CatalystAI — AI-Powered Molecular Discovery Platform

AI FOR BHARAT 2026

An AI-powered platform for de novo catalyst and enzyme discovery — replacing trial-and-error with a continuous, self-improving Lab-in-the-Loop.

---

🎯 Problem Statement

The global push toward sustainable aviation fuel and bio-based chemicals is fundamentally constrained by a discovery problem:

• Design space: >10^60 potential molecules
• Lab throughput: ~hundreds per month
• Time to discovery: Years of manual screening
• GPS Renewables' need: Optimize 2G Ethanol-to-SAF pathway catalysts

CatalystAI compresses discovery from years to weeks using generative AI + Bayesian active learning.

---

🚀 Core Innovation: The Lab-in-the-Loop

┌─────────────┐     ┌──────────────┐     ┌─────────────┐     ┌──────────────┐
│  Generate   │ ──▶ │   Predict    │ ──▶ │  Rank (EI)  │ ──▶ │  Synthesize  │
│  (DiffCSP)  │     │  (GNN/FBA)   │     │  (Bayesian) │     │  (Wet Lab)   │
└─────────────┘     └──────────────┘     └─────────────┘     └──────────────┘
                                                                       │
┌─────────────┐     ┌──────────────┐     ┌─────────────┐            │
│   Update    │ ◀── │   Validate   │ ◀── │   Measure   │ ◀──────────┘
│ Model Weights│     │ (Human-in-  │     │   Results   │
│             │     │   the-Loop)  │     │             │
└─────────────┘     └──────────────┘     └─────────────┘

---

🧪 Comprehensive Feature Set

Direction 1: Chemical Catalysis

1. Generative Design (Mocked for MVP)

• DiffCSP for novel crystal lattice generation
• Conditional cDVAE for material structure design
• Novel IP generation beyond database retrieval

2. RAG-Based Literature Retrieval ✅ NEW

• Vector database semantic search (Milvus-style)
• Extract baseline performance from scientific literature
• Knowledge-grounded candidate generation
• Literature context for explainability

3. Predictive Ranking

• GNN surrogate models (trained on Open Catalyst Project data)
• Activity, selectivity, and stability prediction
• Epistemic uncertainty quantification via deep ensembles

4. Advanced Bayesian Optimization ✅ ENHANCED

• 5 Acquisition Functions:
  • Expected Improvement (EI) — Default, balanced
  • Probability of Improvement (PI) — Conservative
  • Upper Confidence Bound (UCB) — Optimistic exploration
  • EI per Cost — Resource-aware optimization
  • Knowledge Gradient (KG) — Finite-horizon optimal
• Multi-objective Pareto front optimization
• Side-by-side strategy comparison

Direction 2: Synthetic Biology

5. Metabolic Flux Balance Analysis ✅ NEW

• Genome-scale metabolic modeling (COBRApy-style)
• Lignocellulosic biomass → ethanol pathway simulation
• Bottleneck identification (cellulase, xylanase, etc.)
• Enzyme impact prediction
• Optimal enzyme cocktail design

6. Protein Engineering Pipeline (Simulated)

• RFdiffusion backbone generation
• ProteinMPNN sequence design
• ESM3 zero-shot fitness prediction

Safety & Compliance

7. Biosecurity TEVV Layer ✅ NEW

• Mandatory screening against:
  • CDC Select Agents Registry
  • Known toxin databases (diphtheria, botulinum, ricin, anthrax)
  • Virulence factor motifs
• Automatic blocking of restricted sequences
• Compliance reporting (NIH Guidelines, DURC, NSABB)
• Audit logging for regulatory documentation

8. Active Learning Feedback Loop

• ELN webhook integration (eLabFTW/Labtools.AI)
• Human-in-the-loop validation
• Transfer learning retraining simulation
• Model versioning with provenance tracking

---

💻 System Architecture

Backend (FastAPI)

backend/
├── app/
│   ├── main.py                    # API endpoints
│   ├── schemas.py                 # Pydantic models
│   └── services/
│       ├── bayesian_service.py    # Expected Improvement
│       ├── acquisition_service.py  # Multi-acquisition functions 
│       ├── rag_service.py          # Literature retrieval 
│       ├── fba_service.py          # Metabolic modeling 
│       ├── biosecurity_service.py  # TEVV screening
│       └── mock_db.py             # In-memory storage
└── requirements.txt

Frontend (Streamlit → Next.js)

streamlit_app/
└── app.py                         # 8-tab comprehensive dashboard
    ├── Generate & Rank
    ├── Pareto Front
    ├── Why EI? (Comparison)
    ├── Multi-Acquisition Comparison  
    ├── ELN Feedback Loop
    ├── RAG Literature Retrieval      
    ├── Metabolic FBA (Enzyme Track)  
    └── Biosecurity TEVV              

Visualizations

• 3D Pareto Front (Plotly) — Activity × Selectivity × Stability
• Molstar Molecular Viewer — WebGL crystal lattices
• Metabolic Flux Diagrams — Cytoscape.js network graphs (planned)
• Acquisition Function Heatmaps — Exploration vs exploitation trade-offs

---

🏗️ Technology Stack

Domain	Technology	Justification
Backend API	FastAPI	Async performance; native ML compatibility
Deep Learning	PyTorch + PyTorch Geometric	Industry standard for GNNs
Vector DB	Milvus (simulated)	Sub-millisecond semantic search
Metabolic Modeling	COBRApy	SBML parsing; Python native
Cheminformatics	RDKit	Valency validation; SMILES parsing
Frontend (MVP)	Streamlit	Rapid scientific prototyping
Frontend (Production)	Next.js	Scalable multi-tenant SaaS
Molecular Viz	Molstar (React)	WebGL-accelerated; PDB-grade
ELN Integration	eLabFTW API	Open-source; webhook support

---

🚀 Quick Start

Prerequisites

• Docker & Docker Compose
• Python 3.10+ (if running locally)

Run with Docker (Recommended)

# Clone repository
cd CatalystAI

# Start services
docker-compose up --build

# Access endpoints
# - Streamlit Dashboard: http://localhost:8501
# - FastAPI Docs: http://localhost:8000/docs

Run Locally (Development)

# Backend
cd backend
pip install -r requirements.txt
uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload

# Frontend (separate terminal)
cd streamlit_app
pip install -r requirements.txt
streamlit run app.py --server.port=8501

Deploy on Render

The repository includes render.yaml for a two-service Render Blueprint:

• catalystai-backend: FastAPI service from backend/
• catalystai-streamlit: Streamlit service from streamlit_app/

Create a new Render Blueprint from the repo root and Render will use the correct subdirectory requirements files, start commands, Python version, and health checks. If you create the services manually instead, use:

# Backend
Root Directory: backend
Build Command: pip install --upgrade pip && pip install -r requirements.txt
Start Command: uvicorn app.main:app --host 0.0.0.0 --port $PORT
Health Check Path: /health

# Streamlit
Root Directory: streamlit_app
Build Command: pip install --upgrade pip && pip install -r requirements.txt
Start Command: streamlit run app.py --server.address 0.0.0.0 --server.port $PORT --server.headless true
Health Check Path: /_stcore/health

For a manual Streamlit service, set API_URL to the backend URL. For Blueprint deploys, API_HOSTPORT is wired automatically through Render's private network.

---

📊 API Endpoints

Core Discovery

• POST /api/sessions — Create new discovery session
• POST /api/generate — Generate catalyst/enzyme candidates
• POST /api/rank — Rank by Expected Improvement
• POST /api/rank-by-score — Rank by predicted activity (baseline)
• POST /api/webhook/mock-eln — Log experimental results

RAG Literature (NEW)

• GET /api/literature/search — Semantic literature search
• GET /api/literature/baseline — Extract baseline performance

Metabolic FBA (NEW)

• POST /api/fba/simulate — Run flux balance analysis
• POST /api/fba/enzyme-impact — Predict enzyme engineering impact
• GET /api/fba/design-cocktail — Design optimal enzyme cocktail

Biosecurity (NEW)

• POST /api/biosecurity/screen — Screen single sequence
• POST /api/biosecurity/batch-screen — Batch screen candidates
• GET /api/biosecurity/compliance-report — TEVV compliance report

Advanced Optimization (NEW)

• POST /api/rank/acquisition — Rank with custom acquisition function
• GET /api/rank/compare-acquisition — Compare multiple strategies

---

🎯 Key Differentiators for Finals

1. Breadth: Both Directions Covered

• Direction 1 (Chemical Catalysis): Generative design, GNN prediction, Bayesian optimization
• Direction 2 (Synthetic Biology): FBA, enzyme engineering, metabolic pathway design

2. Depth: Technical Sophistication

• 5 acquisition functions (not just EI)
• Uncertainty quantification with epistemic variance
• Multi-objective Pareto optimization
• Literature-grounded generation (RAG)

3. Responsibility: Safety-First

• Mandatory biosecurity screening (TEVV)
• Toxin/pathogen homology detection
• Regulatory compliance documentation
• Audit trail for restricted agents

4. Scientific Rigor

• Literature baseline extraction
• Physics-informed heuristic filters
• Stoichiometric constraint modeling (FBA)
• Experimental cost/time optimization

5. Production-Ready Architecture

• Modular microservice design
• ELN webhook integration
• Model versioning system
• Multi-tenant data isolation (design)

---

📈 Demo Flow (Load Demo Button)

1. Start: Pre-loaded session with 10 catalyst candidates
2. Literature: Extract baseline performance (70% activity from literature)
3. Rank (EI): Bayesian ranking highlights high-uncertainty candidates
4. Compare: Show how EI differs from greedy score ranking
5. Log Result: Simulate lab measurement (85% activity achieved)
6. Re-rank: Active learning updates best_so_far, reprioritizes candidates
7. Improvement: Show +15% gain over literature baseline

Optional: Switch to enzyme track → run FBA → screen biosecurity

---

Why CatalystAI is Best

• Direct Impact: Accelerates 2G Ethanol-to-SAF catalyst discovery

• Economic: Reduces screening costs by prioritizing high-EI candidates

• Scalable: Handles both catalyst and enzyme tracks (paddy straw pretreatment + conversion)

• Completeness: Covers all 4 ML layers (Retrieval, Generation, Prediction, Biology)

• Innovation: First to combine RAG + Bayesian BO + FBA + TEVV in one platform

• Explainability: Side-by-side comparisons show why AI picks certain candidates

• Safety: Only submission with biosecurity screening layer

• MVP → Production Path: Clear architecture for scaling

• Scientific Validation: Built on proven methods (DiffCSP, GNoME, COBRApy)

• Regulatory Awareness: TEVV compliance from day one

---

📚 References

• Open Catalyst Project (OCP): 260M+ DFT calculations
• Materials Project: Crystal structure database
• COBRApy: Constraint-based metabolic modeling
• RFdiffusion: Protein backbone generation
• BRENDA, KEGG, BiGG: Enzyme/pathway databases

---

🔧 Future Roadmap

Phase 2 (Months 1-4)

• Live GNN surrogate on OCP data
• Real eLabFTW integration
• Multi-user collaboration + RBAC
• GPU cluster for live diffusion

Phase 3 (Months 5-12)

• Next.js enterprise frontend
• Robotic lab hardware API
• Full audit trail + IP provenance
• CO₂ capture catalyst expansion

---