feat(genomic-intelligence-nim): hosted DNA-sequence prediction skill

[boldakov]

Summary

Adds a nim-skills/genomic-intelligence-nim skill that wraps Genomic Intelligence's hosted DNA language models, following the existing nim-skills/ pattern (a SKILL.md driving a bearer-authed REST call). One skill, six tasks over the shared POST /v1/tasks/{task}/predict contract:

• promoter - promoter regions
• splice - splice donor/acceptor sites
• enhancer - developmental & housekeeping enhancer activity
• chromatin - chromatin state across hundreds of tracks
• expression - sequence -> expression, log(TPM+1)
• annotation - de-novo gene/transcript structure

Inference is hosted - no local weights or GPU; the only dependency is requests.

Third-party hosted endpoint

Unlike the existing nim-skills/ (which target NVIDIA-operated endpoints), this skill calls an endpoint operated by Genomic Intelligence. The interface is identical to the hosted-NIM shape (HTTPS + Authorization: Bearer + JSON); it is hosted-only. Flagging it explicitly so you can weigh in, happy to adjust naming. Context in #5.

Why a scripts/ runner (not inline-only)

The published nim-skills/ inline their Python in SKILL.md. This skill instead ships a small, self-contained requests-only runner, because the surface (six tasks, async jobs, and an exact 9,198 bp window for expression task) does not inline cleanly. Shipping scripts/ is not prohibited by CONTRIBUTING, and other categories ship scripts too. SKILL.md still includes a minimal inline example for the simple sync case.

What's included

• SKILL.md - frontmatter + task table + runner usage + a minimal inline example.
• scripts/ - gi_predict.py (one CLI, six tasks, sync + async, writes report + JSON), gi_client.py (the /v1 client), and optional gi_fetch.py / gi_ensembl.py (gene/region -> FASTA via public Ensembl, incl. TSS-centring).
• references/{api,tasks,authentication,sequence-acquisition,errors}.md.
• evals/{evals,trigger_evals}.json - matching the repo's eval schema (promoter, expression-window, and annotation-async cases).
• assets/demo/*.fa - one real reference FASTA per task (e.g. TP53 promoter, HBB splice, HBB-K562 9,198 bp expression window).

Auth

Users set GI_API_KEY (partner key, request at contact@genomicintelligence.ai). No key is committed; the skill resolves it from the environment.

Licensing / DCO

Contribution dual-licensed Apache-2.0 OR CC-BY-4.0; commits are DCO signed-off.

[boldakov]

Updated the branch: merged current main and brought the skill in line with the layout changes that landed after this PR was opened.

• registered the skill in skills.sh.json and committed the plugin_sync --write payload, so the plugin-sync check should now pass
• added evals/config.yml and trimmed evals.json to a single bounded case, with the network-heavy expression/annotation cases moved to deferred_evals (same shape as evo2-nim)
• a few contract corrections in the references: the enhancer minimum is the user-visible 50 bp, the exact 9,198 bp expression window is documented as a client-side guard rather than an API limit, default model ids are no longer hardcoded, and an off-by-one in the expression fixture header is fixed

Two questions we didn't want to guess on:

1. Placement. The repo has since moved to components.d entries sourcing from product-owned repos, with nim-skills/ described as BioNeMo NIM bundles. If you'd rather take this through components.d, we can publish the skill in a public genomicintelligence-owned repo and register an entry there instead.
2. Eval credentials. The runnable eval needs GI_API_KEY (free partner key) instead of NGC_API_KEY, which would make it the first skill here whose eval uses a third-party credential. If that doesn't fit the harness, we can move that case to deferred_evals too.

Also glad to add a row to the README skill catalog, or leave that to you. Broader scope question is still #5.

[greptile-apps]

Greptile Summary

The PR adds a hosted Genomic Intelligence skill for six DNA prediction tasks, including sequence acquisition, synchronous and asynchronous inference, validation, reporting, evaluations, and a synchronized plugin payload.

• Adds strict single-record FASTA parsing and task-specific input validation.
• Adds canonical-TSS expression-window acquisition with strand-aware coordinate handling.
• Adds response-envelope and nested-field validation with stable hosted-service diagnostics.
• Publishes the skill through skills.sh.json and mirrors it into the generated plugin tree.

Confidence Score: 5/5

The PR appears safe to merge because no blocking failure remains in the previously reported paths.

No blocking failure remains.

Important Files Changed

Filename	Overview
nim-skills/genomic-intelligence-nim/scripts/gi_client.py	Implements bearer-authenticated prediction and polling, strict response-envelope checks, and single-record FASTA validation; the displayed parser and response-handling threads are addressed.
nim-skills/genomic-intelligence-nim/scripts/gi_ensembl.py	Resolves Ensembl loci and constructs canonical-transcript, strand-aware expression windows; the prior fallback and off-by-one issues are addressed.
nim-skills/genomic-intelligence-nim/scripts/gi_predict.py	Provides task validation, sync/async execution, nested response validation, reporting, and stable diagnostics for expected hosted-service failures.
nim-skills/genomic-intelligence-nim/tests/test_input_validation.py	Covers the previously reported FASTA, TSS-centering, envelope, nested-field, and CLI error-handling regressions.
plugins/bionemo-agent-toolkit/skills/genomic-intelligence-nim/SKILL.md	Adds the generated plugin-facing copy of the hosted genomic prediction skill.
skills.sh.json	Adds the new skill to the reviewed distribution catalog.

Sequence Diagram

sequenceDiagram
  participant User
  participant Fetch as gi_fetch.py
  participant Ensembl
  participant Predict as gi_predict.py
  participant GI as Hosted Genomic Intelligence API
  User->>Fetch: Gene, region, or expression request
  Fetch->>Ensembl: Resolve locus and fetch sequence
  Ensembl-->>Fetch: FASTA sequence
  Fetch-->>User: Single-record FASTA
  User->>Predict: Task, FASTA, and task options
  Predict->>GI: Bearer-authenticated prediction request
  alt Annotation async
    GI-->>Predict: Job identifier
    loop Until completion
      Predict->>GI: Poll job
      GI-->>Predict: Progress or result
    end
  else Synchronous task
    GI-->>Predict: Prediction envelope
  end
  Predict-->>User: report.md, result.json, and JSON summary

Loading

_{Reviews (12): Last reviewed commit: "chore(genomic-intelligence-nim): include..." | Re-trigger Greptile}

[boldakov]

Good catch, that one was mine. The previous commit wrapped only wait_for_job; the edit intended for predict() matched no text and silently did nothing, so the sync path stayed unguarded while the async path was fixed.

3a95615 wraps predict() and tightens the check itself, since a null or non-object data passed the presence test and then failed in the summarizer. Both are pinned by tests, including one asserting predict() calls the validator so the same no-op cannot recur.

[greptile-apps]

Want your agent to iterate on Greptile's feedback? Try greploops.

[boldakov]

Could a maintainer approve the workflow runs on this PR? As a fork PR they need a manual approval, and all 36 runs since the PR opened are sitting at action_required, so plugin-sync, skill-security and Request NVSkills CI have never executed against this branch.

Locally the skill's suite is 150 passed / 1 skipped and scripts/plugin_sync.py --check reports the plugins/ payload matching source, but that is my evidence rather than yours.