ICLR 2026 · Benchmark

VenusX Leaderboard

Curated baseline tables for residue-level, fragment-level, and pairwise functional tasks. Higher is better unless noted.

Overview Full interactive board Code

Numbers below are transcribed from the public VenusX board (ai4protein.github.io/venusx). For the complete interactive view across all annotation tabs, use the full board.

Residue binary · Cross-family · Active sites

Out-of-distribution (cross-family). Metric suite: AUPR / Precision / Recall / F1-Positive / Macro-F1.

Model	Type	AUPR	Precision	Recall	F1-Pos	Macro-F1
SaProt (AF_650M)	Seq-Structure	0.185	0.241	0.072	0.110	0.538
Ankh (Base)	Sequence-only	0.166	0.190	0.025	0.045	0.507
ProtSSN (k20_h512)	Seq-Structure	0.156	0.241	0.014	0.026	0.498
ESM2 (t30)	Sequence-only	0.143	0.278	0.060	0.098	0.533
ESM2 (t33)	Sequence-only	0.143	0.126	0.031	0.050	0.507
ProtBert	Sequence-only	0.131	0.131	0.020	0.035	0.501
SaProt (AF_35M)	Seq-Structure	0.114	0.132	0.036	0.056	0.510
GVP-GNN	Structure-only	0.101	0.019	0.001	0.002	0.485

Residue binary · Cross-family · Binding sites (InterPro)

Model	Type	AUPR	Precision	Recall	F1-Pos	Macro-F1
SaProt (AF_35M)	Seq-Structure	0.230	0.634	0.135	0.223	0.599
SaProt (AF_650M)	Seq-Structure	0.182	0.661	0.135	0.224	0.600
ESM2 (t33)	Sequence-only	0.159	0.581	0.108	0.181	0.579
Ankh (Base)	Sequence-only	0.145	0.437	0.086	0.144	0.559
ESM2 (t30)	Sequence-only	0.133	0.525	0.078	0.136	0.556
ProtBert	Sequence-only	0.112	0.416	0.048	0.086	0.530
ProtSSN (k20_h512)	Seq-Structure	0.095	0.379	0.029	0.053	0.514
GVP-GNN	Structure-only	0.040	0.000	0.000	0.000	0.488

Residue binary · Cross-family · Conserved sites

Model	Type	AUPR	Precision	Recall	F1-Pos	Macro-F1
Ankh (Base)	Sequence-only	0.275	0.387	0.169	0.235	0.595
SaProt (AF_650M)	Seq-Structure	0.274	0.456	0.111	0.178	0.568
SaProt (AF_35M)	Seq-Structure	0.272	0.382	0.172	0.238	0.596
ESM2 (t33)	Sequence-only	0.262	0.403	0.122	0.187	0.572
ProtBert	Sequence-only	0.243	0.482	0.009	0.017	0.489
ESM2 (t30)	Sequence-only	0.235	0.374	0.097	0.154	0.555
ProtSSN (k20_h512)	Seq-Structure	0.227	0.452	0.034	0.062	0.511
GVP-GNN	Structure-only	0.101	0.176	0.035	0.058	0.506

Residue binary · Cross-family · Functional motif

Model	Type	AUPR	Precision	Recall	F1-Pos	Macro-F1
ESM2 (t33)	Sequence-only	0.456	0.566	0.384	0.457	0.704
SaProt (AF_650M)	Seq-Structure	0.441	0.504	0.350	0.414	0.680
ESM2 (t30)	Sequence-only	0.433	0.510	0.432	0.467	0.707
SaProt (AF_35M)	Seq-Structure	0.408	0.485	0.411	0.445	0.695
Ankh (Base)	Sequence-only	0.394	0.499	0.303	0.377	0.662
ProtSSN (k20_h512)	Seq-Structure	0.390	0.390	0.365	0.412	0.678
ProtBert	Sequence-only	0.348	0.472	0.231	0.310	0.628
GVP-GNN	Structure-only	0.329	0.329	0.453	0.399	0.661

Residue binary · Cross-family · Functional domain

Model	Type	AUPR	Precision	Recall	F1-Pos	Macro-F1
SaProt (AF_650M)	Seq-Structure	0.564	0.572	0.444	0.500	0.632
SaProt (AF_35M)	Seq-Structure	0.525	0.548	0.349	0.427	0.594
ProtBert	Sequence-only	0.508	0.588	0.138	0.223	0.501
ESM2 (t33)	Sequence-only	0.506	0.530	0.367	0.433	0.593
ESM2 (t30)	Sequence-only	0.470	0.496	0.360	0.417	0.578
GVP-GNN	Structure-only	0.468	0.519	0.087	0.149	0.462
Ankh (Base)	Sequence-only	0.449	0.494	0.280	0.357	0.552
ProtSSN (k20_h512)	Seq-Structure	—	—	—	—	—

Residue binary · Mixed-family · Active sites

In-distribution (mixed-family). Strong sequence PLMs often lead here.

Model	Type	AUPR	Precision	Recall	F1-Pos	Macro-F1
Ankh (Base)	Sequence-only	0.873	0.862	0.700	0.773	0.883
ESM2 (t30)	Sequence-only	0.855	0.826	0.676	0.744	0.868
ESM2 (t33)	Sequence-only	0.852	0.845	0.682	0.755	0.874
ProtBert	Sequence-only	0.764	0.791	0.565	0.659	0.825
SaProt (AF_650M)	Seq-Structure	0.745	0.812	0.511	0.627	0.808
SaProt (AF_35M)	Seq-Structure	0.688	0.818	0.408	0.544	0.767
GVP-GNN	Structure-only	0.523	0.735	0.362	0.485	0.736
ProtSSN (k20_h512)	Seq-Structure	0.465	0.523	0.209	0.329	0.658

Fragment multi-class · MF50 · Active sites

Accuracy / Precision / Recall / Macro-F1 / MCC.

Model	Type	Accuracy	Precision	Recall	Macro-F1	MCC
SaProt (AF_650M)	Seq-Structure	0.928	0.830	0.830	0.825	0.926
GVP-GNN	Structure-only	0.907	0.826	0.833	0.822	0.906
SaProt (AF_35M)	Seq-Structure	0.928	0.810	0.823	0.807	0.926
ProtSSN (k20_h512)	Seq-Structure	0.891	0.773	0.774	0.764	0.889
Ankh (Base)	Sequence-only	0.824	0.661	0.665	0.647	0.821
ESM2 (t30)	Sequence-only	0.819	0.659	0.670	0.647	0.815
ESM2 (t33)	Sequence-only	0.814	0.603	0.634	0.605	0.810
ProtBert	Sequence-only	0.736	0.618	0.636	0.609	0.731

Pairwise similarity · F50 · Active sites

Metric: AUC (%).

Model	Type	AUC (%)
ESM-IF	Seq-Structure	96.5
Foldseek (3Di-AA)	Alignment	96.1
Foldseek (3Di)	Alignment	96.0
SaProt (AF2_35M)	Seq-Structure	95.8
TM-align	Alignment	94.6
TM-VEC	Seq-Structure	93.6
ProtT5 (xl-uniref50)	Seq-Enc-Dec	91.8
ProstT5	Seq-Structure	90.8
ProtSSN (k20_h512)	Seq-Structure	79.1
ESM2 (t30)	Sequence-only	69.4
BLAST	Alignment	52.9

Submit / reproduce

Baselines and training scripts are in ai4protein/VenusX. Datasets: Hugging Face collection. Read the project overview for task definitions and citation.

Yang Tan (谭扬)