Habitat Prediction for Microbial Genomes Using Textual Metadata

A dataset of 13,637 microbial genomes with predicted habitat classifications for deep marine, shallow marine, or terrestrial environments. The labels were generated using an XGBoost classifier trained on a manually curated set of 865 genomes, achieving an average AUC of 0.987. The dataset was created by Zhanghan Ni and last updated in May 2026.

Use Cases

• Training habitat classification models based on textual metadata features.
• Benchmarking machine learning algorithms for biological text classification.
• Studying the distribution of microbial life across marine and terrestrial habitats.
• Validating automated habitat inference pipelines against manual curation.

Strengths

• Training set of 865 manually curated genomes provides a high-quality reference.
• Model performance is quantified with an average AUC of 0.987.
• Feature extraction generated 3,801 vectors from textual habitat descriptions.

Limitations

• Row count for the main dataset is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• Habitat definitions are broad (three categories), which may obscure finer ecological gradients.

Provenance

Source

Data aggregated from NCBI BioProject, BioSample, GOLD, GTDB, Bergey's Manual, and published papers.

Collection Method

Habitat information was manually curated from textual descriptions and used to train an XGBoost classifier for prediction on an expanded genome set.

Freshness

Last updated 2026-05-14 00:10:09; freshness should be verified.

Geography

Global, covering marine (deep and shallow) and terrestrial environments.