation for Tomato QTL

Turn a fine-mapped QTL interval into a ranked, mechanistically-reasoned set of candidate genes plus marker-assisted-selection (MAS) markers — the manual triage a geneticist does with a genome browser, an annotation database, and the literature, done as a repeatable workflow.

Target users: anyone who has localized a trait (e.g. via haploqtl local-ancestry clustering or traditional QTL mapping) — geneticists, molecular biologists, and breeders — who needs to know which genes are in the interval, what they do, and which markers track it.

When to use this skill

Use when the user:

• Has a genomic interval on the SL4.0 tomato assembly (e.g. ch09:62,452,852-63,002,852) plus the trait it was mapped for, and wants the candidate genes.
• Asks for candidate or causal genes, candidate-gene hypotheses, or "what's in this QTL" for a given trait.
• Wants MAS / KASP markers diagnostic of a trait-associated haplotype.
• Has haploqtl cluster output and wants to interpret a locus biologically.

The trait can be anything mapped in tomato — disease resistance, fruit size/quality, color, architecture, abiotic-stress tolerance. It determines which gene families are weighted as candidates, so it is required, not assumed.

Inputs

• Required: an interval — chromosome + start + end (SL4.0 bp).
• Required: the trait the QTL was mapped for (free text, e.g. "early-blight resistance", "fruit weight", "soluble solids content"). This drives candidate ranking.
• Optional (for MAS markers): a VCF of the relevant accessions plus two contrasting phenotype groups — e.g. resistant donors vs susceptible controls, or high- vs low-trait lines.

Workflow

[ ] Step 1  Genes in the interval        (offline, ITAG4.1)
[ ] Step 2  Protein function             (live UniProt)        -- needs network
[ ] Step 3  Retrieve literature          (live PubMed)         -- needs network
[ ] Step 4  Diagnostic MAS markers       (offline, from VCF)   -- if accessions given
[ ] Step 5  Synthesize candidate report  (reasoning)
[ ] Step 6  Verify + stamp               (PubMed + ITAG4.1)    -- gate before delivery

Step 1 — Genes in the interval

List the protein-coding genes overlapping the interval, with ITAG4.1 functional descriptions:

python scripts/genes_in_interval.py --chrom ch09 --start 62452852 --end 63002852

Returns JSON (gene_id, seqid, start, end, strand, description). The bundled source is an ITAG4.1 slice (SL4.0), so IDs and coordinates match the published annotation. For an interval outside the bundled EB-5/EB-9 regions, pass a full ITAG4.1 GFF3 with --gff (see references/data-sources.md).

Step 2 — Protein function (UniProt)

Enrich the most promising genes with protein-level function from UniProtKB (live):

python scripts/gene_function.py Solyc09g074790 Solyc09g074510

DECISION POINT: this requires network access. If offline, skip and rely on the ITAG4.1 descriptions from Step 1. Not every Solyc gene has a UniProt entry; null fields are expected and fine.

Step 3 — Retrieve literature (PubMed)

Search PubMed for evidence linking the candidate gene families (Step 1–2) to the trait — e.g. for disease resistance "tomato F-box protein defense Alternaria"; for fruit size "tomato cell number regulator fruit weight":

python scripts/pubmed.py "tomato F-box protein defense Alternaria" --retmax 5

Returns real {pmid, title} records from NCBI E-utilities. Cite only PMIDs this returns — never one recalled from memory (recalled PMIDs are frequently fabricated). The Step 6 gate strips any cited PMID that doesn't resolve, so memory-citations get caught — but retrieve up front so the claim keeps its evidence. Needs network; if truly offline, cite no PMIDs rather than guessing — the report still stands on the ITAG4.1 + UniProt evidence.

Step 4 — Diagnostic MAS markers (if accessions provided)

Find variants whose allele is present in one phenotype group and absent from the contrasting group — diagnostic markers for marker-assisted selection. The two groups are defined by the trait (resistant vs susceptible for a disease; high vs low for a quantitative trait). The script's flags are named --resistant/--susceptible for the canonical disease case, but they simply mean "carries the trait allele" vs "does not":

python scripts/diagnostic_variants.py --vcf <accessions.vcf.gz> \
    --chrom ch09 --start 62452852 --end 63002852 \
    --resistant "<trait-positive lines>" --susceptible "<trait-negative lines>"

With no group flags the defaults reproduce the EB-9 contrast from Anderson et al. (2024) (Devon Surprise pathway vs NC EBR 1 / NC 84173 / Brandywine). Override with --resistant/--susceptible (comma-separated accession names) and relax with --max-exceptions.

Step 5 — Synthesize the candidate-gene report

Using references/interpretation-guide.md, reason over the assembled evidence and produce a report:

1. Ranked candidate genes — name only genes from Step 1 and cite only PMIDs retrieved in Step 3; weight gene families plausibly linked to the stated trait above housekeeping genes; justify each with its function + literature. Match the families to the trait (don't default to disease): e.g. disease resistance → R-genes/NBS-LRR, receptor-like kinases, F-box/ubiquitin signaling, defense transporters/TFs; fruit size/quality → cell-number/expansion regulators, invertases and sugar/acid transport, hormone signaling; abiotic stress → transporters, LEA/dehydrins, stress TFs. See references/interpretation-guide.md.
2. Mechanistic hypothesis per top candidate (how it could plausibly affect the trait).
3. MAS markers — the diagnostic marker table from Step 4, ready for assay design.
4. Caveats — overlapping loci, annotation version, that interval co-location is not causation.

Step 6 — Verify the report (gate before delivery)

Run the deterministic gate on your draft before returning it. Save the draft as a JSON object (the candidates / determinable / rationale structure) and check it against the interval and PubMed:

python scripts/verify_report.py --report draft.json --chrom ch09 --start 62452852 --end 62950075

It drops any candidate whose Solyc ID is not in the interval, strips any cited PMID that doesn't resolve on PubMed (flagging that claim to re-retrieve — it never invents a replacement), flags an over-confident un-hedged call, and prints a verification stamp. Append the stamp to the report. If PMIDs were stripped, return to Step 3, retrieve real ones for those claims, and re-verify.

For the strongest check, add --support-model <model> (e.g. claude-haiku-4-5-20251001): for every PMID that resolves, the gate fetches its abstract and asks the model whether it actually supports the specific claim — stripping citations that are real but off-topic, and recording the supporting quote for the ones that hold. Without the flag the gate only confirms a PMID resolves, not that it backs the claim. This needs network, the anthropic package, and an ANTHROPIC_API_KEY; the rest of the gate stays deterministic and offline-capable.

python scripts/verify_report.py --report draft.json --chrom ch09 --start 62452852 --end 62950075 \
    --support-model claude-haiku-4-5-20251001

The stamp then also reports PMIDs supporting the claim, and each surviving citation carries the abstract quote that backs it — put that quote in the report so a reader can verify it at a glance. (To read an abstract yourself: python scripts/pubmed.py --abstract <PMID>.)

Output

A markdown report: ranked candidates with evidence, mechanistic hypotheses, a MAS marker table, explicit caveats, and the Step 6 verification stamp (genes in interval · PMIDs resolved · PMIDs supporting the claim · calibration) confirming the report was grounded and self-checked. See EXAMPLE.md for a worked EB-9 run.

References

• references/data-sources.md — data provenance (ITAG4.1, UniProt), coordinates, regenerating the gene slice, why Ensembl Plants REST is not used.
• references/interpretation-guide.md — how to rank candidate genes for the trait at hand (disease resistance worked through in detail, other traits sketched), and MAS-marker conventions.