ICLR 2026
VenusX: Unlocking Fine-Grained Functional Understanding of Proteins
Full Paper →1Shanghai Jiao Tong University 2Shanghai Innovation Institute 3East China University of Science and Technology 4Shanghai Matwings Technology Co., Ltd.
* Equal contribution. † Corresponding authors.
Protein function is rarely a whole-molecule property. Catalysis, binding, conservation, motifs, domains, and epitopes are decided by residues and local fragments. VenusX is the first large-scale benchmark built to test whether protein models actually see those parts.
The benchmark covers three task families and six annotation types, with more than 878,000 curated samples from InterPro, BioLiP, and SAbDab. Mixed-family and cross-family splits at 50 / 70 / 90% identity separate in-distribution recall from out-of-family generalization.
Three resolutions, one question
Each VenusX task asks a model to recover function below the protein label. Residue tasks are imbalanced binary classification. Fragment tasks assign InterPro families to contiguous regions. Pair tasks score whether two proteins or fragments share a family, without training.
Residue
Mark active, binding, conserved, motif, domain, or epitope residues. Primary metric is AUPR, with precision, recall, F1-positive, and macro-F1.
Fragment
Classify functional fragments into InterPro families. Primary metrics are accuracy and macro-F1; we also report precision, recall, and MCC.
Pair
Rank same-family vs different-family pairs from embeddings or alignments. Metric is AUC (%) on F50 fragment and P50 protein settings.
Leaderboard
Numbers follow the ICLR 2026 paper. Aggregate view averages the primary metric across targets in the current split. Detail view opens every reported metric for one target. Top cells in each column are highlighted. Cross-family residue prediction remains the hardest setting: the best active-site AUPR drops from 0.873 in-distribution to 0.185 out-of-family.
Primary scores
Citation Yang Tan, Wenrui Gou, Bozitao Zhong, Huiqun Yu, Liang Hong, and Bingxin Zhou. VenusX: Unlocking Fine-Grained Functional Understanding of Proteins. In The Fourteenth International Conference on Learning Representations, 2026. https://openreview.net/forum?id=zcmL592XRG
@inproceedings{tan2026venusx,
title = {{VenusX}: Unlocking Fine-Grained Functional Understanding of Proteins},
author = {Yang Tan and Wenrui Gou and Bozitao Zhong and Huiqun Yu and Liang Hong and Bingxin Zhou},
booktitle = {The Fourteenth International Conference on Learning Representations},
year = {2026},
url = {https://openreview.net/forum?id=zcmL592XRG}
}