title	ProtEngine Labs
emoji	🧬
colorFrom	indigo
colorTo	blue
sdk	docker
app_port	7860

ProtEngine Labs

Precision Medicine & ProtEngine Labs Pipeline A 22-agent AI orchestrator for novel lead discovery, synthesis planning, and clinical validation.

---

Overview

ProtEngine Labs is an enterprise-grade drug discovery pipeline designed to identify, optimize, and validate therapeutic candidates for specific protein mutations (e.g., EGFR T790M).

The system leverages a hierarchical 22-agent architecture to move from mutation parsing to synthesis-ready lead compounds in as little as 90 seconds (standard) to 6 hours (with full Molecular Dynamics validation).

System Architecture

graph TD
    %% Global Styling
    classDef acquisition fill:#eff6ff,stroke:#1d4ed8,stroke-width:2px,color:#1e3a8a
    classDef analysis fill:#fefce8,stroke:#a16207,stroke-width:2px,color:#713f12
    classDef design fill:#fff7ed,stroke:#c2410c,stroke-width:2px,color:#7c2d12
    classDef validation fill:#fef2f2,stroke:#b91c1c,stroke-width:2px,color:#7f1d1d
    classDef context fill:#f5f3ff,stroke:#6d28d9,stroke-width:2px,color:#4c1d95
    classDef output fill:#f0fdf4,stroke:#15803d,stroke-width:2px,color:#14532d
    classDef start_end fill:#f8fafc,stroke:#334155,stroke-width:2px,color:#0f172a

    Start([Mutation Query]):::start_end --> Data[Stage 1: Multi-Source Data Acquisition]:::acquisition
    Data --> Struct[Stage 2: Structural & Variant Analysis]:::analysis
    Struct --> Design[Stage 3: Generative Molecule Design]:::design
    Design --> Docking[Stage 4: Docking & Selectivity Filtering]:::design
    Docking --> Validation[Stage 5: GNN & Molecular Dynamics Validation]:::validation
    Validation --> Synthesis[Stage 6: ASKCOS Retrosynthesis Planning]:::output
    Synthesis --> Report([Final Enterprise Drug Report]):::start_end

Loading

Key Metrics & Precision

• Affinity Scoring: Multi-method validation using Vina, Gnina CNN, DimeNet++ GNN, and MM-GBSA.
• Confidence Tiers: Grounded in pLDDT and ESM-1v scores (WELL_KNOWN, PARTIAL, NOVEL).
• Stability Labels: Ranked by RMSD trajectories from 50ns MD simulations.
• Selectivity: Dual-docking against 10+ off-target proteins to ensure therapeutic windows.

---

Technical Differentiation

Intelligent Lead Filtration (The Funnel)

The pipeline employs a sophisticated filtration funnel to optimize computational resources while maintaining maximum empirical rigor:

1. Scaffold Hopping: Generates ~150 candidates using RDKit-driven bioisostere replacements and 3D diffusion.
2. GNN Pre-Selection: Utilizes a DimeNet++ GNN to filter the top 30 leads down to exactly two high-confidence finalists.
3. Molecular Dynamics: Performs 50ns OpenMM simulations on finalists only, calculating precise MM-GBSA binding free energies (ΔG) and RMSD stability.

Selectivity & Toxicity Benchmarking

ProtEngine Labs dual-docks top leads against a panel of 10+ off-target proteins. This computes a Selectivity Ratio (Target Affinity / Off-target Affinity), identifying potential side-effects and ensuring a wide therapeutic window early in the discovery phase.

Retrosynthetic Feasibility

Every discovered lead is validated through ASKCOS retrosynthesis planning. Candidates are scored by Synthetic Accessibility (SA) and mapped to specific reagents and steps, ensuring that the top leads are practical for experimental synthesis.

---

The 22-Agent Architecture

The pipeline is organized into 10 specialized stages, each managed by autonomous agents.

Stage	Mission	Agents
1	Data Acquisition	MutationParser, Planner, FetchAgents
2-3	Structure & Variant	StructurePrep (ESMFold), VariantEffect (ESM-1v), PocketDetection
4-5	Design & Docking	MoleculeGen (Pocket2Mol), Docking (Gnina/Vina), Selectivity, ADMET
6-7	Ranking & Validation	GNNAffinity (DimeNet++), MDValidation (OpenMM), Resistance Forecasting
8-9	Context Analysis	SimilaritySearch, SynergyAgent, ClinicalTrialAgent
10	Output Generation	SynthesisAgent (ASKCOS), ReportAgent

---

Quick Start (Enterprise Setup)

The project includes a comprehensive automation script that handles system-level dependencies including Miniconda, Python 3.11, AutoDock Vina, and Node.js.

1. One-Command Setup & Launch

./start.sh

This script automatically detects missing dependencies, installs Miniconda if needed, creates the environment, and launches both frontend and backend.

2. Manual Prerequisites

• OS: Linux (preferred) or macOS.
• Conda: Miniconda3 recommended.
• Node.js: v20 or later.
• Bio-Tools: AutoDock Vina, fpocket, Open Babel (automated via setup script).

---

Project Architecture

protengine/
├── backend/            # FastAPI + LangGraph Orchestrator
│   ├── agents/         # 22 Pipeline Agents
│   ├── pipeline/       # LangGraph state machine
│   ├── tests/          # Stress & Validation suites
│   └── data/           # Cache & Structure storage
├── frontend/           # Next.js 16 + Tailwind v4 + GSAP
│   ├── app/            # Feature-first routing
│   ├── components/     # Analysis, 3D Mol, & GNN Visuals
│   └── lib/            # SSE streaming & API hooks
├── tools/              # (Auto-generated) Local Miniconda & Binaries
└── docs/               # Technical specifications & Architecture

---

Environment Configuration

Backend (backend/.env)

Variable	Required	Purpose
OPENAI_API_KEY	Yes	Primary reasoning (GPT-4o)
GROQ_API_KEY	Recommended	Fast Llama 3.3 orchestration
DATABASE_URL	Optional	Persistent discovery library (Neon/Postgres)

---

Enterprise Standards

• Observability: Fully integrated with LangSmith for real-time agent tracing and audit logs.
• Predictive Reliability: All scores include uncertainty ranges (e.g., -9.1 ± 1.2 kcal/mol).
• Safety Protocols: No clinical claims. Predictive outputs only. Mandatory disclaimers on all reports.
• Scalability: Async SSE streaming for real-time pipeline progress updates.

---

License

Apache License 2.0 — Computational Predictions Only. Experimental synthesis and binding validation required before biological testing.

Notice: Not for clinical use.