Intelligent pesticide preparation mixed toxicity prediction method and system based on knowledge graph

Abstract

The application discloses a pesticide preparation mixed toxicity intelligent prediction system based on a knowledge graph, comprising a priori knowledge graph construction module, which utilizes unstructured books or manuals and structured databases to establish the correlation of dosage forms and adjuvant types through natural language processing technology; the dosage form perception double-flow deep learning prediction model comprises a structure flow, a dosage form correction flow and a fusion layer; an intelligent report and suggestion generation engine calls a preset rule base and knowledge base based on the final toxicity prediction value output by the model, and comprises the following functions: registration report, toxicity grading, risk assessment suggestion, first aid suggestion and registration suggestion; the application has the advantages that the structure information, dosage form information and adjuvant correlation knowledge of the pesticide preparation are fused and modeled to predict the acute toxicity of the pesticide preparation, and a system for automatically generating registration / risk assessment / first aid reports is generated.

Description

Technical Field

[0001] This invention relates to the fields of computational toxicology and pesticide risk assessment, specifically to a knowledge graph-based intelligent prediction method and system for pesticide formulation mixture toxicity. Background Technology

[0002] Pesticide formulations are complex mixtures containing active ingredients and various adjuvants (solvents, surfactants, synergists, etc.). Their final toxicity is often the result of synergistic or antagonistic effects among the components.

[0003] QSAR technology has been widely used to predict the toxicity of a single active ingredient (such as oral LD50 in rats) and has been recognized and applied by organizations such as the US EPA and the European ECHA.

[0004] Currently, some studies have attempted to estimate the toxicity of mixtures using "concentration summation" or "independent action" models, but these models cannot capture the nonlinear effects caused by adjuvants.

[0005] In practice, testing units have accumulated a large amount of experimental data on "effective ingredients + content + dosage form + corresponding toxicity", but lack specific information on the ratio of adjuvants.

[0006] Traditional QSAR models (such as TEST, ECOSAR, etc.) predict toxicity based solely on the molecular structure (SMILES) of a single active ingredient, without distinguishing between dosage forms and adjuvants.

[0007] The OECD QSAR Toolbox-based categorical / read-across method extrapolates toxicity by finding structurally similar analogs. While this method can handle some mixtures, it heavily relies on experimental data from analogs and also lacks a systematic consideration of adjuvant effects. The conclusion is that current technologies are ineffective in predicting pesticide formulation toxicity when adjuvant / formulation information is lacking.

[0008] The existing technology has the following drawbacks: 1. Low prediction accuracy: Numerous animal experimental data on pesticide formulations show that the LD50 values ​​of different formulations of the same active ingredient (e.g., EC vs. SC) vary greatly. Existing QSAR models often fail to provide accurate predictions because they ignore the influence of adjuvants.

[0009] 2. Limited applicability: Existing models require complete chemical structure or formulation information as input. A large amount of data accumulated by users in actual production (with known active ingredients and dosage forms, but with confidential adjuvant formulations) cannot be effectively utilized, resulting in data waste.

[0010] 3. Lack of business automation: Existing models only provide toxicity values ​​or classification labels and cannot automatically generate standardized reports that meet pesticide registration requirements (such as the requirements of the Pesticide Health Toxicology Database) (such as including toxicity classification, risk management recommendations, and first aid recommendations).

[0011] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0012] The technical problem this invention aims to solve is to overcome the aforementioned technical deficiencies and provide a knowledge graph-based intelligent prediction method and system for pesticide formulation mixture toxicity. This addresses the low accuracy of existing quantitative structure-activity relationship (QSAR) models in predicting the acute toxicity of actual pesticide formulations (formulated products). Most existing QSAR models are designed for single chemical substances (pesticide active ingredient / technical grade) and cannot handle complex mixture systems, especially when adjuvant / formulation information is missing, failing to accurately predict the final toxicity. This invention uses a fusion modeling approach, integrating structural information, formulation information, and adjuvant-related knowledge of pesticide formulations, to predict the acute toxicity of pesticide formulations and automatically generate a system for registration / risk assessment / emergency reports. This technology can be widely applied to new product development by pesticide companies, safety assessments by pesticide registration agencies, and toxicological prediction services by third-party testing laboratories.

[0013] To address the aforementioned problems, the technical solution of this invention is a knowledge graph-based intelligent prediction system for pesticide formulation mixed toxicity, comprising: The prior knowledge graph construction module utilizes unstructured books or manuals and structured databases to establish the association between dosage form and excipient type through natural language processing technology. A dosage form-aware dual-stream deep learning prediction model is proposed, comprising a structural stream, a dosage form correction stream, and a fusion layer. The structural stream takes the 2D molecular fingerprint of the active ingredient as input, trains a graph neural network, and outputs the baseline toxicity of the active ingredient. The dosage form correction stream takes the percentage content of the active ingredient, the dosage form tag, and a set of potential adjuvant types corresponding to the dosage form retrieved from a knowledge graph as input. The fusion layer adds or concatenates the baseline toxicity with a dosage form correction factor, passes it through a fully connected layer and a Softmax / Sigmoid activation function, and outputs the final toxicity value of the formulation. The intelligent report and suggestion generation engine, based on the final toxicity prediction value output by the model, calls a preset rule base and knowledge base, and includes the following functions: registration report, toxicity classification, risk assessment suggestion, first aid suggestion, and registration suggestion.

[0014] Preferably, the prior knowledge graph is stored as a graph database structure, where nodes include active ingredients, dosage forms, and adjuvant types, and edges represent relationships such as membership and use.

[0015] Preferably, the registration report automatically generates a structured report including predicted values, confidence intervals, applicable domains, and comparison results of toxicity with existing original drugs.

[0016] Preferably, the toxicity classification automatically matches the pesticide toxicity classification standard to generate an acute toxicity level.

[0017] Preferably, the risk assessment recommendations are based on the exposure route and predicted values, and automatically generate risk management recommendations such as: "This preparation has high oral toxicity, and it is recommended to add warnings such as wearing protective clothing and gloves to the label."

[0018] Preferably, the first aid recommendations are automatically generated based on the predicted toxicity mechanism and the first aid knowledge base, such as: "If accidentally ingested, seek medical attention immediately. Do not induce vomiting to prevent suffocation."

[0019] Preferably, the registration recommendations are based on the predicted toxicity level, generating intelligent registration recommendations such as "This formulation is expected to require supplemental toxicological testing".

[0020] This invention also discloses a knowledge graph-based intelligent prediction method for pesticide formulation mixed toxicity, which, when applied to the above system, includes the following steps: Step 1: The user enters the CAS number, content percentage, and formulation code of the active ingredient in the pesticide formulation; Step 2: Extract the molecular fingerprint of the active ingredient using the system; Step 3: The system retrieves a list of standard adjuvant types for the dosage form from the knowledge graph based on the dosage form code; Step 4: Input all the information from Steps 1 to 3 into the two-stream deep learning model; Step 5: The model outputs the predicted LD50 value and toxicity classification of the formulation; Step Six: The prediction results are automatically transferred to the report generation engine to generate a professional report that includes risk assessment, first aid, and registration recommendations.

[0021] The advantages of this invention compared to existing technologies are: 1. Existing technologies require the complete chemical structure of the active ingredient, and sometimes specific excipients. This invention, however, only requires the structure, content, and dosage form code of the active ingredient, perfectly suited to enterprise data.

[0022] 2. Existing technologies perform well in predicting pure components, but have extremely low accuracy in predicting formulations (complex mixtures) and cannot reflect the effects of excipients. This invention significantly improves the prediction accuracy for formulations with unknown proportions by implicitly learning the "dosage form-excipient" association.

[0023] 3. Existing technologies typically only output numerical values ​​or a single classification label. This invention outputs numerical values ​​and toxicity classifications, and automatically generates professional reports covering risk assessment, first aid, and registration guidance.

[0024] 4. Existing technologies cannot utilize historical experimental data lacking detailed formulation information. This invention can efficiently utilize the large amount of historical data accumulated by enterprises regarding "active ingredients + dosage form + toxicity". Attached Figure Description

[0025] Figure 1 This is a system structure diagram of the present invention.

[0026] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0027] To make the content of this invention easier to understand, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings.

[0028] like Figure 1-2 As shown, the knowledge graph-based intelligent prediction system for pesticide formulation mixed toxicity includes: Prior knowledge graph construction module: This module utilizes unstructured books or manuals and structured databases, employing Natural Language Processing (NLP) technology to establish a "dosage form - excipient type" relationship. For example, it automatically identifies: Dosage form: emulsifiable concentrate (EC) → Potential adjuvants: solvents (xylene, toluene), emulsifiers (calcium alkylbenzene sulfonate, nonylphenol polyoxyethylene ether).

[0029] Dosage form: Water-dispersible granules (WG) → Potential adjuvants: Dispersants (polycarboxylate, naphthalene sulfonate), wetting agents (alkyl sulfate).

[0030] The graph is stored as a graph database structure, with nodes including "active ingredient", "dosage form" and "adjuvant type", and edges representing relationships such as "belonging to" and "used for".

[0031] Dosage form-aware dual-flow deep learning prediction model: This model consists of two parallel networks.

[0032] Flow 1 (Structure Flow): Input the 2D molecular fingerprint of the active ingredient (such as ECFP4, Morgan fingerprint), train a graph neural network (GNN), and output the "benchmark toxicity" of the active ingredient.

[0033] Flow 2 (Dosage Form Correction Flow): Input the percentage content of the active ingredient, the dosage form label (one-hot encoded), and the set of "potential adjuvant types" corresponding to the dosage form retrieved from the knowledge graph above. Using a graph attention network (GAT), the active ingredient features and dosage form features are used as queries to apply attention weights to the "potential adjuvant type" nodes, generating a "dosage form correction factor".

[0034] Fusion layer: The “baseline toxicity” and the “dosage form correction factor” are added or concatenated, and after passing through a fully connected layer and a Softmax / Sigmoid activation function, the final toxicity value of the formulation (such as LD50 value or toxicity grade) is output.

[0035] Intelligent report and suggestion generation engine: Based on the final toxicity prediction value output by the model, the preset rule base and knowledge base are invoked: Registration Report: Automatically generates a structured report that includes predicted values, confidence intervals, applicable domains, and comparisons with existing drug toxicity results.

[0036] Toxicity rating: Automatically matches pesticide toxicity rating standards to generate an "acute toxicity level".

[0037] Risk assessment recommendations: Based on the exposure route and predicted values, the system automatically generates risk management recommendations such as "This preparation has high oral toxicity, and it is recommended to add warnings such as 'wear protective clothing and gloves' to the label."

[0038] First aid advice: Based on the predicted toxicity mechanism (such as neurotoxicity), the system matches the first aid knowledge base and automatically generates first aid guidance such as "If accidentally ingested, seek medical attention immediately and do not induce vomiting to prevent choking."

[0039] Registration recommendations: Based on the predicted toxicity level, generate intelligent registration recommendations such as "This formulation is expected to require supplementary toxicological testing of class XX".

[0040] This invention also discloses a knowledge graph-based intelligent prediction method for pesticide formulation mixed toxicity, which, when applied to the above system, includes the following steps: Step 1: The user enters the CAS number, content percentage, and formulation code of the active ingredient in the pesticide formulation; Step 2: Extract the molecular fingerprint of the active ingredient using the system; Step 3: The system retrieves a list of standard adjuvant types for the dosage form from the knowledge graph based on the dosage form code; Step 4: Input all the information from Steps 1 to 3 into the two-stream deep learning model; Step 5: The model outputs the predicted LD50 value and toxicity classification of the formulation; Step Six: The prediction results are automatically transferred to the report generation engine to generate a professional report that includes risk assessment, first aid, and registration recommendations.

[0041] The present invention and its embodiments have been described above, and such description is not restrictive. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A knowledge graph-based intelligent prediction system for pesticide formulation mixed toxicity, characterized in that, include: The prior knowledge graph construction module utilizes unstructured books or manuals and structured databases to establish the association between dosage form and excipient type through natural language processing technology. A dosage form-aware dual-flow deep learning prediction model, comprising a structural flow, a dosage form correction flow, and a fusion flow, wherein the structural flow takes the 2D molecular fingerprint of the active ingredient as input, trains a graph neural network, and outputs the baseline toxicity of the active ingredient; The dosage form correction stream input includes the percentage of active ingredient content, the dosage form tag, and the set of potential adjuvant types corresponding to the dosage form retrieved from the knowledge graph; The fusion layer adds or splices the baseline toxicity with the dosage form correction factor, and outputs the final toxicity value of the formulation after passing through a fully connected layer and a Softmax / Sigmoid activation function. The intelligent report and suggestion generation engine, based on the final toxicity prediction value output by the model, calls a preset rule base and knowledge base, and includes the following functions: registration report, toxicity classification, risk assessment suggestion, first aid suggestion, and registration suggestion.

2. The intelligent prediction system for pesticide formulation mixed toxicity based on knowledge graphs according to claim 1, characterized in that: The prior knowledge graph is stored as a graph database structure, with nodes including active ingredients, dosage forms, and adjuvant types, and edges representing relationships such as membership and use.

3. The intelligent prediction system for pesticide formulation mixed toxicity based on knowledge graphs according to claim 1, characterized in that: The registration report automatically generates a structured report containing predicted values, confidence intervals, applicable domains, and comparisons with existing drug toxicity results.

4. The intelligent prediction system for pesticide formulation mixed toxicity based on knowledge graphs according to claim 1, characterized in that: The toxicity classification automatically matches the pesticide toxicity classification standard to generate an acute toxicity level.

5. The intelligent prediction system for pesticide formulation mixed toxicity based on knowledge graphs according to claim 1, characterized in that: The risk assessment recommendations are based on the exposure route and predicted values, and automatically generate risk management suggestions such as: "This preparation has high oral toxicity, and it is recommended to add warnings such as wearing protective clothing and gloves to the label." 6. The intelligent prediction system for pesticide formulation mixed toxicity based on knowledge graphs according to claim 1, characterized in that: The first aid recommendations are automatically generated based on the predicted toxicity mechanism and the first aid knowledge base, such as "If accidentally ingested, seek medical attention immediately. Do not induce vomiting to prevent suffocation." 7. The intelligent prediction system for pesticide formulation mixed toxicity based on knowledge graphs according to claim 1, characterized in that: The registration recommendations are based on the predicted toxicity level and generate intelligent registration suggestions such as "This formulation is expected to require supplemental toxicological testing".

8. A knowledge graph-based intelligent prediction method for pesticide formulation mixed toxicity, using the system described in any one of claims 1-7, characterized in that, Includes the following steps: Step 1: The user enters the CAS number, content percentage, and formulation code of the active ingredient in the pesticide formulation; Step 2: Extract the molecular fingerprint of the active ingredient using the system; Step 3: The system retrieves a list of standard adjuvant types for the dosage form from the knowledge graph based on the dosage form code; Step 4: Input all the information from Steps 1 to 3 into the two-stream deep learning model; Step 5: The model outputs the predicted LD50 value and toxicity classification of the formulation; Step Six: The prediction results are automatically transferred to the report generation engine to generate a professional report that includes risk assessment, first aid, and registration recommendations.

Images