A Systematic Review on the Generative AI Applications in Human Medical Genomics

Technical Summary: A Systematic Review on the Generative AI Applications in Human Medical Genomics

Problem Statement
Traditional statistical techniques and machine learning methods, while effective for well-defined problems with abundant labeled data, struggle with the high-dimensional complexity and long-range dependencies inherent in genomic data and unstructured medical records. The diagnostic pipeline for inherited diseases involves three interconnected stages—pre-analytical (patient preparation and data organization), analytical (variant calling and annotation), and post-analytical (clinical decision-making and reporting)—each requiring the integration of diverse data modalities. Existing reviews on AI in genetics often lack systematic rigor, focus narrowly on specific models (e.g., ChatGPT) or diseases, or fail to specifically address transformer-based architectures and Large Language Models (LLMs). There is a need for a comprehensive assessment of how LLMs and generative AI are transforming hereditary disease diagnostics and genetic education.

Methodology
The authors conducted a systematic review following PRISMA 2020 guidelines. The search strategy targeted articles from 2023 through January 2025 across four databases: PubMed, bioRxiv, medRxiv, and arXiv, utilizing automated Python scripts to access APIs. The initial search yielded 57,558 records.

The filtering process involved:

1. Automated Deduplication: Reducing the dataset to 51,613 records.
2. Semantic Filtering: Using TF-IDF scores and regular expressions to identify articles containing terms from three semantic groups: LLM-related (e.g., transformer, GPT, BERT), clinical (e.g., syndrome, phenotype), and genetics-related (e.g., NGS, variant interpretation). An exclusion list was applied to prevent false positives from structurally similar but irrelevant terms.
3. Manual Curation: Two researchers evaluated articles for relevance, removing duplicates and classifying them as irrelevant, partially relevant, or fully relevant.
4. Final Selection: From an initial pool of 304 articles (including manually added relevant studies), 172 articles were ultimately used in the review. These were categorized by application area: diagnostics (pre-analytical, analytical, post-analytical) and education.

Key Contributions and Results
The review classifies LLM applications across the diagnostic pipeline and education:

• Pre-Analytical Stage:

  • Knowledge Navigation & Literature Review: LLMs are used for Named Entity Recognition (NER) and Relation Extraction (RE) to identify gene-disease and phenotype-variant relationships from biomedical literature. Encoder-based models (e.g., BERT) dominate extraction tasks, while decoder-based models (e.g., GPT) are increasingly used for question answering and hypothesis generation. Tools like PubTator 3.0 and domain-specific Q&A systems (e.g., BioMedLM) facilitate evidence retrieval.
  • Risk Stratification: LLMs assess genetic predispositions by interpreting clinical notes and phenotypic descriptions. Models integrate multimodal data (EHRs, imaging, genetics) to predict disease trajectories and identify high-risk patients, though challenges regarding hallucinations and prompt sensitivity remain.

• Analytical Stage:

  • Medical Imaging Analysis: Vision Transformers (ViTs) and multimodal frameworks link visual features to genomic data. Applications include detecting structural abnormalities (e.g., cysts in polycystic kidney disease), predicting genetic mutations from histology or facial images (e.g., DeepGestalt extensions), and karyotyping from microscopic images.
  • Analysis of Variant Effects: Foundation models (e.g., Nucleotide Transformer, GENA-LM) learn from raw DNA sequences to predict functional consequences, such as protein structure impacts (AlphaMissense), regulatory activity, and drug response. These models bridge genotype and phenotype by modeling complex relationships, including polygenic risk scores.
  • Clinical Variant Interpretation: LLMs emulate expert reasoning to identify and classify causal variants in rare diseases. Tools like Genetic Transformer (GeneT) and MAVERICK prioritize variants based on phenotypic evidence, reducing the search space compared to traditional filtering pipelines.

• Post-Analytical Stage:

  • Patient Clustering & Subtyping: Transformers integrate imaging, textual, and longitudinal data to subtype diseases (e.g., Alzheimer’s, cancer) and identify patient subgroups with divergent survival trajectories.
  • Data & Results Aggregation: LLMs structure diagnostic outputs and predict clinical outcomes from EHR data. Generative models are beginning to support personalized treatment recommendations, with GPT-4 showing improved accuracy and guideline compliance over earlier versions in oncology contexts.
  • Clinical Report Generation & Decision Support: LLMs automate report generation and support clinical decision-making. While useful for drafting narratives and answering queries, studies highlight limitations in reliability, specificity, and the risk of hallucinations, necessitating expert oversight.

• Education:

  • LLMs serve as conversational learning systems for genetics and rare diseases.
  • Generative AI creates synthetic visual materials (e.g., facial images for syndromes) that preserve patient privacy while maintaining diagnostic training efficacy.

Significance and Discussion
The paper claims that transformer-based models significantly advance disease risk stratification, variant interpretation, medical imaging analysis, and report generation. However, it emphasizes that major challenges persist in integrating multimodal data into unified, clinically robust pipelines. Key limitations include issues with generalizability, benchmark leakage, racial and demographic biases in training data, and the "English-centric" nature of most models.

The authors argue that the effectiveness of LLMs in genetics depends not just on model scale but on task alignment, prompt structure (e.g., Retrieval-Augmented Generation), and access to real-time knowledge. They highlight that while LLMs are powerful supportive tools, human expertise remains essential for clinical decision-making. The review serves as a comprehensive classification and guide for navigating the rapidly evolving landscape of LLM applications in genetic diagnostics and education.