GEMINI MOLECULAR RANKER: AI FOR DRUG DISCOVERY

Gemini Molecular Ranker: Democratizing Drug Discovery with AI

The Problem We're Solving

Drug discovery is broken. Developing a single drug costs $2.6 billion and takes 10-15 years. Meanwhile, millions suffer from diseases without treatments.

The critical bottleneck? Early-stage molecular docking:

Traditional tools give conflicting results (DiffDock vs Vina disagree 40% of the time)
No explainability—researchers can't trust AI black boxes
Requires expensive infrastructure ($50K/year for commercial software)
Manual analysis takes weeks of expert time

Result: Small labs in developing countries can't compete. Rare diseases go untreated. Promising candidates are missed.

Our Solution: AI Agent That Thinks Like a Scientist

Gemini Molecular Ranker is an autonomous AI agent that orchestrates molecular docking with multi-method consensus validation.

What Makes It Different:

1. Multi-Method Consensus

Runs both ML-based (DiffDock) and physics-based (Vina) methods
Only reports results when methods agree (RMSD < 2Å)
Eliminates 40% of false positives from single-method approaches

2. Explainable Reasoning

Shows step-by-step thinking: "Running Vina to validate DiffDock results"
Natural language analysis: "Rank 1 has strong H-bonding but potential toxicity concerns"
Researchers can trust and learn from the AI

3. Accessible to Everyone

Free web interface—no installation, no coding required
Runs on university servers (no cloud costs)
Open-source for global research community

Social Impact: Who Benefits?

Academic Labs in Developing Countries

No access to $50K/year commercial software → Our tool is free

Rare Disease Foundations

Patient advocacy groups can now screen candidates themselves, attracting pharma investment

Antibiotic Resistance Crisis

10x faster initial screening → more candidates reach clinical trials

Pandemic Preparedness

Upload viral protein → get ranked candidates in 30 minutes instead of weeks

How We Built It

Architecture:

Frontend (Next.js)
   ↓ REST API
Backend (FastAPI)
   ↓ Python
Gemini AI Agent
   ├─→ DiffDock (ML-based docking)
   ├─→ AutoDock Vina (physics-based)
   ├─→ Consensus Analysis
   ├─→ ADMET Scoring
   └─→ Natural Language Explanation

Our Innovation: Agentic Workflow

The Gemini agent autonomously:

Plans: "I need both DiffDock and Vina for validation"
Executes: Runs tools in parallel
Validates: Calculates consensus RMSD
Decides: "Weak consensus → run refinement"
Explains: Generates human-readable analysis

Consensus Scoring:

$$\text{Score} = 2.0 \times H_{bonds} + \frac{Contacts}{20} + 1.5 \times Shape - 2.0 \times Lipinski_{violations}$$

Challenges We Overcame

1. Coordinate System Mismatch
DiffDock and Vina output different coordinate systems → implemented Kabsch RMSD alignment

2. Gemini Context Limits
Molecular data too large → designed hierarchical summarization (pose → summary → Gemini analysis)

3. University Server Constraints
No root access, firewall restrictions → containerized backend with Tailscale VPN

4. Real-Time Progress for Long Jobs
Docking takes 5-15 minutes → background job queue with WebSocket-like polling

What We Learned

Technical:

Agentic AI design: Tool calling, reasoning loops, decision trees
Computational chemistry: RMSD calculations, force fields, Lipinski rules
Full-stack development: FastAPI async, Next.js, 3D visualization (3Dmol.js)

Domain:

Why consensus matters: Single methods are 40% unreliable
ADMET properties: Drug-likeness vs binding affinity trade-offs
Explainability in science: Trust requires transparency

Impact:

Accessibility > Features: Free tools democratize research
Open source accelerates science: Build on others' work, share yours
User needs drive design: Researchers need explanations, not just predictions