Can large language models understand molecules?

Area of Science:

• Cheminformatics
• Artificial Intelligence
• Computational Chemistry

Background:

• Large Language Models (LLMs) are emerging as powerful tools in cheminformatics.
• Simplified Molecular Input Line Entry System (SMILES) is a crucial format for representing chemical structures.
• LLMs can decode SMILES strings into meaningful vector representations for downstream tasks.

Purpose of the Study:

• To evaluate the performance of Generative Pre-trained Transformer (GPT) and Large Language Model Meta AI (LLaMA) in generating SMILES embeddings.
• To compare LLM-based embeddings against traditional pre-trained models on SMILES.
• To assess the utility of these embeddings in molecular property and drug-drug interaction (DDI) prediction.

Main Methods:

• Investigated the performance of GPT and LLaMA models.
• Generated SMILES embeddings using both LLMs.
• Evaluated embeddings on molecular property prediction and DDI prediction tasks.
• Compared LLM-based embeddings with pre-trained SMILES models.

Main Results:

• LLaMA-generated SMILES embeddings outperformed GPT-generated embeddings in both molecular property and DDI prediction.
• LLaMA embeddings achieved performance comparable to pre-trained models in molecular prediction.
• LLaMA embeddings surpassed pre-trained models in DDI prediction tasks.

Conclusions:

• LLMs demonstrate significant potential for generating effective molecular embeddings.
• LLaMA shows particular promise for molecular representation tasks.
• Further research into LLMs for molecular embedding is warranted to bridge the gap between AI and cheminformatics.