Add PyEnsemblAdapter as offline alternative to MartsAdapter

[jonasscheid]

Problem

The MartsAdapter relies on live HTTP queries to Ensembl BioMart, which is a known source of flakiness — server outages, rate limiting, and malformed responses (see #103). This is especially problematic for downstream tools like nf-core/epitopeprediction, where epaa.py depends on MartsAdapter for two calls:

1. get_transcript_information() — CDS, strand, gene name retrieval (blocking — no peptides without it)
2. get_protein_ids_from_transcripts() — transcript → protein/RefSeq/UniProt ID mapping (annotation)

While #103 proposes migrating to the Ensembl REST API, that still requires live network queries and is subject to rate limits (15 req/s).

Proposal

Add a PyEnsemblAdapter that implements ADBAdapter using PyEnsembl (by OpenVax). PyEnsembl downloads Ensembl GTF + FASTA files once and indexes them into a local SQLite database — all subsequent queries are entirely offline.

This covers the critical path:

ADBAdapter method	PyEnsembl equivalent
get_transcript_information()	transcript.coding_sequence, transcript.strand, gene.name
get_transcript_sequence()	transcript.coding_sequence
get_product_sequence()	genome.protein_sequence()

What it does not cover:

• Cross-database ID mapping (Ensembl → RefSeq/UniProt) — but this is already being addressed in nf-core/epitopeprediction via VCF/VEP CSQ extraction

Advantages over REST/BioMart

• No network dependency after initial download — eliminates the Fix CI, move to GH actions #1 failure mode
• Reproducible — pinned to a specific Ensembl release
• Fast — local SQLite queries instead of HTTP round-trips
• Non-breaking — new adapter class alongside existing MartsAdapter

Scope

A new epytope/IO/PyEnsemblAdapter.py implementing the three ADBAdapter abstract methods:

• get_product_sequence()
• get_transcript_sequence()
• get_transcript_information()

PyEnsembl would be an optional dependency (e.g. pip install epytope[pyensembl]).

Notes

• PyEnsembl supports Ensembl releases 54–114, human + mouse + other species
• Initial download is ~1–2 GB per release/species (one-time cost, cacheable in containers/CI)
• Users need to match Ensembl release to genome build (e.g. GRCh37 → release 75, GRCh38 → release 110+)
• This complements Consider migrating MartsAdapter from BioMart to Ensembl REST API #103 — PyEnsembl for the offline-capable subset, REST API for cross-references if needed