PolyLLM: polypharmacy side effect prediction via LLM-based SMILES encodings.

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in medicine

Background:

• Polypharmacy, the concurrent use of multiple drugs, is prevalent in managing complex diseases but increases the risk of drug-drug interactions (DDIs) and adverse side effects.
• Predicting these adverse effects is vital for patient safety and effective treatment strategies.

Purpose of the Study:

• To evaluate the efficacy of various Large Language Models (LLMs) in predicting polypharmacy-related side effects.
• To determine the optimal combination of LLM-derived drug representations and machine learning models for accurate side effect prediction.

Main Methods:

• Drug chemical structures were vectorized using multiple LLMs, including ChemBERTa and GPT.
• Vectorized representations of drug pairs were generated and input into Multilayer Perceptron (MLP) and Graph Neural Network (GNN) models.
• Performance was evaluated based on the accuracy of predicted side effects.

Main Results:

• Integrating Deepchem ChemBERTa embeddings with a GNN architecture significantly outperformed other tested methods in predicting polypharmacy side effects.
• LLMs utilizing only drug chemical structures proved effective for polypharmacy side effect prediction, even without incorporating protein or cell line data.

Conclusions:

• LLMs, particularly ChemBERTa embeddings coupled with GNNs, offer a powerful and efficient approach for predicting polypharmacy side effects.
• This method provides a valuable tool for clinicians, especially when detailed biological data is unavailable, enhancing drug safety.