REINFORCEMENT LEARNING FROM VERIFIABLE REWARDS IN CNS DRUG OPTIMIZATION
DAVID SCOTT LEWIS, ENRIQUE ZUECO, HALEY YI
AIXC BIO, ZARAGOZA, SPAIN | REPORTS@AIXCBIO.COM

ALZHEIMER'S ASSOCIATION
AAIC '26

KEY TAKEAWAY
Verifiable Physics-Based Rewards Enable True Autonomous Self-Improvement for Multi-Parameter CNS Drug Optimization.

INTRODUCTION
Designing Brain-Penetrant Therapeutics requires balancing multi-parameter constraints:
- Solubility vs. Blood-Brain Barrier (BBB) permeability
- Lipophilicity

Why current methods (like pure LLMs) fail:
Pure LLMs fail without verifiable biological feedback rewards.

METHODS
Primary Objectives:
- Autonomous CNS optimization goals
- Multi-target BBB crossing and optimal lead generation

Reward Calculation Detail:
- Tau dock score -> Deterministic Physics-Based Reward Function -> Verifiable Scalar Reward
- BBB value -> Deterministic Physics-Based Reward Function -> Verifiable Scalar Reward

This is superior to scalar reward models based on human preference.

CLOSED-LOOP SYSTEM FOR TAU INHIBITORS
SMILES Strings -> AI AGENT -> TEST-TIME RECURSIVE THINKING -> VERIFIABLE REWARD EVALUATION (SIMULATION)
VERIFIABLE REWARD EVALUATION (SIMULATION) -> MOLECULAR DOCKING (Tau Aggregation) + In Silico BLOOD-BRAIN BARRIER (BBB) PERMEABILITY SIMULATION -> HARD SCALAR REWARDS -> FEEDBACK -> AI AGENT (Loop repeats)

RESULTS
SUCCESS RATE IMPROVEMENT
(Bar chart showing +42% improvement)

BINDING AFFINITY vs. BBB PERMEABILITY
(Scatter plot showing Binding Affinity vs. BBB Permeability)

Rapid Autonomous Self-Improvement
(Arrow pointing right indicating Time, with molecular structures shown)

CONCLUSIONS
AUTONOMOUS SELF-DRIVING DRUG DISCOVERY PIPELINE
AUTOMATED SYNTHESIS -> TESTING -> AI-LED DESIGN (Loop)

- Significance: Establishes a verifiable path towards faster, more effective Alzheimer's drug discovery.
- Validates physics-grounded AI learning as scalable for complex multi-objective optimization.
- Future Directions: Generalization to other target classes; Regulatory pathways for autonomous discovery frameworks.