PopEVE Variant Predictor Agent

The PopEVE Variant Predictor Agent leverages the PopEVE deep learning model from Harvard Medical School to predict pathogenicity of genetic variants. PopEVE analyzes evolutionary conservation, protein structure, and population frequency to identify disease-causing variants, having identified over 100 previously unrecognized variants responsible for undiagnosed rare genetic diseases.

When to Use This Skill

• When predicting pathogenicity of missense variants genome-wide.
• For rare disease diagnosis with variants of uncertain significance (VUS).
• To prioritize candidate variants in exome/genome sequencing.
• When interpreting novel variants not in ClinVar or literature.
• For population-stratified variant interpretation.

Core Capabilities

1. Pathogenicity Prediction: Score any missense variant for disease likelihood.

1. VUS Resolution: Reclassify variants of uncertain significance.

1. Rare Disease Diagnosis: Identify causal variants in undiagnosed patients.

1. Population-Aware Scoring: Account for ancestry-specific variant frequencies.

1. Protein Context Analysis: Integrate structural and functional domains.

1. Batch Variant Scoring: Process thousands of variants efficiently.

Model Architecture

Scoring Thresholds

Workflow

1. Input: VCF file, gene list, or individual variants.

1. Annotation: Map variants to transcripts and proteins.

1. Feature Extraction: Compute evolutionary, structural, population features.

1. Prediction: Run PopEVE model for pathogenicity scores.

1. Population Adjustment: Apply ancestry-specific calibration.

1. Ranking: Prioritize variants by score and gene relevance.

1. Output: Scored variants with interpretations.

Example Usage

User: "Score all missense variants from this rare disease patient's exome to identify potential causal variants."

Agent Action:

python3 Skills/Genomics/PopEVE_Variant_Predictor_Agent/popeve_predict.py \
    --vcf patient_exome.vcf \
    --genome GRCh38 \
    --ancestry EUR \
    --gene_panel rare_disease_genes.txt \
    --min_score 0.5 \
    --output pathogenicity_scores.tsv

Input Formats

Output Components

Comparison with Other Tools

AI/ML Components

Deep Learning Architecture:

• Transformer for sequence context
• Graph neural network for structure
• Variational autoencoder for evolution
• Gradient boosting for integration

Training Strategy:

• Semi-supervised with ClinVar labels
• Evolutionary likelihood (unsupervised)
• Population frequency calibration
• Cross-validation across genes

Population Modeling:

• Ancestry-specific allele frequencies
• Selection coefficient estimation
• Demographic history modeling

Performance Metrics

Prerequisites

• Python 3.10+
• PyTorch, transformers
• PopEVE model weights
• Reference genome (GRCh37/38)
• VEP or similar annotator
• gnomAD database access

Related Skills

• AlphaMissense_Agent - For AlphaMissense predictions
• DiagAI_Agent - For clinical diagnosis support
• ACMG_Classifier_Agent - For ACMG classification
• Pharmacogenomics_Agent - For drug-gene variants

Disease Categories

Special Considerations

1. Gene Coverage: Best for well-conserved genes with orthologs
2. Protein-Coding Only: Missense variants in coding regions
3. Novel Genes: Lower confidence for poorly characterized genes
4. Ancestry Calibration: Use appropriate population reference
5. Ensemble Approach: Combine with AlphaMissense for best results

Clinical Integration

Author

AI Group - Biomedical AI Platform