Genomic Intelligence - DNA Sequence Models

Genomic Intelligence (GI) serves transformer DNA language models over six sequence-analysis tasks on managed GPUs. Give it a gene symbol, a genomic region, or a DNA/FASTA sequence; it returns structured predictions. Nothing runs locally - no model weights, no GPU. It is a thin client over a hosted, versioned inference API.

Docs: https://docs.genomicintelligence.ai | REST contract at https://api.genomicintelligence.ai/v1/openapi.json | hosted MCP server at https://mcp.genomicintelligence.ai/mcp.

Research and development use. Not for clinical or diagnostic decisions.

When to Use

Use GI when the user has DNA and wants a model prediction:

• promoter - promoter regions in a genomic region (sliding window)
• splice - donor/acceptor splice sites
• enhancer - developmental & housekeeping enhancer activity (DeepSTARR)
• chromatin - chromatin state across hundreds of tracks (DeepSEA)
• expression - sequence-to-expression, log(TPM+1), with a cell-type context
• annotation - de-novo gene/transcript annotation (async recommended)
• composite - find the genes in a region and predict each one's expression

Not for local alignment, variant calling, or file I/O - use a local tool (BioPython, bcftools) for those. GI is for model inference from sequence.

Access and Authentication

• The hosted MCP server (https://mcp.genomicintelligence.ai/mcp, Streamable HTTP) works keyless against a rate- and concurrency-limited public demo tier - prefer it on hosts that support MCP. An optional gi_ bearer key raises those limits.
• The REST /v1 API requires a GI_API_KEY (a gi_ bearer), sent as Authorization: Bearer <key>. Request one at contact@genomicintelligence.ai.
• Never hardcode the key. Read it from the GI_API_KEY environment variable.

export GI_API_KEY="gi_yourkeyhere"     # optional for MCP; required for REST

The Six Tasks

Each task is its own published operation - POST /v1/tasks/promoter/predict, /v1/tasks/splice/predict, /v1/tasks/enhancer/predict, /v1/tasks/chromatin/predict, /v1/tasks/annotation/predict, /v1/tasks/expression/predict - with its own request schema, its own minimum length, and its own closed options object. There is no shared PredictRequest. The URLs are the same strings clients already POST to, so no URL construction changes. Body {sequence, sequence_name?, model?, options?}, returning a {data, meta} envelope. expression is the strictest: alone among the six it requires options too - see the rules below.

Recommended mode is latency guidance, not a constraint - every task accepts both. Omit Prefer for a synchronous 200; send Prefer: respond-async for a 202 plus GET /v1/tasks/jobs/{job_id}. Only the composite workflow enforces a mode, rejecting sync above 50,000 bp with 413 sync_too_large.

The minimum is admission control, not a scoring regime. A request above the floor but shorter than the selected model's bio_spec.context_window_bp is accepted and scored - against a window padded out to the context window. Enhancer is the sharp case: the floor is 50 bp but the context window is 249 bp, so 50-248 bp is scored mostly on padding. Compare your length against context_window_bp from GET /v1/tasks/{task}/models to know whether the model saw real sequence. Longer input is fine - the scanner steps a prediction window at a time and pads only the final partial window.

Under the floor and over the 500,000 bp cap are both 422 validation_failed at loc ["body","sequence"] - over-length is not a 413. All lengths are measured after whitespace is stripped, so a line-wrapped FASTA body pastes verbatim (a > header line still fails the alphabet check).

options is typed and closed (additionalProperties: false) per task - an unknown key is a hard 422 (type: "extra_forbidden"), never ignored: promoter threshold; splice threshold, site_types; enhancer none; chromatin threshold; annotation batch_size, shift_coordinates, reverse_complement; expression description (required, and the only key).

Three hard rules the API enforces for expression (every violation is a 422; nothing is padded, clamped, or truncated, and there is no opt-out flag):

• The model always scores exactly one 9,198 bp TSS-centred window - sequence[tss_index-4599 : tss_index+4599]. The endpoint accepts 9,198-500,000 bp; below 9,198 bp is a hard 422.
• tss_index is required unless the sequence is exactly 9,198 bp. It is the 0-based TSS offset into the whitespace-stripped sequence, and must satisfy 4599 <= tss_index <= len(sequence) - 4599. At exactly 9,198 bp it defaults to 4,599 - the only legal value there. Hand over a whole locus (up to 500 kb) plus a tss_index and the server cuts the window for you; it does not discover the TSS itself, and it does not reverse-complement - submit gene-sense.
• options.description - a cell-type / assay string (e.g. "K562 cells") - is required, and is the only key expression accepts inside options. Unknown top-level body fields are also rejected.

Trap: the legal tss_index range is wide, so a wrong but in-range offset (e.g. counted over the raw FASTA including newlines, or relative to a locus start instead of the submitted slice) returns a confident 200 for the wrong window. Always assert on meta.task_specific_counts.scored_window (and .tss_index) in the response. Note also that data.input.sequence_length is the scored length (always 9,198); the length you submitted is data.input.submitted_sequence_length / meta.sequence_length.

Both tss_index failures - "required unless exactly 9,198 bp" and the range check - come from a whole-model validator, so they report at loc: ["body"], never body.tss_index. Match on error.code == "validation_failed"; never branch on loc.

Splice: submit gene-sense. Strand-specific, and the wrong strand fails silently. Submit transcript orientation, reverse-complementing minus-strand genes. A reverse-complemented sequence does not return zeros or an empty result: it returns plausible sites at different positions, often still at high confidence, and the site count can hold or collapse depending on the locus. Neither the score nor the count tells you the orientation was wrong, so there is no post-hoc check - get the orientation right on input. A gene-symbol fetch follows the gene's own strand; a coordinate fetch does not, so a minus-strand gene pulled by coordinates arrives antisense unless you reverse-complement it yourself. expression never reverse-complements either; annotation is strand-insensitive.

A splice coordinate is a token span, not a junction. Each site's start/end bounds one variable-width tokenizer token - 4-10 bp across the sequences measured so far - reported with a token_index. The exon/intron junction sits somewhere inside that span, so do not reduce the pair to a single base position, and do not intersect it against reference annotation as though it marked a boundary.

Omit model and the API uses the task's default - that is the recommended call. Default model IDs are intentionally not documented here: defaults change and retired IDs fail hard, so never hardcode one. To pin a model, or to pick a non-human one (Drosophila, yeast, and Arabidopsis models exist for several tasks - match the species), discover IDs at call time with GET /v1/tasks/{task}/models (REST) or list_models (MCP). Never invent a model ID.

How to Execute (REST)

Called synchronously - no Prefer header - a prediction is one request:

import os, requests

BASE = os.environ.get("GI_BASE_URL", "https://api.genomicintelligence.ai")
HEADERS = {"Authorization": f"Bearer {os.environ['GI_API_KEY']}"}

def gi_predict(task, sequence, sequence_name, model=None, options=None, tss_index=None):
    body = {"sequence": sequence, "sequence_name": sequence_name}
    if model:   body["model"] = model
    if options: body["options"] = options
    if tss_index is not None: body["tss_index"] = tss_index   # expression only
    r = requests.post(f"{BASE}/v1/tasks/{task}/predict", headers=HEADERS, json=body)
    # 422 validation_failed - under the task floor OR over 500,000 bp, bad
    #   tss_index, missing options.description, or ANY unknown body/options key
    # 401 no/bad key | 404 unknown task | 413 body over 16 MiB | 429 rate limit
    r.raise_for_status()
    return r.json()               # {"data": {...}, "meta": {...}}

# Promoter:
out = gi_predict("promoter", seq, "TP53_region")
print(out["data"]["summary"])

# Expression - a pre-cut 9,198 bp TSS-centred window (tss_index defaults to 4,599):
out = gi_predict("expression", tss_window_9198bp, "HBB",
                 options={"description": "K562 cells"})
print(out["data"]["prediction"]["expression_log_tpm"])

# Expression - a whole locus, server cuts the window around the TSS you name.
# tss_index is 0-based into the whitespace-stripped sequence.
out = gi_predict("expression", locus_seq, "HBB", options={"description": "K562 cells"},
                 tss_index=tss_offset_in_locus)
print(out["meta"]["task_specific_counts"]["scored_window"])   # verify the right window

Prefer: respond-async is a declared header on all six predict operations and on the composite - a 202 carries the same {data, meta} envelope with data = {job_id, status, links} (job id also in Content-Location / X-Job-Id). Async is JSON-only and available on every task; annotation is the one that usually needs it: send Prefer: respond-async, get a job_id, then poll GET /v1/tasks/jobs/{job_id} until it returns 200 (a 202 means still running):

import time
r = requests.post(f"{BASE}/v1/tasks/annotation/predict",
                  headers={**HEADERS, "Prefer": "respond-async"},
                  json={"sequence": seq, "sequence_name": "TP53"})
r.raise_for_status()
job_id = r.json()["data"]["job_id"]
while True:
    j = requests.get(f"{BASE}/v1/tasks/jobs/{job_id}", headers=HEADERS)
    if j.status_code == 200: break
    j.raise_for_status(); time.sleep(5)
transcripts = j.json()["data"]["transcripts"]

Sequence Acquisition

You rarely start from a raw sequence. Fetch reference sequence from Ensembl REST (rest.ensembl.org, public, no key) for a gene symbol or region, then feed it to GI. For expression, either build the 9,198 bp TSS-centred window from the gene's canonical transcript (expand=1): 4,599 bp upstream + 4,598 bp downstream on the gene's strand - or fetch a wider locus and pass the TSS offset as tss_index and let the server cut it. Default species is human, GRCh38; use the Ensembl production name for others (mus_musculus, drosophila_melanogaster). Fetch at least the task floor (promoter 300, splice 100, enhancer 50, chromatin 200, annotation 1,000 bp) - and ideally at least the model's context_window_bp, or the score reflects padding rather than sequence.

Hosted MCP (keyless, preferred on MCP hosts)

Acquire a sequence handle (sequence_ref), then predict against it, so large sequences never enter the context:

load_demo_sequence(name="promoter_tp53")  # keyless smoke test -> handle; name REQUIRED
fetch_ensembl_sequence(gene="TP53")       # gene symbol or Ensembl ID -> handle
fetch_region(region="chr11:5,225,000-5,235,000")   # coordinates -> handle
fetch_gene_for_expression(gene="HBB")     # TSS-centred 9,198 bp handle
predict_promoter(sequence_ref=<ref>)      # + predict_splice/_enhancer/_chromatin
predict_expression(sequence_ref=<ref>, description="K562 cells")
find_genes(sequence_ref=<ref>)                                   # annotation task
find_genes_and_predict_expression(sequence_ref=<ref>, description=...)  # composite

Note the hosted server exposes 15 tools and has no predict_annotation: the annotation task is find_genes, which takes a handle rather than a region and runs async internally (wait=True by default returns the result; wait=False returns a job_id to poll with get_job). There is no load_local_fasta on the hosted server either; use store_inline_sequence. Reference context lives in the gi://models, gi://docs/tasks, and gi://account MCP resources.

Key Endpoints

bio_spec carries request_max_bp (the enforced cap, 500,000 everywhere), context_window_bp (the sliding window; null for annotation/expression) and trained_window_bp (9,198 for the expression model). The legacy max_seq_length_bp has been withdrawn and no longer appears in bio_spec; request_max_bp is the cap. There is no strand_sensitive flag.

Notes

• Errors: 422 validation_failed is the catch-all - sequence under the task floor or over 500,000 bp, expression below 9,198 bp, missing/out-of-range tss_index, missing options.description, or any unknown body/options key. 401/403 missing/invalid key (REST). 404 not_found unknown task or job. 413 means only payload_too_large (the 16 MiB raw-body cap) or sync_too_large (the composite above 50,000 bp synchronous - retry async), never an over-long sequence. 415 unsupported_format for a bad format query value - there is no silent fallback to JSON. 429 rate cap (honour Retry-After; ask GI to raise the tier). 5xx retry.
• error.code is a closed 21-value enum; treat an unlisted value as a generic failure, not a parse error. Branch on code, never on details or loc: details matches the declared schema - a validation failure carries {errors: [{loc, msg, type}, ...]} - so read it defensively all the same. error.request_id mirrors the X-Request-Id header and both are always populated; success envelopes carry meta.request_id. Every response carries RateLimit-* headers.
• The API serves the six literal predict operations, typed options, per-task floors, the published composite, the Prefer parameter, the code enum and the bio_spec fields described here. The served schema at https://api.genomicintelligence.ai/v1/openapi.json is the authority; check it if a detail here does not match.
• GI is a hosted service; nothing here ships weights or runs local inference.

Follow-up Suggestions

• For "what genes are here and how are they expressed?", use the composite: POST /v1/workflows/find-genes-and-predict-expression (REST) or find_genes_and_predict_expression (MCP). sequence 1,000-500,000 bp and options are required, and options.description (cell type / assay) is required too - a missing or empty value is a 422. It cuts a TSS-centred 9,198 bp window per discovered gene (padding with N at region edges rather than dropping a gene) and returns a prediction each; meta.task_specific_counts = {genes_found, genes_predicted, genes_skipped}, per-gene causes in data.expression_predictions[].skip_reason. Above 50,000 bp a synchronous call is 413 sync_too_large - retry the same body with Prefer: respond-async.
• To explore available models per task, call list_models / GET .../models before predicting.