<!-- # COPYRIGHT NOTICE # This file is part of the "Universal Biomedical Skills" project. # Copyright (c) 2026 MD BABU MIA, PhD <md.babu.mia@mssm.edu> # All Rights Reserved. # # This code is proprietary and confidential. # Unauthorized copying of this file, via any medium is strictly prohibited. # # Provenance: Authenticated by MD BABU MIA -->

---

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

Component	Function	Innovation
Input Layer	Gene-level summaries	Aggregated variant effects
Network Encoder	PPI graph convolution	Biological structure
Attention Layer	Gene importance	Interpretability
Predictor	Disease/trait prediction	Non-linear mapping
Explanation	Pathway enrichment	Biological insights

Comparison to Traditional PRS

Aspect	Linear PRS	PRS-Net
Gene Interactions	Not modeled	GNN captures
Network Biology	Ignored	Integrated
Interpretability	Limited (SNP weights)	Pathway-level
Cross-Ancestry	Often poor	Improved
Computational Cost	Low	Moderate
Training Data Needed	Low	Moderate

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

Input	Format	Purpose
Genotypes	VCF/PLINK	SNP data
PPI Network	GraphML, edge list	Gene relationships
Gene Mapping	BED	SNP-to-gene
Training Labels	Phenotype file	Model training
GWAS Summary	Optional	Initialization

Output Components

Output	Description	Format
PRS-Net Score	Non-linear polygenic score	.csv
Risk Percentile	Population ranking	.csv
Gene Importance	Attention weights	.csv
Pathway Enrichment	Top pathways	.csv
Module Visualization	Network subgraphs	.png
Uncertainty	Prediction confidence	.json

Network Biology Integration

Network	Source	Genes	Edges
STRING PPI	String-db	19,000	5.5M
BioGRID	BioGRID	18,000	1.2M
Reactome	Reactome	10,000	250K
GO Biological Process	Gene Ontology	18,000	Hierarchical

Performance Benchmarks

Disease	Linear PRS AUC	PRS-Net AUC	Improvement
Type 2 Diabetes	0.65	0.72	+7%
Coronary Artery Disease	0.70	0.76	+6%
Schizophrenia	0.62	0.68	+6%
Alzheimer's Disease	0.68	0.74	+6%

Cross-Ancestry Portability

Ancestry	Linear PRS Drop	PRS-Net Drop
EUR → EAS	-15%	-8%
EUR → AFR	-30%	-18%
EUR → SAS	-20%	-12%
EUR → AMR	-18%	-10%

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

Interpretation Level	Output	Clinical Use
Gene	Top contributing genes	Target identification
Pathway	Enriched pathways	Mechanism understanding
Module	Network subgraphs	Biological insight
Hub Genes	Central genes	Druggable targets

Training Considerations

Factor	Recommendation	Rationale
Sample Size	>10,000	Deep learning needs data
Class Balance	Oversample or weight	Avoid bias
Validation	Cross-validation	Avoid overfitting
Regularization	Dropout, L2	Generalization

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

Application	PRS-Net Advantage	Benefit
Risk Stratification	Higher accuracy	Better prediction
Biological Insight	Pathway interpretation	Mechanism
Drug Targets	Hub gene identification	Therapeutic targets
Ancestry Equity	Better portability	Fairer prediction

Limitations

Limitation	Impact	Future Direction
Training Data	EUR-dominated	Diverse cohorts
Network Completeness	Missing edges	Multi-network integration
Rare Variants	Not well captured	WGS + rare variant methods
Clinical Validation	Limited trials	Prospective studies

Author

AI Group - Biomedical AI Platform

<!-- AUTHOR_SIGNATURE: 9a7f3c2e-MD-BABU-MIA-2026-MSSM-SECURE -->