Residency · Residency · Medical Genetics Genomics

Artificial Intelligence in Variant Interpretation and Phenotyping

Introduction

Artificial intelligence (AI) and machine learning are transforming clinical genetics by improving variant interpretation, automating phenotype recognition, accelerating diagnosis, and enabling analysis at scales impossible for human experts alone. Understanding these tools, their capabilities, and their limitations is increasingly important for clinical geneticists who will encounter AI-derived outputs in daily practice.

AI in Variant Interpretation

Deep Learning for Variant Pathogenicity Prediction

Several deep learning models have reshaped how variant pathogenicity is assessed. AlphaMissense, released by DeepMind in 2023, predicts the pathogenicity of all approximately 71 million possible human missense variants using protein structure and evolutionary information, classifying roughly 89% of missense variants as likely benign or likely pathogenic. CADD (Combined Annotation Dependent Depletion) integrates multiple annotations into a single deleteriousness score and is widely used as supporting evidence in ACMG classification. REVEL is an ensemble method combining 13 individual pathogenicity predictors, optimized for rare missense variants in Mendelian disease. PrimateAI-3D is a deep learning model trained on primate variation and three-dimensional protein structure to predict variant effects. EVE (Evolutionary model of Variant Effect) uses unsupervised learning from protein sequence alignments to predict variant effects without relying on labeled training data.

Splice Prediction

SpliceAI, developed by Illumina, is a deep neural network that predicts splice-altering effects of any variant with high accuracy, providing delta scores for acceptor and donor gain and loss. Pangolin offers context-dependent splice effect prediction that considers tissue-specific splicing patterns. These tools identify cryptic splice variants that are missed by traditional in-silico predictors.

Automated ACMG Classification

Tools such as InterVar, Varsome, and Franklin (Genoox) apply ACMG criteria to variants automatically, incorporating population frequency data, in-silico predictions, ClinVar classifications, and functional annotations. However, these tools cannot fully replicate expert judgment, and gene-specific and disease-specific nuances require human oversight. They are best used as decision support rather than autonomous classifiers, and discordance between automated tools and expert classifications remains significant for certain variant types.

AI in Phenotyping

Facial Analysis and Syndrome Recognition

Face2Gene and GestaltMatcher (developed by FDNA) use deep learning on facial photographs to suggest syndromic diagnoses. Trained on thousands of photographs of individuals with confirmed genetic diagnoses, these tools achieve top-10 accuracy of approximately 60-90% depending on the condition. They are particularly useful for conditions with a characteristic facial gestalt, such as Down syndrome, Williams syndrome, Noonan syndrome, and Cornelia de Lange syndrome. Important limitations include variable performance across ethnicities due to training bias, reduced accuracy for neonatal faces which tend to be less distinctive, and poor performance for conditions with subtle or absent facial features. Ethical concerns include facial data privacy, consent for image use, and potential for surveillance or discrimination.

Automated HPO Term Extraction

Natural language processing (NLP) tools can extract phenotype information directly from clinical notes, converting free-text clinical descriptions into standardized Human Phenotype Ontology (HPO) terms. Tools such as Doc2HPO, ClinPhen, and PhenoTagger enable phenotype-driven gene prioritization at scale without manual HPO coding, though their accuracy depends on the quality and completeness of the clinical documentation.

Phenotype-Driven Gene Prioritization

Exomiser integrates phenotype data expressed as HPO terms, variant data, and cross-species phenotype comparisons to rank candidate genes. LIRICAL uses a likelihood ratio-based approach to phenotype-driven variant prioritization, while Phevor combines phenotype information with gene ontology and pathway data. These tools significantly reduce the candidate variant list for manual review, improving diagnostic efficiency.

AI in Genomic Data Analysis

Variant Calling

DeepVariant, developed by Google, is a deep learning-based variant caller that outperforms traditional statistical methods like GATK for single nucleotide variants and small insertions and deletions. PEPPER-Margin-DeepVariant extends this approach to long-read sequencing data, while Clair3 is a deep learning variant caller optimized for Oxford Nanopore data. AI-based callers have reduced error rates and improved sensitivity, particularly in difficult genomic regions.

Structural Variant Detection

Neural network approaches such as SVcnn and DeepSV improve structural variant detection and filtering. AI methods enhance sensitivity for complex structural variants while reducing false positives, and integration of multiple callers with machine learning-based filtering is becoming standard practice.

Copy Number Variant Analysis

Machine learning approaches improve CNV calling from exome data by reducing noise and batch effects. Tools such as DECoN and ExomeDepth with ML filtering offer improved specificity for clinical CNV detection.

AI for Rare Disease Diagnosis

Integrated Diagnostic Systems

AMELIE (Automatic Mendelian Literature Evaluation) is an NLP system that ranks candidate genes by mining the medical literature for phenotype-gene associations. Fabric GEM is a commercial platform integrating phenotype, variant, and literature data for automated rare disease diagnosis. Mastermind (Genomenon) is an AI-powered genomic search engine that identifies literature supporting variant-disease associations.

Performance in Clinical Settings

AI-assisted interpretation has been shown to reduce analysis time by 50-80% in some studies, with diagnostic concordance between AI systems and expert geneticists reaching approximately 80-90% for well-characterized conditions. AI systems are particularly valuable for reanalysis of previously negative cases, enabling automated screening of updated databases. Ultra-rapid genome sequencing in NICUs relies heavily on AI-assisted analysis to achieve sub-24-hour turnaround times.

Challenges and Limitations

Bias and Fairness

AI models trained predominantly on European-descent populations perpetuate and amplify existing biases. Facial analysis tools perform better on populations well represented in training data, and variant pathogenicity predictions may be less accurate for variants not represented in training datasets, which disproportionately affects non-European populations. Mitigation strategies include diversifying training datasets, evaluating performance across populations, and transparent reporting of limitations.

Interpretability and the Black Box Problem

Deep learning models often function as black boxes, producing predictions without explaining their reasoning. Clinicians need to understand why a variant is predicted to be pathogenic, not just receive the prediction itself. Explainable AI (XAI) represents an emerging set of approaches that provide interpretable features contributing to predictions, such as which protein structural features drive a pathogenicity call. Regulatory agencies are beginning to require interpretability for AI-based medical devices.

Validation and Regulatory Standards

Clinical-grade AI tools require analytical and clinical validation before deployment. The FDA has approved or cleared several AI-based medical devices, but standards for genomic AI tools are still evolving. Continuous validation is essential because AI models must be monitored for performance drift as genomic databases evolve, and differences between research-grade and clinical-grade implementations must be recognized.

Clinical Integration

AI tools should augment, not replace, expert clinical judgment. A human-in-the-loop approach ensures that final variant classifications and clinical decisions always involve qualified human review. Training clinicians to use and critically evaluate AI outputs is essential, and liability and responsibility when AI-assisted decisions lead to adverse outcomes remain legally undefined.

Clinical Pearls

AlphaMissense and similar deep learning tools represent a step-change in variant pathogenicity prediction but should be used as supporting evidence within the ACMG framework, not as standalone classifiers. AI-powered facial analysis tools like GestaltMatcher can suggest syndromic diagnoses from photographs but perform variably across ethnicities and should be considered hypothesis-generating rather than diagnostic. Phenotype-driven prioritization tools such as Exomiser and LIRICAL significantly improve diagnostic efficiency by integrating HPO-coded phenotypes with genomic data. All AI tools in clinical genetics require transparent validation across diverse populations, explainability of predictions, and human expert oversight of final clinical decisions.

References

1. Cheng J, Novati G, Pan J, et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science. 2023;381(6664):eadg7492.
2. Hsieh TC, Bar-Haim A, Moosa S, et al. GestaltMatcher facilitates rare disease matching using facial phenotype descriptors. Nature Genetics. 2022;54(3):349-357.
3. Smedley D, Jacobsen JOB, Jager M, et al. Next-generation diagnostics and disease-gene discovery with the Exomiser. Nature Protocols. 2015;10(12):2004-2015.
4. Poplin R, Chang PC, Alexander D, et al. A universal SNP and small-indel variant caller using deep neural networks. Nature Biotechnology. 2018;36(10):983-987.

Read this lecture as Markdown