INTERNATIONAL CONGRESS OF
PharmacoGenetics
Beyond Single-Gene Pharmacogenetics: AI-Driven Polygenic and Epigenetic Dose Prediction for Cancer Therapeutics
-
Zahra Ghanibeygi,1,*
1. Department of Biology, Faculty of Biology, Falavarjan Branch, Islamic Azad University, Isfahan, Iran
- Introduction: Conventional pharmacogenetics (PGx) based on variants of single genes (most notably the genetic polymorphisms in CYP2D6 and TPMT) have made a significant contribution toward personalizing therapy yet, in terms of its application to complicated chemotherapeutics and targeted agents, is limited when drug response or toxicity involve polygenic contributions and dynamic epigenetics. The subsequent review takes into consideration a transition from a reductionist scope of PGx, toward that of an integrated and holistic AI based perspective that integrates polygenic risk scores (PRS) with epigenomic data to develop predictive models for precision dosing of cancer therapeutics with the objectives of maximizing efficacy and minimizing the risk of life threatening adverse events.
- Methods: This study shows that AI can develop more effective polygenic risk/predictive scores compared to traditional, single-gene methodologies. We conducted a systematic literature review using peer-reviewed articles and pre prints emphasizing the years (2018-2024) from sources including PubMed, IEEE Xplore, and pre-print servers such as arXiv. We searched two distinct literatures, focusing on search terms employing only machine learning (ML) and deep learning (DL) models (e.g., random forests, support vector machines, neural networks, gradient-boosting) for any applicability to predictive pharmacokinetics, pharmacodynamics, and toxicity findings. We also considered studies that went beyond studies detailing at-a-single-gene association to construct polygenic risk scores (PRS) in polynomial risk models combining discoveries from genome-wide association studies and information on epigenomes (e.g., DNA methylation arrays, chromatin accessibility data) and phenomes (i.e., clinical features), which created multimodal AI architectures.
- Results: PGx recommendations are important in determining dose and toxicity for a wide range of cancer medications including fluoropyrimidines, irinotecan, and tyrosine kinase inhibitors. If polygenic components measured by a PRS are added to the prediction of drug metabolism phenotypes, the results are improved estimates. If epigenetic markers are also used, that is the methylation states of various genes such as DPYD and UGT1A1 promoters, we add another dynamic to predictions when we think about gene expression being regulated and not just rely on what's in the DNA sequence. We include some studies as examples of multimodal-AI that resulted in clinically actionable dose recommendations and importantly we see how this type of multimodal-AI is the way to personalize chemotherapy. Nevertheless, there are still a number of hurdles to overcome in broader clinical use, especially the interpretability of the models (black box issue), the multiplicity of data, and the requirement for sufficiently large and diverse training samples.
- Conclusion: The development of precision oncology is poised to transition from the singular era of pharmacogenetics toward a new era of artificial intelligence-based models that properly utilize polygenic and epigenetic data, pose a challenging way to develop personalized cancer therapeutic doses. This review outlines the evidence that indicates this approach can yield higher predictive power with the ability to influence clinical guidelines. Future developments must prioritize and utilize explainable approaches (XAI), facilitate international data-sharing agreements, and conduct prospective clinical trials examining the effectiveness and cost-effectiveness of these models within real world oncology contexts.
- Keywords: Artificial Intelligence, Precision Oncology, Polygenic Risk Score
