name: 'prs-net-deep-learning-agent' description: 'Geometric deep learning-based polygenic risk score prediction using PRS-Net for modeling gene interactions, enhanced disease prediction, and cross-ancestry portability.' measurable_outcome: Execute skill workflow successfully with valid output within 15 minutes. allowed-tools:

• read_file
• run_shell_command

PRS-Net Deep Learning Agent

The PRS-Net Deep Learning Agent implements interpretable geometric deep learning for polygenic risk score prediction. PRS-Net models non-linear gene-gene interactions and biological network relationships to enhance disease prediction accuracy and improve cross-ancestry portability compared to traditional linear PRS methods.

When to Use This Skill

• When linear PRS methods show limited performance.
• For modeling complex gene-gene interactions.
• To improve PRS portability across ancestries.
• When biological interpretability of PRS is needed.
• For integrating pathway and network information.

Core Capabilities

1. Non-Linear PRS: Capture gene-gene interactions via deep learning.

2. Network Integration: Incorporate protein-protein interaction networks.

3. Interpretability: Identify important pathways and gene modules.

4. Cross-Ancestry Transfer: Improved portability via learned biology.

5. Multi-Task Learning: Joint modeling of related traits.

6. Uncertainty Quantification: Provide prediction confidence.

PRS-Net Architecture

Comparison to Traditional PRS

Workflow

1. Input: Individual genotypes, PPI network, training phenotypes.

2. Gene Summarization: Aggregate SNPs to gene-level scores.

3. Network Encoding: Learn representations on PPI graph.

4. Prediction: Non-linear disease risk prediction.

5. Interpretation: Extract important genes and pathways.

6. Cross-Ancestry: Apply to diverse populations.

7. Output: Risk scores, uncertainty, biological explanations.

Example Usage

User: "Calculate PRS-Net scores for Type 2 Diabetes with pathway-level interpretation."

Agent Action:

python3 Skills/Precision_Medicine/PRS_Net_Deep_Learning_Agent/prs_net_predict.py \
    --genotypes cohort_genotypes.vcf.gz \
    --ppi_network string_ppi.graphml \
    --trait type2_diabetes \
    --model_weights prs_net_t2d_v1.pt \
    --interpret_pathways true \
    --ancestry_calibration multi \
    --output prs_net_results/

Input Requirements

Output Components

Network Biology Integration

Performance Benchmarks

Cross-Ancestry Portability

AI/ML Components

Graph Neural Networks:

• Graph convolutional networks (GCN)
• Graph attention networks (GAT)
• Message passing neural networks

Interpretability:

• Attention visualization
• Integrated gradients
• Pathway enrichment analysis

Transfer Learning:

• Pre-training on EUR
• Fine-tuning on diverse
• Domain adaptation

Prerequisites

• Python 3.10+
• PyTorch, PyTorch Geometric
• NetworkX, igraph
• Scanpy (optional for visualization)
• GPU recommended

Related Skills

• Multi_Ancestry_PRS_Agent - Traditional multi-ancestry PRS
• PopEVE_Variant_Predictor_Agent - Variant interpretation
• Pharmacogenomics_Agent - Drug-gene interactions
• Pathway_Analysis - Pathway enrichment

Biological Interpretation

Training Considerations

Special Considerations

1. Interpretability Trade-offs: More complex = less interpretable
2. Computational Requirements: GPU accelerates training
3. Network Quality: PPI accuracy affects results
4. Gene Mapping: SNP-to-gene assignment matters
5. Overfitting: Regularization essential

Clinical Applications

Limitations

Author

AI Group - Biomedical AI Platform