We searched the Web of Science and PubMed databases using keywords such as “algicidal,” “antialgal,” “algal bloom control,” and “harmful algae mitigation.” From the retrieved literature, records of algicidal activity were extracted, including information on target algae, algicides, algicidal effects, experimental conditions, and the mechanisms (Figure 1). Details of these 204 articles are provided in the Supporting Information file. Additional information on target algae, such as their classifications and general environment, was supplemented using the AlgaeBase (Guiry and Guiry, 2024) (https://www.algaebase.org/). The phylogeny was constructed based on taxonomic classification data from AlgaeBase for each algal species in our database, which established a hierarchical tree structure representing taxonomic relationships, and was then visualized using the ggtree v3.5.3 package (Yu et al., 2017). Details of algicides, including their form, source, and structural information, were curated from the literature and further analyzed using NPClassifier (Kim et al., 2021) and ClassyFire (Djoumbou Feunang et al., 2016) tools for chemical classification. To organize the curated data, we created five interconnected database tables (Supplementary Figure S1) to store information on algae species, algae strains, algicides, algicidal activity records, and literature references. The data processing workflow involved extracting experimental records from literature, collecting additional information on algae from AlgaeBase and chemical information from the literature and PubChem. Finally, all collected data was systematically checked for consistency and completeness to ensure data quality.

The concept of quantitative estimate of drug-likeness (QED), introduced by Bickerton et al. (2012), provides a scoring system to evaluate drug-likeness by describing the distribution of molecular descriptors, with scores ranging from zero (low drug-likeness) to one (high drug-likeness). Based this concept, Sorin et al. developed the quantitative estimate of pesticide-likeness (QEP) to assess pesticide potential (Avram et al., 2014). Inspired by these approaches, we established a quantitative estimate of algicide-likeness (QEA) to evaluate the algicidal potential of compounds. First, six molecular descriptors were calculated using RDKit (Landrum, 2006) including molecular weight (MW), octanol-water partition coefficient (logP), the number of hydrogen bond acceptors (HBA), the number of hydrogen bond donors (HBD), the number of rotatable bonds (nRotB), and the number of aromatic rings (nArR) (Avram et al., 2014). The selection of these descriptors was guided by principles established in drug and agrochemical discovery (Bickerton et al., 2012; Avram et al., 2014). For each descriptor, a desirability function was fitted (Equation 1), with values ranging from 0 to 1, representing the likelihood of a molecule being an algicide (Supplementary Figure S2). The coefficients (o, a, b, and c) used in the desirability function were derived from the distribution of physicochemical properties, as detailed in Supplementary Table S1. For the continuous descriptors (MW and logP), the optimal bin size for creating these frequency counts was determined using Shimazaki-Shinimoto histogram binning method (Shimazaki and Shinomoto, 2007). For the discrete descriptors (HBA, HBD, nRotB, and nArR), the frequency of each integer value was directly counted. The resulting frequency counts and the corresponding values for each molecular descriptor then served as input for curve fitting, which was processed using the https://findcurves.com/ Curve Fitting and Surface Fitting Web Site. Finally, the individual desirability functions, df(i), for each molecular descriptor were combined into the QEA score using geometric means and logarithms, as described in Equation 2. We calculated the algicide-likeness scores for algicides in the database (positive set) and 2,000 randomly selected compounds from the ChEMBL dataset (negative set). A receiver operating characteristic (ROC) curve was generated based on these scores, and the area under the curve (AUC) was computed to evaluate the performance of the QEA in distinguishing algicides from non-algicides.

The simplified molecular input line entry specification (SMILES) representations of the algicides were obtained using the Dicimer tool (Rajan et al., 2023) and the PubChem database (https://pubchem.ncbi.nlm.nih.gov/). These SMILES were input into the ADMET Evaluation functionality of ADMETlab 3.0 (Fu et al., 2024) to predict toxicity values for three aquatic species: Tetrahymena pyriformis (protozoa), Daphnia magna (zooplankton), and Pimephales promelas (fish). The predictions included comprehensive toxicity endpoints under standardized experimental conditions, i.e., the 48 h median effective concentration (EC_{50, 48h}) for T. pyriformis (50% growth inhibition), the 48 h median lethal concentration (LC_{50, 48h}) for D. magna (50% lethality), and the 96 h median lethal concentration (LC_{50, 96h}) for P. promelas (50% lethality). According to the guidelines established by the US Environmental Protection Agency (EPA) (US EPA, 2015), the toxicity levels were defined as follows: practically non-toxic (LC₅₀ or EC₅₀ > 100 mg/L), slightly toxic (10~100 mg/L), moderately toxic (1~10 mg/L), highly toxic (0.1~1 mg/L), and very highly toxic (<0.1 mg/L).

We constructed a knowledge base using RAGFlow (https://github.com/infiniflow/ragflow), an open-source Retrieval-Augmented Generation (RAG) engine (Figure 2). A total of 208 review articles on HABs published between 2010 and 2025 were manually collected. These articles cover topics related to the formation mechanisms, ecological impacts, monitoring methods, and management strategies of HABs. The collected articles were split by RAGFlow, resulting in 14,954 semantically meaningful text chunks. These chunks were then vectorized with the embedding model BAAI/bge-base-en-v1.5 (https://huggingface.co/BAAI/bge-base-en-v1.5) and stored in the retrieval database. This retrieval database was further integrated with the LLM (DeepSeek-v3), enabling retrieval-augmented generation for HABs-related question answering. To evaluate the performance of the RAG-based LLM, we used DeepSeek to automatically generate 1,040 high-quality question-answer pairs, where questions were derived from randomly selected literature chunks. The generated questions were then used as input for the RAG-based LLM to produce responses. We collected the generated answers and the retrieved references, and evaluated their quality using metrics (Supplementary Table S2) provided by Ragas (https://github.com/explodinggradients/ragas).

The AlgicideDB platform is hosted on an Nginx (v.1.16.1) web server (http://nginx.org/) running on a CentOS 7.4.1,708 system. The backend is built using a SQLite database (https://www.sqlite.org/) for efficient data management, and developed with the Django (v.4.2.8) framework (https://www.djangoproject.com/). The frontend is built with Vue (v. 3.5.13) JavaScript framework (https://v3.vuejs.org/) to provide interactive user interface. The platform integrates functionalities including searching algicide records, uploading and downloading data, calculating algicide-likeness scores, predicting aquatic toxicity values. Additionally, a RAG-based LLM chatbot is embedded to offer a question-and-answer service, enabling real-time responses to HAB-related queries.

In total, our database includes 110 algal species and 246 algal strains, with the majority belonging to the Cyanobacteria (37 species), Chlorophyta (23 species), Dinoflagellata (22 species), and Heterokontophyta (21 species) (Figure 3). Cyanobacteria and dinoflagellates are prominent targets due to their toxic and harmful nature, which poses significant challenges in ecological management. Notably, Microcystis aeruginosa (Cyanobacteria), linked to 260 algicides, along with harmful species such as Heterosigma akashiwo (Heterokontophyta), associated with 88 algicides, and Cochlodinium polykrikoides (Dinoflagellata), related to 50 algicides, are extensively studied. In contrast, some widely distributed green algae and diatoms are commonly used as controls to evaluate the selectivity of algicides, including species such as Chlorella pyrenoidosa (Chlorophyta), Raphidocelis subcapitata (Chlorophyta) and Phaeodactylum tricornutum (Heterokontophyta). Moreover, green macroalgae such as Ulva are well-studied for their association with biofouling and the formation of green tides, large-scale blooms that cause significant environmental and economic impacts. The distribution of algal species in the database highlights current research priorities, with a focus on managing harmful and toxic species while exploring the selectivity of algicides for broader ecological applications.

The AlgicideDB database includes a total of 542 algicides, of which 58.5% are derived from natural sources and 41.5% are classified as non-natural sources (Figure 4a). Natural sources include a diverse range of plants, microorganisms, animals (summarized in Supplementary Tables S3–S5), showing the important role of natural products in algicide discovery. Common examples include aquatic plants such as Myriophyllum spicatum (Zhu et al., 2010) and Arundo donax (giant reeds) (Hong et al., 2010), macroalgae like Sargassum (Bazes et al., 2009; Cho, 2013), terrestrial plant materials such as barley straw (Murray et al., 2010; Xiao et al., 2014) and rice hulls (Park et al., 2009), and marine bacteria including Vibrio sp. (Wang et al., 2012; Li et al., 2014), and Bacillus sp. (Zhao et al., 2014; Quan et al., 2021).

The most abundant natural product classes are alkaloids, fatty acids, and terpenoids (Figure 4b), which exhibit broad-spectrum activity against diverse algal taxa. Especially, alkaloids in the database target all seven algal phyla represented, highlighting their effectiveness as algicidal agents. Non-natural sources also contribute significantly to the diversity of the database, with numerous synthetic derivatives and analogs inspired by natural compounds (Supplementary Figure S3; Supplementary Table S6), and commercial algicides (Supplementary Table S7). These products expand the chemical diversity of available algicides, offering enhanced efficacy and environmental adaptability. Besides, our analysis of the 61 algicide-algae pairs with known mechanisms reveals mechanistic diversity across different chemical classes. Photosynthesis inhibition and membrane disruption were the most prevalent mechanisms, each accounting for 25.3% of the entries, followed closely by oxidative stress induction (20.5%), and metabolic inhibition (6.0%). Fatty acids primarily disrupted photosynthesis (40.0%) and cell membranes (20.0%), while alkaloids tended to induce oxidative stress (37.5%) and membrane disruption (31.2%). Notably, the mechanisms of action for over 90% of the algicide-algae pairs in AlgicideDB remain to be determined, indicating a significant area for future investigation.

We analyzed six key molecular descriptors for the algicides in the database: molecular weight (MW), octanol-water partition coefficient (logP), number of rotatable bonds, number of aromatic rings, number of hydrogen bond acceptors (HBA), and number of hydrogen bond donors (HBD). These properties are crucial for evaluating pesticide-like molecules and their environmental behavior (Avram et al., 2014). The distributions of these descriptors (Figure 4c) indicate that the MW of algicides ranges from approximately 100 to 1,000 g/mol, with the majority falling between 200 and 500 g/mol. The logP distribution shows a peak around 2–5, reflecting moderate hydrophobicity favorable to algicidal activity. Most compounds have fewer than five rotatable bonds, ensuring a balance between molecular flexibility and stability. Additionally, the majority of algicides have 1–2 aromatic rings and fewer than five HBA and two HBD.

Using these molecular descriptors, we developed a desirability function to quantify algicide-likeness. The algicide-likeness scores were computed (Figure 5a), and their ability to distinguish algicides from non-algicides was evaluated using a negative set of 2,000 randomly selected compounds from the ChEMBL dataset. The receiver operating characteristic (ROC) analysis yielded an area under the curve (AUC) of 0.717 (Figure 5b), demonstrating that the algicide-likeness metric effectively differentiates algicides, which aligns with the performance observed in similar quantitative likeness estimation methods within the broader agrochemical domain (e.g., herbicide-likeness (QEH) AUC = 0.721, insecticide-likeness (QEI) AUC = 0.668, and fungicide-likeness (QEF) AUC = 0.677) (Avram et al., 2014). The algicide-likeness evaluation has been integrated into our web-based platform to facilitate high-throughput screening and algicide discovery.

Besides algicide-likeness evaluation, we also considered the potential toxicity of algicides to non-target aquatic organisms. Among the 204 studies we collected, only 20 have evaluated the toxicological effects of algicides on non-algal aquatic species. To address this limitation, we predicted the toxicity of algicides across three trophic levels: T. pyriformis (protozoa), D. magna (zooplankton), and P. promelas (fish) (Figure 5c). The predictions indicate that algicides generally exhibit relatively lower toxicity toward T. pyriformis. However, more than half of the compounds were categorized as moderately toxic or higher for D. magna and P. promelas. Further analysis of toxicity levels across chemical classifications revealed that amino acid-derived algicides show relatively lower toxicity to all three aquatic species, suggesting their potential as environmentally friendly algicides (Figure 5d). This lower toxicity may be attributed to their natural occurrence in ecosystems, which enables safer interactions with aquatic organisms, as well as their potential for safe environmental degradation. Conversely, certain classes, such as thiazolidinedione and rhodanine derivatives, were associated with higher toxicity levels across multiple species, possibly due to their distinctive sulfur-containing heterocyclic structures. These findings underscore the importance of incorporating toxicity assessments into the development of safer and more sustainable algicides. Considering that the confidence of these predictions may vary across different chemical scaffolds, especially for novel structures underrepresented in toxicological databases, experimental validation remains essential to confirm these toxicity patterns.

To explore innovative approaches in HABs management, we applied a RAG-based LLM to construct a domain-specific knowledge base for answering HABs-related queries. When a user submits a question, the system first retrieves relevant literature chunks from the database, which were then combined with the user's query as input for the LLM to enhance the contextuality and factuality of the generated response. The replies not only address the query but also provide references to the retrieved literature as supporting evidence (Supplementary Figure S4). This RAG-based LLM has been integrated into the AlgicideDB platform to provide users with support for HABs-related research and decision-making in HABs management.

To evaluate model performance, we calculated four metrics: context recall, context precision, faithfulness, and answer relevancy (Supplementary Table S2) based on the retrieved references and generated responses. The first two metrics assessed the precision and relevance of the retrieved references, while the latter two evaluated the relevance and factuality of the generated answers. The model achieved scores of 0.732, 0.754, 0.834, and 0.896 (Figure 6), respectively, demonstrating robust performance in both retrieval and generation tasks. To further illustrate the system's capabilities, we also assessed its performance on complex queries that require synthesizing information from multiple sources within the knowledge base (see Supplementary Figure S4 for examples). For instance, one such query was “Compare the efficacy of ultrasonic treatment and chemical algicides for controlling Microcystis aeruginosa blooms in drinking water reservoirs, considering potential side effects on water quality.” The system demonstrated the ability to retrieve relevant information from different articles and provide a synthesized response, supported by citations.