AI and Chemists Co-Develop TYR Inhibitors via Dual-Track Optimization

Background

Tyrosinase (TYR) inhibitors often show activity in enzyme assays but fail in cellular or physiological models due to stability, toxicity, or selectivity issues, making drug discovery for melanin‑related disorders costly and complex.

Reinforcement‑Learning‑Driven Molecular Generation

A deep reinforcement‑learning (RL) framework was built starting from the 3‑D structure of TYR. The framework uses fragment‑based generation and interacts with a virtual environment to create a large virtual library. After filtering for synthetic accessibility and drug‑likeness, 28 candidate molecules were selected for synthesis and biological evaluation. The lead compound AI10 displayed enzymatic activity but only moderate potency and noticeable cytotoxicity.

Dual‑Track Lead Optimization

Optimization proceeded in two parallel tracks:

Expert‑guided track : Medicinal‑chemistry knowledge guided the design of 57 AI10‑derived derivatives, focusing on improving solubility and pharmacological properties. Representative compounds include AI10‑m15 , AI10‑m52 , and AI10‑a2 . AI10‑m15 achieved nanomolar TYR inhibition and strong anti‑melanogenesis effects in vitro and in vivo; activity ranking was AI10‑m15 > AI10‑m52 > AI10‑a2.

RL‑driven track : The same RL framework continued to explore chemical space using a fragment‑growth strategy based on predefined reaction templates. Starting from the AI10 scaffold, fragments were recombined to generate 32 candidates prioritized for structural similarity and physicochemical properties.

Compound Synthesis and Biological Evaluation

Enzymatic inhibition assays, cellular melanin‑generation suppression tests, and in‑vivo efficacy studies were performed on selected candidates. Dose‑dependent reductions in melanin levels were observed for AI10‑m15, AI10‑m52, and AI10‑a2, confirming the potency of the optimized leads.

Comparison of AI‑Driven and Expert‑Guided Strategies

The AI‑driven pathway expanded the searchable chemical space, uncovering unconventional scaffolds and non‑intuitive structure‑activity patterns. The expert‑guided pathway provided focused selection and convergence toward developable molecules. Together, the two paradigms broadened chemical‑space exploration while ensuring robustness of the candidates.

Key References

Science Advances article: https://www.science.org/doi/10.1126/sciadv.aeg0376

Figures

Code example

来源：ScienceAI
本文
约2000字
，建议阅读
5
分钟
人机竞赛最终都要回到人机协作上来。