Geniml Genomic Interval ML Workflow Guide

What it does

This skill acts as the safety gate and planning layer for Geniml, the machine-learning toolkit for genomic interval sets.

• BED validation: checks the 0-based half-open convention, negative coordinates, end <= start, unknown contigs, ends beyond contig length, and assembly/naming mismatches (GRCh38 vs GRCh38.p14, chr1 vs 1). The validator only proposes actions — it never rewrites your BED.
• Reproducible installs: pins geniml==0.8.4 with gtars==0.9.2 via uv, and pushes you to commit the lockfile plus artifact digests.
• Current API map: points at module paths that actually work in 0.8.4 (gtars.tokenizers.Tokenizer, geniml.region2vec.main.Region2VecExModel, geniml.scembed.main.ScEmbed) and flags stale docs (geniml assess → assess-universe, geniml scembed is a no-op, embedding_size → embedding_dim).
• Compatibility auditing: verifies that config.yaml vocab_size/embedding_dim, exact universe.bed bytes and row order, special-token IDs, and checkpoint tensor shapes all agree before loading anything.
• Leakage control: requires splits by patient, donor, or biological replicate rather than by BED row or cell, with a manifest auditor to catch violations.
• Network discipline: BEDbase (api.bedbase.org) and Hugging Face downloads require explicit approval; bundled helpers are stdlib-only and offline.

Who it's for

• Computational biologists building embeddings from ATAC-seq/ChIP-seq peak sets
• Teams running scEmbed on single-cell ATAC AnnData objects
• Pipeline engineers merging many BED files into consensus universes
• Groups that must document versions, checksums, and coordinate conventions for reproducibility

Example uses

1. Pre-flight BED check: "Does data/peaks.bed satisfy the GRCh38 contract?" → bed_validator.py reports contig naming, out-of-bounds ends, duplicates, and a normalization plan.
2. Plan a Region2Vec run: give it a token Parquet (list-valued tokens column) plus universe.bed; embedding_plan.py bounds epochs, workers, and output size before any training command is approved.
3. Verify an inference bundle: before loading models/region2vec, run model_artifact_inspector.py and tokenizer_compatibility.py to confirm vocab size, special tokens, and assembly all line up.

Install the skill I found on Claude Skill Mart.
Copy the skills/geniml folder from the GitHub repo K-Dense-AI/scientific-agent-skills into my ~/.claude/skills/geniml/.
When it's done, tell me in one line what this skill can do.