metadata

pretty_name: Human–Virus Protein Mistake Predictions
license: cc-by-4.0
tags:
  - biology
  - proteins
  - classification
  - viruses
  - tabular
task_categories:
  - tabular-classification
size_categories:
  - 10K<n<100K

Human–Virus Protein Mistake Predictions (Parquet)

This dataset provides per-sequence labels, predictions, and lightweight descriptors used in the paper:

Protein Language Models Expose Viral Immune Mimicry — Viruses 2025, 17(9):1199.
DOI: 10.3390/v17091199 · Article: https://www.mdpi.com/1999-4915/17/9/1199

What’s included: a single file HumanVirus_Protein_mistakes.parquet with 25,117 rows and 20 columns.
Not included: PLM model embeddings (paper used Swiss-Prot T5 static embeddings or trained ESM2).
Code: https://github.com/ddofer/ProteinHumVir
Workshop poster: ICML 2024 (ML4LMS): https://openreview.net/forum?id=gGnJBLssbb

Summary

We trained and analyzed protein language model classifiers, and interpretable tabular models in distinguishing viral from human proteins. The dataset focuses on DL model errors (mistake=True), which are enriched for proteins implicated in immune mimicry / immune evasion. Use this table to reproduce error-profiling, build new classifiers, or explore biological correlates of misclassification. e.g. Explaining the mistakes of the DL/PLM models, using tabular models and explainable features, as we do in the paper.

Schema

column	type	brief description
`Sequence`	string	amino-acid sequence
`Length`	int	Protein sequence length
`virus`	int	ground truth (1=viral, 0=human)
`model_pred_proba_vir`	float	P(viral)
`model_pred_vir`	int	model prediction (1/0)
`mistake`	bool	prediction ≠ label
`Baltimore`	string?	viral group (null for human)
`Family`	string?	viral family
`Genome Composition`	string	e.g., dsDNA
`Genus`	string?	viral genus
`Keywords`	string	UniProt keywords
`Mass`	int	molecular mass (Da)
`Organism`	string	source organism
`Protein names`	string	protein name(s)
`Taxonomic lineage`	string	taxonomy path
`UR50_Cluster ID`	string	UniRef/UR50 ID
`Virus hosts`	string	known hosts
`av_mw`	float?	avg AA mass feature
`genus_human_host`	bool	Virus's genus has human host indicator
`human_host`	bool	Virus with human host indicator

Loading

Direct (single Parquet):

from datasets import load_dataset
ds = load_dataset(
    "parquet",
    data_files={"train": "https://huggingface.co/datasets/<user>/<repo>/resolve/main/HumanVirus_Protein_mistakes.parquet"},
)["train"]