GEAR: Guided Exploration for Aptamer Retrieval

A self-play neural network guided Monte Carlo Tree Search framework for de novo aptamer design

🧬 Overview

GEAR is a computational pipeline for designing high-affinity aptamers (DNA/RNA sequences that bind to specific protein targets). This project applies reinforcement learning techniques, similar to AlphaGo, to the problem of aptamer design.

The Problem

Current MCTS-based aptamer design tools (Apta-MCTS, AptaTrans) use random rollouts to explore sequence space. This is inefficient - it's like playing chess by making random moves until the game ends, then evaluating the position.

The Solution

GEAR uses a learned policy-value network trained through self-play to intelligently guide which nucleotides to add at each step, making the search process smarter and potentially discovering better aptamers.

Project Structure

Research Objectives

1. Baseline Models - Build traditional ML aptamer-protein interaction classifiers
2. Model Distillation - Create a faster, distilled version of AptaTrans (target: 2x speedup)
3. Self-Play Training - Train policy-value network through iterative self-play
4. Evaluation - Benchmark against existing methods using ZDOCK molecular docking
5. Deployment - Package as Docker-based pipeline with web interface
6. Open Source - Release as accessible research tool

Development Setup

# Clone repository
git clone https://github.com/yourusername/GEAR.git
cd GEAR

# Python environment
conda create -n gear python=3.8
conda activate gear
pip install -r requirements.txt

# Rust MCTS engine
cd rust-mcts
cargo build --release

Current Status

This project is in active development as part of a research project at Minnetonka High School.

Technology Stack

• PyTorch - Deep learning framework
• Rust - High-performance MCTS engine
• ONNX - Model inference and deployment
• Docker - Containerization
• React - Web interface

Research Background

This work builds upon:

• Apta-MCTS (Lee et al., 2021) - First MCTS approach for aptamer design
• AptaTrans (Shin et al., 2023) - Transformer-based binding prediction
• AlphaGo (Silver et al., 2016) - Self-play reinforcement learning

Acknowledgments

• Minnetonka Research Program - Research support
• University of Cincinnati - Mentorship and data

Contact

Emilio Moreno
Minnetonka High School
Email: 008774@mtka.org

---

This is an active research project. Documentation and features will be added as development progresses.