Method for cleaving enzyme recognition based on integrated sampling class prior estimation and related products

Abstract

This application relates to the field of weakly supervised learning technology, and in particular to a lyase identification method and related products based on ensemble sampling prior estimation. The method first constructs a known set of positive lyase samples P and a set of unlabeled samples U to be identified; then, based on the unlabeled sample set U, it constructs multiple training subsets U. des For each training subset U des Each base classification model is independently constructed and trained; each base classification model is used to predict the positive sample set P and the unlabeled sample set U, and the class prior probability is independently estimated based on the predicted score distribution; the multiple estimation results π are then fused together. des The final global class prior probability π is obtained; the basic classification model is optimized based on the global class prior probability π, and the optimized model is used to score and rank the unlabeled sample set U to obtain the candidate lyase identification results; effectively ensuring the stability and reliability of candidate lyase sequence screening in the lyase identification task.

Description

Technical Field

[0001] This invention relates to the field of weakly supervised learning technology, and in particular to a lyase identification method and related products based on ensemble sampling prior estimation. Background Technology

[0002] In the fields of artificial intelligence and machine learning, traditional supervised learning models heavily rely on large-scale, precisely bidirectionally labeled datasets, requiring a sufficient number of both positive and negative samples. However, in real-world applications of massive data mining, obtaining accurate and genuine negative samples often faces extremely high technical barriers and cost constraints. Typically, researchers only have access to a small number of validated, known positive samples and massive amounts of unlabeled data containing unknown proportions of positive and negative samples. To address this data acquisition dilemma, Positive Unlabeled Learning (PU Learning) has emerged. PU Learning aims to train high-performance binary classification models using only positive and unlabeled samples. In the Positive Unlabeled Learning framework, due to the lack of accurate negative sample labeling information, accurate estimation of the class prior probability (usually denoted as π) becomes a crucial prerequisite for constructing an unbiased loss function. The class prior probability refers to the proportion of samples in the unlabeled dataset that truly belong to the target class. The accuracy of this parameter directly affects the training effect and prediction performance of the subsequent classification model. When the class prior probability is underestimated, the model over-penalizes misclassification of positive samples, causing a large number of potential true positive samples to be incorrectly classified as negative samples. Conversely, when the class prior probability is overestimated, the model over-tolerates misclassification of negative samples, resulting in a large number of false positive predictions. Therefore, an accurate π value can effectively correct the serious statistical bias caused by directly equating unlabeled samples with negative samples, thereby ensuring the fairness of model training and the reliability of prediction results.

[0003] In existing technologies, the estimation of class prior probability π typically employs methods based on statistical distributions, such as the BBE (Best Bin Estimation) algorithm. The basic principle of this approach is to initially train a base classifier, obtaining its predicted score distributions on both the positive and unlabeled sample sets. Then, by comparing the differences between these two distributions, the proportion of positive samples in the unlabeled set can be inferred. The theoretical basis of this method is that if the classifier can effectively distinguish between positive and negative samples, then the predicted score distributions of positive and unlabeled samples should show a significant difference. Ideally, the predicted scores of positive samples should be concentrated in high-scoring regions, while the predicted scores of negative samples should be concentrated in low-scoring regions. Since the proportion of positive and negative samples in the unlabeled set is π, the value of π can be estimated by comparing the differences between the two distributions.

[0004] However, in the process of implementing the inventive technical solutions in the embodiments of this application, the inventors of this application discovered that the above-mentioned prior art solutions have at least the following technical problems:

[0005] Due to the massive scale and extremely complex composition of unlabeled data in real-world scenarios, the global data distribution exhibits strong heterogeneity. Existing techniques struggle to capture the true global distribution characteristics of the data through a single, limited random sampling. They are extremely sensitive to the initial sampling distribution of the input data; once random sampling introduces local data bias, the estimated π value fluctuates drastically and uncontrollably. Furthermore, the estimation error of π value has a severe cascading amplification effect in deep networks. If the initial π value estimation is flawed, the unbiased loss function (PU) built upon it will directly fail, causing a shift in the deep feature representations learned by the model. Therefore, under different sampling batches or random perturbations, the class prior probability estimates output by the algorithm are extremely discrete, failing to provide a stable and reliable single numerical benchmark for downstream tasks. Simultaneously, existing methods focus on optimizing the loss function, neglecting the statistical robustness modification of the underlying class prior estimation process. This leads to significant fluctuations in the global ranking ability of the basic classification model when dealing with minor changes in the distribution of unlabeled data, lacking inherent robustness. Therefore, existing class prior probability estimation methods suffer from high variance, low stability, and low robustness. Summary of the Invention

[0006] This application aims to provide a lyase identification method and related products based on ensemble sampling prior estimation. By introducing multiple rounds of random sampling to construct data perturbation and combining multi-model independent estimation with a distribution effective interval extraction mechanism, the variance of the prior probability estimation is fundamentally reduced, and the stability and robustness of the estimation results are improved. This allows for high-precision identification of candidate lyase sequences in unlabeled data even without real negative samples. Therefore, the objective of this application is to effectively reduce the variance of the prior probability estimation, improve the stability and robustness of the estimation results, and enhance the reliability of candidate lyase sequence screening results in lyase identification tasks. To achieve this objective, the following technical solution is used:

[0007] In a first aspect, this application provides a lyase identification method based on ensemble sampling prior estimation, including:

[0008] Construct a set of known positive lyase samples P and a set of unlabeled samples to be identified U;

[0009] Multiple training subsets U are constructed based on the unlabeled sample set U. des ;

[0010] For each of the training subsets U des Independently build and train the basic classification model;

[0011] Each of the aforementioned basic classification models is used to predict the positive sample set P and the unlabeled sample set U. Based on the predicted score distribution, the class prior probability is independently estimated, resulting in multiple estimation results π. des ;

[0012] Integrating multiple estimation results π des , thus obtaining the final global class prior probability π;

[0013] The basic classification model is optimized based on the global class prior probability π, and the optimized model is used to score and rank the unlabeled sample set U to obtain the candidate lyase identification results.

[0014] Secondly, this application provides a lyase recognition system based on ensemble sampling prior estimation, comprising:

[0015] The sample construction module is used to construct a set of known lyase positive samples P and a set of unlabeled samples to be identified U; and to construct multiple training subsets U based on the unlabeled sample set U. des ;

[0016] The classification model module is used for each of the training subsets U des Independently build and train the basic classification model;

[0017] The class prior probability estimation module is used to predict the positive sample set P and the unlabeled sample set U using each of the basic classification models, and to independently estimate the class prior probability based on the predicted score distribution, obtaining multiple estimation results π. des ;

[0018] A prior probability fusion module is used to fuse multiple estimated results π. des , thus obtaining the final global class prior probability π;

[0019] The lyase identification module is used to optimize the basic classification model based on the global class prior probability π, and to use the optimized model to score and rank the unlabeled sample set U to obtain the candidate lyase identification results.

[0020] Thirdly, this application provides an electronic device comprising: at least one processor; a memory communicatively connected to the at least one processor; and a computer program stored in the memory and running on the at least one processor; wherein, when the at least one processor executes the computer program, it is used to implement the lyase identification method based on integrated sampling prior estimation described in the first aspect above.

[0021] Fourthly, this application provides a computer-readable storage medium, wherein a computer program present in the computer-readable storage medium, when executed by a computer device, is used to implement the lyase identification method based on integrated sampling prior estimation described in the first aspect.

[0022] Fifthly, this application provides a computer program product comprising a computer program, which, when executed by a computer device, is used to implement the lyase identification method based on ensemble sampling prior estimation described in the first aspect.

[0023] As can be seen from the above, this application has the following beneficial effects:

[0024] This application constructs multiple training subsets U from an unlabeled sample set U. des Then, for each training subset U des Each base classification model is independently constructed and trained; then, each base classification model is used to predict the positive sample set P and the unlabeled sample set U. Based on the predicted score distribution, the class prior probability is independently estimated, resulting in multiple independently estimated class prior probability estimates π. des Finally, by fusing multiple independent estimates of π des The final global class prior probability π is obtained, and the basic classification model is optimized based on the global class prior probability π. The optimized model is then used to score and rank the unlabeled sample set U to obtain the candidate lyase identification results. It can be seen that the technical solution of this invention essentially utilizes the core idea of ​​ensemble learning in machine learning, that is, by constructing and combining multiple learners to complete the learning task, thereby improving the performance of a single learner; and by using the idea of ​​ensemble learning to improve the existing method of single-time estimation of class prior probability. This invention applies this idea of ​​ensemble learning to the application scenario of class prior probability estimation. By using this ensemble sampling strategy, controllable distribution perturbation is introduced at the data level. Combined with multi-model independent estimation and distribution effective interval extraction techniques, the single-point estimation paradigm in the existing technology is fundamentally changed, transforming it into a statistical estimation paradigm. This effectively reduces the variance of class prior probability estimation, provides a solid data foundation for the reliable training of downstream classification models, and solves the technical problems of high variance and low stability faced by traditional single-time class prior estimation methods.

[0025] Furthermore, this ensemble sampling strategy can effectively smooth the prediction scores of the base classification model on the sample set, increasing the number of effective intervals, thereby improving the reliability and stability of class prior probability estimation; and, the ensemble learning concept inherently possesses high fault tolerance, even if individual training subsets U desEven if the estimation of the basic classification model constructed and trained is biased, it will not significantly affect the estimation result of the final fused global class prior probability π. Therefore, it can effectively improve the stability and robustness of the class prior probability estimation result, so that the technical solution of this invention can maintain stable and reliable class prior probability estimation performance in complex and ever-changing real data environments, even if the data distribution changes with factors such as time, space and collection conditions. This is conducive to promoting the widespread application of PU Learning technology.

[0026] Furthermore, different training subsets U des Independently constructed and trained base classification models with different bias characteristics are combined by fusing multiple independent estimation results π. des This can precisely cancel out the bias, thus obtaining a more accurate global prior probability π.

[0027] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 A flowchart illustrating a lyase identification method based on ensemble sampling prior estimation provided in this application embodiment;

[0030] Figure 2 Scatter plot comparing the stability of the traditional single estimation method Single BBE provided for the embodiments of this application with the lyase identification method Bagging-BBE based on ensemble sampling class prior estimation in this application on the class prior probability estimation results;

[0031] Figure 3 A robustness diagram of the basic classification model provided in the embodiments of this application under different class prior probability estimates;

[0032] Figure 4 A schematic diagram of the structure of a lyase recognition system based on integrated sampling prior estimation provided in this application embodiment;

[0033] Figure 5 This is a schematic diagram of the structure of a computer electronic device provided in an embodiment of this application. Detailed Implementation

[0034] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] Example 1

[0036] In this embodiment, the target sequence can be a lyase protein sequence, the target category can be a lyase category, the positive sample set P can be composed of known lyase sequences, and the unlabeled sample set U can be composed of unlabeled protein sequences to be identified. The subsequent term "target sequence screening" specifically refers to the identification and screening of candidate lyase sequences in this embodiment.

[0037] The technical solution of this invention is mainly applicable to lyase identification tasks under conditions where no real negative samples are available. For example, in metagenomic data, environmental sequencing data, or public protein databases, there are a large number of protein sequences whose functions have not yet been validated, which may contain potential lyase sequences. Since the number of experimentally validated positive lyase samples is limited, and real non-lyase negative samples are difficult to rigorously determine, this task is suitable for modeling using a positive unlabeled learning framework. The lyase identification method based on ensemble sampling prior estimation provided in this application can achieve stable identification of candidate lyase sequences under conditions where only known positive lyase samples and a large number of unlabeled protein sequences are available. Without departing from the concept of this invention, the ensemble sampling prior estimation mechanism in this method can also be extended to other positive unlabeled learning scenarios.

[0038] To address the issues of poor stability, high sensitivity to data distribution perturbations, and model overconfidence in existing class prior probability estimation methods when dealing with large-scale unlabeled data, especially in scenarios where unlabeled samples are directly treated as negative samples for training, traditional single-shot estimation methods easily lead to breakages in the support interval of the predicted score distribution, significantly reducing the reliability and accuracy of class prior probability estimation. Therefore, this application's embodiments transform traditional class prior probability estimation methods using the concept of ensemble learning. Ensemble learning is an important technique in machine learning, its basic principle being to improve the performance of a single learner by constructing and combining multiple learners to complete the learning task. This application's embodiments creatively apply the concept of ensemble learning to the field of class prior probability estimation. By performing multiple rounds of random sampling on unlabeled data, multiple different training subsets are constructed; a classification model is independently trained for each training subset, obtaining multiple independent class prior probability estimates; finally, these independent estimates are aggregated using a statistical fusion method to obtain a stable global class prior probability estimate. This ensemble strategy fundamentally changes the traditional single-point estimation paradigm, transforming it into a statistical estimation paradigm, thereby significantly reducing the variance of the estimate.

[0039] Specifically, such as Figure 1 As shown, Figure 1 This is a flowchart illustrating a lyase identification method based on ensemble sampling prior estimation, provided in an embodiment of this application. The embodiment of this application provides a lyase identification method based on ensemble sampling prior estimation, comprising the following steps:

[0040] 101. Construct a set of known positive lyase samples P and a set of unlabeled samples to be identified U.

[0041] In this embodiment, the positive sample set includes rigorously functionally validated lyase protein sequences, while the unlabeled sample set originates from large-scale environmental sequencing datasets. The positive sample set consists of lyase protein sequences with known functions. In protein function prediction tasks, experimentally validated proteins can constitute a high-quality positive sample set, whose sequence information, functional annotations, and structural features can serve as the basis for model learning. The selection of positive samples should follow strict standards to ensure the accuracy and reliability of their functional annotations. For example, in lyase identification tasks, positive samples should be protein sequences that have been experimentally validated to possess lyase activity or lyase functional annotations. The unlabeled sample set originates from environmental sequencing data, metagenomics, or public genome databases. Environmental sequencing data is characterized by its large sample size and wide range of sources, representing complex data distributions in the real world; however, the proportion of truly positive samples it contains is often unknown and needs to be determined through prior probability estimation. Public genome databases such as NCBI (National Center for Biotechnology Information) and UniProt (Universal Protein Resource) provide abundant sequence resources and can serve as sources of unlabeled data.

[0042] Specifically, firstly, functionally validated or highly reliable lyase samples are obtained from known lyase databases or functional annotation databases to construct an initial positive sample set P. To avoid feature homology data leakage during subsequent model training and to ensure that the feature spaces of subsequent training and testing data are independent, a feature clustering method is used to cluster the positive sample set P, and a strict feature similarity threshold is set to ensure that samples within the same cluster have highly consistent underlying features. It should be noted that the quality of positive samples directly determines the learning upper limit of the subsequent classification model. Therefore, the selection of target samples in this embodiment should follow strict functional validation standards to ensure that the selected positive samples have clear biological significance or application value. The feature clustering method can be hierarchical clustering, K-means clustering, or DBSCAN density clustering, etc., and the specific selection can be determined according to the distribution characteristics of the sample features. The setting of the feature similarity threshold for clustering needs to balance maintaining sample diversity and avoiding feature homology. In this embodiment, the sequence consistency threshold is set to 0.5, and the coverage is set to 0.4, which can ensure that proteins within the same cluster have highly consistent sequence features and functional characteristics.

[0043] Based on the results of the above clustering process, the data is partitioned into clusters. Some clusters are divided into independent test sets, and the remaining clusters are used for training set partitioning, thus obtaining the training set P. trainand test set P test This approach ensures that test data and training data are relatively independent in feature space, thus preventing the model from relying solely on shallow feature similarity for discrimination. This cluster-based data partitioning strategy effectively avoids the information leakage problems that may arise from traditional random partitioning methods, ensuring the objectivity and reliability of model evaluation results. In practical applications, typically 80% of clusters are used for training and 20% for testing; the specific ratio can be adjusted according to the size of the dataset.

[0044] For example, the above similarity classification can be calculated based on biological sequence identity and coverage. Sequence similarity is typically calculated using sequence alignment tools (such as BLAST, Basic Local Alignment Search Tool), while coverage reflects the degree of length matching between two sequences. By considering both metrics, the evolutionary relationships and functional similarities between proteins can be assessed more accurately. Coverage reflects the proportion of the aligned region to the entire sequence; excessively low coverage may indicate incomplete alignment or false positive matches.

[0045] Furthermore, the training set can be further divided into groups to maintain a relatively balanced number of samples, thereby improving the stability of subsequent model training. A balanced subset partition helps ensure that each training batch has similar statistical characteristics, reducing variance fluctuations during training. Grouping should consider the uniformity of sample feature distribution to avoid some batches containing too many similar samples while other batches have too few.

[0046] Then, large-scale unlabeled raw data is obtained from complex open domain and / or public databases, and the unlabeled raw data is screened using a quality assessment algorithm to obtain unlabeled samples; noisy samples with low integrity or low confidence are removed. In the embodiments of this application, the quality control method adopts integrity assessment methods based on CheckM and CheckV to quantify the integrity of genome or protein sequences, and combines preset thresholds for screening.

[0047] Finally, key features are extracted from the unlabeled samples, and feature redundancy is removed using a high threshold to obtain a widely distributed and non-redundant set of unlabeled samples, U. The purpose of feature redundancy removal is to eliminate the interference of redundant features on model learning and improve the model's generalization ability. Setting a high threshold ensures that the retained features have sufficient information and avoids introducing too many noisy features. Commonly used feature extraction methods include Principal Component Analysis (PCA), feature selection algorithms, and embedding learning.

[0048] Key feature extraction is performed on the unlabeled samples. For example, coding region prediction and target protein sequence extraction are performed on the unlabeled sequences. Coding region prediction identifies regions in the sequence that may encode proteins or functional RNA (ribonucleic acid). Commonly used tools include GeneMark and FragGeneScan. Target protein sequence extraction extracts sequence fragments related to the target protein from these regions, providing input features for subsequent classification model training. Feature extraction can also include advanced analyses such as sequence structure prediction and functional domain identification.

[0049] 102. Construct multiple training subsets U based on the unlabeled sample set U. des .

[0050] First, the number of ensemble sampling rounds, N, is set; where N is a positive integer greater than 1. The number of ensemble sampling rounds needs to be determined by a trade-off between the estimated variance and computational cost. Too few rounds will result in insignificant ensemble effects, while too many rounds will increase computational overhead. Those skilled in the art can determine the optimal number of sampling rounds through methods such as cross-validation, depending on the specific application scenario and data scale. Experiments show that when the number of sampling rounds reaches 20-30, the ensemble effect is already quite significant, and the improvement brought by further increasing the number of rounds gradually diminishes. Therefore, this embodiment uses N=20 as an example for illustration.

[0051] Then, N rounds of random sampling without replacement are performed on the unlabeled sample set U. In each round of sampling, sampling without replacement is performed on the unlabeled sample set to obtain the unlabeled sub-sample set for the current round, ensuring that there are no duplicate samples within a single training dataset. The sampling without replacement strategy guarantees the independence of samples within each subset, avoiding the potential impact of duplicate samples on model training. The sampling ratio is set considering both computational efficiency and estimation accuracy: too small a sampling size will lead to overly sparse subsets, affecting model training performance; too large a sampling size will increase the similarity between subsets, reducing ensemble performance. Therefore, the sampling ratio is typically set between 60% and 90%, with the specific value determined based on the size and heterogeneity of the dataset.

[0052] Finally, the unlabeled subsample set obtained from each round of sampling is combined with the positive sample set P to form multiple training subdata sets U. desUnder each random seed condition, the unlabeled sample set U is sampled multiple times without replacement. Each random seed corresponds to the construction of 20 training subsets for independently training the base classification model. This positive-negative sample pairing method ensures that each training subset contains complete positive sample information, providing a balanced sample foundation for subsequent independent model training. Since the positive sample set is relatively small and of high quality, adding it entirely to each subset ensures that the model can always learn the discriminative features of positive samples. Furthermore, since each round uses sampling without replacement, a certain degree of sample overlap is allowed between different subsets, thus simulating the random perturbation characteristics of the real unlabeled data distribution. This helps increase the diversity of subsequent estimation results and improves the stability and robustness of the class prior probability estimation results. Further, through the above multi-round random sampling mechanism, each training subset corresponds to a different data distribution, thus providing a basis for diversified estimation of subsequent class prior probabilities. Sample overlap between different subsets is an important component of the ensemble learning strategy; it increases the diversity of estimation results and helps improve the stability of the final fusion estimation. Meanwhile, because sampling without replacement is used, each subset of data represents a different perspective of the unlabeled data, and this diversity is the fundamental reason why this invention can effectively reduce variance.

[0053] Through multiple rounds of random sampling as described in the embodiments of this application, the distribution characteristics of the data can be explored from different perspectives, a more comprehensive data view can be obtained, the heterogeneity of the real data distribution can be effectively simulated, and the estimation results can be more robust.

[0054] 103. For each of the aforementioned training subsets U des Independently build and train the basic classification model.

[0055] For each of the training subsets U des Construct a binary classification model; use the training subset U des Unlabeled samples are considered as negative samples and used in loss calculation. Multiple binary classification models are trained independently. The model parameters of the binary classification models are iteratively updated by optimizing the loss function to obtain N independently trained binary classification models.

[0056] Specifically, for each of the training subsets U desConstruct a binary classification model to distinguish between positive samples and unlabeled samples. The choice of binary classification model should be based on the specific application scenario and data characteristics. Different types of classifiers can be used, such as logistic regression, support vector machines, decision trees, random forests, or deep neural networks. Different types of classifiers have different learning capabilities and generalization characteristics. Choosing an appropriate classifier is crucial for obtaining high-quality class prior probability estimates. For high-dimensional sparse data, such as text data or biological sequence data, deep neural networks are usually a better choice; for low-dimensional structured data, such as tabular data, gradient boosting trees or support vector machines may perform better.

[0057] During training, unlabeled samples are temporarily treated as negative samples in the loss calculation. While this approach introduces label noise to some extent, its impact can be effectively offset by subsequent class prior probability correction. Experiments show that when the class prior probability estimation is sufficiently accurate, this simplified training strategy can achieve classification performance comparable to or even better than complex loss functions. Furthermore, since this invention ultimately incorporates the class prior probability into the PU (Positive-Unlabeled) unbiased loss function for retraining, the simplified training in the first stage is primarily aimed at obtaining the initial class prior probability estimate.

[0058] Then, the model parameters are iteratively updated by optimizing the loss function to obtain multiple independently trained classification models. Appropriate regularization strategies should be employed during model training to prevent overfitting to the limited training data. Regularization methods include L1 regularization, L2 regularization, Dropout, and early stopping; the specific choice should be determined based on the model type and data characteristics. Hyperparameters during training, such as learning rate, batch size, and number of training epochs, should be tuned using a validation set.

[0059] Each classification model is trained independently. By training and fusing multiple models independently, the potential bias and overfitting risks associated with individual models are reduced. Their parameter spaces and predictive behaviors differ, thus creating diversity at the model level. This model diversity is a key factor in the success of ensemble learning strategies, ensuring that different models provide complementary predictions for the same problem. Fusion further improves the stability and accuracy of the estimation. Sources of model diversity include: differences in training data (generated by ensemble sampling), differences in model architecture, and differences arising from random initialization.

[0060] It should be noted that the basic classification model employs a deep neural network structure consisting of an encoder and a classification head based on a pre-trained protein language model (ESM) to extract features and predict the classification of protein sequences. Multiple basic classification models are trained independently on different training subsets, thus creating diversity at the model level. During training, a stochastic gradient descent optimizer is used, with a learning rate of 0.01, a momentum coefficient of 0.9, and a weight decay coefficient of 5 × 10⁻⁶. -4 The batch size was set to 32, and each model underwent 8 training epochs. A cosine annealing learning rate scheduling strategy was used during training to optimize convergence performance, and a fixed number of training epochs was employed during the model training phase. The validation set comprised 20% of the model.

[0061] 104. Using each of the basic classification models, predict the positive sample set P and the unlabeled sample set U, and independently estimate the class prior probability based on the predicted score distribution to obtain multiple estimation results π. des .

[0062] By acquiring the positive sample prediction score of each binary classification model on the positive sample set P and the unlabeled sample prediction score on the unlabeled sample set U; selecting a decision threshold one by one based on the positive sample prediction score; counting the number M of unlabeled sample prediction scores greater than or equal to the decision threshold; determining whether M is greater than 0; if so, determining the interval corresponding to the decision threshold as the valid interval, and accumulating the number L of the valid intervals; calculating the local estimate of the class prior probability corresponding to the valid interval based on the ratio of the number of positive samples to the number of unlabeled samples in the valid interval; traversing all the decision thresholds, and combining with the preset numerical constraint range, statistically processing the local estimate of the class prior probability corresponding to all valid intervals to obtain the class prior probability estimation result π independently estimated by each binary classification model. des .

[0063] Specifically, each of the basic classification models is used to predict the positive sample set P and the unlabeled sample set U, obtaining corresponding prediction scores. Based on the distribution of the prediction scores, the class prior probability is estimated under the judgment threshold that satisfies the effective interval condition. Further, the prediction scores of each classification model on the positive sample set P and the unlabeled sample set U are obtained separately, and the two sets of prediction scores are sorted. Obtaining the prediction scores is the foundation for effective interval extraction, and the distribution characteristics of the scores directly determine the recognition effect of the effective interval. By sorting the two sets of scores, the changes in classification performance under different thresholds can be systematically analyzed. Sorting facilitates the subsequent threshold scanning process, allowing each threshold to be examined sequentially in order of score from low to high or from high to low.

[0064] The specific method for extracting the effective interval is as follows:

[0065] Using the predicted scores of positive samples as a benchmark, decision thresholds are selected sequentially according to the sorted scores. For each threshold, the number of unlabeled sample predicted scores greater than or equal to that threshold is counted. When the count of unlabeled samples is greater than zero, the threshold is determined as the boundary of the valid interval, i.e., the interval corresponding to that threshold is determined as the valid interval, and the number of valid intervals is accumulated. This decision strategy based on positive sample score thresholds can effectively characterize the distribution of unlabeled samples within the threshold region corresponding to the positive sample distribution, providing data support for subsequent class prior probability estimation. The selection range of the thresholds should cover the entire distribution range of positive sample scores. The valid interval refers to the interval where, under that threshold, there are unlabeled samples with predicted scores greater than or equal to that threshold. This interval has statistical significance and can be used to reliably estimate the class prior probability. The number of valid intervals reflects the availability of predicted scores under different thresholds.

[0066] Then, based on the ratio of positive samples to unlabeled samples in the effective interval, the local estimate of the class prior probability corresponding to that interval is calculated. The calculation formula is shown below:

[0067] π _local =Z _in_bin / W _in_bin ;

[0068] Among them, Z _in_bin W represents the number of samples whose positive predicted scores are greater than or equal to the current threshold. _in_bin This represents the number of unlabeled samples whose predicted scores are greater than or equal to the threshold; π _local This represents the local estimate of the class prior probability corresponding to the threshold. The principle behind this formula is that, above the threshold, the predicted score distributions of positive samples and unlabeled samples satisfy the consistency assumption. That is, above the threshold, the proportion of positive samples among unlabeled samples can be estimated from the ratio of positive to unlabeled samples, thus obtaining the local estimate π of the class prior probability. _local In some implementations, the local estimate can be normalized by combining the ratio of the total number of positive samples to the total number of unlabeled samples.

[0069] It should be noted that in traditional single-model estimation, the excessive suppression of unlabeled samples by the model can easily lead to a concentrated collapse of the predicted score distribution, causing a large number of samples to cluster in low-scoring regions. This results in a significant reduction in the number of effective intervals, thus affecting the reliability of the estimation results. This phenomenon is particularly pronounced in datasets with extremely imbalanced positive and negative samples, severely limiting the application effectiveness of traditional methods. Specifically, when the ratio of positive to negative samples is severely imbalanced, the classifier tends to predict most samples as negative, resulting in a severe bimodal structure in the predicted score distribution: positive samples are concentrated in high-scoring regions, while unlabeled samples are concentrated in low-scoring regions. This leads to a very small number of effective intervals in high-scoring regions, making reliable class prior probability estimation impossible. Therefore, this embodiment introduces the aforementioned multi-round random sampling mechanism to introduce distribution perturbation at the data level. Combined with multi-model prediction results, this effectively smooths the predicted score distribution, maintains a high distribution resolution, and significantly increases the number of effective intervals, ensuring the stability and theoretical unbiasedness of the class prior probability estimation. Through training on different subsets of datasets, the model learns different aspects of the data, thereby generating different prediction behaviors. When the predictions from multiple models are combined, the distribution of predicted scores is smoothed, and samples that were originally concentrated in low-score regions are dispersed across a wider range of score intervals. This increases the number of effective intervals, making the class prior probability estimation more reliable.

[0070] Then, by iterating through all possible decision thresholds, and following the method described above, local estimates π of the prior probabilities of multiple classes can be obtained. _local Furthermore, by combining preset numerical constraints, such as the class prior probability being between 0 and 1, statistical processing is performed on the local estimates corresponding to all effective intervals using methods such as mean, weighted average, or median, and then fused to obtain the class prior probability estimate π corresponding to the current single model. des The statistical processing method can be determined based on the distribution characteristics of the local estimates. For example, if the distribution of the local estimates is relatively symmetrical and there are no obvious outliers, the mean can be used as the class prior probability estimate π corresponding to the current single model. des If outliers exist, it is recommended to use the median as the final class prior probability estimate π for the current single model. des To improve robustness, in some feasible implementations, the estimated result π can also be affected by the range of numerical constraints. des Post-processing is performed, such as adjusting values ​​less than 0 to 0 and values ​​greater than 1 to 1.

[0071] The embodiments of this application utilize the effective interval extraction algorithm to fully leverage the global information of the predicted score distribution, thereby improving the availability of the predicted score distribution under different thresholds.

[0072] 105. Integrate multiple estimation results πdes This yields the final global prior probability π.

[0073] Collect the class prior probability estimates π corresponding to all training subsets. des The estimated value π generated for each subset of data. des All datasets contain a certain amount of random error, but these errors are independent of each other in different subsets of the dataset. These random fluctuations can be effectively offset by appropriate fusion strategies.

[0074] The estimated values ​​are statistically fused to obtain the global prior probability. Statistical fusion methods can employ strategies such as simple averaging, weighted averaging, or median fusion. Experiments show that simple averaging fusion achieves significant performance improvements in most scenarios, but may not be robust enough for extreme cases. In such cases, more robust methods such as median fusion or truncated averaging can be used. The principle of truncated averaging is to remove the highest and lowest estimated values ​​and then take the average, which can further reduce the impact of extreme values. The global prior probability is output as a stable estimation result. The fused global estimate has lower variance and higher reliability than any single estimate, providing a reliable parameter benchmark for downstream tasks. The variance of the global estimate can be estimated using the bootstrap method, thus providing users with a quantitative indicator of estimation uncertainty.

[0075] In summary, the embodiments of this application significantly reduce the random error caused by a single estimation by fusing multiple independent estimation results, transforming the prior probability estimation from a single-point estimation to a statistical estimation, thereby achieving a significant reduction in variance and a significant improvement in stability. This paradigm shift from single-point estimation to statistical estimation is the core objective of this invention and a creative application of ensemble learning methods in the field of prior probability estimation. When multiple independent but biased estimates exist, appropriate fusion can offset their respective biases, resulting in a more accurate estimate. In the embodiments of this application, the estimation of each sub-model is affected by the sampling of training data, but since different sub-models use different subsets of training data, these effects are relatively independent. Fusing multiple estimates can effectively reduce the impact of such random fluctuations.

[0076] Based on the above embodiments, the lyase identification method based on ensemble sampling prior estimation described in this application includes the following:

[0077] 106. The basic classification model is optimized based on the global class prior probability π, and the optimized model is used to score and rank the unlabeled sample set U to obtain the candidate lyase identification results.

[0078] First, the global class prior probability π is introduced into the PU unbiased loss function in the positive unlabeled learning framework to optimize the basic classification model; the calculation formula of the PU unbiased loss function is as follows: ;

[0079] in, This represents the distribution constraint term for positive samples. This represents the distribution term for unlabeled samples. This represents a function that measures the difference in distribution.

[0080] By introducing accurate prior probability estimates, the bias caused by directly treating unlabeled samples as negative samples can be effectively corrected. The optimization training process can refer to step 103 above, and will not be repeated here.

[0081] Finally, the optimized model is used to score and rank the unlabeled sample set U, achieving high-precision screening of candidate lyase sequences. Unlabeled samples are input into the optimized classification model to obtain a predicted score for each sample; samples are then sorted from highest to lowest predicted score, with higher-scoring samples being more likely to be candidate lyase sequences. In some implementations, a prediction score threshold (e.g., 0.5 or higher) can be set according to specific task requirements to screen high-confidence candidate lyase sequences. Precision, recall, and F1 scores at different thresholds are also reported to evaluate the quality of the screening results.

[0082] The technical effectiveness of the lyase identification method based on ensemble sampling prior estimation described in the embodiments of this application is illustrated below using experimental data:

[0083] like Figure 2 As shown, Figure 2 A scatter plot comparing the stability of the traditional single-experiment estimation method Single BBE and the lyase identification method Bagging-BBE based on ensemble sampling class prior estimation provided in this application in terms of class prior probability estimation results; where the horizontal axis represents different numbers of random experiments. Figure 2 Five different random seeds were used, and the vertical axis represents the class prior probability estimate. Figure 2 The number of valid bins is also marked, which is used to measure the resolution and stability of the predicted score distribution.

[0084] from Figure 2The results clearly show that the estimates obtained using the Single BBE method fluctuate wildly across different experiments, exhibiting an extremely discrete distribution. The standard deviation of the five estimates is large, with a mean of only 0.0161. The number of valid intervals (ValidBins) oscillates significantly between 1245 and 1883, a fluctuation exceeding 50%, indicating that this method is highly sensitive to the sampling process and its predicted score distribution is unstable. The main reasons for this are: data perturbation caused by a single random sampling, and overfitting of a single model to the training data. These two factors together lead to the instability of the predicted score distribution, causing drastic changes in the number of valid intervals. Specifically, when the sampling process accidentally includes more easily classified samples, the model learns an overly optimistic classification boundary, resulting in higher predicted scores; conversely, when the sampling process includes more difficult-to-classify samples, the model learns an overly conservative classification boundary, resulting in lower predicted scores.

[0085] The estimates obtained using the Bagging-BBE method are highly stable, with the five estimates almost completely overlapping, the mean remaining stable at 0.0282, and the number of effective intervals remaining stable within a narrow range from 2011 to 2019, with fluctuations of less than 1%. The variance is reduced by nearly 20 times compared to a single estimate, significantly improving the stability of the class prior probability estimation and the robustness of the downstream classification model. This significant improvement in stability demonstrates the effectiveness of the ensemble sampling strategy of this invention. Compared to the Single BBE method, this invention constructs a diverse data view through multiple rounds of random sampling and offsets the bias of a single model through multi-model fusion, thereby achieving smoothing of the predicted score distribution and stabilization of the estimation results. This stable and reliable class prior probability estimation provides a solid data foundation for subsequent unbiased learning of PU, enabling deep learning models to learn high-quality feature representations and classification decision boundaries even with only positive samples and unlabeled data, thus playing an important role in many fields such as bioinformatics, drug discovery, and medical diagnostics.

[0086] like Figure 3 As shown, Figure 3 The robustness performance of the basic classification model provided in this application embodiment under different class prior probability estimates is shown in the graph; including curves showing the changes of three core global ranking indicators—the area under the precision-recall curve (AUPRC), the area under the ROC curve (AUC), and the F1 score—with the class prior probability. Figure 3 (a), (b), and (c) in the example. Figure 3 The smoothness of the curve reflects the model's tolerance to errors in prior probability estimation. From... Figure 3 As can be observed, all three curves exhibit a highly stable straight-line shape, indicating that the downstream model has excellent robustness.

[0087] In the experiment, the class prior probabilities estimated by this invention were used to perform unbiased PU training on the downstream classification model, with small perturbation values ​​around 0.0282, such as 0.025, 0.026, 0.027, 0.028, 0.029, 0.030, 0.031, and 0.032. The model performance was evaluated on an independent test set. The test set was divided into clusters completely independent of the training set to ensure the objectivity and reliability of the evaluation results.

[0088] Experimental results show that the fine-tuned classification model of this invention maintains extremely high levels in the three core global ranking metrics: AUPRC, AUC, and F1-Score. Figure 3 The curves in all three subplots exhibit highly stable linear shapes, remaining almost unaffected by minor perturbations in the class prior probability. Specifically, the AUPRC metric remains stable around 0.96, the AUC metric around 0.95, and the F1-Score around 0.91. This highly stable performance has significant practical value, meaning that even with a certain error in the class prior probability estimation, such as an absolute error of 0.003, the performance of the downstream model will not be significantly affected. This greatly reduces the requirement for high accuracy in class prior probability estimation and enhances the overall practical value of the technical solution.

[0089] As global metrics independent of fixed classification thresholds, AUPRC and AUC demonstrate that the stable performance of the deep feature representations learned by the model will not fluctuate drastically due to minor evaluation biases at the lower levels. This is because the stable class prior probability estimation provided in this invention ensures the correct construction of the PU unbiased loss function, enabling the model to learn discriminative deep features in the correct optimization direction. This stability of feature representation is particularly important for practical applications that require handling large-scale unknown data, such as screening target protein-coding genes in metagenomic data, identifying potential active molecules in drug candidate molecule libraries, and detecting abnormal behavior in network traffic. These applications share common characteristics: massive data scale, complex data distribution, and difficulty in obtaining labeled data, but all require stable and reliable classification models to support business decisions.

[0090] In summary, the lyase identification method based on ensemble sampling prior estimation proposed in this invention achieves significant performance improvements in both the stability of class prior probability estimation and the robustness to downstream mining tasks. Experimental results show that the method described in this invention effectively eliminates the variance surge problem caused by evaluation discontinuities in traditional positive unlabeled learning, and, combined with a deep classification model, successfully achieves high-precision and robust screening of candidate lyase sequences in scenarios without real negative samples. This technical solution provides reliable technical support for the practical application of ensemble sampling learning in bioinformatics, protein function identification, and enzyme function mining, and is expected to promote the transition of positive unlabeled learning technology from theoretical research to practical application.

[0091] Example 2

[0092] The above combination Figures 1 to 3 This application provides a detailed description of a lyase identification method based on ensemble sampling prior estimation, as illustrated in its embodiments. The following is in conjunction with... Figure 4 This document describes a lyase identification system based on ensemble sampling prior estimation, which is used to implement the lyase identification method based on ensemble sampling prior estimation provided in the embodiments of this application. Figure 4 A schematic diagram of a lyase recognition system based on ensemble sampling prior estimation provided in an embodiment of the present invention; see reference. Figure 4 The system includes:

[0093] The sample construction module 100 is used to construct a known set of positive lyase samples P and a set of unlabeled samples to be identified U; and to construct multiple training subsets U based on the unlabeled sample set U. des ;

[0094] Classification model module 200, for each of the training subsets U des Independently build and train the basic classification model;

[0095] The class prior probability estimation module 300 is used to predict the positive sample set P and the unlabeled sample set U using each of the basic classification models, and to independently estimate the class prior probability based on the prediction score distribution, obtaining multiple estimation results π. des ;

[0096] Prior probability fusion module 400 is used to fuse multiple estimated results π. des , thus obtaining the final global class prior probability π;

[0097] The lyase identification module 500 is used to optimize the basic classification model based on the global class prior probability π, and to score and rank the unlabeled sample set U using the optimized model to obtain the candidate lyase identification results.

[0098] In one feasible implementation, the sample construction module 100 is specifically used for:

[0099] Obtain known lyase samples from a known lyase database or a functional annotation database, and construct a positive sample set P;

[0100] The positive sample set P is clustered using a feature-based clustering method;

[0101] Based on the clustering results, the data is divided into clusters to obtain the training set P. train and test set P test ;

[0102] Unlabeled raw data is obtained from the open domain and filtered using a quality assessment algorithm to obtain unlabeled samples;

[0103] Key features are extracted from the unlabeled samples to obtain an unlabeled sample set U;

[0104] Set the number of rounds N for ensemble sampling; where N is a positive integer greater than 1;

[0105] Perform N rounds of random sampling without replacement on the unlabeled sample set U;

[0106] The unlabeled subsample set obtained from each round of sampling is combined with the positive sample set P to construct N training subdatasets U. des .

[0107] In one feasible implementation, the classification model module 200 is specifically used for:

[0108] For each of the training subsets U des Construct a binary classification model;

[0109] The training subset U des Unlabeled samples are treated as negative samples and used in loss calculation; the binary classification models are trained independently of each other.

[0110] By iteratively updating the model parameters of the binary classification model through optimizing the loss function, N independently trained binary classification models are obtained.

[0111] In one feasible implementation, the prior probability estimation module 300 is specifically used for:

[0112] Obtain the positive sample prediction score of each of the binary classification models on the positive sample set P, and the unlabeled sample prediction score on the unlabeled sample set U;

[0113] Based on the predicted scores of the positive samples, a judgment threshold is selected one by one;

[0114] The number M of unlabeled samples whose predicted scores are greater than or equal to the judgment threshold is counted.

[0115] Determine if M is greater than 0;

[0116] If so, the interval corresponding to the determination threshold is determined as the valid interval, and the number of valid intervals L is accumulated;

[0117] Based on the ratio of positive samples to unlabeled samples in the effective interval, calculate the local estimate of the class prior probability corresponding to the effective interval;

[0118] By iterating through all the aforementioned decision thresholds and combining them with a preset numerical constraint range, statistical processing is performed on the local estimates of the class prior probability corresponding to all valid intervals to obtain the class prior probability estimate π independently estimated by each binary classification model. des .

[0119] In one feasible implementation, the lyase recognition module 500 is specifically used for:

[0120] The global class prior probability π is introduced into the unbiased loss function in the positive unlabeled learning framework to optimize and train the basic classification model;

[0121] The optimized model was used to score and rank the unlabeled sample set U to obtain the candidate lyase identification results.

[0122] The specific implementation method and technical effects of this system are described in the aforementioned method for lyase identification based on integrated sampling prior estimation, and will not be repeated here.

[0123] Example 3

[0124] This invention also provides a computer electronic device. Figure 5 This is a structural diagram of an electronic device according to an embodiment of the present invention, such as... Figure 5As shown, this computer electronic device includes a central processing unit (CPU) 501, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 502 or a program loaded from a storage section 508 into a random access memory (RAM) 503. The RAM 503 also stores various programs and data required for system operation. The CPU 501, ROM 502, and RAM 503 are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504. The following components are connected to the I / O interface 505: an input section 506 including a keyboard, mouse, etc.; an output section 507 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN card, modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A driver 510 is also connected to the I / O interface 505 as needed. Removable media 511, such as disks, optical discs, magneto-optical discs, semiconductor memories, etc., are installed on drive 510 as needed so that computer programs read from them can be installed into storage section 508 as needed.

[0125] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0126] Example 4

[0127] This invention also provides a computer-readable storage medium, which may be the computer-readable storage medium included in the lysozyme identification system based on ensemble sampling prior estimation described in the above embodiments; or it may be a standalone computer-readable storage medium not assembled into an electronic device. The computer-readable storage medium stores one or more programs, which are used by one or more processors to execute the lysozyme identification method based on ensemble sampling prior estimation described in this invention.

[0128] Example 5

[0129] This invention also provides a computer program product, which includes a computer program that, when executed by a computer device, implements a lyase identification method based on ensemble sampling prior estimation according to this invention.

[0130] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention is not limited to the specific combination of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A lyase identification method based on ensemble sampling prior estimation, characterized in that, include: Construct a set of known lyase positive samples P and a set of unlabeled samples to be identified U; constructing a plurality of training sub-data sets U based on the unlabeled sample set U des ; For each of the training subsets U des Independently build and train the basic classification model; Each of the aforementioned basic classification models is used to predict the positive sample set P and the unlabeled sample set U. Based on the predicted score distribution, the class prior probability is independently estimated, resulting in multiple estimation results π. des ; Integrating multiple estimation results π des , thus obtaining the final global class prior probability π; The basic classification model is optimized based on the global class prior probability π, and the optimized model is used to score and rank the unlabeled sample set U to obtain the candidate lyase identification results.

2. The lyase identification method based on ensemble sampling prior estimation as described in claim 1, characterized in that, Construct a set of known positive lyase samples P and a set of unlabeled samples to be identified U, including: Obtain known lyase samples from a known lyase database or a functional annotation database, and construct a positive sample set P; The positive sample set P is clustered using a feature-based clustering method; Based on the clustering results, the data is divided into clusters to obtain the training set P. train and test set P test ; Unlabeled raw data is obtained from the open domain and filtered using a quality assessment algorithm to obtain unlabeled samples; Key features are extracted from the unlabeled samples to obtain the unlabeled sample set U.

3. The lyase identification method based on ensemble sampling prior estimation as described in claim 2, characterized in that, Multiple training subsets U are constructed based on the unlabeled sample set U. des ,include: Set the number of rounds N for ensemble sampling; where N is a positive integer greater than 1; Perform N rounds of random sampling without replacement on the unlabeled sample set U; The unlabeled subsample set obtained from each round of sampling is combined with the positive sample set P to construct N training subdatasets U. des .

4. The lyase identification method based on ensemble sampling prior estimation as described in claim 3, characterized in that, For each of the training subsets U des Independently build and train the basic classification model, including: For each of the training subsets U des Construct a binary classification model; The training subset U des Unlabeled samples are considered as negative samples and participate in the loss calculation. The binary classification models are trained independently. The model parameters of the binary classification models are iteratively updated by optimizing the loss function to obtain N independently trained binary classification models.

5. The lyase identification method based on ensemble sampling prior estimation as described in claim 4, characterized in that, Each of the aforementioned basic classification models is used to predict the positive sample set P and the unlabeled sample set U. Based on the predicted score distribution, the class prior probability is independently estimated, resulting in multiple estimation results π. des ,include: Obtain the positive sample prediction score of each of the binary classification models on the positive sample set P, and the unlabeled sample prediction score on the unlabeled sample set U; Based on the predicted scores of the positive samples, a judgment threshold is selected one by one; The number M of unlabeled samples whose predicted scores are greater than or equal to the judgment threshold is counted. Determine if M is greater than 0; If so, the interval corresponding to the determination threshold is determined as the valid interval, and the number of valid intervals L is accumulated; Based on the ratio of positive samples to unlabeled samples in the effective interval, calculate the local estimate of the class prior probability corresponding to the effective interval; By iterating through all the aforementioned decision thresholds and combining them with a preset numerical constraint range, statistical processing is performed on the local estimates of the class prior probability corresponding to all valid intervals to obtain the class prior probability estimate π independently estimated by each binary classification model. des .

6. The lyase identification method based on ensemble sampling prior estimation as described in any one of claims 1 to 5, characterized in that, The basic classification model is optimized based on the global class prior probability π, and the optimized model is used to score and rank the unlabeled sample set U to obtain the candidate lyase identification results, including: The global class prior probability π is introduced into the unbiased loss function in the positive unlabeled learning framework to optimize and train the basic classification model; The optimized model was used to score and rank the unlabeled sample set U to obtain the candidate lyase identification results.

7. A lyase identification system based on ensemble sampling prior estimation, characterized in that, include: The sample construction module is used to construct a set of known lyase positive samples P and a set of unlabeled samples to be identified U; Furthermore, multiple training subsets U are constructed based on the unlabeled sample set U. des ; The classification model module is used for each of the training subsets U des Independently build and train the basic classification model; The class prior probability estimation module is used to predict the positive sample set P and the unlabeled sample set U using each of the basic classification models, and to independently estimate the class prior probability based on the predicted score distribution, obtaining multiple estimation results π. des ; A prior probability fusion module is used to fuse multiple estimated results π. des , thus obtaining the final global class prior probability π; The lyase identification module is used to optimize the basic classification model based on the global class prior probability π, and to use the optimized model to score and rank the unlabeled sample set U to obtain the candidate lyase identification results.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; A memory that is communicatively connected to the at least one processor; A computer program stored in the memory and running on the at least one processor; Wherein, when the at least one processor executes the computer program, it is used to implement the lyase identification method based on integrated sampling prior estimation as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer program present in the computer-readable storage medium, wherein when executed by a computer device, the computer program is used to implement the lyase identification method based on ensemble sampling prior estimation as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes: Computer programs; The computer program, when executed by a computer device, is used to implement the lyase identification method based on integrated sampling prior estimation as described in any one of claims 1 to 6.

Images