Use Nucleotide Transformer v3 for long-context multispecies workflows with Hugging Face Transformers as the primary path, with JAX helper APIs as secondary compatibility, including pre-trained MLM embeddings, post-trained species-conditioned track/annotation inference, dual-mode fine-tuning workflows (prep + full training), length-divisibility checks, and gated-repo troubleshooting. Use when Codex needs to write, fix, explain, or review code or notebooks involving `AutoTokenizer.from_pretrained(..., trust_remote_code=True)`, `AutoModelForMaskedLM`, `AutoModel`, `encode_species`, `num_downsamples`, `keep_target_center_fraction`, NTv3 model names, bigwig/bed outputs, NTv3 fine-tuning notebooks, or NTv3 install/auth issues.
Install, configure, and use Evo 2 for DNA sequence scoring, embeddings, and generation across local GPU inference, Docker, and Nvidia hosted API or NIM deployments. Use when Codex needs to write, fix, explain, or review Evo 2 Python code, shell commands, notebooks, or deployment steps involving checkpoint selection, hardware compatibility, `Evo2(...)`, tokenization, forward passes, embeddings extraction, generation, or Evo 2 installation and troubleshooting.
Build and debug AlphaGenome Python API workflows for variant-effect and track-prediction tasks, including API key setup, package installation, `dna_client` creation, selecting `requested_outputs`, adding `ontology_terms`, plotting results, and troubleshooting environment or response issues. Use when Codex needs to write, fix, explain, or review code and notebooks that use `alphagenome`, `dna_client`, `genome.Interval`, `genome.Variant`, `predict_variant`, `predict_interval`, AlphaGenome plotting helpers, or AlphaGenome API prediction workflows.
Use Calico Borzoi workflows for RNA-seq coverage prediction from DNA, including environment setup, model download, data processing, mini-model training, variant scoring, sequence-interpretation gradients, and SV/STR analysis. Use when Codex needs to write, fix, explain, or review Borzoi commands, scripts, or notebooks from the official `borzoi` tutorials (`make_data`, `train_model`, `score_variants`, `interpret_sequence`, `analyze_sv`) or example QTL notebooks. Prefer this skill for Borzoi repository workflows; use `segment-nt` only for SegmentBorzoi JAX segmentation APIs from Nucleotide Transformer docs.
Use SegmentNT, SegmentEnformer, and SegmentBorzoi for JAX-based genomic segmentation at nucleotide resolution, including model-family selection, `pmap` inference setup, SegmentNT Yarn rescaling, feature-probability extraction, and sequence/token-length troubleshooting. Use when Codex needs to write, fix, explain, or review code or notebooks involving `get_pretrained_segment_nt_model`, `get_pretrained_segment_enformer_model`, `get_pretrained_segment_borzoi_model`, `rescaling_factor`, `hk.transform_with_state`, segmentation logits, or sequence constraints for nucleotide-resolution annotation.
Build production-style Codex skills from a JSON spec, including SKILL.md, skill.yaml, agents/openai.yaml, references, and runnable scripts. Use when Codex needs to scaffold a new tool-specific skill quickly and consistently with this repository's conventions.
Use DNABERT-2 for DNA sequence embeddings, GUE benchmark evaluation, and supervised fine-tuning on custom classification datasets with the official Hugging Face + `finetune/train.py` workflow. Use when Codex needs to write, fix, explain, or review Python, shell, notebook, or `torchrun` code involving `zhihan1996/DNABERT-2-117M`, `AutoTokenizer`, `AutoModel`, `AutoModelForSequenceClassification`, `trust_remote_code=True`, DNABERT2 dataset formatting (`train.csv`/`dev.csv`/`test.csv`), or DNABERT2 training hyperparameters.
Choose and use the Song Lab GPN model family, including GPN, GPN-MSA, PhyloGPN, and GPN-Star, for genomic language model loading, embeddings, training, and variant-effect workflows. Use when Codex needs to write, fix, explain, or review Python, Hugging Face, `torchrun`, Snakemake, notebook, or shell workflows involving `gpn`, `gpn.model`, `gpn.star.model`, `AutoModelForMaskedLM`, `AutoModel`, GPN single-sequence training, embeddings extraction, or GPN variant-effect prediction.