gene-ctx-2412/GeneContext-Dataset

GeneContext Dataset

The GeneContext dataset is a subsample of the BacBench Essential Genes Dataset designed to evaluate the ability of Genome Language Models (gLMs) to interpret genomic context using an essentiality-based benchmark. The primary dataset contains genes from "mixed essentiality" orthologous groups (i.e. groups containing at least one essential and one non-essential gene) with essentiality labels as predictive targets for a linear probe.

Dataset Structure

The repository contains four datasets to ensure full reproducibility of the paper's experiments:

• —`main`: The primary subsample of BacBench mixed-essentiality groups used for probing experiments.
• —`bacbench_full`: The full BacBench dataset before any filtering was applied, augmented with additional annotations
• —`f4_og_pairs`: Pairwise identifiers used to for context swapping experiments.
• —`og_metadata`: Comprehensive mapping of genes to eggNOG orthologous groups.

Dataset Creation & Provenance

• —Source Data: Subsampled from the BacBench Essential Genes Dataset.
• —Annotation: Orthology labels assigned using eggNOG-mapper v2.
• —Curation: Specifically filtered for orthologous groups containing at least one essential and one non-essential representative to isolate context-dependent signals.

This dataset is part of a submission to NeurIPS 2026.