lhallee/virulence-c3

Balanced C3 virulence dataset

This folder contains a balanced proteome-derived, pathogen-associated protein-sequence classification dataset. It contains 100 positive proteomes labeled 1 and 100 negative-control proteomes labeled 0, with 50 bacteria, 40 viruses, and 10 eukaryotes per label. Selection is deterministic, uses one assembly per NCBI taxid, and restricts both labels to shared broad-lineage and eukaryote-scope strata.

The positive class comes from the curated human-pathogen inventory in virulence/. The negative class comes from the taxonomy-matched candidate mirror in virulence_negative/. The negative label means absence from the current positive inventory after the mirror's taxonomic exclusion rules. It is not a claim of universal non-virulence. The labels apply to the source proteome, not to individual proteins as experimentally confirmed virulence factors.

The current build contains 550,143 downloaded protein entries and 548,317 unique, non-conflicting sequence strings after exact-sequence filtering. The final split has 493,485 training, 27,416 validation, and 27,416 test sequences across 315,374 MMseqs2 clusters.

Build

py -m virulence_c3.build_dataset select
py -m virulence_c3.build_dataset download
py -m virulence_c3.build_dataset prepare

The downloader retrieves one FASTA per selected UniProt reference proteome from the UniProt REST API and writes data/sequence_records.csv and data/all_sequences.fasta.

Cluster the complete sequence set before splitting. The command below uses MMseqs2 easy-cluster with 50% minimum sequence identity and 80% coverage:

docker run --rm -v "${PWD}/virulence_c3:/app/virulence_c3" ghcr.io/soedinglab/mmseqs2@sha256:41b12b0d5f41432fa1b9976123da6e2e06e7fab49a34964f3b54ec038e5845d9 easy-cluster /app/virulence_c3/data/cluster_input.fasta /app/virulence_c3/mmseqs/easy_cluster /app/virulence_c3/mmseqs/tmp --min-seq-id 0.5 -c 0.8 --cov-mode 0 --cluster-mode 0 --threads 8
Copy-Item mmseqs/easy_cluster_cluster.tsv mmseqs/cluster.tsv
py -m virulence_c3.build_dataset split

The preparation step excludes exact sequence strings observed under both labels and retains one row for each remaining exact sequence. The cluster input contains 548,317 unique sequence strings from 550,143 downloaded records in the current snapshot. The split command assigns whole MMseqs2 clusters to train, validation, or test with a deterministic greedy assignment targeting 90/5/5 overall and within each label. data/sequence_dataset.csv and split-specific CSV files contain the sequence, label, proteome_id, taxonomic metadata, cluster_id, and split.

Attribution and terms

The protein sequences and UniProt proteome metadata are derived from UniProt REST API downloads. UniProt states that copyrightable parts of its databases are available under CC BY 4.0. This release preserves UniProt proteome IDs and source URLs for attribution. The pathogen labels and inventory metadata are derived from the repository's FDA-ARGOS/HIVE and taxonomy-backed source files and may have separate upstream terms. This card does not grant a blanket license for upstream material; users should review the cited source terms before redistribution or commercial use.

Release contents

• selected_proteomes.csv: auditable 200-proteome selection manifest.
• data/all_sequences.fasta: downloaded input to MMseqs2.
• data/cluster_input.fasta: deduplicated, conflict-filtered input to MMseqs2.
• data/sequence_records.csv: parsed sequence records before clustering.
• data/cluster_sequence_records.csv: exact-sequence-filtered records used for clustering.
• data/sequence_dataset.csv: final labeled dataset after cluster assignment.
• data/train.csv, data/validation.csv, data/test.csv: split files.
• selection_metadata.json, split_metadata.json: source hashes, counts, and split diagnostics.

The Hugging Face release should include the CSV split files and metadata. The raw per-proteome FASTAs, full downloaded FASTA, and MMseqs2 database files are local build intermediates and should not be uploaded unless a separate archival release is desired. UniProt attribution and applicable data terms are documented in the release card.