Antimicrobial Peptide Generator

Overview

This project leverages deep learning to address the critical issue of antibiotic resistance. By integrating Reinforcement Learning (RL) into a Generative Adversarial Network (GAN) framework, we developed a model capable of de novo design of Antimicrobial Peptides (AMPs). The core innovation lies in embedding a Monte Carlo search within the generator to optimize sequence generation through policy gradients.

Key highlights

Technical Architecture

• Implemented a Generative Adversarial Network (GAN) framework to generate novel peptide sequences.
• Enhanced the generator by embedding a Monte Carlo search to estimate intermediate action values, effectively treating sequence generation as a Reinforcement Learning problem.
• Utilized Policy Gradient methods where the discriminator provides the final reward signal to update the generator.
• Trained on experimentally verified peptide sequences curated from major databases including dbAMP, DADP, DBAASP, and APD.

Validation Pipeline

• Candidates were screened by comparing key properties (e.g., hydrophobicity, net charge, alpha-helical propensity, and Boman index) against known active peptides.
• High-confidence sequences were analyzed using AlphaFold to predict 3D structures, followed by Molecular Dynamic (MD) simulations to assess stability.