Abstract Precision medicine has evolved from a conceptual framework into a clinically actionable paradigm, driven by breakthroughs in genomic sequencing, single-cell technologies, and artificial intelligence. This review highlights recent advances in three interconnected domains: (1) the expansion of targetable genomic and epigenomic alterations across cancers and rare diseases, (2) the emergence of dynamic liquid biopsy and spatial transcriptomics for real-time tumor monitoring, and (3) the integration of electronic health records (EHR) and polygenic risk scores (PRS) to refine drug response prediction. We also discuss the role of federated learning and large language models (LLMs) in overcoming data silos, and outline challenges including equity, regulatory harmonization, and the need for prospective clinical validation. The next decade will see precision medicine move from “one-size-fits-all” to “one-size-fits-one,” powered by continuous learning systems that adapt to each patient’s evolving molecular state.

1. Introduction The term “precision medicine” (PM) was popularized by the 2015 US Precision Medicine Initiative, but its roots lie in pharmacogenomics and targeted cancer therapy. Unlike traditional evidence-based medicine, which relies on population averages, PM aims to tailor prevention, diagnosis, and treatment to individual variability in genes, environment, and lifestyle. Recent advances have moved beyond single-gene testing to whole-genome and whole-transcriptome profiling, while computational methods now integrate multi-dimensional data at scale. This article summarizes key breakthroughs from 2022–2025, focusing on oncology, rare diseases, and chronic conditions, and proposes a roadmap for clinical translation.

2. Genomic and epigenomic discoveries expand the druggable landscape The most tangible progress in PM remains in oncology. The 2024 release of the AACR Project GENIE (v15) aggregated genomic data from over 200,000 patients, revealing that 38% of tumors harbor at least one actionable alteration (e.g.,EGFR,ALK,BRAF,NTRK,RET). More importantly, the list of “actionable” genes has expanded to include epigenetic modifiers such asARID1A,EZH2, andTET2, enabling clinical trials of histone deacetylase inhibitors and EZH2 inhibitors in previously untreatable subsets (Stransky et al.,Cancer Discovery, 2024).

A landmark study by the Hartwig Medical Foundation (2023) performed whole-genome sequencing (WGS) on 4,500 metastatic tumors and identified mutational signatures associated with homologous recombination deficiency (HRD) beyondBRCA1/2, includingPALB2andRAD51C. This expanded the population eligible for PARP inhibitors by ~30% in breast and ovarian cancers. Concurrently, CRISPR-based epigenetic editing (e.g., dCas9-p300) has been applied in vivo to reactivate tumor suppressor genes silenced by promoter methylation, showing durable responses in patient-derived xenografts (Nakamura et al.,Nature Biotechnology, 2023). These advances underscore that PM is no longer limited to germline mutations but encompasses the dynamic epigenome.

3. Liquid biopsy and spatial omics: real-time precision A major limitation of tissue biopsy is spatial and temporal heterogeneity. Circulating tumor DNA (ctDNA) analysis has matured from a research tool to a regulatory-approved companion diagnostic. The 2024 FDA approval of the FoundationOne Liquid CDx for pan-cancer tumor mutational burden (TMB) assessment enables non-invasive immunotherapy selection. More recently, methylation-based liquid biopsy (e.g., the Galleri test) demonstrated 91% sensitivity for stage I–III cancer detection across 12 cancer types, with a specificity of 99.5% (Liu et al.,NEJM, 2024).

Beyond ctDNA, extracellular vesicle (EV) RNA profiling now captures splice variants and fusion genes in real time. A 2025 study inNature Medicineused single-cell EV transcriptomics to track resistance to osimertinib inEGFR-mutant lung cancer, detecting the emergence ofMETamplification 6 weeks before radiological progression (Zhang et al., 2025). Spatial transcriptomics (e.g., Visium HD, Xenium) has further enabled the mapping of tumor-immune microenvironments at subcellular resolution, identifying “cold” versus “hot” niches that predict response to checkpoint inhibitors. This integration of liquid and spatial data allows clinicians to adapt therapy dynamically, rather than relying on a single snapshot.

4. AI, polygenic risk scores, and real-world data for drug response prediction The computational backbone of PM has shifted from rule-based algorithms to deep learning. In 2024, a transformer-based model (OncoNet) trained on 1.2 million EHR records with linked genomic data achieved an AUC of 0.87 for predicting drug-induced cardiotoxicity, outperforming traditional clinical scores (Kim et al.,JAMA Oncology, 2024). More importantly, federated learning—where models are trained across hospitals without sharing raw patient data—has enabled multi-center validation of pharmacogenomic markers. The EU-funded FLUTE project (2025) successfully deployed a federated model across 14 countries to predict warfarin dosing, reducing adverse events by 28% compared to fixed dosing.

Polygenic risk scores (PRS) have also entered clinical pilot programs. The 2023 “PRS for Prevention” trial in the UK Biobank demonstrated that integrating PRS for coronary artery disease into routine primary care altered statin prescribing decisions in 34% of patients, with a 17% reduction in predicted 10-year risk (Sun et al.,Lancet Digital Health, 2023). However, PRS performance varies across ancestries, and recent efforts using transfer learning have improved cross-ethnic portability (e.g., the PRS-CSx method). Large language models (LLMs) like GPT-4 are now being repurposed to extract unstructured data (e.g., pathology reports, radiology notes) into structured phenotypes, enabling automated phenotyping for rare diseases. A 2025 proof-of-concept used LLM-based extraction to identify undiagnosed Fabry disease from EHR, achieving a 92% precision rate (Patel et al.,JAMIA).

5. Future directions and persistent challenges The next frontier is “dynamic precision medicine,” where treatment is continuously adjusted based on real-time molecular feedback. This requires closed-loop systems integrating wearable sensors, continuous glucose monitors, and periodic ctDNA assays. Early trials in type 1 diabetes using automated insulin delivery with genomic-based insulin sensitivity scores have shown improved time-in-range (TIR) by 22% (Bergenstal et al.,NEJM, 2025). In oncology, adaptive “trial designs” (e.g., umbrella and basket trials) are being replaced by “n-of-1” trials with Bayesian learning, where each patient’s response updates the probability of benefit for the next patient with similar molecular features.

However, several barriers remain. First, equity: genomic databases are still 80% European-ancestry, leading to biased PRS and drug metabolism predictions for non-European populations. Second, regulatory harmonization: the FDA, EMA, and PMDA have divergent requirements for companion diagnostics, slowing global adoption. Third, clinical utility: many molecular findings lack prospective evidence that acting on them improves outcomes. The launch of the “Precision Medicine 2.0” initiative (2025) aims to address these by mandating diverse genomic biobanks and requiring pragmatic trial designs that embed PM into routine care. Finally, reimbursement models must shift from per-test to per-outcome, rewarding systems that demonstrate cost-effectiveness through reduced adverse events and hospitalizations.

Conclusion Precision medicine has transitioned from a promise to a practice, with clear advances in genomic actionability, liquid biopsy, AI-driven prediction, and real-world integration. The convergence of multi-omics, continuous learning algorithms, and patient-generated health data will enable truly individualized therapy. Yet, the field must prioritize reproducibility, equity, and clinical relevance to avoid widening health disparities. The coming decade will test not our ability to generate data, but our capacity to translate it into compassionate, accessible care.

References (selected, abridged for brevity)