Allergy, Immunology, and Medical Genetics
Medical genetics generates thousands of variants per patient from exome and genome sequencing. Determining which variants cause disease requires integrating population frequency, functional predictions, literature, and clinical phenotype. AI automates ACMG criteria application, achieving 85-95% concordance with expert manual review for clearly pathogenic or benign variants. Facial recognition AI (Face2Gene) shows 90%+ top-10 accuracy for recognizable syndromes. Pharmacogenomics clinical decision support through CPIC guidelines is well-established. But population bias in variant databases, rare disease data scarcity, and polygenic risk score equity challenges remain substantial barriers.
After reading this chapter, you will be able to:
- Evaluate AI systems for genetic variant classification and interpretation
- Understand AI applications in allergic disease phenotyping
- Assess immunodeficiency diagnostic support tools
- Navigate the role of AI in precision medicine and pharmacogenomics
- Recognize the unique challenges of rare disease AI
- Apply evidence-based frameworks for genetics and immunology AI
Part 1: Genetic Variant Interpretation AI
The Classification Challenge
Exome and genome sequencing identify thousands of variants per patient. Determining which variant(s) cause disease requires integrating:
- Population frequency (rare variants more likely pathogenic)
- Functional impact predictions (protein truncation, missense effects)
- Published literature and case reports
- Clinical phenotype match
- Segregation in family members
The ACMG/AMP guidelines (Richards et al., 2015) standardized variant classification into five categories: pathogenic, likely pathogenic, variant of uncertain significance (VUS), likely benign, and benign. AI automates criteria application and evidence integration.
AI-Assisted Classification Tools
Guidelines-based AI:
| Tool | Approach | Evidence |
|---|---|---|
| InterVar | Automated ACMG criteria application | Widely used in clinical labs |
| Franklin (Genoox) | AI-powered evidence aggregation | Commercial, EHR-integrated |
| Illumina Emedgene | ML-enhanced variant prioritization | FDA-registered |
| VarSome | Automated ACMG classification | Free tier available |
Performance: Automated ACMG classification tools show 85-95% concordance with expert manual review for clearly pathogenic or benign variants. Discordance highest for VUS classification (Li et al., 2022).
In silico predictors:
Pathogenicity prediction algorithms (CADD, REVEL, AlphaMissense) provide scores that inform the PP3/BP4 ACMG criteria (computational evidence):
- AlphaMissense (DeepMind, 2023): Trained on evolutionary conservation and protein structure
- REVEL: Ensemble of 13 prediction algorithms
- CADD: Integrates diverse annotations
Limitation: Computational predictors lack clinical validation for standalone classification. They support, not replace, comprehensive variant analysis.
ACMG/AMP Position on AI
The foundational ACMG/AMP variant classification guidelines (Richards et al., 2015) established standards that AI tools now implement:
Core principles that apply to AI:
- Classification requires integrating multiple evidence types
- Population database frequency informs but does not determine pathogenicity
- Functional studies provide strong evidence when available
- Clinical correlation remains essential
Implications for AI tools:
- AI-assisted classification should follow ACMG criteria, not proprietary algorithms
- Novel variants not in training databases require caution
- AI output requires review by certified molecular geneticists or pathologists
- Patient consent should address AI-assisted analysis
2024 Updates: Gene-specific guidelines (e.g., ATM, BRCA1/2) refine criteria application. AI tools must update as guidelines evolve. The UK ACGS released 2024 practice guidelines incorporating AI considerations.
Population Bias in Variant Databases
A critical equity issue: most variant databases overrepresent European ancestry populations.
Database composition (approximate):
| Database | European Ancestry |
|---|---|
| ClinVar | ~70% |
| gnomAD | ~55% (improving) |
| UK Biobank | ~95% |
Clinical consequence: Variants benign in non-European populations may be misclassified as pathogenic when absent from databases. Patients of African, Asian, and Hispanic ancestry face higher VUS rates.
Mitigation: gnomAD v4 expanded global representation. Population-matched databases (e.g., All of Us) are improving equity, but gaps persist.
Part 2: Facial Recognition for Genetic Syndromes
Face2Gene and DeepGestalt
Face2Gene uses the DeepGestalt deep convolutional neural network to identify syndromic patterns from facial photographs. The system compares uploaded photos against ~300 syndrome models and generates a ranked differential diagnosis.
Performance data:
| Study | Setting | Top-10 Accuracy |
|---|---|---|
| Gurovich et al., 2019 | Multi-center validation | 91% |
| Carrer et al., 2024 | Italian pediatric clinic | 98% correct in differential |
| Egyptian study, 2024 | Genetic counseling unit | 56% top-3 accuracy |
Use case: Face2Gene is most useful for: - Generating differential diagnosis for unfamiliar dysmorphic features - Supporting non-geneticists in recognizing syndromic patterns - Ultra-rare conditions that non-specialists may not consider
Current deployment: Face2Gene is used in 2,000+ clinics across 130 countries according to FDNA.
Limitations
Facial recognition AI for syndromes faces significant bias concerns:
Training data skew: Most training images are from European populations. Performance may decrease for: - African and African-American patients - East Asian patients - Middle Eastern patients - Mixed-ancestry patients
Clinical implication: Do not rely solely on Face2Gene for patients from underrepresented populations. Clinical gestalt and genetic testing remain essential.
Photo quality dependence: Performance varies with: - Lighting conditions - Image resolution - Patient age (some syndromes more recognizable in childhood) - Presence of glasses, facial hair, or other features
Appropriate Use
Face2Gene should: - Generate hypotheses, not diagnoses - Supplement, not replace, clinical dysmorphology assessment - Be used with awareness of population limitations - Lead to confirmatory genetic testing, not standalone management changes
Part 3: Pharmacogenomics Clinical Decision Support
CPIC Guidelines: The Evidence Base
The Clinical Pharmacogenetics Implementation Consortium (CPIC) provides peer-reviewed, evidence-based guidelines for translating pharmacogenomic test results into prescribing recommendations.
Scope (2024): - 34 genes covered - 164 drugs with recommendations - 28 active guidelines - 85% of published PGx implementation studies reference CPIC
Key gene-drug pairs with strong evidence:
| Gene | Drugs | Recommendation |
|---|---|---|
| CYP2C19 | Clopidogrel | Poor metabolizers: alternative antiplatelet |
| CYP2D6 | Codeine, tramadol | Poor/ultrarapid metabolizers: alternative analgesic |
| HLA-B*57:01 | Abacavir | Positive: do not prescribe |
| TPMT/NUDT15 | Azathioprine, 6-MP | IM/PM: dose reduction |
| CYP2C9/VKORC1 | Warfarin | Genotype-based initial dosing |
AI-Powered Pharmacogenomics CDS
Electronic health record integration enables automated PGx alerts:
Implementation models:
- Pre-emptive testing: Genotype patients before first prescription; CDS fires when relevant drug ordered
- Reactive testing: Test when high-risk drug ordered; results guide dosing
- Hybrid: Pre-emptive panel for common genes; reactive for specialized drugs
Performance: - CPIC API supports 80,000+ monthly queries - Epic’s foundational genomics module integrates CPIC content - Studies show improved adherence to PGx-guided prescribing with CDS
AAAAI Framework for Augmented Intelligence
The American Academy of Allergy, Asthma and Immunology published a framework for AI in the specialty (Shaker et al., 2022):
Key recommendations:
- Allergist-immunologists should be involved in design, validation, and implementation of AI tools for the specialty
- Data science and bioinformatics training should be incorporated into fellowship programs
- AI applications should be evaluated for bias before clinical deployment
- Electronic health records provide rich data sources for AI in allergy
Identified applications:
- Allergen component-resolved diagnostics
- Asthma phenotype classification
- Drug allergy de-labeling decision support
- Immunodeficiency pattern recognition
Current status: Most applications remain research-stage. The AAAAI continues to develop educational resources on AI for members.
Part 4: Allergic Disease AI Applications
Asthma Phenotyping
AI enables classification of asthma into clinically relevant subtypes:
T2-high vs. T2-low asthma: - T2-high: Eosinophilic, responds to biologics (omalizumab, mepolizumab, dupilumab) - T2-low: Neutrophilic or paucigranulocytic, limited biologic options
AI approaches: - Cluster analysis of clinical features, biomarkers, lung function - Treatment response prediction based on phenotype - Biomarker integration (FeNO, blood eosinophils, IgE)
Clinical status: Research-stage. Phenotyping tools not yet integrated into routine practice.
Food Allergy Prediction
AI models attempt to predict: - Which sensitized patients will react on oral food challenge - Severity of reactions - Development of tolerance
Inputs: - Skin prick test wheal size - Specific IgE levels and component testing - Clinical history - Age and other demographics
Performance: Research models show AUC 0.75-0.85 for predicting challenge outcomes. Not yet validated for clinical use.
Drug Allergy Assessment
AI tools for: - Penicillin allergy de-labeling risk stratification - Cross-reactivity prediction (beta-lactam, sulfonamide families) - Distinguishing IgE-mediated from non-immune reactions
Clinical relevance: 90% of patients labeled penicillin-allergic are not truly allergic. AI could identify low-risk candidates for direct challenge without testing.
Part 5: Primary Immunodeficiency AI
The Diagnostic Challenge
Primary immunodeficiencies (PIDs) are underdiagnosed: - >450 distinct genetic conditions - Many present with common infections in early stages - Delayed diagnosis: median 6+ years from symptom onset
AI Approaches
Pattern recognition from clinical data: - Infection types, frequency, pathogens - Autoimmune features - Family history patterns - Growth and development
Genetic prioritization: - Phenotype-driven gene panel selection - Variant interpretation in immune-related genes - Immunologic phenotype-genotype correlation
Jeffrey Modell Foundation Warning Signs
AI could operationalize PID warning signs: - 4+ ear infections in one year - 2+ serious sinus infections in one year - 2+ months on antibiotics with little effect - 2+ pneumonias within one year - Failure to thrive - Recurrent deep skin or organ abscesses - Persistent thrush after age 1 - Need for IV antibiotics to clear infections - 2+ deep-seated infections - Family history of PID
Status: Research-stage. Pattern recognition tools for PID not validated for clinical deployment.
Part 6: Polygenic Risk Scores
Current Evidence
Polygenic risk scores (PRS) aggregate the effects of thousands of genetic variants to estimate disease risk. The 2024 evidence base:
Cardiovascular disease: - PRS validated in multiple large cohorts - HEART study (2024) showed feasibility of clinical integration - PRS combined with traditional risk factors may improve prediction - Several commercial labs offer CAD-PRS
Breast cancer: - PRS could stratify screening intensity - Prospective trials needed before screening guideline changes - Debate continues on PRS for enhancing vs. de-intensifying screening
Implementation status: The eMERGE Network selected 10 conditions for PRS implementation based on performance, actionability, and evidence in diverse populations (Lennon et al., 2024).
Critical Limitations
Population transferability: - Most PRS developed in European populations - Predictive accuracy drops 30-50% when applied to African ancestry populations - Calibration (absolute risk estimation) often poor outside development populations
Clinical utility uncertainty: - Does knowing PRS change patient behavior or outcomes? - How should PRS integrate with established risk factors? - Threshold for “high risk” classification varies by study
Current guidance: - PRS should not be used as standalone risk assessment - Population-matched validation required before clinical use - Genetic counseling should accompany PRS results - Equity review essential: PRS may widen rather than narrow health disparities
Part 7: Rare Disease Diagnosis
The Diagnostic Odyssey
- 7,000+ rare diseases
- Average 4-7 years from symptom onset to diagnosis
- Many conditions seen by non-specialists unfamiliar with presentations
AI Solutions
Phenotype-driven prioritization: - HPO (Human Phenotype Ontology) standardizes phenotype description - Exomiser, Phenomizer: Match patient phenotypes to disease databases - LIRICAL: Likelihood ratio-based phenotype matching
Facial analysis: - Face2Gene (discussed above) - GestaltMatcher: Open-source alternative
Multi-omic integration: - AI combines genomic, transcriptomic, methylation data - Helps identify variants when WES/WGS inconclusive
Undiagnosed Diseases Programs
The NIH Undiagnosed Diseases Program and Network use systematic approaches including AI-assisted analysis. Success rates for diagnosis: 25-35% of referred cases.
AI contribution: Prioritizes variants, identifies candidate genes, suggests additional testing.
Part 8: Professional Society Positions
ACMG Position
The American College of Medical Genetics emphasizes: - AI-assisted variant classification should follow established ACMG/AMP criteria - Certified clinical laboratory geneticists/pathologists must review AI output - Population diversity in training data is essential - Patient consent should address AI-assisted analysis
AAAAI Position
The American Academy of Allergy, Asthma and Immunology: - Published framework for augmented intelligence in the specialty (2022) - Emphasizes allergist involvement in AI tool development - Supports training fellows in data science fundamentals - Issued 2024 position statement on EHR allergy documentation terminology
CPIC Implementation
CPIC provides implementation resources: - Machine-readable guidelines (JSON format) - EHR integration support - Standardized terminology for CDS - Regular guideline updates
Clinical Scenarios
Case: A 3-year-old with developmental delay undergoes exome sequencing. A variant in a developmental disorder gene is classified as VUS by the AI-assisted analysis. Parents ask if this explains their child’s condition.
Question: How should the geneticist counsel about this VUS?
Discussion
Understanding VUS classification:
- VUS means insufficient evidence to classify as pathogenic or benign
- It does NOT mean “probably causes disease”
- Many VUS are eventually reclassified as benign
Counseling approach:
- Explain that VUS cannot be used for clinical decision-making
- Describe factors that could lead to reclassification (more affected individuals, functional studies)
- Offer periodic reanalysis (1-2 years) as databases grow
- Consider additional family testing if feasible (segregation)
- Do not change management based on VUS
AI limitation: AI classified this as VUS because evidence is genuinely insufficient. The appropriate classification is uncertain, not “probably pathogenic.”
Teaching point: VUS is the most common challenging classification. AI cannot resolve genuine uncertainty. Time and additional evidence, not AI refinement, will reclassify most VUS.Case: A pediatrician sees a 2-year-old with dysmorphic features and developmental delay. They upload a photo to Face2Gene, which suggests Cornelia de Lange syndrome in the top 3 matches. The pediatrician has never seen this syndrome.
Question: What should the pediatrician do with this suggestion?
Discussion
Appropriate next steps:
- Review the syndrome features: Does the patient have cardinal features (synophrys, limb anomalies, growth restriction)?
- Refer to genetics: Face2Gene suggestions require expert evaluation
- Do not diagnose based on Face2Gene alone: The tool generates hypotheses, not diagnoses
- Genetic testing: If features are consistent, targeted gene panel or exome sequencing can confirm
Face2Gene value here:
- Helped a non-specialist consider a rare diagnosis
- Generated a hypothesis that would otherwise require dysmorphology expertise
- Appropriate use: hypothesis generation leading to referral
What not to do:
- Tell parents “the computer says your child has Cornelia de Lange syndrome”
- Change management without genetic confirmation
- Rely on Face2Gene for populations underrepresented in training data
Case: A cardiologist orders clopidogrel for a 65-year-old post-PCI. The EHR displays a pharmacogenomics alert: “Patient is CYP2C19 poor metabolizer. Per CPIC guidelines, consider alternative antiplatelet therapy.”
Question: How should the cardiologist respond?
Discussion
Clinical context:
- CYP2C19 poor metabolizers have reduced conversion of clopidogrel to active metabolite
- Higher rates of cardiovascular events on clopidogrel vs. normal metabolizers
- CPIC recommends alternative antiplatelet (prasugrel or ticagrelor) for poor metabolizers
Appropriate response:
- Review the genotype result in the EHR
- Confirm patient has no contraindications to alternatives (bleeding risk, prior ICH)
- Switch to ticagrelor or prasugrel per CPIC recommendation
- Document rationale for genotype-guided prescribing
If cardiologist disagrees:
- Document clinical reasoning for overriding alert
- Consider that poor metabolizer status may have contributed to events in patients labeled “clopidogrel non-responders”
AI/CDS value:
- Alert appeared at point of prescribing
- Linked to evidence-based CPIC guideline
- Actionable recommendation provided
Case: A 45-year-old woman asks about a direct-to-consumer (DTC) genetic test that reported her “coronary artery disease polygenic risk score is in the 90th percentile.” She is concerned about her heart disease risk.
Question: How should the physician counsel this patient?
Discussion
Understanding PRS results:
- 90th percentile means her PRS is higher than 90% of the reference population
- This does NOT mean 90% chance of heart disease
- Relative risk increase is typically 1.5-2x compared to average, not absolute risk
Counseling points:
Contextualize: Ask about the reference population. If patient is non-European and score was developed in Europeans, interpretation is limited.
Integrate with traditional factors: What is her BP, lipids, smoking status, family history? These matter more than PRS.
Actionability: What would she do differently? Standard cardiovascular prevention applies regardless of PRS.
Limitations of DTC testing:
- Validation varies by company
- Quality of genetic counseling often inadequate
- May not use most validated PRS algorithms
Recommended response:
- Perform comprehensive cardiovascular risk assessment
- Do not order additional testing based on PRS alone
- Recommend standard primary prevention based on established guidelines
- Note that elevated PRS may provide additional motivation for lifestyle modification
Key Takeaways
Genetic variant interpretation:
- AI tools consistently apply ACMG criteria; 85-95% concordance with experts for clear cases
- VUS classification remains challenging; AI cannot resolve genuine uncertainty
- Population bias in databases affects patients of non-European ancestry
- All AI classifications require certified geneticist review
Facial recognition for syndromes:
- Face2Gene achieves 90%+ top-10 accuracy for recognizable syndromes
- Useful for hypothesis generation, not diagnosis
- Performance varies across ethnic groups
- Most valuable for non-geneticists considering rare diagnoses
Pharmacogenomics CDS:
- CPIC guidelines are the evidence-based standard
- EHR-integrated CDS improves PGx-guided prescribing
- Pre-emptive genotyping enables point-of-care alerts
- 34 genes and 164 drugs have CPIC recommendations
Allergy and immunology AI:
- Most applications remain research-stage
- Asthma phenotyping and food allergy prediction are active research areas
- AAAAI framework supports allergist involvement in AI development
Polygenic risk scores:
- Cardiovascular and breast cancer PRS approaching clinical utility
- Population transferability and equity are critical concerns
- PRS should supplement, not replace, standard risk assessment
Rare disease diagnosis:
- AI accelerates phenotype-to-gene matching
- Facial analysis and HPO-based tools support diagnosis
- Definitive diagnosis still requires genetic confirmation