DruGNNosis-MoA: Elucidating Drug Mechanisms as Etiological or Palliative With Graph Neural Networks Employing a Large Language Model

Understanding the complex mechanisms of drugs' therapeutic effects is essential for advancing precision medicine and optimizing treatment strategies. However, systematically distinguishing between drugs that address root causes (etiological mechanisms) and those that alleviate symptoms (palliative mechanisms) using modern Artificial Intelligence (AI)-based strategies remains underexplored. We present a novel computational framework for classifying drug Mechanisms of Action (MoA) as etiological or palliative, comparing three approaches: (i) Fine-tuning Science Bidirectional Encoder Representations from Transformers (SciBERT) with drug descriptions; (ii) Training various Graph Neural Networks (GNNs) on a constructed heterogeneous network of drugs, genes, and diseases; and (iii) Developing DruGNNosis-MoA, which integrates GNN with our fine-tuned SciBERT embeddings as node features. DruGNNosis-MoA excelled (F1-score 0.94) at identifying drug MoA. DruGNNosis-MoA characterizes drug mechanisms for subsequent pharmacological studies, thereby advancing precision medicine and therapeutic development.