Skip to main content

Table 12 Performance of fine-tuned BERT on polaris benchmark datasets

From: Positional embeddings and zero-shot learning using BERT for molecular-property prediction

Sequence

\({\text {Data}}^{\text{cls}}\)

PE

Loss

Accuracy

Precision

Recall

F1-score

SMILES

BBBP

Absolute

0.5189

0.8054

0.8074

0.9970

0.8922

RKQ

0.4732

0.8079

0.8079

1.0000

0.8937

Tox21

Absolute

0.3327

0.8952

0.8952

1.0000

0.9447

RKQ

0.3111

0.8952

0.8952

1.0000

0.9447

Sequence

\({\text {Data}}^{\text{reg}}\)

PE

Loss

RMSE

   

SMILES

Lipophilicity

Absolute

0.4573

0.6763

   

RKQ

0.4297

0.6555

   
  1. RKQ Relative_key_query, cls Classification, reg Regression
Back to article page

Journal of Cheminformatics

ISSN: 1758-2946

Contact us