VarDrug: Predicting Variant-Drug Interactions to Enhance Personalized Drug Safety

Adverse drug reactions (ADRs) are a common problem in personalized medicine, where genetic differences—particularly single nucleotide polymorphisms (SNPs)—can significantly impact how people respond to medications. Since some SNPs are widespread in the population, it's critical to identify any harmful drug-SNP interactions early in the drug development process. Doing so can help reduce side effects by allowing scientists to adjust a drug’s molecular structure, making it safer for patients. In our study, we focus on developing a deep learning approach to predict how genetic variants, such as SNPs, affect individual drug responses. Initially, we worked with an XGBoost model that combined Morgan fingerprints (which capture the molecular structure of drugs) and co-expression profiles of genes (to reflect gene activity). Using this model on a dataset of about ۳,۳۰۰ samples from PharmaKGB, we achieved an ۸۵% weighted F۱-score. This indicated that our approach could successfully capture important patterns. However, given the small sample size and the complexity of drug-variant interactions, we anticipated that traditional models like XGBoost might struggle to generalize to new, unseen data. To address this, we turned to a deep learning model enhanced with self-supervised learning, which allows the system to generalize better. Our model has two main parts: a Drug Encoder and a Variant Encoder. The Drug Encoder learns to predict the ۳D structure of target proteins (from PDB data) using drug structures represented by SMILES strings, helping the model understand how drugs interact with biological targets. The Variant Encoder predicts how genetic variants (e.g., SNPs) alter protein structures. This part uses information like chromosome location, reference and alternate alleles, and variant impact to make its predictions. Both encoders were pre-trained on large datasets, so the model could develop a deep understanding of drug and variant data before fine-tuning on our smaller dataset. After pre-training, we fine-tuned the model on the PharmaKGB data using a multi-layer perceptron (MLP) to predict how specific variants might affect drug efficacy or toxicity. This deep learning approach performed well, achieving a CrossEntropy loss of around ۹۹% on smaller sample sizes, suggesting it could make accurate predictions even with limited data. To make this research accessible, we developed VarDrug, an online platform where researchers can input drug structures (like SMILES) and receive predictions about which SNPs or variants are most likely to affect drug responses. The platform supports both individual patient profiles and broader population studies, helping to flag potential side effects early in the drug development process. This study highlights the importance of predicting drug-variant interactions early on. By using a deep learning-based method with self-supervised learning, we tackle the challenges of small datasets and complex biological data. With VarDrug, we hope to contribute to safer, more effective drug development and personalized medicine.

کلیدواژه ها:

Personalized Medicine ، Variant-Drug Interactions ، Pharmacogenomics ، Adverse Drug Reactions ، Deep Learning