SpliceAI Lookup through GeneFoundry

Router invocation: spliceai_get_server_capabilities()

Use this first in a cold session to learn the tools, parameters, score glossary, recommended workflows, error codes, and limitations. detail='lean' returns a trimmed doc (tool list + verdicts + error codes + capabilities_version) that omits per-parameter prose already in the tool schemas. Full ~4kB, lean ~1-2kB.

Router invocation: spliceai_warmup()

Pre-warm the SpliceAI + Pangolin Cloud Run containers before a burst so the first real call does not eat the 10-40s cold start. Warms the (basic gene_set, chosen mask) path per model; pass mask='both' to warm raw and masked together. Cloud Run scales per-instance, so other param combos or concurrent calls may still cold-start and warmth decays after minutes idle. Returns per-(model,mask) elapsed_ms, coverage, and stay_warm_estimate_s. Returns <1kB.

Router invocation: spliceai_resolve_variant()

Use this when the caller's variant is HGVS, an rsID, or loosely formatted, and you need the canonical CHROM-POS-REF-ALT that the prediction tools require. Coordinate inputs are normalized locally; HGVS/rsIDs are resolved via Ensembl VEP, which also returns the most-severe consequence and gene symbol. Then call predict_splicing. Returns <1kB. Coordinate inputs are normalized; by default the REF base is also checked against the requested build (one Ensembl lookup) and a ref_warning + ref_validated:false is returned on mismatch (set check_ref=false to skip). HGVS/rsIDs are resolved and validated via Ensembl VEP.

Router invocation: spliceai_predict_spliceai()

ONE model only (SpliceAI); use predict_splicing for BOTH models with an agreement verdict. Use this for the SpliceAI delta scores (acceptor/donor gain/loss, each 0-1 with a position) of a single variant, optionally with the SpliceAI-10k consequence prediction (exon skipping / intron retention / frameshift). For a quick raw-vs-masked or single-model question; use predict_splicing to also get Pangolin. Δ>=0.5 is high-confidence. Returns ~1-4kB (full/all larger). Note: cold calls take 10-30s. Supports MCP background tasks (execution.taskSupport=optional): augment the call with a task to fire-and-continue instead of blocking 15-40s.

Router invocation: spliceai_predict_pangolin()

ONE model only (Pangolin); use predict_splicing for BOTH models with an agreement verdict. Use this for the Pangolin splice gain/loss scores of a single variant. Pangolin is an independent splice model; agreement with SpliceAI strengthens a prediction, disagreement warrants caution. Use predict_splicing to get both models in one call. Returns ~1-3kB. Note: cold calls take 10-30s. Supports MCP background tasks (execution.taskSupport=optional): augment the call with a task to fire-and-continue instead of blocking 15-40s.

Router invocation: spliceai_predict_splicing()

BOTH models (SpliceAI + Pangolin) in one call -- the default "what does this variant do to splicing?" answer. Use this as the default one-call answer for "what does this variant do to splicing?". It resolves HGVS/rsIDs, runs SpliceAI and Pangolin (two independent models), includes the SpliceAI-10k consequence prediction, and reports whether the models agree. Read the top-level headline first. For a single model use predict_spliceai / predict_pangolin. Returns ~3-6kB. Note: cold calls take 15-40s (two model calls). Supports MCP background tasks (execution.taskSupport=optional): augment the call with a task to fire-and-continue instead of blocking 15-40s.

Router invocation: spliceai_predict_splicing_batch()

Score a list of variants in ONE call. The server fans out under its concurrency cap and returns a single envelope with per-variant results (+ per-item errors that do not fail the batch) and a summary. Use this for gene panels instead of N predict_splicing calls. Accepts 1-25 variants (more than max_items=25 returns invalid_input, not a truncated result); each item returns about one compact predict_splicing result, so a full batch is ~25x a single compact response, and _meta echoes items_submitted and max_items. Supports MCP background tasks (execution.taskSupport=optional): augment the call with a task to fire-and-continue instead of blocking 15-40s.