m42-health/variant-benchmark

Variant Benchmark

This benchmark is designed to evaluate how effectively models leverage variant information across diverse biological contexts. Unlike conventional genomic benchmarks that focus primarily on region classification, our approach extends to a broader range of variant-driven molecular processes.

Existing assessments, such as BEND and the Genomic Long-Range Benchmark (GLRB), provide valuable insights into specific tasks like noncoding pathogenicity and tissue-specific expression. However, they do not fully capture the complexity of variant-mediated effects across multiple biological mechanisms. This benchmark addresses that gap by incorporating a more comprehensive set of evaluations, enabling a deeper assessment of functional genomics models.

Quick Start

from datasets import load_dataset

DATASET_SUBSET = "ancestry_prediction"

# Dataset subset should be from one of the available tasks:
# ['ancestry_prediction', 'non_coding_pathogenicity', 'expression',
#                         'common_vs_rare', 'coding_pathogenicity', 'meqtl', 'sqtl']

ds = load_dataset(
      "m42-health/variant-benchmark",
      DATASET_SUBSET,
)

print(ds)

# Example output:
# (Each dataset has a default 'train' split, but we recommend using k-fold cross-validation for better evaluation)
# DatasetDict({
#     train: Dataset({
#         features: ['sequence', 'id', 'start_idx', 'chromosome', 'variants', 'parents', 'label'],
#         num_rows: 14085
#     })
# })

Benchmark Tasks

• Coding pathogenicity assessment: subset: coding_pathogenicity
  Accurate prediction of pathogenic coding variants is fundamental to precision medicine and clinical genomics. For this task, we use the AlphaMissense dataset, which provides a comprehensive catalog of coding variants annotated for pathogenicity.

• Noncoding pathogenicity assessment: subset: non_coding_pathogenicity
  Pathogenic variants in noncoding regions significantly impact gene regulation, influencing many complex traits and diseases. We assess this using the BEND dataset, which contains 295,000 annotated single-nucleotide variants (SNVs) in noncoding genomic regions.

• Expression effect prediction: subset: expression
  Variant-driven changes in gene expression contribute to phenotypic diversity and disease processes. To quantify these effects, we use gene expression data from DeepSea, which provides variant-associated regulatory annotations.

• Alternative splicing: subset: sqtl
  Variant-induced alternative splicing contributes significantly to human proteomic diversity and affects biological processes and diseases. We evaluate splicing-related variant effects using an sQTL dataset derived from sqtlSeeker2, containing over one million variant-tissue pairs.

• DNA methylation: subset: meqtl
  Genomic variants can influence DNA methylation patterns, affecting gene regulation and disease susceptibility. For this task, we utilize meQTL data from the GRASP database, which links genetic variants to methylation changes.

• Ancestry classification: subset: ancestry_prediction
  Genomic variation encodes population structure, informing studies in evolutionary biology and disease susceptibility. To evaluate this capability, we used genomic segments labeled by five major superpopulations from the 1000 Genomes Project.

• Common vs synthetic variants: subset: common_vs_rare
  This task evaluates the model’s ability to recognize biologically conserved genomic contexts characteristic of authentic common variants. To create this deataset, we randomly sampled 100K common variants (MAF > 0.05) from GnomAD~\citep{chen2024gnomad} and paired each with a synthetic control variant generated by randomly substituting a nucleotide within a ±20-nucleotide local context window.

Citation

@article {Medvedev2025.03.27.645711,
    author = {Medvedev, Aleksandr and Viswanathan, Karthik and Kanithi, Praveenkumar and Vishniakov, Kirill and Munjal, Prateek and Christophe, Clement and Pimentel, Marco AF and Rajan, Ronnie and Khan, Shadab},
    title = {BioToken and BioFM - Biologically-Informed Tokenization Enables Accurate and Efficient Genomic Foundation Models},
    elocation-id = {2025.03.27.645711},
    year = {2025},
    doi = {10.1101/2025.03.27.645711},
    publisher = {Cold Spring Harbor Laboratory},
    URL = {https://www.biorxiv.org/content/early/2025/04/01/2025.03.27.645711},
    eprint = {https://www.biorxiv.org/content/early/2025/04/01/2025.03.27.645711.full.pdf},
    journal = {bioRxiv}
}