GeneProv v1.1

Benchmarking gene embeddings

How much can a gene embedding predict when its input already contains clues to the answer? GeneProv compares text, sequence and expression models across four tasks, tracking the source of each input.

Explore the leaderboard Read the preprint

GO-generated gene summaries

Human genes with an NCBI summary generated from Gene Ontology annotations, by linked publication count.

NCBI Gene6 October 2026
Human protein-coding genes
19,256
Open-track representations in v1
20
Fixed-text encoders in v1
6
Evaluation tasks
4

Open-track leaderboard

Top five of the original 20 representations, ranked by the GeneProv Index with a linear probe.

Open-track ranking for genes with 0–10 linked publications
RankRepresentationInput or training modalityIndex95% interval

The original GO scorer mishandled tied predictions. GO scores and indices shown here await rescoring. Correction record

Full leaderboard

Two ways to evaluate a model

Choose a track based on whether you can change the model’s input text.

Benchmark design
Track O

Bring a representation

Compare an existing embedding with other models. Report gene coverage, the origin of any text it embeds, and whether text entered its training.

Open-track instructions
Track F

Fix the text, compare encoders

Encode the same GenePT corpus twice. Removing Alliance-generated summaries in one version measures how much each encoder’s GO score depends on those summaries.

Fixed-text instructions

The index averages GO recovery, non-GO endpoints and co-essentiality retrieval. A fourth task evaluates later experimental annotations for models that meet the historical input rules. Task definitions

Research paper Preprint · 2026

When gene summaries reveal the answers: Gene Ontology circularity in gene embedding benchmarks

Jiawei Wu, Huishu Wang, Junlin Chen and Hongyi Wang

Read the paper Benchmark data
Cite this paper 
@misc{wu2026genesummaries,
  title  = {When gene summaries reveal the answers: Gene Ontology circularity in gene embedding benchmarks},
  author = {Wu, Jiawei and Wang, Huishu and Chen, Junlin and Wang, Hongyi},
  year   = {2026},
  note   = {Preprint},
  url    = {https://hwang595.github.io/gene-embedding-exposure/}
}