Can large language models understand molecules?

BMC Bioinformatics

Table 5 Results on regression tasks

Dataset	FreeSolv		Lipophilicity		ESOL
# Compounds	642		4200		1128

Models	RMSE	R\(^2\)	RMSE	R\(^2\)	RMSE	R\(^2\)
Morgan FP	0.534 ± 0.101	0.712 ± 0.101	0.817 ± 0.025	0.331 ± 0.025	0.703 ± 0.020	0.502 ± 0.020
BERT	0.425 ± 0.031	0.816 ± 0.031	0.752 ± 0.013	0.434 ± 0.013	0.382 ± 0.015	0.854 ± 0.015
ChemBERTa	0.331 ± 0.034	0.888 ± 0.034	0.716 ± 0.022	0.486 ± 0.022	0.365 ± 0.007	0.866 ± 0.007
MolFormer-XL	0.545 ± 0.047	0.690 ± 0.047	0.740 ± 0.012	0.451 ± 0.012	0.493 ± 0.027	0.754 ± 0.027
GPT	0.567 ± 0.087	0.675 ± 0.087	0.852 ± 0.010	0.273 ± 0.010	0.562 ± 0.030	0.681 ± 0.030
LLaMA	0.483 ± 0.036	0.758 ± 0.036	0.785 ± 0.015	0.382 ± 0.015	0.425 ± 0.013	0.818 ± 0.013
LLaMA2	0.422 ± 0.051	0.814 ± 0.051	0.790 ± 0.026	0.375 ± 0.026	0.420 ± 0.023	0.821 ± 0.023

The reported performance metrics are the mean and standard deviation of the RMSE and R\(^2\), calculated across the five-folds. The Best Performance is Highlighted in Bold

ISSN: 1471-2105