Language models as QSPR predictors: Unleashing potential through in-context learning and instruction tuning

Abstract

Quantitative structure–property relationship (QSPR) modeling is a fundamental approach for property-oriented fuel design because it establishes quantitative relationships between molecular structures and physicochemical properties. Conventional QSPR models generally rely on manually engineered molecular descriptors, are usually developed to predict a single property, and often cannot readily incorporate newly available data, limiting their scalability and long-term applicability. Recent advances in large language models (LLMs) have demonstrated remarkable capabilities in text understanding and parsing knowledge in specialized fields such as chemistry. This study proposes FuelProp-LM, a QSPR framework based on instruction tuning and in-context learning that predicts multiple fuel properties with only the SMILES notation of a fuel molecule as input. Four open-source language models were fine-tuned using 100000 data entries obtained from the PubChem database. To enhance contextual reasoning, high-information prompts were constructed through MACCS molecular fingerprint similarity retrieval. The proposed framework was evaluated on 10 fuel physicochemical property datasets and compared with the widely-adopted LLM DeepSeek-V3.2 and existing QSPR approaches. FuelProp-LM consistently achieved competitive or superior predictive performance across downstream tasks. Quantitative analysis further revealed how instruction tuning and in-context learning improve molecular property prediction by enhancing chemically relevant attention patterns. The accuracy, generalizability, data scalability, and ease of use of this work demonstrate that FuelProp-LM can serve as a practical and flexible tool for fuel property prediction and play a significant role in future AI-assisted fuel design.