Skip to main content

Table 9 Summary on F1 score and RMSE comparisons among position encoding/PEs using BERT

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

Task

Data

Sequence

Relative_key_query

Sinusoidal

Relative_key

Absolute

Classification (F1 score)

Malaria

SMILES

0.7439

0.6811

0.7254

0.6685

DeepSMILES

0.6532

0.6866

0.6913

0.6346

COVID

SMILES

0.7568

0.7805

0.8000

0.7733

DeepSMILES

0.7568

0.7671

0.8180

0.7733

COVID-19

SMILES

0.7718

0.7950

0.7819

0.8049

DeepSMILES

0.7179

0.7417

0.7654

0.7407

Cocrystals

SMILES

0.5538

0.6713

0.6466

0.5538

DeepSMILES

0.5755

0.7285

0.6711

0.5152

BBBP\(^{c_w}\)

SMILES

0.8092

0.8967

0.8498

0.8542

DeepSMILES

0.8456

0.7626

0.8512

0.8477

BBBP

SMILES

0.8555

0.8531

0.9061

0.8400

DeepSMILES

0.8483

0.8473

0.9196

0.8483

ClinTox

SMILES

0.9617

0.9577

0.8800

0.9580

DeepSMILES

0.9617

0.9617

0.9231

0.9617

Tox21\(^{c_w}\)

SMILES

0.9493

0.9058

–

0.8811

DeepSMILES

0.9139

0.9647

–

0.8348

Tox21

SMILES

0.9688

0.9680

0.9680

0.9672

DeepSMILES

0.9688

0.9680

0.9680

0.9672

Regression (RMSE)

ESOL

SMILES

0.6185

0.5883

0.7878

0.5983

DeepSMILES

0.6557

0.6256

0.8431

0.5584

FreeSolv

SMILES

1.8858

2.0491

2.6242

2.4169

DeepSMILES

2.1103

2.1572

2.0209

1.9840

Lipophilicity

SMILES

0.5704

0.5732

0.5716

0.6025

DeepSMILES

0.6333

0.6707

0.6857

0.6747

  1. Bold values denote the best-achieved performance for clarity and emphasis
  2. \({c_w}\) class-weighted function, DeepSMILES zero-shot learning analysis of BERT, F1 score classification tasks, RMSE regression tasks