Mutation Cost penalizes disruptive or overly aggressive alterations.
Rewards are sparse and often delayed, emphasizing the importance of temporal credit assignment, a hallmark strength of RL methods.
3. Learning Algorithm and Architecture
We suggest employing Deep Reinforcement Learning methods for high-dimensional control, such as:
Deep Q-Networks (DQN) for discrete mutation actions
Proximal Policy Optimization (PPO) for policy-gradient-based control
Multi-Agent RL where agents specialize (e.g., one focuses on active sites, another on hydrophobic core)
The RL agent may be further enhanced by:
Attention mechanisms to focus on structurally or functionally critical residues
Curriculum learning: agents start with simple tasks (e.g., preserve folding), then advance to complex ones (e.g., evolve novel substrate binding)
4. Feedback Loop with CAS Dynamics
