Intelligent screening method for stability of antidepressants in wastewater

Abstract

The present application relates to a kind of intelligent screening method for the stability of antidepressant in wastewater, the molecular fingerprint data of antidepressant is calculated by RDkit tool, correlation test is carried out using Spearman correlation coefficient to screen input independent variable, based on the objective score of the physicochemical indexes of the biodegradability, water solubility and adsorption distribution of antidepressant in the relevant guide of economic cooperation and development organization, then based on total score, three classification is carried out as input dependent variable, based on Mixup data enhancement method, a kind of semi-supervised learning process suitable for small sample amount antidepressant is proposed, MSSL-RealFormer architecture is constructed, the rapid intelligent screening of the stability of antidepressant in wastewater is realized.The present application aims at, current traditional method can not be carried out on the comprehensive evaluation of the biomarker stability of the multiple kinds of antidepressants related to depressive symptoms in sewer, and the data overfitting problem existing in small sample classification, the optimization of antidepressant wastewater stability screening process is realized.

Description

Technical Field

[0001] This invention relates to an intelligent screening method for the stability of antidepressants in wastewater. Background Technology

[0002] Depression is projected to become the leading cause of global disease burden by 2030, and antidepressants (ADs) have significant therapeutic effects on the clinical symptoms of depression. ADs have been detected in aquatic environments (e.g., surface water, groundwater, drinking water, wastewater, or sewer systems) globally, in Europe, the Americas, Africa, and Oceania. Numerous studies have also highlighted that environmental concentrations of ADs and their metabolites (TPs) can significantly impact key life events in aquatic organisms, including reproduction, growth and development, oxidative stress, metabolism, neural function, endocrine function, and physiological behavior.

[0003] Wastewater-based epidemiology (WBE) is a method based on the analysis of specific biomarkers in sewage wastewater and has been successfully applied to retrospectively estimate regional consumption and exposure patterns of common pharmaceutical products. Studying the wastewater stability of adrenal inhibitors (ADs) associated with depressive symptoms helps reflect public mental health conditions and can further guide environmental regulatory strategies by assessing the aquatic biotoxicity of the parent drug and its total drug substances (TPs). WBE requires that the selected biomarkers remain stable in the sewage system. However, pharmaceutical products, as organic active ingredients, are significantly affected by the bioactivity of biofilms and sediments in the sewage system, and their loss in wastewater can be attributed to a combination of biodegradation, water solubility, and adsorption. Currently, the concentration of ADs in wastewater is often measured using bioanalytical methods such as solid-phase extraction (SPE), reversed-phase liquid chromatography, and tandem mass spectrometry (RPLC-MS / MS). However, these methods only focus on a limited number of drug types and have not yet provided a comprehensive assessment of the biomarker stability of multiple types of ADs associated with depressive symptoms in the sewage system. Furthermore, some potential biomarkers may exhibit false-positive bias during screening.

[0004] Studies indicate that applying deep learning (DL) methods to predict the stability of biomarkers in pharmaceutical wastewater systems may effectively improve screening efficiency and reduce false positive bias. Recently, deep learning (DL) has become one of the most powerful and promising tools in environmental chemical screening and public health applications. For example, using multilayer perceptrons (MLP) and one-way gated recurrent units (GRUs), the predictive power of community-level demographic data and wastewater characteristics combinations on COVID-19 and kidney disease can be investigated, with the best average Spearman correlation coefficient being 0.546. Using random forests (RF), weekly new hospitalizations due to COVID-19 in 45 US states can be predicted based on WBE wastewater monitoring data, with the model's normalized mean absolute error within 0.3. With the rapid development of attention mechanisms, the RealFormer method has more stable and regularized training effects compared to the original Transformer method, and can be used in the field of AI-assisted drug discovery. However, the few-shot classification learning involved in this study suffers from overfitting problems due to limited data and underfitting problems caused by the direct addition of a large number of unlabeled samples. Building high-performance models with limited data has become a research hotspot. Therefore, how to integrate WBE concepts with optimized DL methods to achieve rapid screening of the stability of ADs in the wastewater environment of sewer systems, explore ADs with biomarker potential suitable for WBE assessment, and optimize the drug stability screening process is a technical problem that researchers in this field urgently need to solve. Summary of the Invention

[0005] The technical problem this invention aims to solve is to provide an intelligent screening method for the stability of antidepressants in wastewater based on the MSSL-RealFormer deep learning architecture. The method uses the RDkit tool to calculate 210 molecular fingerprint data of 66 commonly used addicts (ADs) in clinical practice. Spearman correlation coefficient is used as the input independent variable for correlation testing. Based on OECD guidelines and relevant literature, a three-step screening framework for biomarker stability, encompassing biodegradability, water solubility, and adsorption / partition properties, is constructed. An independent additive method is used to objectively score and classify the above physicochemical property indicators as the input dependent variable, establishing a data matrix. Based on a semi-supervised learning (SSL) classification process, and by coupling a data augmentation (DE) strategy based on the Mixup method with a RealFormer structure based on multi-head attention, an MSSL-RealFormer deep learning architecture is proposed. This effectively avoids the problem of insufficient sample data and saves computational costs while maintaining good robustness, achieving rapid and intelligent screening of the stability of antidepressants in wastewater. This invention is implemented as follows: an intelligent screening method for the stability of antidepressants in wastewater based on the MSSL-RealFormer deep learning architecture:

[0006] Includes the following steps:

[0007] 1) Identify the types and molecular structures of commonly used adrenaline-addictive drugs (ADs) associated with depressive symptoms in clinical practice, using the EPI Suite developed by the U.S. Environmental Protection Agency (EPA). TM Software, the Eawag-PPS pathway prediction system, and VEGA QSAR software were used to predict biodegradability, with EPI Suite employed. TM Software predicts water solubility, and the EPA CompTox ChemicalsDashboard system is used to predict adsorption and partitioning.

[0008] 2) Based on relevant guidelines and literature from the Organization for Economic Cooperation and Development (OECD), establish screening criteria for assessable stability, objectively score the physicochemical indicators of ADs’ biodegradability, water solubility, and adsorption partitioning, calculate the total score of each physicochemical property using an independent additive method, and classify the total score into high stability, medium stability, and low stability as input dependent variables.

[0009] 3) The RDkit tool was used to calculate 210 molecular fingerprints of 66 ADs, and the Spearman correlation coefficient was used to test the correlation and evaluate the degree of correlation between different molecular fingerprints (<0.65). Forty non-linear molecular fingerprints were selected as input variables.

[0010] 4) Construct a data matrix for the deep learning architecture. Based on the DE strategy of the Mixup method and the RealFormer structure based on the multi-head attention mechanism, the fingerprint data of the original sample ADs is input into the RealFormer_1 structure for training. At the same time, Mixup is used to expand the initial fingerprint, and the expanded fingerprint data is input into the trained RealFormer_1 for label prediction and combined into a new training set to complete one MSSL process.

[0011] 5) Further, update the RealFormer_2 structure and repeat step 4), where the fingerprint data after secondary expansion is re-inputted into RealFormer_2 for label prediction and the training set is recombined to complete the secondary MSSL process;

[0012] 6) Repeat the MSSL process n times, output the model parameters and label prediction values, and complete the training. It is worth noting that after each update of the RealFormer model, the labels of the generated training set should be re-predicted to ensure the accuracy of the predicted labels.

[0013] In step 4), the specific process of Mixup is as follows:

[0014]

[0015] In the formula, x i With x j The original input fingerprint data of the i-th and j-th ADs are respectively, and the mixing ratio λ∈[0,1];

[0016] The first sub-layer of the Encoder in the RealFormer structure contains a multi-head attention mechanism transfer module, which aggregates the embedding layers (V) of a set of answers (K) to calculate the embedding layer (Q) of the output question. Based on the Transformer, a residual module (A) is added, as shown in the following formula:

[0017] ResidualMultiHead(Q,K,V,A)=Concat(head1,…,head h W O (2)

[0018] head i =ResidualAttention(QW i Q ,KW i K VW i V A i (3)

[0019] In the formula, W i Q W i K With W i V It is a matrix that linearly maps queries, keys, and values ​​to the "attention space" of the i-th head; A i It corresponds to head i A slice of A; W O It is a matrix that performs a linear transformation on the outputs of all heads; the residual attention module calculates a weighted sum of these values ​​as follows:

[0020]

[0021] In the formula, Q', K', V', and A' are the query, key, value, and residual scores of the current layer; finally, the new attention score... It is passed to the next sub-layer;

[0022] The gradient descent process can be accelerated by passing attention scores, but the risk of gradient vanishing is increased. Therefore, different learning rates should be selected in different MSSL processes to ensure the robustness of the model. The MSSL-RealFormer deep learning architecture proposed above effectively alleviates the problems of overfitting and underfitting caused by small sample size and improves the applicability of deep learning models.

[0023] Compared with existing technologies, the advantages of this invention are as follows: it can overcome the problems of overfitting with limited data and underfitting caused by the direct addition of a large number of unlabeled samples in the current traditional methods during the few-sample classification learning process. Through the MSSL-RealFormer deep learning architecture, a high-performance classification model can be built, realizing the rapid screening and comprehensive evaluation of the stability of ADs and other drugs related to mental illnesses in the sewage system wastewater environment, and optimizing the drug stability screening process. Attached Figure Description

[0024] Figure 1 It is an intelligent screening method for the stability of antidepressants in wastewater;

[0025] Figure 2 This is a schematic diagram of the MSSL-RealFormer deep learning architecture for intelligent screening of the stability of antidepressants in wastewater;

[0026] Figure 3 This is the evaluation result of the ADs wastewater stability classification model; Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0028] Example

[0029] See Figure 1 A smart screening method for the stability of antidepressants in wastewater, comprising:

[0030] 1) Identify the types and molecular structure information of commonly used adrenal insufficiency (AD) drugs associated with depressive symptoms, including molecular structural formulas and SMILES numbers. Biodegradability data are sourced from the EPI Suite developed by the U.S. Environmental Protection Agency (USEPA) and Syracuse Corporation (SRC Inc.). TM The anaerobic biodegradation index predicted by the software, the aerobic biodegradation probability predicted by the Eawag-PPS pathway prediction system (http: / / eawag-bbd.ethz.ch / ), and the water and sediment half-lives predicted by the VEGAQSAR software (https: / / www.vegahub.eu / portfolio-item / vega-qsar / ) were used. Water solubility data were obtained using EPI Suite. TM The water solubility, adsorption, and partitioning properties predicted by the software are the absorption coefficients of organic compounds in soil and sewage sludge and the partition coefficients of n-octanol and buffer solution at physiological pH 7.4, predicted using the EPACompTox Chemicals Dashboard system (http: / / www.comptox.epa.gov / dashboard / ).

[0031] 2) Based on relevant guidelines and literature from the Organization for Economic Cooperation and Development (OECD), establish screening criteria for assessable stability, objectively score the physicochemical indicators of ADs’ biodegradability, water solubility and adsorption partitioning, calculate the total score of each physicochemical property using an independent additive method, and classify the total scores into high stability, medium stability and low stability from high to low as input dependent variables;

[0032] 3) The RDkit tool (https: / / www.rdkit.org) was used to calculate 210 molecular fingerprints of 66 ADs, including molecular fingerprints representing information such as molecular physicochemical structure, spatial topology, chemical image and chemical fragments. The Spearman correlation coefficient was used to test the correlation. The correlation between different molecular fingerprints was evaluated by the correlation coefficient (<0.65) and its significance (P<5%). Forty nonlinear molecular fingerprints were selected as input variables.

[0033] 4) Using 40 molecular features of 66 ADs and their corresponding classification labels as a data matrix, the fingerprint data of the original sample ADs is input into the RealFormer_1 structure based on the multi-head attention mechanism for training. At the same time, the initial fingerprint is expanded by the Mixup method to increase sample diversity and improve the data generalization ability of the deep learning architecture. The expanded fingerprint data is then input into the trained RealFormer_1 for label prediction and combined into a new training set to complete one MSSL process.

[0034] 5) Further, update the RealFormer_2 structure and repeat step 4), where the fingerprint data after secondary expansion is re-inputted into RealFormer_2 for label prediction and the training set is recombined to complete the secondary MSSL process;

[0035] 1)6) Repeat the MSSL process n times, outputting model parameters and label predictions to complete training. It is worth noting that after each update of the RealFormer model, the labels of the generated training set must be re-predicted to ensure the accuracy of the predicted labels and the universality of the deep learning architecture.

[0036] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart screening method for the stability of antidepressants in wastewater, characterized in that, Includes the following steps: 1) Identify the types and molecular structures of commonly used adrenal insufficiency (AD) drugs associated with depressive symptoms in clinical practice, using the EPI Suite developed by the U.S. Environmental Protection Agency (EPA). TM Software, the Eawag-PPS pathway prediction system, and VEGA QSAR software were used to predict biodegradability, employing the EPI Suite. TM Software predicts water solubility, and the EPA CompTox ChemicalsDashboard system is used to predict adsorption and partitioning. 2) Based on relevant guidelines and literature from the Organization for Economic Cooperation and Development (OECD), establish screening criteria for assessable stability, objectively score the physicochemical indicators of ADs’ biodegradability, water solubility, and adsorption partitioning, calculate the total score of each physicochemical property using an independent additive method, and classify the total score into high stability, medium stability, and low stability as input dependent variables. 3) The RDkit tool was used to calculate 210 molecular fingerprints of 66 ADs, and the Spearman correlation coefficient was used to test the correlation between different molecular fingerprints to evaluate the degree of correlation between them. Forty non-linear molecular fingerprints were selected as input variables. 4) Construct a data matrix for the deep learning architecture. Based on the DE strategy of the Mixup method and the RealFormer structure based on the multi-head attention mechanism, the fingerprint data of the original sample ADs is input into the RealFormer_1 structure for training. At the same time, Mixup is used to expand the initial fingerprint, and the expanded fingerprint data is input into the trained RealFormer_1 for label prediction and combined into a new training set to complete one MSSL process. 5) Further, update the RealFormer_2 structure and repeat step 4), where the fingerprint data after secondary expansion is re-inputted into RealFormer_2 for label prediction and the training set is recombined to complete the secondary MSSL process; 6) Repeat the MSSL process n times, output the model parameters and label prediction values, and complete the training. It is worth noting that after each update of the RealFormer model, the labels of the generated training set should be re-predicted to ensure the accuracy of the predicted labels. In step 4), the specific process of Mixup is as follows: （1） In the formula, x i With x j The original input fingerprint data of the i-th and j-th ADs are respectively, and the mixing ratio λ∈[0,1]; The first sub-layer of the Encoder in the RealFormer structure contains a multi-head attention mechanism transfer module, which aggregates the embedding layers (V) of a set of answers (K) to calculate the embedding layer (Q) of the output question. Based on the Transformer, a residual module (A) is added, as shown in the following formula: ResidualMultiHead(Q,K,V,A)=Concat(head1 ,… ,head h )W O (2) （3） In the formula, , and It is a matrix that linearly maps queries, keys, and values ​​to the "attention space" of the i-th head; A i It corresponds to head i A slice of A; W O It is a matrix that performs a linear transformation on the outputs of all heads; the residual attention module calculates a weighted sum of these values ​​as follows: （4） In the formula, Q', K', V', and A' are the query, key, value, and residual scores of the current layer; finally, the new attention score... It is passed to the next sub-layer; The gradient descent process can be accelerated by passing attention scores, but the risk of gradient vanishing is increased. Therefore, different learning rates should be selected in different MSSL processes to ensure the robustness of the model. The MSSL-RealFormer deep learning architecture proposed above effectively alleviates the problems of overfitting and underfitting caused by small sample size and improves the applicability of deep learning models.

Images