Supplement for "Regional sequence and embedding divergence among five grass OsNAC25-like proteins"

Contents
- input_sequences.fasta: five-protein analysis panel.
- sequence_provenance.tsv: source databases, URLs, accessions, checksums, and candidate-selection notes.
- source_archives.sha256: SHA-256 values for the source collections retained from the original screen.
- candidate_screen/: original top-hit table, Pfam results, and source download manifest.
- alignment/: Galaxy-hosted MAFFT alignment.
- residue_embeddings_f32/: 15 ESM C residue-embedding tensors, little-endian float32, one row per residue and 1,152 columns.
- embedding_manifest.tsv: tensor dimensions and SHA-256 values.
- pooled_embedding_distances/: full-length, residues 1-180, and residues >180 cosine-distance matrices.
- structures/: Biohub-hosted ESMFold2 PDB predictions. Per-residue normalized pLDDT is stored in the B-factor field.
- provenance_logs/ and analysis_summary.json: Godsgate model provenance and run summary.
- pairwise_metrics.csv, structure_confidence_summary.csv, and quantitative_summary.json: derived numerical tables.
- figure1_embedding_distances.png and figure2_structure_confidence.png: manuscript figures.
- scripts/: execution and analysis scripts.
- original_question.txt: the question recorded before analysis. Its use of "homologs" predates the manuscript's more conservative "OsNAC25-like candidates" terminology.

Validation
- The five input sequences match the retained source records exactly.
- The retained source archive URLs were reachable on 2026-07-24; PlantTFDB and Ensembl archive hashes matched the local copies.
- Averaging each retained residue-embedding tensor and recalculating cosine distance reproduced the archived matrices with a maximum absolute difference of 4.8e-09.
- Every PDB contains one C-alpha record per input residue.
- Non-rice sequences are candidates from a one-way screen, not confirmed orthologs.
- Model outputs are predictions, not experimental measurements.
