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

Table 10 Averaged performance of fine-tuned BERT using K-fold cross-validation on newly proposed datasets

PE	Data	Sequence	Aver. loss	Aver. accuracy	Aver. precision	Aver. recall	Aver. F1-score
Absolute	Malaria	SMILES	0.5364 ± 0.01	0.7332 ± 0.01	0.7227 ± 0.04	0.6203 ± 0.01	0.6672 ± 0.02
	Malaria	DeepSMILES	0.5662 ± 0.01	0.7130 ± 0.02	0.7189 ± 0.02	0.5510 ± 0.02	0.6234 ± 0.02
	COVID	SMILES	0.5640 ± 0.02	0.7973 ± 0.04	0.8472 ± 0.06	0.7530 ± 0.06	0.7959 ± 0.05
	COVID	DeepSMILES	0.5316 ± 0.01	0.7716 ± 0.04	0.8010 ± 0.04	0.7560 ± 0.05	0.7770 ± 0.04
	Cocrystals	SMILES	0.5795 ± 0.01	0.6953 ± 0.02	0.6888 ± 0.03	0.6286v0.04	0.6560 ± 0.02
	Cocrystals	DeepSMILES	0.5723 ± 0.02	0.6807 ± 0.03	0.6692 ± 0.04	0.6188 ± 0.05	0.6418 ± 0.03
RKQ	Malaria	SMILES	0.5025 ± 0.01	0.7616 ± 0.01	0.7542 ± 0.02	0.6647 ± 0.02	0.7062 ± 0.01
	Malaria	DeepSMILES	0.5546 ± 0.01	0.7263 ± 0.01	0.7669 ± 0.04	0.5288 ± 0.04	0.6244 ± 0.02
	COVID	SMILES	0.4777 ± 0.02	0.8432 ± 0.02	0.9079 ± 0.03	0.7818 ± 0.03	0.8398 ± 0.03
	COVID	DeepSMILES	0.4408 ± 0.04	0.8068 ± 0.03	0.8752 ± 0.01	0.7385 ± 0.06	0.8002 ± 0.04
	Cocrystals	SMILES	0.5247 ± 0.03	0.7197 ± 0.02	0.7127 ± 0.02	0.6633 ± 0.03	0.6867 ± 0.02
	Cocrystals	DeepSMILES	0.5154 ± 0.03	0.7416 ± 0.02	0.7323 ± 0.03	0.7041 ± 0.07	0.7151 ± 0.03

Bold values denote the best-achieved performance for clarity and emphasis
Aver. Averaged, RKQ Relative_key_query, DeepSMILES zero-shot learning analysis of BERT

ISSN: 1758-2946