Discovery of target-biased natural products for algal control by integrating deep learning and large language models

Harmful algal blooms (HABs) pose an escalating global threat to water security. Conventional chemical agents remain a widely used approach for HAB control, but pose ecological risks by harming non-target aquatic organisms. To address this challenge, we developed a novel computational framework integrating safety-first multi-task learning with large language model (LLM)-powered agents for candidate screening and experimental validation. Within this framework, a multi-task directed message passing neural network (DMPNN) first screens for compounds that balance high algicidal efficacy with low non-target toxicity, followed by the LLM agent that performs automated multi-source validation to finalize candidate prioritization. From a library of 18,895 natural compounds, we identified hinokitiol, a monoterpenoid natural compound, as the leading candidate. Experimental validation demonstrated hinokitiol's broad-spectrum algicidal potency (e.g., IC 50 = 0.690 mg/L for Microcystis aeruginosa and 4.286 mg/L for Skeletonema costatum ) and low non-target toxicity (LC 50 > 10 mg/L for Danio rerio and Daphnia magna ), making it superior to traditional copper-based algicides by offering effective HAB control with reduced ecological risks to aquatic organisms. Mechanistic investigation revealed that hinokitiol disrupts iron homeostasis in algal cells, triggering iron-induced intracellular oxidative stress. Our study presents a novel computation-driven framework for natural product discovery that inherently balances efficacy and ecological safety, which could serve as a transferable candidate discovery paradigm for pesticides and bactericides.