Finalist
Alpha_BioAgent
🔥 The Problem: AI "Hallucinations" in Biology Normal AIs (like ChatGPT) are great at talking but bad at physics.
Core trick
The agent runs Bio.PDB distance calculations on a protein structure before accepting each LLM-proposed mutation, retrying on steric clashes.
Description
🔥 The Problem: AI "Hallucinations" in Biology
Normal AIs (like ChatGPT) are great at talking but bad at physics. If you ask an AI to fix a protein, it often suggests adding a big molecule into a small space. In text, it sounds perfectly logical. But in a real lab, that protein would crash into itself and instantly fall apart (Steric Clash). We call this "Structural Hallucination."
✅ The Solution: Teaching AI to Use Physics Tools
Alpha-Agent solves this by forcing the AI to use physical tools (like the Bio.PDB Python library):
It Calculates First: Before the AI proposes a mutation, it secretly runs a Python script to measure the exact 3D distance between atoms.
It Corrects Itself: If the physical measurement shows that the AI's idea won't fit, the AI realizes its mistake and automatically recalculates until it finds a design that actually works in the real world.
In short: Alpha-Agent anchors the AI to reality, giving researchers protein designs they can actually trust in the lab.
README excerpt
# 🧬 Alpha-Agent: Autonomous Multimodal Protein Design Loop ## Gemini 3 Seoul Hackathon 2026 Finalist Project (Alpha_BioAgent) [🇬🇧 English](#english) | [🇰🇷 한국어](#korean) --- ## <a name="english"></a> 🇬🇧 English ### 1. Project Overview & The Problem We Solve **Alpha-Agent** is a **Physics-Grounded Autonomous Agent** designed to eliminate LLM hallucinations in protein engineering. **The Problem: LLM Hallucinations in Biology** While Large Language Models (LLMs) can generate text about protein design, they lack intrinsic knowledge of physical laws. When asked to stabilize a protein, an LLM might propose a bulky mutation that sounds logical but physically causes massive steric clashes, instantly unfolding the protein in a real lab setting. **Our Solution: Grounded Verification** Alpha-Agent transforms the LLM from a text-generator into a tool-using director. Before proposing a mutation, it executes backend Python scripts (like Bio.PDB) to calculate actual 3D distances (e.g., verifying a salt bridge is exactly 3.78Å apart). If a proposed design fails physics-based simulation checks, the Agent autonomously rejects it and triggers a **Self-Correction Loop** to redesign a physically viable alternative. ### 2. Core Features * **Grounded Physics Engine (`tools/analyzer.py`):** Calculates B-factors and Euclidean distances using `Bio.PDB` to identify structural instability hotspots rather than relying on LLM hallucinations. * **Multimodal Discovery Loop:** Generates and analy