The Drug Discovery Challenge

Traditional drug discovery is notoriously slow and expensive, typically requiring 10-15 years and over $1 billion to bring a single drug to market, with failure rates exceeding 90%. Pharmaceutical researchers must screen millions of compounds to identify promising candidates, then spend years optimizing their properties—improving efficacy while minimizing toxicity and ensuring favorable pharmacokinetics. A critical challenge is the trade-off between potency and drug-likeness: compounds that bind strongly to targets often have poor ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) properties, while drug-like molecules frequently lack sufficient binding affinity. Traditional approaches optimize these properties sequentially rather than simultaneously, leading to suboptimal compounds that fail in later development stages. Our CADD-driven active learning approach addresses this fundamental limitation by using iterative computational cycles that simultaneously optimize binding affinity and drug-like properties, ensuring that generated compounds maintain the delicate balance needed for successful therapeutics from the earliest design stages.

---

Our CADD-Driven Active Learning Process:
Data Integration & Substructure Identification: Starting with large chemical databases (ChEMBL, ZINC20), algorithmic methods identify valuable molecular substructures with expert feedback, creating a foundation of pharmacologically relevant chemical fragments. Generative Model Training: AI systems map desirable properties within chemical space using fine-tuning on known active compounds and CADD evaluation data, learning to generate molecules that balance multiple competing objectives. Iterative CADD Evaluation: Each generated compound undergoes computational assessment including molecular docking and MM-GBSA binding calculations, while machine learning models simultaneously evaluate toxicity, IC50 potency, drug-likeness (QED), pharmacokinetics, and novelty scores. Active Learning Refinement: Expert evaluation of the most promising candidates feeds back into the generative model, creating an iterative cycle where each round produces compounds with improved binding affinity and ADMET properties. Optimized Synthesis Targeting: Final candidates represent computationally optimized molecules that maintain high target binding while exhibiting favorable drug-like characteristics, ready for experimental synthesis and validation.

---

Our Research Areas:
CADD-Driven Active Learning Systems Iterative computational workflows that combine molecular docking, MMGBSA calculations, and ADMET prediction with generative AI to simultaneously optimize binding affinity and drug-like properties.
Multi-Objective Molecular Optimization AI systems that balance competing requirements—maximizing target binding while maintaining favorable toxicity, pharmacokinetic, and drug-likeness profiles throughout the design process.
Expert-Guided Generative Models Machine learning approaches that incorporate human expertise and computational evaluation into the training loop, improving the quality and relevance of generated molecular candidates.

---

Featured Publications:
"Improving Drug Discovery with a Hybrid Deep Generative Model Using Reinforcement Learning Trained on a Bayesian Docking Approximation" - Journal of Computer-Aided Molecular Design (2023) Advanced AI system demonstrating our CADD-driven active learning approach, showing how reinforcement learning can optimize both binding affinity and drug-like properties simultaneously. DOI: 10.1007/s10822-023-00523-3