AMR PPO Solution: Reinforcement Learning for Multiple Sequence Alignment (MSA)

--Note: Documentation Generated by ChatGPT--

📌 Overview

Antimicrobial Resistance (AMR) is a global health crisis, with 5 million deaths linked to AMR in 2019.
Multiple Sequence Alignment (MSA) helps identify AMR genes, but AI models struggle to match human accuracy.
We developed a Proximal Policy Optimization (PPO) RL model to optimize MSA, outperforming traditional AI approaches.

🔍 Understanding the Problem

Why Multiple Sequence Alignment (MSA)?

• MSA aligns multiple DNA sequences to detect conserved regions.
• Crucial for phylogenetic analysis and AMR gene identification.
• Superior to pairwise alignment, which only compares two sequences.

Why AI and ML Struggle With MSA?

• High computational complexity.
• Lack of domain-specific knowledge—humans still outperform AI.
• Traditional ML (e.g., XGBoost) fails to capture sequence dependencies.

🛠️ Solution Strategy

1️⃣ Exploring MSA and BLS

• Borderlands Science (BLS) crowdsourced MSA solutions from players.
• This dataset provided real-world human alignments for training AI.

2️⃣ Why Traditional Models Failed

• Regression models (e.g., XGBoost) failed due to:
  • No sequential decision-making.
  • Inability to handle structured DNA gaps.

3️⃣ Why Reinforcement Learning (RL)?

• RL suits MSA’s step-based structure.
• BLS was already a game, making RL the best fit.
• Transformer models were too expensive to train, while RL offered scalability.

Built With

• data-preprocessing
• machine-learning
• python
• reinforcement-learning