BioLatent | Representation Registry & Measured Benchmark

Compare representations on equal terms

BioLatent combines a registry of published results with a controlled benchmark of 18 representations across 9 public datasets.

No universal winner

The leading representation changes with the biological or chemical endpoint.

Classical methods remain competitive

Fingerprints and calculated descriptors remain strong baselines for molecular prediction.

Many rankings remain uncertain

19 of 54 eligible comparisons were distinguishable after correction.

How the benchmark works

The measured results were generated under one prespecified procedure rather than transcribed from papers. Each compatible representation was calculated once and evaluated with the same standardised linear model, regularisation search, folds and seed. The comparison method was chosen using validation data before the test set was examined.

Difference intervals use paired cluster bootstrap, p-values use paired randomisation, and eligible comparisons are adjusted together using Holm correction. Molecular tests resample Bemis–Murcko scaffolds and DeepLoc resamples MMseqs2 homology clusters. Fluorescence is descriptive because its test variants do not form independent clusters at the prespecified threshold.

Literature comparison. 11 registry records matched by representation family, endpoint and metric. All used documented scaffold splits. The measured value was higher in 7 pairs and the literature value in 4. This describes how results transfer across protocols; it does not test random-split inflation.

14/47

molecular comparisons resolved

5/5

eligible protein comparisons resolved

0/2

genomic comparisons resolved

68

measured model–dataset evaluations

Results by dataset

Choose a domain, then open a dataset to inspect its scores.

BBBP9 representations · ROC-AUC · test n = 205

balanced Murcko scaffold split (seed 42) · DeepChem MoleculeNet BBBP.csv

Dataset and statistical details

Variant: source SMILES retained; 11 RDKit-invalid rows excluded. Reference selected on held-out data: ChemBERTa-77M. Difference intervals use paired murcko scaffold cluster bootstrap (2,000 resamples); p-values use 2,000 paired randomisations with study-wide Holm correction. Molecular dependence is represented by complete Murcko scaffolds. Pretraining input-exposure proxy: ZINC 30.2%, PubChem 44.4%.

Representation	Score	95% interval
ChemBERTa-ZINC	0.9713	[0.9504, 0.9870]
MoLFormer-XL	0.9541	[0.9213, 0.9774]
ChemBERTa-77M	0.9538	[0.9173, 0.9807]
GROVER Large	0.9340	[0.8916, 0.9664]
RDKit2Dno pretraining	0.9227	[0.8756, 0.9548]
GROVER Base	0.9182	[0.8665, 0.9561]
ECFP4no pretraining	0.9141	[0.8567, 0.9525]
Uni-Mol v1	0.9090	[0.8609, 0.9457]
MolCLR GIN	0.8839	[0.8141, 0.9285]

ClinTox (FDA approval and clinical toxicity)8 representations · Macro ROC-AUC · test n = 148

balanced Murcko scaffold split (seed 42) · DeepChem MoleculeNet clintox.csv.gz

Dataset and statistical details

Variant: two-endpoint task; macro-average over FDA_APPROVED and CT_TOX; 4 RDKit-invalid rows excluded. Reference selected on held-out data: ChemBERTa-ZINC. Difference intervals use paired murcko scaffold cluster bootstrap (2,000 resamples); p-values use 2,000 paired randomisations with study-wide Holm correction. Molecular dependence is represented by complete Murcko scaffolds. Pretraining input-exposure proxy: ZINC 31.8%, PubChem 43.2%.

Representation	Score	95% interval
ChemBERTa-ZINC	0.9851	[0.9373, 1.0000]
ChemBERTa-77M	0.9751	[0.9180, 1.0000]
MoLFormer-XL	0.9718	[0.8974, 1.0000]
GROVER Large	0.8889	[0.7897, 0.9710]
GROVER Base	0.8879	[0.7994, 0.9578]
RDKit2Dno pretraining	0.8529	[0.7206, 0.9627]
ECFP4no pretraining	0.8482	[0.6797, 0.9739]
Uni-Mol v1	0.6755	[0.4670, 0.8554]

BACE9 representations · ROC-AUC · test n = 152

balanced Murcko scaffold split (seed 42) · DeepChem MoleculeNet bace.csv

Dataset and statistical details

Variant: source SMILES retained; binary Class endpoint. Reference selected on held-out data: ChemBERTa-ZINC. Difference intervals use paired murcko scaffold cluster bootstrap (2,000 resamples); p-values use 2,000 paired randomisations with study-wide Holm correction. Molecular dependence is represented by complete Murcko scaffolds. Pretraining input-exposure proxy: ZINC 1.3%, PubChem 2.6%.

Representation	Score	95% interval
ECFP4no pretraining	0.8904	[0.8130, 0.9467]
GROVER Base	0.8850	[0.8019, 0.9500]
GROVER Large	0.8817	[0.7943, 0.9493]
ChemBERTa-77M	0.8391	[0.7473, 0.9111]
RDKit2Dno pretraining	0.8315	[0.7233, 0.9182]
ChemBERTa-ZINC	0.8265	[0.7369, 0.9178]
MolCLR GIN	0.8203	[0.7287, 0.8953]
Uni-Mol v1	0.8156	[0.7239, 0.8907]
MoLFormer-XL	0.8102	[0.7174, 0.8861]

ESOL9 representations · Spearman rho · test n = 114

balanced Murcko scaffold split (seed 42) · DeepChem MoleculeNet delaney-processed.csv

Dataset and statistical details

Variant: source SMILES retained; measured log-solubility endpoint. Reference selected on held-out data: GROVER Large. Difference intervals use paired murcko scaffold cluster bootstrap (2,000 resamples); p-values use 2,000 paired randomisations with study-wide Holm correction. Molecular dependence is represented by complete Murcko scaffolds. Pretraining input-exposure proxy: ZINC 48.3%, PubChem 86.0%.

Representation	Score	95% interval
GROVER Large	0.9153	[0.8392, 0.9503]
GROVER Base	0.9128	[0.8336, 0.9499]
RDKit2Dno pretraining	0.9049	[0.8343, 0.9399]
MoLFormer-XL	0.8928	[0.7908, 0.9459]
Uni-Mol v1	0.8906	[0.8120, 0.9283]
ChemBERTa-77M	0.8469	[0.7242, 0.9181]
ChemBERTa-ZINC	0.8428	[0.7244, 0.9097]
ECFP4no pretraining	0.6161	[0.4489, 0.7321]
MolCLR GIN	0.6052	[0.4131, 0.7294]

Lipophilicity9 representations · Spearman rho · test n = 420

balanced Murcko scaffold split (seed 42) · DeepChem MoleculeNet Lipophilicity.csv

Dataset and statistical details

Variant: source SMILES retained; experimental logD endpoint. Reference selected on held-out data: GROVER Large. Difference intervals use paired murcko scaffold cluster bootstrap (2,000 resamples); p-values use 2,000 paired randomisations with study-wide Holm correction. Molecular dependence is represented by complete Murcko scaffolds. Pretraining input-exposure proxy: ZINC 22.9%, PubChem 38.8%.

Representation	Score	95% interval
GROVER Base	0.7811	[0.7105, 0.8329]
GROVER Large	0.7787	[0.7145, 0.8305]
RDKit2Dno pretraining	0.6730	[0.5835, 0.7569]
Uni-Mol v1	0.6055	[0.4985, 0.7060]
ChemBERTa-77M	0.6029	[0.5018, 0.6991]
MoLFormer-XL	0.5988	[0.4775, 0.7004]
ECFP4no pretraining	0.5868	[0.4864, 0.6674]
MolCLR GIN	0.5863	[0.4940, 0.6634]
ChemBERTa-ZINC	0.5483	[0.4449, 0.6362]

CYP3A4 Substrate9 representations · ROC-AUC · test n = 68

TDC scaffold split (seed 42) · TDC CYP3A4_Substrate_CarbonMangels

Dataset and statistical details

Variant: canonical-SMILES deduplicated before TDC scaffold splitting. Reference selected on held-out data: RDKit2D. Difference intervals use paired murcko scaffold cluster bootstrap (2,000 resamples); p-values use 2,000 paired randomisations with study-wide Holm correction. Molecular dependence is represented by complete Murcko scaffolds. Pretraining input-exposure proxy: ZINC 41.2%, PubChem 73.5%.

Representation	Score	95% interval
ChemBERTa-77M	0.7285	[0.5850, 0.8581]
RDKit2Dno pretraining	0.6873	[0.5186, 0.8256]
ECFP4no pretraining	0.6838	[0.5187, 0.8195]
MolCLR GIN	0.6702	[0.5236, 0.7861]
ChemBERTa-ZINC	0.6579	[0.5206, 0.7859]
GROVER Large	0.6579	[0.4903, 0.7875]
MoLFormer-XL	0.6518	[0.4979, 0.7871]
GROVER Base	0.6140	[0.4443, 0.7567]
Uni-Mol v1	0.5772	[0.3882, 0.7314]

Additional analyses and interpretation

How often each representation is not resolved as worse than the reference

Deliberately a count, not a ranking. The comparator is selected on held-out validation data, so the test set is not used to choose the reference it judges.

GROVER Large

6 of 6

ChemBERTa-77M

5 of 6

ChemBERTa-ZINC

5 of 6

GROVER Base

5 of 6

MoLFormer-XL

5 of 6

RDKit2Dbaseline

4 of 6

Uni-Mol v1

4 of 6

ECFP4baseline

3 of 6

MolCLR GIN

2 of 5

Fixed-test-set subsampling stability

Across 400 repeated subsets, this shows the median consistency of each reference comparison's direction. It is conditional on the observed test set and is not a causal estimate of how adding data would change a task. Molecular subsets retain whole scaffold groups and DeepLoc subsets retain whole MMseqs2 homology clusters, so their effective n can exceed the target.

Task	Test n	Direction at smaller n	Direction at larger n	Median central-range width
BBBP	205	95.8% at n ≈ 100	100.0% at n ≈ 153	0.0372
ClinTox	148	100.0% at n ≈ 100	100.0% at n ≈ 111	0.0838
BACE	152	73.0% at n ≈ 101	100.0% at n ≈ 150	0.0118
ESOL	114	99.8% at n ≈ 101	99.8% at n ≈ 101	0.0506
Lipophilicity	420	99.5% at n ≈ 100	100.0% at n ≈ 315	0.0977
CYP3A4	68	85.5% at n ≈ 51	85.5% at n ≈ 51	0.1326

Pretraining input-exposure audit

This measures input familiarity, not downstream label leakage. Molecular values are exact identity / near-duplicate / shared-scaffold fractions against random 200,000-molecule database samples, which are proxies rather than the checkpoints' exact dated training subsets. Protein values are homology proxies from MMseqs2 at ≥ 50% identity and ≥ 50% coverage against Swiss-Prot.

Task	Test n	ZINC	PubChem	Swiss-Prot	Note
BBBP	205	0.0% / 0.0% / 30.2%	0.0% / 3.4% / 44.4%	—	proxy applies to declared corpora for ChemBERTa-ZINC, MoLFormer-XL, Uni-Mol v1, GROVER Base, GROVER Large, ChemBERTa-77M, MolCLR GIN; ChEMBL: unmeasured for GROVER Base, GROVER Large
ClinTox	148	0.0% / 3.4% / 31.8%	0.0% / 4.0% / 43.2%	—	proxy applies to declared corpora for ChemBERTa-ZINC, MoLFormer-XL, Uni-Mol v1, GROVER Base, GROVER Large, ChemBERTa-77M, MolCLR GIN; ChEMBL: unmeasured for GROVER Base, GROVER Large
BACE	152	0.0% / 0.0% / 1.3%	0.0% / 0.0% / 2.6%	—	proxy applies to declared corpora for ChemBERTa-ZINC, MoLFormer-XL, Uni-Mol v1, GROVER Base, GROVER Large, ChemBERTa-77M, MolCLR GIN; ChEMBL: unmeasured for GROVER Base, GROVER Large
ESOL	114	0.0% / 0.0% / 48.3%	0.0% / 11.4% / 86.0%	—	proxy applies to declared corpora for ChemBERTa-ZINC, MoLFormer-XL, Uni-Mol v1, GROVER Base, GROVER Large, ChemBERTa-77M, MolCLR GIN; ChEMBL: unmeasured for GROVER Base, GROVER Large
Lipophilicity	420	0.0% / 0.2% / 22.9%	0.2% / 0.9% / 38.8%	—	proxy applies to declared corpora for ChemBERTa-ZINC, MoLFormer-XL, Uni-Mol v1, GROVER Base, GROVER Large, ChemBERTa-77M, MolCLR GIN; ChEMBL: unmeasured for GROVER Base, GROVER Large
CYP3A4	68	0.0% / 2.9% / 41.2%	0.0% / 2.9% / 73.5%	—	proxy applies to declared corpora for ChemBERTa-ZINC, MoLFormer-XL, Uni-Mol v1, GROVER Base, GROVER Large, ChemBERTa-77M, MolCLR GIN; ChEMBL: unmeasured for GROVER Base, GROVER Large
DeepLoc	5,963	—	—	99.6%	proxy applies to declared corpora for ESM-2 8M, ESM-2 35M, ESM-2 150M, ESM-2 650M, ProtBERT
Fluorescence	27,217	—	—	100.0%	proxy applies to declared corpora for ESM-2 8M, ESM-2 35M, ESM-2 150M, ESM-2 650M, ProtBERT
Promoters	1,571	—	—	—	human reference genome: input exposure by construction; labels not implied

How to read these numbers

“Not resolved” does not mean equal. It means this test set could not separate the model from the validation-selected reference at the stated multiplicity correction.
“Descriptive only” means no formal population claim. The score and interval describe the observed panel, but no p-value is reported when suitable independent resampling units are unavailable.
The numerical test best is descriptive. It is shown for orientation but is not selected and tested on the same outcomes.
Intervals cover the test set only, not variance from the split itself.
The MLP column is a diagnostic, never ranked. Ranking it would measure the MLP's capacity rather than the representation; the informative part is its gap to the linear model.
Values are internally comparable, not externally. Every dataset variant and split is stated beside its result. These values must not be mixed with literature values produced under different protocols.