Screening method and system for macrolide antibacterial candidate compounds based on probability calibration and graph neural network integration

Abstract

The present application relates to the technical field of bioinformatics, and discloses a screening method and system for macrolide antibacterial candidate compounds based on probability calibration and graph neural network integration, which comprises the following steps: obtaining molecular structure data of a known active compound set and dividing the data set; converting the training set into molecular fingerprints, training by using a random forest and introducing probability calibration to obtain a first prediction model; converting the training set into molecular graphs, training based on a directed message passing neural network to obtain a second prediction model; fusing the two models to obtain an integrated prediction model, determining a classification threshold by maximizing F1 of a validation set; inputting unknown compounds into the integrated model, outputting activity probability and comparing with the threshold to determine Go or No-Go candidates. The present application improves the activity prediction reliability and screening decision practicality under the small structural difference of homologues by integrating heterogeneous models and probability calibration, and is suitable for efficient discovery of veterinary macrolide antibacterial lead compounds.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, and in particular to a method and system for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration. Background Technology

[0002] Antimicrobial resistance has become a significant challenge in global public health and a core issue urgently needing resolution within the "One Health" framework. In veterinary clinical practice, macrolide antibiotics have long been used to prevent and treat respiratory infections and other bacterial diseases in animals due to their favorable tissue distribution, strong intracellular penetration, and certain immunomodulatory effects. However, with the continued use of macrolides in livestock farming and veterinary clinical practice, resistance to these drugs in animal-derived bacteria is constantly increasing, and the emergence and spread of resistant strains threaten the effectiveness of existing treatment regimens. Therefore, developing novel macrolide candidate compounds that can overcome existing resistance mechanisms and possess superior antimicrobial activity and application potential has become an important direction in veterinary antimicrobial drug research and development. Simultaneously, the global antimicrobial resistance crisis is accelerating the demand for new antimicrobial drugs. It is estimated that drug-resistant bacterial infections directly caused approximately 1.27 million deaths in 2019. However, the discovery of traditional antimicrobial drugs typically requires long research and development cycles and substantial financial investment, making it difficult to address the real-world pressure brought about by the rapid spread of resistance in a timely manner. Therefore, introducing computer-aided screening methods in the lead compound discovery stage, and using low-cost, high-throughput virtual screening techniques to preferentially discover potential active compounds in a large chemical space, has become an important technical approach for antibacterial drug development.

[0003] Quantitative structure-activity relationship (QSAR) models provide theoretical and technical support for the rapid screening of antibacterial candidate compounds by establishing a mapping relationship between molecular structure characterization and biological activity. In recent years, QSAR modeling methods have evolved from early linear models and shallow machine learning models to nonlinear prediction models that integrate molecular fingerprints, complex descriptors, and deep learning characterization, significantly improving the performance of molecular activity prediction. However, QSAR models still face several key challenges in screening macrolide derivatives. First, macrolide compounds typically have similar macrocyclic skeletons, and activity differences are often caused by subtle variations in local substituents, glycosyl modifications, polarity distribution, and spatial configuration. In these homologues or compounds with similar skeletons, minute structural differences can cause significant activity cliffs, making it difficult for models to stably predict their antibacterial activity. Second, publicly available activity data for veterinary antibacterial compounds often suffer from limited sample sizes, uneven sources, differences in experimental conditions, and unstable positive / negative ratios. These factors limit the model's ability to extrapolate to novel candidate compounds. Furthermore, many QSAR studies still focus primarily on overall evaluation metrics such as ROC-AUC and accuracy, while paying insufficient attention to issues such as the reliability of model output probabilities, whether prediction results can be directly translated into experimental selection decisions, and how to determine priority validation targets under limited experimental resources.

[0004] Therefore, it is necessary to develop a screening scheme for veterinary macrolide antibacterial candidate compounds that can overcome the above limitations and balance the reliability of prediction with the practicality of screening decisions. Summary of the Invention

[0005] This invention provides a method and system for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration, in order to overcome the shortcomings of the prior art.

[0006] This invention provides a method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration, comprising: Molecular structure data of a set of macrolide antibacterial candidate compounds with known antibacterial activity were obtained and divided into training set, validation set and independent test set. The antibacterial activity status included those with antibacterial activity and those without antibacterial activity. The molecular structure data of the training set is converted into molecular fingerprints. Using random forest decision trees, the model is trained to learn the molecular structure features of macrolide antibacterial candidate compounds with and without antibacterial activity. The model output is probabilistically calibrated during the model training process to obtain the first antibacterial activity prediction model. The molecular structure data of the training set is converted into molecular graphs, and a graph model based on a directed message passing neural network is constructed. The training model learns the structural characterization of macrolide antibacterial candidate compounds with and without antibacterial activity from the molecular graphs, thus obtaining a second antibacterial activity prediction model. The first and second antimicrobial activity prediction models are combined to obtain the third antimicrobial activity prediction model, and the classification threshold of the third antimicrobial activity prediction model is determined using the validation set. The molecular structure data of macrolide antimicrobial candidate compounds with unknown antimicrobial activity are obtained. The antimicrobial activity prediction results of macrolide antimicrobial candidate compounds with unknown antimicrobial activity are output through a third antimicrobial activity prediction model and its classification threshold.

[0007] The present invention provides a method for screening macrocyclic lactone antibacterial candidate compounds based on probabilistic calibration and graph neural network integration, which further includes the following steps: Molecular structure data of macrolide antibacterial candidate compounds with known antibacterial activity were preprocessed, including: The molecular structure data is standardized to obtain molecular structure data in SMILES format; Deduplication of molecular structure data in SMILES format; Structural resolvability testing and normalization were performed on molecular structure data in SMILES format.

[0008] According to the present invention, a method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration is provided. Specifically, the method for converting the molecular structure data of the training set into molecular fingerprints involves converting the molecular structure data of the training set into Morgan molecular fingerprints.

[0009] According to the present invention, a method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration is provided. The method involves converting the molecular structure data of the training set into molecular fingerprints, using a random forest decision tree to train a model to learn the molecular structure features of macrolide antibacterial candidate compounds with and without antibacterial activity, and performing probabilistic calibration on the model output during model training to obtain a first antibacterial activity prediction model. The method includes: During the model training phase, hierarchical K-fold cross-validation was used to generate outlier prediction probabilities for the first antibacterial activity prediction model. Using outlier prediction probability as input and actual antibacterial activity as output, a mapping relationship between the original prediction probability and the calibrated prediction probability is established. This is used to perform a forward transformation on the original prediction probability output by the first antibacterial activity prediction model during the model prediction stage, so that the final output of the first antibacterial activity prediction model is the calibrated prediction probability.

[0010] According to the present invention, a method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration is provided. The step of converting the molecular structure data of the training set into a molecular graph specifically involves converting the molecular structure data of the training set into a graph structure composed of atomic nodes and chemical bond edges. Message passing occurs along directed chemical bonds in the graph structure, enabling a second antibacterial activity prediction model to learn the contribution of the local chemical environment, substituent patterns, and topological relationships in the molecule to antibacterial activity.

[0011] According to the present invention, a method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration is provided, wherein the expression of the third antibacterial activity prediction model is: ens =α RF +β D-MPNN In the formula, RF This represents the calibrated predicted antibacterial activity probability of the first antibacterial activity prediction model for a certain macrolide antibacterial candidate compound. D-MPNN This indicates the probability of the second antibacterial activity prediction model predicting the antibacterial activity of the same macrolide antibacterial candidate compound. ens The third antimicrobial activity prediction model represents the probability of predicting the antimicrobial activity of the same macrolide antimicrobial candidate compound, while α and β represent the output weights of the first and second antimicrobial activity prediction models.

[0012] According to the present invention, a method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration, wherein determining the classification threshold of a third antibacterial activity prediction model using a validation set includes: The third antibacterial activity prediction model was used to predict the antibacterial activity of the molecular structure data in the validation set. Combined with the known antibacterial activity of macrolide antibacterial candidate compounds in the validation set, the F1 value of the third antibacterial activity prediction model under different preset candidate thresholds was obtained. The preset candidate threshold corresponding to the maximum F1 value was used as the classification threshold of the third antibacterial activity prediction model.

[0013] According to the present invention, a method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration includes: acquiring molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity; and outputting the antibacterial activity prediction results of macrolide antibacterial candidate compounds with unknown antibacterial activity through a third antibacterial activity prediction model and its classification threshold. The molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity are input into the third antibacterial activity prediction model to obtain the prediction probability of the antibacterial activity of the macrolide antibacterial candidate compounds with unknown antibacterial activity by the third antibacterial activity prediction model. When the third antimicrobial activity prediction model predicts the antimicrobial activity probability of macrolide antimicrobial candidate compounds with unknown antimicrobial activity as greater than or equal to the classification threshold, the antimicrobial activity prediction result of the macrolide antimicrobial candidate compounds with unknown antimicrobial activity is determined to be a Go candidate; otherwise, it is a No-Go candidate.

[0014] This invention also provides a screening system for macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration, comprising: The test data acquisition module is used to: acquire molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity; The screening module is used to: input the molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity into the third antibacterial activity prediction model obtained by any of the above methods, and obtain the antibacterial activity prediction results of macrolide antibacterial candidate compounds with unknown antibacterial activity based on the classification threshold of the third antibacterial activity prediction model.

[0015] The present invention also provides an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the computer program to implement any of the above-described methods for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described methods for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration.

[0017] The present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer is able to execute any of the above-described methods for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration.

[0018] The present invention provides a method and system for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration, which can bring at least the following beneficial effects: This invention integrates two heterogeneous models: random forest (based on Morgan's molecular fingerprint) and directed message-passing neural network (based on molecular graph). The former captures the distribution of substructure features of the molecule as a whole through global structural encoding in the form of fingerprints, while the latter aggregates atomic neighborhood information layer by layer through a message-passing mechanism along directed chemical bonds, enabling precise perception of the contribution of substituent connection patterns, local chemical environments, and topological relationships to antibacterial activity. The differences and complementarities between the two models in terms of representational dimensions make the integrated model (the third antibacterial activity prediction model) more sensitive to local changes in key pharmacophore groups in macrolide derivatives, effectively alleviating the prediction instability problem caused by activity mutations due to structural similarity in homologous systems, and improving the discriminability and reliability of screening results among structurally similar compounds.

[0019] This invention introduces a probability calibration mechanism during the training process of the first antibacterial activity prediction model, correcting the model output to a posterior estimate that more closely reflects the true positive rate. This allows the predicted probability to not only be used for ranking but also to provide a confidence assessment with practical reference value for experimental screening. In veterinary scenarios with limited and imbalanced data, the calibrated predicted probability output helps researchers more rationally determine the priority of candidate compound validation, avoiding misleading decisions due to inflated or understated probabilities.

[0020] This invention utilizes F1 scores at different preset candidate thresholds on a validation set to determine the optimal classification threshold with the goal of maximizing F1, transforming continuous probability outputs into explicit "Go / No-Go" binary decisions. This design closely aligns with the actual workflow of early screening for veterinary antimicrobial drugs—experimental validation resources are limited, and researchers need a clear and actionable list of recommended candidates to arrange subsequent wet experiments such as minimum inhibitory concentration (MIC) determination and cytotoxicity assessment. Through F1 threshold optimization, the model achieves a balance between precision and recall for practical screening scenarios. This avoids wasting resources by sending a large number of inactive compounds to subsequent validation due to excessively low thresholds, while also avoiding missing potential active lead compounds due to excessively high thresholds, significantly improving the engineering practicality and decision-making efficiency of the screening scheme.

[0021] The second antibacterial activity prediction model constructed in this invention uses a molecular diagram as input and performs directed message passing along the chemical bond direction, enabling each atomic node to receive directional information from adjacent chemical bonds and their connected atoms. This design allows the model to more specifically learn the influence of local structural patterns such as glycosyl side chains, ester substitution sites, and polar modifications on the macrolide ring in macrolide compounds on antibacterial activity, rather than relying solely on global statistical features. This directly helps in understanding the structure-activity relationship of veterinary macrolide compounds and provides an interpretable model basis for subsequent rational design and structural optimization.

[0022] This invention integrates a weighted model of antibacterial activity prediction based on molecular fingerprints with a second model based on molecular graphs, fusing global substructure counting features and local topological environment features to achieve complementarity at the representation level. The first antibacterial activity prediction model can quickly capture cyclic substructures and surrounding substitution patterns in molecules using extended connectivity fingerprints, while the second antibacterial activity prediction model excels at sensing the directionality of chemical bonds and the directional interactions between atoms. The integration of these two models retains the robustness of traditional machine learning in high-dimensional sparse features while leveraging the advantages of deep learning in local structure representation, forming a dedicated representation strategy for the skeletal characteristics of macrocyclic lactone molecules. This approach effectively balances prediction accuracy and computational efficiency, making it suitable for medium-scale virtual screening scenarios.

[0023] In summary, this invention addresses the core problems in the screening of candidate macrolide antibacterial compounds for veterinary use, such as difficulties in predicting the structure-activity relationship of homologues, unreliable models under limited data, and a disconnect between predicted results and experimental decisions. It provides a complete technical solution encompassing data preprocessing, probability calibration, graph neural network characterization, and integrated threshold decision-making. This solution significantly improves the reliability and operability of virtual screening results without relying on additional experimental data, providing a powerful computational tool for the early discovery of macrolide antibacterial drugs for veterinary use. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is a schematic flowchart of a method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration, provided by the present invention.

[0026] Figure 2 The overall discriminative and ranking capabilities of the three models on the independent test set are shown. RF represents the Random Forest model after Platt calibration, Ensemble represents the ensemble model, and Chemprop represents the D-MPNN model.

[0027] Figure 3 The relationship between the reliability curves of the three models and the ideal diagonal (red dashed line) is shown. RF represents the random forest model after Platt calibration, Ensemble represents the ensemble model, and Chemprop represents the D-MPNN model.

[0028] Figure 4The similarity graphs of the top 12 high-probability samples are shown (red indicates a positive contribution to the 'activity' prediction; blue indicates a negative contribution. The area of ​​the circle is proportional to the size of the contribution. In many cases, a consistent positive contribution pattern appears around aromatic rings, nitrogen-containing fragments, and hydrogen bond donor / acceptor fragments; prediction probability p≈0.95).

[0029] Figure 5 The similarity map of representative false positive samples (Top-FP) is shown (red indicates a positive contribution to the 'activity' prediction; blue indicates a negative contribution. The area of ​​the circle is proportional to the size of the contribution).

[0030] Figure 6 The similarity map of representative false negative samples (Top-FN) is shown (red indicates a positive contribution to 'activity' prediction; blue indicates a negative contribution). Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0032] Figure 1 This is a schematic flowchart illustrating a method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration, as provided by the present invention. The execution entity of this method can be any applicable terminal-side device or network-side device, such as a macrolide antibacterial candidate compound screening device based on probabilistic calibration and graph neural network integration.

[0033] See Figure 1 The present invention provides a method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration, which may include: S110. Obtain the molecular structure data of a set of macrolide antibacterial candidate compounds with known antibacterial activity, and divide them into a training set, a validation set, and an independent test set. Antibacterial activity is categorized as either having antibacterial activity (positive) or not having antibacterial activity (negative). Antibacterial activity is represented by a binary label. For the same compound with multiple activity records under different experimental conditions, this embodiment adopts the rule of "showing antibacterial activity under any experimental condition is judged as positive," constructing a unified activity label denoted as is_potent_any. This label reflects whether the candidate compound has the potential to enter subsequent priority validation.

[0034] In one embodiment, the molecular structure data of a set of macrolide antibacterial candidate compounds with known antibacterial activity can be preprocessed, wherein the preprocessing includes: The molecular structure data is standardized to obtain molecular structure data in SMILES format; Deduplication of molecular structure data in SMILES format; Structural resolvability testing and normalization were performed on molecular structure data in SMILES format.

[0035] To ensure data consistency and traceability, this embodiment uses standardized SMILES as the unique index for compounds. After data import, SMILES is first deduplicated to ensure that each compound corresponds to a unique molecular identifier in subsequent modeling, prediction, threshold determination, and result merging processes. Subsequently, cheminformatics tools are used to perform structural resolvability detection and normalization on SMILES. Structure records that cannot be resolved or cannot form valid molecular objects are removed before model training and evaluation. All subsequent operations, including model prediction probability merging, threshold classification, probability integration, and result backtracking, are performed based on the standardized SMILES index to reduce the risk of sample mismatch and information leakage. After data processing, the standardized dataset is divided into a training set, a validation set, and an independent test set. The training set is used for model parameter learning, the validation set is used for model selection and threshold determination, and the independent test set is used only for final performance evaluation. This division method avoids introducing test set information into the threshold selection process, making the model evaluation closer to real-world candidate compound screening scenarios.

[0036] S120. Convert the molecular structure data of the training set into Morgan molecular fingerprints. Use random forest decision trees to train the model to learn the molecular structure features of macrolide antibacterial candidate compounds with and without antibacterial activity. During the model training process, perform probability calibration on the model output to obtain the first antibacterial activity prediction model (also known as the random forest model).

[0037] To construct a baseline for traditional machine learning predictions, this embodiment first uses Morgan molecular fingerprints to encode the structure of candidate compounds. Morgan fingerprints can convert local molecular substructures into fixed-length binary vectors, making them suitable for characterizing macrocyclic lactones with similar structures but different substituents. This embodiment uses binary Morgan fingerprints with a radius of 2 and a length of 2048 bits as input features to the random forest model. Because these features inherently possess sparse binary properties, no additional feature scaling is performed.

[0038] Random forest classification models are used to learn the nonlinear relationship between molecular fingerprints and antibacterial activity tags. Compared to single decision tree models, random forests reduce model variance through the ensemble of multiple decision trees and typically exhibit better stability in tasks with limited sample sizes, imbalanced data distribution, and high molecular structural similarity. Therefore, this embodiment uses the random forest model as one of the foundational models for probability calibration and subsequent ensemble.

[0039] While conventional random forest models can output predicted probabilities of candidate compounds belonging to the positive category, these probabilities may not necessarily align with the actual frequency of activity. Directly using uncalibrated probabilities for candidate compound screening could result in insufficient actual hit rates for high-probability samples, thus impacting experimental resource allocation. Therefore, this embodiment performs Platt probability calibration on the output of the random forest model. Specifically, hierarchical K-fold cross-validation is used during the training phase to generate outlier prediction probabilities for the random forest model. Outlier prediction probabilities refer to the prediction results obtained for each training sample when it does not participate in the training of its corresponding sub-model, and can be used to reduce the risk of information leakage during probability calibration. Subsequently, using the outlier prediction probabilities as input and the actual activity labels as output, a logistic regression calibration model is trained to establish a mapping relationship between the original predicted probabilities and the calibrated probabilities.

[0040] On the validation set, independent test set, and subsequent candidate compound set, only the calibration mapping obtained during the training phase is used to forward transform the original predicted probabilities of the random forest to obtain the calibrated random forest activity probability, denoted as: p RF,cal , where p RF,cal This represents the predicted probability of antimicrobial activity from the random forest after Platt calibration. This probability is used for subsequent threshold decisions and probabilistic ensembles.

[0041] S130. Convert the molecular structure data of the training set into a molecular graph and construct a graph model based on a directed message passing neural network. Train the model to learn the structural characterization of macrocyclic lactone antibacterial candidate compounds with and without antibacterial activity from the molecular graph, and obtain the second antibacterial activity prediction model (also known as the D-MPNN model).

[0042] In one embodiment, the molecular structure data of the training set can be converted into a molecular graph. Specifically, the molecular structure data of the training set is converted into a graph structure composed of atomic nodes and chemical bond edges, and message passing is performed along the directed chemical bonds in the graph structure, so that the second antibacterial activity prediction model learns the contribution of the local chemical environment, substituent patterns and topological relationships in the molecule to antibacterial activity.

[0043] To further capture the impact of local substituent changes in macrolide compounds on antibacterial activity, this embodiment constructs a graph model based on a directed message-passing neural network. This model represents the molecule as a graph structure composed of atomic nodes and chemical bond edges, and performs message passing along directed chemical bonds to learn the contributions of local chemical environments, substituent patterns, and topological relationships within the molecule to antibacterial activity.

[0044] Compared to traditional models based on fixed fingerprints, directed message-passing neural networks can learn structural representations directly from molecular diagrams, without relying entirely on predefined artificial descriptors. Therefore, when faced with macrolide derivatives with the same macrocyclic skeleton but different substituent patterns, this model helps to identify activity differences caused by subtle structural variations.

[0045] To ensure comparability with the random forest model, the graph neural network model uses the exact same data partitioning and binary classification task settings. The model is trained using a molecular graph obtained from standardized SMILES transformation as input and `is_potent_any` as the prediction target. During training, the validation set loss is used as the model selection criterion, and the parameters of the model with the best validation performance are saved. Finally, the best model is used to predict on the validation set, test set, or candidate compounds to be screened, obtaining the antibacterial activity probability output by the graph neural network, denoted as p. D MPNN , where p D MPNN This represents the predicted probability of a candidate compound belonging to the antibacterial activity category by a directed message-passing neural network.

[0046] S140. The first and second antimicrobial activity prediction models are fused to obtain the third antimicrobial activity prediction model (also known as the ensemble model), and the classification threshold of the third antimicrobial activity prediction model is determined using the validation set.

[0047] Because random forest models and graph neural network models have different structural representation methods and prediction preferences, they are complementary in candidate compound screening. Probabilistically calibrated random forest models can provide more stable probability estimates, while graph neural network models are more sensitive to local structural changes and substituent differences, exhibiting better coverage of active compounds. Based on this consideration, this embodiment employs a probabilistic ensemble strategy to fuse the outputs of the two models.

[0048] Therefore, the expression for the third antibacterial activity prediction model can be: ens =α RF +β D-MPNN In the formula, RF This represents the calibrated predicted antibacterial activity probability of the first antibacterial activity prediction model for a certain macrolide antibacterial candidate compound. D-MPNN This indicates the probability of the second antibacterial activity prediction model predicting the antibacterial activity of the same macrolide antibacterial candidate compound. ens The third antimicrobial activity prediction model represents the predicted antimicrobial activity probability of the same macrolide antimicrobial candidate compound. α and β represent the output weights of the first and second antimicrobial activity prediction models, both of which are 0.5 in this embodiment. The reason for using equal-weighted average is that this setting does not introduce additional trainable parameters, can reduce the risk of overfitting under small sample conditions, and at the same time retains the complementary information of the two models. For subsequent practical applications, the equal-weighted average can also be extended to a weighted average strategy according to the performance of the validation set or the screening target. However, in order to ensure the simplicity and repeatability of the process, this embodiment adopts the equal-weighted fusion method as the main integration strategy.

[0049] To avoid leakage of test set information, the classification thresholds for all models in this embodiment are determined solely based on the validation set, without adjustment using an independent test set. Specifically, for the random forest model, graph neural network model, and ensemble model, candidate thresholds are iterated over on the validation set, and precision, recall, and F1 scores are calculated for different thresholds. The threshold that maximizes the F1 score on the validation set is ultimately selected as the fixed classification threshold for the corresponding model.

[0050] After determining the threshold, it is directly applied to the independent test set. If the predicted probability of a candidate compound is greater than or equal to the corresponding threshold, it is classified as a Go candidate, indicating that the compound has the value of being prioritized for subsequent experimental verification; if the predicted probability is lower than the threshold, it is classified as a No-Go candidate, indicating that it is not considered as a priority for verification. In this way, the model output is no longer limited to simple probability scores, but is further transformed into an executable candidate compound screening decision.

[0051] S150. Obtain the molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity, and output the antibacterial activity prediction results of macrolide antibacterial candidate compounds with unknown antibacterial activity through the third antibacterial activity prediction model and its classification threshold.

[0052] In one embodiment, S150 may include: The molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity are input into the third antibacterial activity prediction model to obtain the prediction probability of the antibacterial activity of the macrolide antibacterial candidate compounds with unknown antibacterial activity by the third antibacterial activity prediction model. When the third antimicrobial activity prediction model predicts the antimicrobial activity probability of macrolide antimicrobial candidate compounds with unknown antimicrobial activity as greater than or equal to the classification threshold, the antimicrobial activity prediction result of the macrolide antimicrobial candidate compounds with unknown antimicrobial activity is determined to be a Go candidate; otherwise, it is a No-Go candidate.

[0053] To comprehensively evaluate the performance of different models in screening macrolide antibacterial candidate compounds, this embodiment evaluates them from four dimensions: ranking ability, probability reliability, threshold classification ability, and early enrichment ability.

[0054] First, the ROC-AUC model was used to evaluate its overall ability to distinguish between positive and negative candidate compounds. Then, PR-AUC, expressed as average precision, was used to evaluate the model's ability to rank active compounds when the proportion of positive samples was high or the class distribution was unbalanced.

[0055] Secondly, the Brier score is used to evaluate the overall error of the predicted probabilities. A lower Brier score indicates a smaller mean squared error between the model's output probabilities and the true labels. Furthermore, Expected Calibration Error and Maximum Calibration Error are used to evaluate the quality of probability calibration. Expected Calibration Error measures the overall deviation between the predicted probabilities and the actual positive rate across different probability intervals, while Maximum Calibration Error reflects the maximum deviation within the worst-case probability interval.

[0056] Next, based on a fixed threshold determined on the validation set, precision, recall, F1 score, accuracy, and confusion matrix (including true negative, false positive, false negative, and true positive) are calculated on the independent test set. These metrics are used to evaluate the model's ability to translate predicted probabilities into actual classification decisions.

[0057] Finally, to simulate a scenario of prioritizing the validation of candidate compounds under limited experimental resources, this embodiment performs early enrichment analysis. The test set compounds are sorted from highest to lowest predicted probability, and the Precision@k and enrichment factor are calculated for the Top 1%, Top 2%, Top 5%, Top 10%, and Top 20% of candidate compounds, respectively. The enrichment factor calculation formula is as follows: EF@k=Precision@k / Prevalence Here, Precision@k represents the proportion of positive compounds among the top k candidate compounds, and Prevalence represents the overall proportion of positive samples in the test set. This metric is used to evaluate whether the model can enrich more truly active compounds at the top of the candidate list, thereby improving screening efficiency under limited experimental conditions.

[0058] To evaluate the stability of the model performance, this embodiment employs the Bootstrap resampling method to perform uncertainty analysis on key metrics. 3000 resampling operations with replacement are performed on the independent test set. Each resampling yields a sample set of the same size as the original test set, and the ROC-AUC, mean precision, and Precision@10% are recalculated. Subsequently, 95% confidence intervals are calculated based on the 2.5th and 97.5th percentiles of the Bootstrap distribution, and the Bootstrap mean and point estimates are reported.

[0059] To visually demonstrate model performance, this embodiment plots ROC curves, precision-recall curves, probability calibration curves, confusion matrix heatmaps, and early enrichment plots. The probability calibration curve uses the average predicted probability as the x-axis and the actual positive rate as the y-axis, with an ideal diagonal line added as a reference. The confusion matrix displays both percentage and absolute sample number, and labels the fixed screening threshold used by the corresponding model. The early enrichment plot shows Precision@k and enrichment factors at different Top-k levels, using the overall positive rate of the test set as a baseline.

[0060] To further improve the interpretability of the model results and provide a basis for subsequent optimization of candidate compound structures, this embodiment performs atomic-level or fragment-level structural contribution analysis on some representative samples. The analysis includes high-confidence true positive samples, representative false positive samples, and representative false negative samples.

[0061] For high-confidence true positive samples, the focus is on identifying their common activity-promoting regions, such as nitrogen-containing groups, amino sugar fragments, hydrophobic aromatic regions, or structural units capable of forming suitable amphiphilic distributions. For false positive samples, the focus is on analyzing whether there is a mismatch between local activity-promoting signals and overall structural inhibitory factors, such as excessive steric hindrance, unbalanced polarity distribution, or overly large side chain structures. For false negative samples, the focus is on analyzing whether they are located at the boundary of the model's applicable domain or whether they contain special structural fragments that are insufficiently covered in the training set.

[0062] Through the structural contribution analysis described above, this embodiment further transforms model prediction errors into interpretable structural information, providing a basis for reducing false positives and false negatives, and designing in-domain probe compounds. This step also enables the screening process established in this embodiment to not only have candidate ranking capabilities but also provide direction for subsequent structural optimization and closed-loop iteration.

[0063] This embodiment focuses on the priority screening requirement of macrolide antibacterial candidate compounds, constructing and comparing three prediction strategies: a random forest model based on Morgan molecular fingerprints and calibrated with Platt probabilities (the first antibacterial activity prediction model), a directed message-passing neural network model based on molecular graph structure (the second antibacterial activity prediction model), and an equal-weighted probability ensemble model of the two (the third antibacterial activity prediction model). All three models use the same data partitioning and the same binary classification labels to ensure that differences in model performance mainly stem from differences in molecular characterization methods, probability processing methods, and model structure, rather than differences in data partitioning.

[0064] Unlike simply evaluating model classification accuracy, this embodiment focuses on whether the model output can further serve the candidate compound screening decision. Therefore, the results analysis is conducted from five aspects: overall ranking ability, probability calibration quality, threshold-dependent classification performance, early enrichment ability, and structural contribution interpretation. The independent test set contained 220 compounds, of which 170 were positive samples, with an overall positive rate of 0.773. This data distribution suggests that when evaluating a model, relying solely on accuracy cannot fully reflect the model's value in practical screening scenarios; it is also necessary to combine accuracy, recall, F1 score, probability calibration error, and Top-k enrichment ability for a comprehensive judgment.

[0065] The overall discrimination and ranking capabilities of the three models on independent test sets are as follows: Figure 2As shown in Table 1, the main evaluation metrics are as follows. From the ROC-AUC perspective, the Platt-calibrated random forest model (RF (sigmoid-calibrated)) achieved the highest value of 0.818, higher than the ensemble model (RF+Chemprop, mean) at 0.784 and the D-MPNN model (Chemprop D-MPNN) at 0.667. This indicates that, under the current data scale and compound structure distribution conditions, the Morgan fingerprint-based random forest model has a stronger overall ability to distinguish between positive and negative compounds.

[0066] Table 1

[0067] TN: true negatives, FP: false positives, FN: false negatives, TP: truepositive The average accuracy also showed a similar trend. The AP value of the random forest model was 0.939, the ensemble model was 0.923, and the D-MPNN model was 0.861. Since AP focuses more on the enrichment of positive compounds at the beginning of the predicted ranking, this result indicates that the calibrated random forest model has a strong advantage in overall ranking, and can better rank truly active compounds at the beginning of the candidate list.

[0068] However, from a screening decision-making perspective, overall ranking ability does not fully represent the practical application value of the model. For screening macrolide antibacterial candidate compounds, the key issue is not only whether the model can distinguish between active and inactive compounds, but also whether the model can capture as many truly active compounds as possible under a given threshold, while controlling the number of false positives. Therefore, this embodiment further determines a fixed classification threshold for each model based on the validation set and evaluates its threshold-dependent classification performance on an independent test set.

[0069] To simulate the decision-making process in real-world candidate compound screening, this embodiment did not readjust the classification thresholds on the test set. Instead, it determined the classification thresholds for each model solely based on the F1 score of the validation set and fixed them for evaluation on the independent test set. The validation set source thresholds for the three models were: Random Forest model 0.690558, D-MPNN model 0.681, and ensemble model 0.617.

[0070] At this threshold setting, the random forest model exhibits high precision, reaching 0.870, indicating a high proportion of truly active compounds among the candidate compounds identified as "Go" by the model. Its recall is 0.829, and its F1 score is 0.849. The confusion matrix results show that it correctly identified 141 positive compounds while generating 21 false positives and 29 false negatives. These results demonstrate that the probability-calibrated random forest model is more suitable as a precision-prioritized screening tool, especially in scenarios with limited experimental resources where minimizing invalid validation samples is desired.

[0071] The D-MPNN model exhibits a different error structure. Its recall reaches 0.912, significantly higher than the random forest model, indicating its ability to identify more truly active compounds. However, its precision is only 0.795, with 40 false positives, suggesting that while improving coverage of active compounds, the model is also more prone to misclassifying some inactive compounds as positive. Its F1 score is 0.849, close to the random forest model, but their decision preferences differ: the random forest model prioritizes high precision, while the D-MPNN model prioritizes high recall.

[0072] The ensemble model achieves a better balance between the two types of models. Its recall is the same as the D-MPNN model, both at 0.912, indicating that the ensemble model retains the high coverage of active compounds by the graph neural network. Meanwhile, its precision is improved to 0.833, higher than the D-MPNN model, indicating that probabilistic fusion can reduce the false positive risk introduced by the deep model to some extent. Ultimately, the ensemble model achieves the highest F1 score of 0.871 and an accuracy of 0.791, demonstrating overall decision-making performance superior to the two individual models.

[0073] From the perspective of candidate compound screening, these results demonstrate that the random forest model and the D-MPNN model are not simply duplicates, but rather complementary. The random forest model provides more stable and conservative probabilistic judgments, the D-MPNN model improves the coverage of active candidates, and the ensemble model combines the advantages of both through probabilistic fusion, making it more suitable for Go / No-Go decision-making for macrolide antibacterial candidate compounds.

[0074] Besides classification performance, the reliability of the predicted probabilities is a crucial foundation for determining whether a model can be used for screening candidate compounds. If a high probability output by the model does not correspond to a high proportion of actual activity, then that probability is difficult to directly guide experimental verification. Therefore, this embodiment further uses the Brier score, Expected Calibration Error, and Maximum Calibration Error to evaluate the probabilistic quality of the model.

[0075] The results show that the calibrated random forest model has the best probabilistic reliability. Its Brier score is 0.138, ECE is 0.046, and MCE is 0.109, all outperforming the D-MPNN model. See [link to relevant documentation] Figure 3 As can be seen from the probability calibration curves, the reliability curve of the random forest model is generally close to the ideal diagonal, indicating a good consistency between its predicted probability and the actual positive rate. Especially in the high probability range, the probability output by the model can reliably reflect the likelihood of candidate compounds becoming active molecules.

[0076] In contrast, the D-MPNN model showed a Brier score of 0.165, an ECE of 0.090, and an MCE of 0.245, indicating a more significant probability bias. This result suggests that while the D-MPNN model has a high recall rate, its output probabilities tend to be overconfident, especially in the medium-to-high probability region, where it may overestimate the true activity probability of candidate compounds. Therefore, directly using uncalibrated or unintegrated D-MPNN probabilities as the screening criterion may lead to a large number of false positives entering subsequent experimental validation.

[0077] The ensemble model achieved a Brier score of 0.144, an ECE of 0.065, and an MCE of 0.139, placing its probabilistic calibration performance between that of the random forest and D-MPNN models. While the probabilistic reliability of the ensemble model was slightly lower than that of the calibrated random forest model, it was significantly better than the standalone D-MPNN model. Furthermore, the ensemble model performed better in terms of recall and F1 score. These results demonstrate that probabilistic ensembles can improve the model's coverage of real active compounds while maintaining good probabilistic reliability.

[0078] Based on the validation set, the ensemble model was fixed at a threshold of 0.617, and its confusion matrix results on the independent test set were further analyzed. The ensemble model correctly identified 155 positive compounds with 15 false negatives; among the 50 negative compounds, it correctly identified 19 negative samples with 31 false positives.

[0079] These results demonstrate that the ensemble model possesses a high capture capability for active compounds, able to identify the vast majority of true positive samples. This characteristic is of practical significance for screening candidate antimicrobial compounds, as missing potential active lead compounds in the early screening stages could result in lost opportunities for subsequent research and development. The high recall of the ensemble model indicates its suitability for expanding the coverage of active candidates.

[0080] Meanwhile, the ensemble model still exhibits a certain number of false positives, indicating that at the current threshold, the model tends to increase the coverage of active compounds rather than maximally excluding all negative samples. This result reflects a common trade-off in screening strategies: increasing recall usually leads to a decrease in precision, while increasing precision may increase the risk of false negatives. Compared to the D-MPNN model, the ensemble model reduces the number of false positives and increases precision while maintaining the same recall; compared to the random forest model, the ensemble model reduces the number of false negatives and increases the coverage of active compounds. Therefore, the ensemble model achieves a better overall balance at the screening decision level.

[0081] To evaluate the stability of the model's performance, this embodiment uses the Bootstrap resampling method to perform 3000 resampling cycles with replacement on the independent test set and repeatedly calculates key performance indicators. The results show that the model's ROC-AUC distribution is mainly concentrated in the range of 0.80–0.85, the average precision is mainly concentrated in the range of 0.93–0.96, and the Precision@10% is close to 1.0 in most resampled datasets.

[0082] The results above demonstrate that the model exhibits good stability in the high-confidence candidate compound ranking region. In other words, although different resampled datasets may cause some fluctuations in the overall evaluation metric, the model's enrichment ability for top-ranked candidate compounds remains relatively stable. This result is particularly important for practical applications because early screening of antibacterial candidate compounds typically does not require immediate validation of all candidate molecules, but rather prioritizes the validation of a small subset of top-ranked compounds. Therefore, the stability of Precision@10% directly reflects the model's screening reliability under resource-constrained conditions.

[0083] To further simulate the scenario of prioritizing the validation of candidate compounds under limited experimental resources, this embodiment performs Top-k early enrichment analysis on different models. After sorting the test set compounds from high to low predicted probabilities, Precision@k and enrichment factors are calculated for the Top 1%, Top 2%, Top 5%, Top 10%, and Top 20% ranges, respectively.

[0084] The results show that the random forest model maintains 100% precision in the top 1% to top 10% ranking range, meaning that all the top-ranked candidate compounds are truly active compounds; in the top 20% range, its precision still reaches approximately 97.7%. Considering the overall positive rate of the test set is 0.773, the enrichment factor of the model in the top 5% to top 10% range is approximately 1.26–1.29. These results indicate that the model can concentrate more truly active compounds at the forefront of the candidate list, thus significantly improving the hit rate of early experimental validation.

[0085] From an application perspective, Top-k early enrichment results demonstrate the practical value of screening methods more effectively than simple AUC or Accuracy. When experimental conditions only allow for the priority validation of a few candidate compounds, the model's ability to enrich truly active molecules in the top 5% or top 10% of the candidate list directly determines the efficiency of experimental resource utilization. The results of this embodiment show that the calibrated random forest model has a strong ability to rank high-confidence candidates, while the ensemble model provides a better balance between Recall and F1 in the overall screening decision. Combining the two can provide a flexible basis for prioritizing the validation of candidate compounds under different experimental resource conditions.

[0086] To further transform statistical predictions into interpretable chemical information, this embodiment performs atomic-level and fragment-level structural contribution analysis on representative samples. The analysis includes true positive samples with high prediction probability and experimentally confirmed activity, representative false positive samples, and representative false negative samples. See also... Figures 4-6 In the structure contribution diagram, the red area represents the structural fragments that make a positive contribution to the prediction of the activity class, and the blue area represents the structural fragments that make a negative contribution to the prediction of the activity class. The size of the circle indicates the intensity of the contribution.

[0087] In the high-confidence true positive samples, the structure contribution map showed relatively consistent and stable activity-enhancing regions. These regions were typically located near basic amine groups, aminoglycoside fragments, nitrogen-containing groups, and amphiphilic structural units adjacent to hydrophobic aromatic regions. This result suggests that a moderate amphiphilic distribution, locally basic structures, and restricted conformations may favor the antibacterial activity of macrolide candidates. These high-confidence true positive samples were mostly concentrated in regions with high model prediction probabilities, consistent with the good predictive reliability of high-probability regions in the probability calibration curve, and also corroborated by the early Top-k enrichment results.

[0088] Representative false-positive samples exhibited a clear "positive-negative contribution conflict." While some compounds contained local activity-enhancing fragments, such as nitrogen-containing aromatic structures or potential pharmacophore regions, they also possessed significant steric hindrance, uneven polarity distribution, or large side-chain structures. This caused the local positive contribution signal to be offset by the negative contribution factors in the overall structure. In such samples, the model may over-rely on the local pharmacophore signal, thereby overestimating its overall antibacterial activity. These results suggest that subsequent structure optimization should prioritize reducing steric hindrance, adjusting polarity distribution, and optimizing local substituent patterns to reduce the risk of similar false positives.

[0089] False negative samples often exhibit extensive negative contribution coverage of key aromatic rings, heterocycles, or local structural regions, while the activity-promoting signals are weak or scattered. Some false negative compounds may be located at chemical spatial boundaries where training data coverage is insufficient, causing the model to underestimate their true activity. For these samples, simply lowering the classification threshold may reduce false negatives, but it will also increase the number of false positives. A more reasonable strategy is to design a small number of in-domain probe compounds around their adjacent structural spaces, and then retrain or recalibrate the model after supplementing experimental data, thereby improving the model's ability to recognize boundary structures.

[0090] Overall, the results of this embodiment demonstrate that the macrolide antibacterial candidate compound screening method based on probabilistic calibration and graph neural network integration has good predictive performance and screening application value. The calibrated random forest model performs best in terms of overall ranking ability and probabilistic reliability, making it suitable for providing a high-precision, low-error probabilistic baseline; the D-MPNN model has high recall ability and can capture more potential active compounds; after probabilistic integration, the two methods improve precision while maintaining high recall, ultimately obtaining the highest F1 score, reflecting a better decision-making balance.

[0091] Furthermore, Bootstrap resampling and Top-k early enrichment analysis further demonstrate that this method exhibits good stability and hit rate in high-ranking candidate compound regions. Structural contribution analysis connects the model prediction results with interpretable chemical structural information, providing a basis for reducing false positives, lowering false negatives, and conducting structural optimization. These results collectively demonstrate that the screening process established in this embodiment can not only predict the activity of candidate compounds but also further support experimental verification of priority ranking, Go / No-Go decision-making, and structural optimization feedback, providing a relatively complete methodological foundation for the intelligent screening of macrolide antibacterial candidate compounds.

[0092] This embodiment addresses the need for efficient screening of macrolide antibacterial candidate compounds by establishing a decision-oriented QSAR screening method based on probabilistic calibration and graph neural network integration. This method, grounded in standardized SMILES structures and antibacterial activity tags, integrates Morgan molecular fingerprinting, random forest models, Platt probabilistic calibration, directed message passing neural networks, probabilistic ensemble, validation set threshold determination, Top-k priority ranking, and structural contribution analysis. This achieves a continuous transformation from molecular structure input to Go / No-Go decision output for candidate compounds.

[0093] The results show that the Platt-calibrated random forest model has good probabilistic reliability and overall ranking ability, with low Brier score and Expected Calibration Error, indicating that the model can provide a relatively stable probabilistic baseline for candidate compound screening. The D-MPNN model exhibits high recall ability for active compounds, capturing more potential positive candidates, but its probabilistic calibration quality and precision are relatively insufficient. By integrating the two models with equal-weighted probabilities, the overall decision balance is improved while maintaining high recall, resulting in the highest F1 score, demonstrating good complementarity between the probabilistic calibration model and the graph neural network model.

[0094] Further bootstrap resampling and Top-k early enrichment analysis showed that the screening process has good stability and hit rate in high-ranking candidate compound regions, concentrating more truly active compounds at the front of the priority validation list, thereby improving screening efficiency under limited experimental resources. Structural contribution analysis further revealed key molecular regions behind the model predictions, which can be used to identify activity-enhancing fragments, explain the sources of false positives and false negatives, and provide a basis for subsequent reduction of steric hindrance, optimization of polarity distribution, and design of in-domain probe compounds.

[0095] In summary, the method proposed in this embodiment can not only predict the activity of macrolide antibacterial candidate compounds, but also further support candidate prioritization, screening threshold decision-making, and structural optimization feedback. Compared with a single QSAR model, this process emphasizes the transformation of predicted probabilities into experimental screening decisions, exhibiting better reproducibility, interpretability, and application expansion potential. Further integration with MIC validation of novel external macrolide derivatives, application domain constraints, and experimental feedback iteration could potentially develop this method into an intelligent screening tool for discovering lead antibacterial compounds for veterinary use, and provide a foundation for the further transformation of related methodological achievements.

[0096] All computational analyses in this embodiment were performed in a separate conda environment. The main software and tools included Python, RDKit, Scikit-learn, Pandas, NumPy, PyTorch, PyTorch Lightning, and Chemprop. RDKit was used for molecular structure resolution, normalization, and Morgan fingerprint calculation; Scikit-learn was used for random forest model construction, probability calibration, and calculation of common evaluation metrics; Chemprop was used for training and prediction of directed message-passing neural network models; Pandas and NumPy were used for data preparation, matrix calculations, and result merging; and PyTorch and PyTorch Lightning were used for deep learning model training.

[0097] To ensure reproducibility, this embodiment uses a fixed random seed during model training and data partitioning, and saves training logs, model parameters, configuration files, prediction probabilities, threshold results, and evaluation metrics. All models are compared using the same training, validation, and independent test sets to ensure that performance differences between models primarily stem from model structure and probability processing methods, rather than differences in data partitioning. This embodiment only performs calculations and analyses based on existing compound structures and antibacterial activity data, and does not involve any new human or animal experiments.

[0098] The following describes the macrolide antibacterial candidate compound screening system based on probabilistic calibration and graph neural network integration provided by the present invention. The macrolide antibacterial candidate compound screening system based on probabilistic calibration and graph neural network integration described below can be referred to in correspondence with the macrolide antibacterial candidate compound screening method based on probabilistic calibration and graph neural network integration described above.

[0099] This invention provides a screening system for macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration, which may include: The test data acquisition module is used to: acquire molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity; The screening module is used to: input the molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity into the third antibacterial activity prediction model obtained by any of the above methods, and obtain the antibacterial activity prediction results of macrolide antibacterial candidate compounds with unknown antibacterial activity based on the classification threshold of the third antibacterial activity prediction model.

[0100] The electronic device provided by this invention may include: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory to execute the following steps: To obtain molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity; The molecular structure data of macrolide antimicrobial candidate compounds with unknown antimicrobial activity are input into the third antimicrobial activity prediction model obtained by any of the above methods, and the antimicrobial activity prediction results of macrolide antimicrobial candidate compounds with unknown antimicrobial activity are obtained according to the classification threshold of the third antimicrobial activity prediction model.

[0101] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0102] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, and the computer program being executed by a processor, enabling the computer to perform the following steps: To obtain molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity; The molecular structure data of macrolide antimicrobial candidate compounds with unknown antimicrobial activity are input into the third antimicrobial activity prediction model obtained by any of the above methods, and the antimicrobial activity prediction results of macrolide antimicrobial candidate compounds with unknown antimicrobial activity are obtained according to the classification threshold of the third antimicrobial activity prediction model.

[0103] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps: To obtain molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity; The molecular structure data of macrolide antimicrobial candidate compounds with unknown antimicrobial activity are input into the third antimicrobial activity prediction model obtained by any of the above methods, and the antimicrobial activity prediction results of macrolide antimicrobial candidate compounds with unknown antimicrobial activity are obtained according to the classification threshold of the third antimicrobial activity prediction model.

[0104] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0105] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration, characterized in that, include: Molecular structure data of a set of macrolide antibacterial candidate compounds with known antibacterial activity were obtained and divided into training set, validation set and independent test set. The antibacterial activity status included those with antibacterial activity and those without antibacterial activity. The molecular structure data of the training set is converted into molecular fingerprints. Using random forest decision trees, the model is trained to learn the molecular structure features of macrolide antibacterial candidate compounds with and without antibacterial activity. The model output is probabilistically calibrated during the model training process to obtain the first antibacterial activity prediction model. The molecular structure data of the training set is converted into molecular graphs, and a graph model based on a directed message passing neural network is constructed. The training model learns the structural characterization of macrolide antibacterial candidate compounds with and without antibacterial activity from the molecular graphs, thus obtaining a second antibacterial activity prediction model. The first and second antimicrobial activity prediction models are combined to obtain the third antimicrobial activity prediction model, and the classification threshold of the third antimicrobial activity prediction model is determined using the validation set. The molecular structure data of macrolide antimicrobial candidate compounds with unknown antimicrobial activity are obtained. The antimicrobial activity prediction results of macrolide antimicrobial candidate compounds with unknown antimicrobial activity are output through a third antimicrobial activity prediction model and its classification threshold.

2. The method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration according to claim 1, characterized in that, It also includes the following steps: Molecular structure data of macrolide antibacterial candidate compounds with known antibacterial activity were preprocessed, including: The molecular structure data is standardized to obtain molecular structure data in SMILES format; Deduplication of molecular structure data in SMILES format; Structural resolvability testing and normalization were performed on molecular structure data in SMILES format.

3. The method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration according to claim 1, characterized in that, The process involves converting the molecular structure data of the training set into molecular fingerprints, using a random forest decision tree to train a model that learns the molecular structure features of macrolide antibacterial candidate compounds with and without antibacterial activity, and performing probability calibration on the model output during training to obtain a first antibacterial activity prediction model, including: During the model training phase, hierarchical K-fold cross-validation was used to generate outlier prediction probabilities for the first antibacterial activity prediction model. Using outlier prediction probability as input and actual antibacterial activity as output, a mapping relationship between the original prediction probability and the calibrated prediction probability is established. This is used to perform a forward transformation on the original prediction probability output by the first antibacterial activity prediction model during the model prediction stage, so that the final output of the first antibacterial activity prediction model is the calibrated prediction probability.

4. The method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration according to claim 3, characterized in that, The process of converting the molecular structure data of the training set into a molecular graph specifically involves converting the molecular structure data of the training set into a graph structure composed of atomic nodes and chemical bond edges, and message passing along directed chemical bonds in the graph structure, so that the second antibacterial activity prediction model learns the contribution of the local chemical environment, substituent patterns and topological relationships in the molecule to antibacterial activity.

5. The method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration according to claim 4, characterized in that, The expression for the third antibacterial activity prediction model is as follows: ens =α RF +β D-MPNN In the formula, RF This represents the calibrated predicted antibacterial activity probability of the first antibacterial activity prediction model for a certain macrolide antibacterial candidate compound. D-MPNN This indicates the probability of the second antibacterial activity prediction model predicting the antibacterial activity of the same macrolide antibacterial candidate compound. ens The third antimicrobial activity prediction model represents the probability of predicting the antimicrobial activity of the same macrolide antimicrobial candidate compound, while α and β represent the output weights of the first and second antimicrobial activity prediction models.

6. The method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration according to claim 5, characterized in that, The step of determining the classification threshold of the third antibacterial activity prediction model using the validation set includes: The third antibacterial activity prediction model was used to predict the antibacterial activity of the molecular structure data in the validation set. Combined with the known antibacterial activity of macrolide antibacterial candidate compounds in the validation set, the F1 value of the third antibacterial activity prediction model under different preset candidate thresholds was obtained. The preset candidate threshold corresponding to the maximum F1 value was used as the classification threshold of the third antibacterial activity prediction model.

7. The method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration according to claim 6, characterized in that, The process involves obtaining molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity, and then using a third antibacterial activity prediction model and its classification threshold to output the antibacterial activity prediction results of the macrolide antibacterial candidate compounds with unknown antibacterial activity, including: The molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity are input into the third antibacterial activity prediction model to obtain the prediction probability of the antibacterial activity of the macrolide antibacterial candidate compounds with unknown antibacterial activity by the third antibacterial activity prediction model. When the third antimicrobial activity prediction model predicts the antimicrobial activity probability of macrolide antimicrobial candidate compounds with unknown antimicrobial activity as greater than or equal to the classification threshold, the antimicrobial activity prediction result of the macrolide antimicrobial candidate compounds with unknown antimicrobial activity is determined to be a Go candidate; otherwise, it is a No-Go candidate.

8. A screening system for macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration, characterized in that, include: The test data acquisition module is used to: acquire molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity; The screening module is used to input the molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity. The third antibacterial activity prediction model obtained by the method according to any one of claims 1-7, and the antibacterial activity prediction results of macrolide antibacterial candidate compounds with unknown antibacterial activity based on the classification threshold of the third antibacterial activity prediction model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the macrolide antibacterial candidate compound screening method based on probabilistic calibration and graph neural network integration as described in any one of claims 1 to 7, or the following steps: To obtain molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity; Input the molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity. The third antibacterial activity prediction model obtained by the method according to any one of claims 1-7, and the antibacterial activity prediction results of macrolide antibacterial candidate compounds with unknown antibacterial activity based on the classification threshold of the third antibacterial activity prediction model.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for screening macrolide antibacterial candidate compounds based on probabilistic calibration and graph neural network integration as described in any one of claims 1 to 7, or the following steps: To obtain molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity; Input the molecular structure data of macrolide antibacterial candidate compounds with unknown antibacterial activity. The third antibacterial activity prediction model obtained by the method according to any one of claims 1-7, and the antibacterial activity prediction results of macrolide antibacterial candidate compounds with unknown antibacterial activity based on the classification threshold of the third antibacterial activity prediction model.

Images