Artificial intelligence for genomic science: a scoping review of concepts, architectures, applications, and open challenges
TL;DR
This scoping review mapped how AI is defined and operationalized in genomic science, including machine learning, deep learning, graph-based methods, foundation models, and large language models, and synthesized their data modalities, applications, evaluation practices, interpretability strategies, and governance challenges.
Abstract
Introduction Artificial intelligence (AI) is becoming central to genomics and multi-omics, but its concepts, architectures, applications, evaluation standards, and translational requirements remain fragmented. This scoping review mapped how AI is defined and operationalized in genomic science, including machine learning, deep learning, graph-based methods, foundation models, and large language models, and synthesized their data modalities, applications, evaluation practices, interpretability strategies, and governance challenges. Methods We conducted a PRISMA-ScR scoping review with Joanna Briggs Institute guidance. Eligible studies applied AI to genomics or closely allied omics in research, clinical, or public health contexts. MEDLINE/PubMed, Embase, and supplementary registers were searched from January 2001 to 3 September 2025 without language restrictions. Records were screened in duplicate, and standardized items were extracted, including AI concept or method family, omics modality, task, metrics, interpretability, governance, and deployment considerations. Methodological reporting and quality were appraised using design-appropriate JBI tools and summarized descriptively as a normalized 0%–100% checklist-fulfillment index. Results From 3,785 records, 1,040 studies were included. Publication remained sparse until 2017 and then expanded steeply, with more than 90% appearing from 2018 onward. The normalized JBI checklist-fulfillment index was modest overall (mean 35.3%, SD 20.1; range 7.5%–87.5%) and was interpreted descriptively, not as a directly comparable quality score across designs. Conceptually, the field has moved from feature-engineered statistical learning toward representation learning systems modeling nucleotide sequences, regulatory context, single-cell states, multi-omics profiles, biomedical text, and clinical-genomic knowledge. Applications concentrated on variant interpretation, regulatory genomics, multi-omics integration, single-cell analysis, pathology/radiology-genomics fusion, and genomic decision support, with increasing use of deep learning, graph models, foundation models, and LLMs. Calibration, external validation, mechanistic interpretability, ancestry-aware fairness, privacy protection, and deployment models for sensitive genomic data were unevenly reported; prospective multisite evaluations were rare. Discussion AI in genomics has scaled rapidly since 2017–2018, but translation remains constrained by heterogeneous concepts, inconsistent benchmarks, incomplete reporting, and limited governance. Priorities include biologically meaningful benchmarks; calibrated uncertainty for genomic decision support; mechanism-linked interpretability; ancestry- and site-aware validation; privacy-preserving analysis of sensitive genomic data; and human oversight for variant interpretation, precision medicine, and public health genomics. Systematic Review Registration https://osf.io/uexzh.