A novel approach to unlocking the synergy of large language models and chemical knowledge in biomedical signal applications

This work explores the potential of using the pre-trained large language model Llama2 to address challenges in biomedical signal processing and control (BSPC), particularly in predicting the electronic and functional properties of organic molecules, an area of growing importance infields such as drug discovery and materials science. Current approaches in BSPC often rely on specialized graph neural network models, which can be limited in their ability to capture the complex relationships inherent in molecular structures. To address this, we demonstrate that a fine-tuned Llama2 model can accurately predict the highest occupied molecular orbital (HOMO) and lowest unoccupied molecular orbital (LUMO) energies of organic semiconductor molecules, with performance comparable to state-of-the-art specialized models. To further enhance the model's robustness and generalization, we propose several key innovations, including optimized simplified molecular input line entry system (SMILES) tokenization, incorporation of chemical knowledge as auxiliary supervised tasks, and a low-rank adaptation (LORA) based fine-tuning strategy. These techniques enable the language model to simultaneously learn SMILES prediction and acquire relevant chemical knowledge, while also improving its handling of incomplete structural information and ability to generalize to "unseen" molecular classes. The work also discusses the limitations of using large language models for molecular property prediction, such as the lack of interpretability and the need for improved handling of non-standard SMILES representations, highlighting the potential of this approach in BSPC while identifying areas for further improvement.