김영찬 (동의대학교)

장동엽 (동의대학교)

김영찬 (동의대학교)

장동엽 (동의대학교)

With the advancement of large language models(LLMs), research applying LLMs to molecular interaction prediction and experimental design has been actively conducted. However, there remains a lack of systematic and quantitative comparative studies regarding the predictive strengths and limitations of individual LLMs across different molecular properties. In this study, we propose a benchmarking methodology to quantitatively compare and analyze the predictive performance and characteristics of various LLMs by molecular property. Each model receives molecular data encoded in the simplified molecular input line entry system(SMILES) format as input and predicts various properties including solubility, partition coefficient(logP), molecular weight, and molecular formula. To quantitatively compare model performance, evaluation metrics such as mean absolute error(MAE), coefficient of determination(R²), as well as accuracy, precision, and recall within pre-defined tolerance thresholds were employed. The methology presented in this study will provide a systematic evaluation basis for training mass spectrometry-specialized LLMs.