Artificial intelligence-assisted variant interpretation: from variant scoring to evidence-centered clinical workflows

The widespread use of exome and genome sequencing has transformed rare disease diagnostics while exposing a persistent bottleneck in clinical genomics: the increasing number of variants requiring interpretation under uncertain and context-dependent evidence. This challenge is often framed as a need for stronger variant-level classifiers. In clinical practice, however, it reflects the broader need to assemble coherent evidence across molecular mechanism, gene-disease validity, penetrance, inheritance architecture, phenotype specificity, and longitudinal context. Variants of uncertain significance (VUS) are a particularly important clinical use case because they represent dynamic evidence states in which biological plausibility and clinical proof remain incomplete. Artificial intelligence (AI) may reshape this interpretive process by expanding the evidence that can be generated, prioritized, and organized. Recent advances in protein language models, long-sequence regulatory models, graph-learning approaches, phenotype-driven diagnostic systems, and multiplexed assays of variant effect are moving the field beyond isolated pathogenicity scoring. Rather than assigning definitive clinical labels, these methods generate complementary and probabilistic evidence across variant-effect prediction, penetrance estimation, gene prioritization, inheritance-architecture inference, functional evidence mining, and workflow-level decision support. Emerging agentic systems may further support evidence retrieval and workflow coordination, but their role in clinical variant classification remains preliminary and requires rigorous clinical validation before routine implementation. In this review, we organize AI-assisted variant interpretation around five areas: variant-effect prediction, penetrance estimation, gene prioritization, inheritance-architecture inference, and clinical decision support, with particular attention to VUS as a setting in which evidence remains incomplete. We focus on how computational outputs can be calibrated and incorporated into ACMG/AMP-based interpretation, and on the validation, evidence traceability, interoperability, and expert review required for clinical use.