GlycoConjVacEpitopeMapper

Computational Strategy for Epitope Mapping of Glycoconjugate Vaccines Using Deep Learning & Virtual Lab Agents

Project Lead: Sandeep Thanna

---

1. Overview

This repository implements a multi-agent research framework for identifying protective conformational epitopes on glycoconjugate vaccines. It bridges high-fidelity structural biology with deep learning through a tiered "Virtual Lab" architecture.

It combines:

• Agent Schools: Structured training of specialized agents (Immuno, Chem, ML, Bioinfo) using RAG.
• Virtual Lab: Multi-agent orchestration for pipeline design and automated scientific execution.

---

2. Research Workflow (The "Swanson" Architecture)

The project follows a two-phase conceptual cycle inspired by the Virtual Lab architecture (Swanson et al. 2024):

2.1 Agent School Phase

Before entering the lab, agents undergo structured mastery:

1. Curriculum Design: Define mastery levels for each domain (Glyco, Chem, ML, Bio).
2. Autonomous Study: Agents retrieve and process literature (PubMed/FAISS) to build localized RAG indexes.
3. Integration: Knowledge is synthesized into "Knowledge Updates" and stored in vector indices.
4. Certification: Agents are assessed by a "Scientific Critic" to ensure conceptual and practical competence.

2.2 Virtual Lab Phase

Certified agents collaborate in a simulated research environment:

1. Planning: Defining vaccine constructs (e.g., MenA + CRM197) and auditing structural feasibility.
2. Pipeline Design: Selecting toolsets (AlphaFold, Rosetta, GNN) for tiered verification.
3. Implementation: Coding the structural modeling, simulation, and feature pipelines.
4. Validation: Ranking candidate epitopes and cross-referencing with IEDB experimental data.

---

3. Methodology & Tiers

The implementation translates the workflow into a multi-tiered validation approach, prioritizing AlphaFold 3 for structural modeling of protein-glycan complexes:

• Tier 1 (Biological Screening): Solvent exposure (SASA) and T-cell epitope (PPZ) masking.
• Tier 2 (Physical Simulation): High-resolution modeling via AlphaFold 3 Server followed by Real Molecular Dynamics (OpenMM) using the Amber14SB (protein) and GLYCAM-06j (glycan) force fields. We perform explicitly solvated simulations to quantify "Phosphate Cloud" shielding via RMSF.
• Tier 3 (ML Modeling): SE(3)-Transformer/GNN using ESM-2 sequence embeddings and structural tensors.

---

4. Educational Resources

We believe complex science should be accessible to everyone.

• Student Reproduction Guide: A simplified, step-by-step guide for middle/high school students to build their own "Vaccine Design Computer."

---

4. Repository Structure

• src/virtual_lab: Core Agent logic and multi-turn meeting orchestration.
• src/agent_schools: Curricula, retrievers, and school orchestration scripts.
• src/virtual_lab/bioinformatics: Screening tools (Bio.PDB).
• src/virtual_lab/chemistry: Simulation engines and chemical protocols (Rosetta/Amber).
• src/virtual_lab/epitope_mapping: ESM-2 feature extraction and GNN training loops.
• dashboard/: React-based visual analysis command center.
• data/knowledge_base: FAISS vector stores for specialist agents.

---

5. Installation & Environment Setup

To ensure all physical simulation (OpenMM) and machine learning (ESM-2) dependencies are correctly linked, use the provided environment definition:

1. Set up Conda Environment:

   # Create the virtual-lab environment
   conda env create -f environment.yml
   
   # Activate the environment
   conda activate virtual-lab

2. Verify Setup:

   python -c "import openmm; import mdtraj; print('Real MD Support: OK')"

3. Configuration: Copy .env.example to .env and fill in your API keys:

   cp .env.example .env
   # Edit .env with your actual API keys

4. Agent Schools Orchestration: Run the full academic cycle to train and certify your agents:

   python src/agent_schools/run_school.py

---

6. Pipeline Execution Guide

Run the following commands in order for a full end-to-end simulation:

A. Initialization

Download the target structure (CRM197) and finalize planning:

python get_pdb_4ae1.py
python src/virtual_lab/main_epitope_mapping.py

B. Tier 1: Screening

Perform biological masking and SASA calculations:

python src/virtual_lab/bioinformatics/crm197_screening.py

C. Tier 2 & 3: Simulation & Training

Generate high-fidelity fluctuational data and train the GNN predictor:

python src/virtual_lab/chemistry/run_amber_simulation.py
python src/virtual_lab/epitope_mapping/extract_md_features.py
python src/virtual_lab/epitope_mapping/train_esm2_gnn.py

D. Analysis & Dashboard

Visualize all findings in the interactive dashboard:

python src/virtual_lab/analysis/export_dashboard_data.py
cd dashboard
npm install
npm run dev

---

7. Success Metrics

• IEDB Recovery: Alignment with documented protective epitopes.
• T-cell Preservation: Zero-overlap with protected presentation zones.
• Phosphate Shielding: Correlation between glycan flexibility and model accessibility indices.

---

AI Assisted Discovery Blueprint (PDF)

9. Collaboration & Future Work

This project is actively maintained by Sandeep Thanna. I am open to collaboration on:

• Advanced Graph Neural Networks (SE3-Transformers) extensions for glycan-protein interfaces.
• High-throughput Molecular Dynamics optimizations for vaccine candidate screening.

Please reach out to discuss integration or potential research partnerships.

10. License

MIT License