AI-designed molecules plus expert chemistry yield nanomolar tyrosinase inhibitors

An artificial intelligence system that designs brand-new drug molecules from scratch, paired with the practiced hands of medicinal chemists, has delivered a family of tyrosinase inhibitors roughly a thousand times more potent than the gold-standard skin-lightening agent kojic acid. In a study published in the Journal of Advanced Research, researchers report that a reinforcement learning algorithm generated a novel lead compound against tyrosinase, the copper enzyme that catalyzes melanin production, and that expert-guided structural optimization subsequently pushed inhibitory activity into the nanomolar range. The best optimized compound, designated V-24, inhibited tyrosinase with an IC50 of just 18 nanomoles against the diphenolase substrate L-dopa and 20 nanomoles against L-tyrosine, representing a 600- to 2,300-fold improvement over the original AI-generated lead.

Tyrosinase sits at the heart of the pigmentation pathway. Human skin color is determined largely by melanin, which exists in two forms: dark eumelanin and reddish pheomelanin. Both absorb ultraviolet radiation and protect the skin from environmental damage, but excessive pigment deposition drives a range of disorders including chloasma, freckles, acanthosis nigricans, Riehl’s melanosis, skin aging and, in severe cases, melanoma. Because tyrosinase is the rate-limiting enzyme in the double-loop oxidation process that converts L-tyrosine into dopachrome, the precursor of melanin, it has long been the primary therapeutic target for controlling hyperpigmentation. Structurally, tyrosinase is a type 3 binuclear copper metalloenzyme, built around six highly conserved histidine residues that coordinate two copper ions within the active site.

Existing inhibitors leave considerable room for improvement. Traditional agents are mostly substrate analogues of L-tyrosine and L-dopa or resorcinol derivatives, and include natural products such as kojic acid, arbutin, resveratrol, polyphenols, flavonoids, stilbenes and lignans. These compounds typically show weak activity and require high concentrations, which raises the risk of skin irritation and, in the case of kojic acid metabolites, liver burden. Many are highly hydrophilic and penetrate skin poorly, and several degrade under light, heat or oxidation. Extraction from natural sources is costly and difficult to standardize. Synthetic second-generation inhibitors, including kojic acid, azole, thiourea, amide, cinnamic acid and benzopentacyclic derivatives, improve activity and stability, but many carry structural safety liabilities, risk of resistance and complex synthetic routes, while their derivation from natural-product scaffolds limits structural novelty.

To break this pattern, the team turned to a de novo molecular generation strategy in which artificial intelligence does not merely screen existing libraries but actively invents new chemical entities. The system is built on the Soft Actor-Critic reinforcement learning algorithm and treats molecular synthesis as a sequential decision-making process. Rather than assembling atoms freely, the model starts from commercially available molecular building blocks and applies established chemical reaction templates through forward reaction prediction, a design choice that guarantees the synthetic accessibility of everything it proposes. In the initial stage, each building block in a predefined library is docked against the tyrosinase crystal structure using AutoDock Vina, and the top 200 fragments by docking score form the starting set.

During the generation phase, each round randomly selects a starting molecule and, guided by the current policy, picks a reaction template to perform a chemical modification. Every product is immediately docked against tyrosinase and assessed for drug-like properties, with the binding affinity feeding a reward function that teaches the model which molecular choices pay off. Each step is logged in a replay buffer, enabling iterative updates of the network parameters through random sampling during training. Generation stops after a maximum of three synthetic steps, when molecular weight exceeds 600 daltons, or when no suitable reaction template can be found. In post-processing, a virtual library of 20,000 generated molecules is filtered through Lipinski’s Rule of Five, ranked by predicted binding affinity, and the top 100 candidates along with their synthetic pathways are submitted to medicinal chemistry experts, who selected roughly 20 molecules for actual synthesis.

Seven generation runs targeting tyrosinase, each yielding the model’s top 100 molecules, produced a pool of 700 candidates from which 17 compounds were synthesized and tested. Ten showed moderate to high inhibition of mushroom tyrosinase, and seven outperformed both alpha-arbutin and beta-arbutin, the clinical reference compounds. The standout was compound V, a piperazine-containing molecule with a monohydroxyphenyl pharmacophore, which inhibited tyrosinase with an IC50 of 18.5 micromoles against L-dopa and 9.6 micromoles against L-tyrosine, already surpassing kojic acid. Molecular docking revealed why: the phenolic hydroxyl group coordinates the catalytic copper ion, while the aromatic ring engages HIS263, ALA286 and VAL283 through pi-stacking, pi-alkyl and pi-sigma interactions.

With compound V in hand, the chemists launched a systematic structure-activity campaign, dividing the scaffold into four modular regions: the aromatic Ar cap, the two linkers, the central piperazine ring, and the phenolic D-ring that chelates copper. Across 34 optimized analogues, clear rules emerged. A monofunctionalized phenyl Ar group, a methylene linker, a vinyl linker and a piperazine core were all favorable, but the decisive change was converting the D-ring to a 2,4-dihydroxyphenyl group. Thirteen compounds in this series, V-22 through V-34, reached nanomolar potency, and compound V-24, which pairs a 4-fluorobenzyl-piperazine cap with a 2,4-dihydroxyphenyl unit, emerged as the most potent inhibitor. In silico property profiling showed V-24 satisfies Lipinski and Veber rules with a high QED of 0.83, strong predicted bioavailability of 87.71 percent and a moderate half-life, supporting its selection as a candidate compound.

Biological validation followed at multiple scales. In B16F10 and A375 melanoma cell lines, both compounds showed no significant cytotoxicity at 100 micromoles and suppressed melanin synthesis in a dose-dependent manner, with V-24 rivaling beta-arbutin in B16F10 cells and clearly outperforming all controls in A375 cells. In zebrafish embryos, a classic whole-organism model for pigmentation, V-24 reduced head melanin signal by 16.32 percent, on par with kojic acid. Most strikingly, in a three-dimensional human skin model incorporating both keratinocytes and melanocytes and subjected to seven days of UVB irradiation, V-24 treatment produced the lightest coloration of any group, with an L-value of 83.38 versus 70.98 for kojic acid at the same concentration. Surface plasmon resonance confirmed direct binding to tyrosinase, and metabolic studies showed the compound is highly stable in human plasma, retaining 89.06 percent after two hours, while being moderately cleared by liver microsomes.

The study demonstrates that pairing an AI de novo generation engine, constrained by reaction templates and rewarded by docking scores, with conventional expert-driven lead optimization can redefine the efficiency of inhibitor discovery, collapsing what is typically a years-long journey from target to nanomolar candidate into a single integrated workflow. The authors suggest this hybrid strategy, in which artificial intelligence proposes structurally novel, synthetically feasible starting points and medicinal chemists refine them through iterative structure-activity analysis, offers a generalizable blueprint not only for anti-pigmentation therapeutics and cosmetics but potentially for drug discovery campaigns against other metalloenzyme targets. All animal experiments were approved by an institutional ethics committee, and the model’s source code and reaction template library have been made publicly available to the research community.

Subject of Research: AI-driven de novo molecular generation and expert-guided optimization of nanomolar tyrosinase inhibitors for treating skin hyperpigmentation

Article Title: Strategy and efficiency-redefined discovery of novel nanomolar tyrosinase inhibitors: AI de novo molecular generation + expert-guided structural optimization

Article References: Sun, Y., Wang, J., Chen, W., Wen, H., Feng, M., Niu, X., Zhi, J., Hu, S., Wang, S., Cai, H., Ju, B., Yang, K., Jiang, X., & Bai, R. (2026). Strategy and efficiency-redefined discovery of novel nanomolar tyrosinase inhibitors: AI de novo molecular generation + expert-guided structural optimization. Journal of Advanced Research, 87, 1079-1104. https://doi.org/10.1016/j.jare.2025.12.041

Image Credits: AI Generated

DOI: 10.1016/j.jare.2025.12.041

Keywords: tyrosinase inhibitors, artificial intelligence, de novo drug design, reinforcement learning, melanin synthesis, skin hyperpigmentation, kojic acid, piperazine derivatives, molecular docking, 3D human skin model, zebrafish assay, medicinal chemistry