LOCAL_README.md

UniProt FoT Graphification Workspace

This directory is the ToricGT landing area for biological graph/FoT training data derived from /home/iska/Documents/amelie/bio/iska-net/data/raw_hf_bio_scale.

The first implementation pass is conservative because the source data is about 26GB and the host filesystem is nearly full. Raw datasets are organized here by symlink under raw_sources/ by default. This keeps the original data reachable from the ToricGT data tree without duplicating it. A true move is supported by the build script, but should only be used after confirming that no iska-net workflow depends on the current raw path.

Record Families

• uniprot_function_text_train: protein sequence plus function and protein-name text, with EC labels extracted when present and AlphaFold DB lookup URLs when accessions allow them.
• uniprot_uniref50_sequence_train: UniRef50 clusters with sequence, taxonomy, GO MF/BP/CC lists, representative accessions, and cluster metadata.
• rfam_sequence_train: RNA family/clan sequence annotations.
• rnacentral_8192_sequence_train: RNAcentral sequence/type/description rows.
• dna_coding_regions_train: genomic sequence, exons, introns, translated proteins, organism, and accession fields.
• pubchem10m_selfies_train: SELFIES molecular strings as graph-tokenizable medicinal-chemistry strings.

Graph/FoT Format

Each derived row has flat Parquet fields for training loaders plus a full graph_json object with:

• id, source, task_family, nodes, edges, targets, metadata, and split_cluster.
• Directed causal edges between source record, sequence, annotation, structure lookup, GO, EC, and feature nodes.
• Stable hash-based latent coordinate proxies for continuous/hybrid GFlowNet metadata. These are curation features, not learned embeddings.
• TokenGT/TropicalGT/ToricGT metadata: node token order, edge token order, tropical active support nodes, a tropical margin proxy, toric phase-basis tags, and ConvexTok/byte-packing compatibility notes.
• GFlowNet reward metadata based on source-field density, sequence presence, GO/EC/structure availability, and directed graph connectivity.
• Leakage-resistant split clusters based on dataset, entry/accession, sequence hashes, or sequence prefixes.

The companion forest_json organizes each source row into four deterministic source-field trees: sequence, annotation, structure lookup, and future design conditions. This is not the authored reasoning dataset yet; it is the graphified raw-data substrate that later authored FoT/ToT trajectories can cite.

Build And Validate

Run from the ToricGT repository root on branch toricblm-data:

/home/iska/miniconda3/envs/iska-net-2/bin/python scripts/build_uniprot_fot_dataset.py build \
  --raw-root /home/iska/Documents/amelie/bio/iska-net/data/raw_hf_bio_scale \
  --output-dir data/uniprot_fot \
  --sample-per-dataset 8 \
  --raw-link-mode symlink

/home/iska/miniconda3/envs/iska-net-2/bin/python scripts/build_uniprot_fot_dataset.py validate \
  --jsonl data/uniprot_fot/derived/uniprot_fot_graphified_sample.jsonl

Outputs:

• raw_sources/: symlink organization layer for the original raw datasets.
• manifests/uniprot_fot_build_manifest.json: raw and derived data manifest.
• derived/uniprot_fot_graphified_sample.jsonl: inspectable JSONL rows.
• derived/uniprot_fot_graphified_sample.parquet: Parquet training sample.

Next Dataset Layer

The next layer should be a separate authored FoT/ToT corpus, not a generator dump. Each record should be written as a technical reasoning artifact grounded in one or more source rows, especially UniProt/UniRef records with rich functional, GO, EC, site, family, pathway, structure, perturbation, and design constraints. Scripts may validate, hash, shard, and publish accepted records, but authored reasoning text should be inspected before acceptance.