Dpp-iv inhibiting peptide screening method based on source perception activity ranking

Abstract

This application relates to a source-aware activity ranking-based method for screening DPP-IV repressive peptides, comprising: constructing a training dataset containing positive samples and at least three levels of negative samples; extracting multimodal features of peptides, including at least sequence deep semantic embedding features and pocket structure features; explicitly aligning the above features using a cross-attention mechanism to generate fusion features; inputting the fusion features into the classification branch and ranking branch of a decoupled prediction architecture, outputting classification probability and activity ranking score respectively, wherein the ranking branch uses an independent ranking model and is trained based on preference pairs constructed from data within the same assay source; and performing virtual screening output by combining classification probability, activity ranking score, multidimensional confidence judgment, and ADMET safety screening results. This application achieves structurally interpretable characterization of substrate-enzyme binding modes and improves the statistical significance and generalization stability of peptide activity ranking under cross-laboratory batch effects.

Description

Technical Field

[0001] This application relates to the fields of bioinformatics and computer-aided drug discovery, and in particular to a method for screening DPP-IV inhibitory peptides based on source-sensing activity sequencing. Background Technology

[0002] Dipeptidyl peptidase IV (DPP-IV, EC 3.4.14.5) plays a crucial role in glucose homeostasis regulation by cleaving and inactivating glucagon-like peptide-1 (GLP-1) and glucose-dependent insulinotropic peptide (GIP). Therefore, DPP-IV inhibitors have become an important strategy for the treatment of type 2 diabetes mellitus (T2DM). Compared to the side effects of synthetic small molecule inhibitors, food-derived DPP-IV inhibitory peptides have become a hot topic in the research and development of novel DPP-IV inhibitors due to their natural safety and low toxicity.

[0003] Traditional screening methods relying on in vitro enzyme activity experiments are time-consuming, costly, and have low throughput, making them unsuitable for the rapid screening of massive amounts of peptides. With the development of computational technology, machine learning-based DPP-IV repressive peptide prediction methods have been proposed, but these methods still suffer from the following significant technical limitations.

[0004] Chinese patent application CN117912595A discloses a machine learning-assisted method for screening and preparing DPP-4 inhibitory peptides. This method extracts pseudo-amino acid composition features from peptide sequences for activity classification and prediction, followed by post-validation via molecular docking. This approach has the following shortcomings: it relies solely on shallow statistical features, failing to capture contextual dependencies and evolutionary conservation information between amino acid residues; molecular docking is only used as a validation tool and not to enhance the prediction model itself; and it completely ignores the IC (independent amino acid) threshold. 50 The Half Maximal Inhibitory Concentration (HMC) activity value information cannot prioritize candidate peptides based on their activity levels.

[0005] Chinese patent application CN117637059A discloses a QSAR method for rapidly predicting the activity of DPP-IV inhibitory peptides based on novel descriptors. The method builds models by manually designing 2D / 3D molecular descriptors. However, this method is primarily designed for short peptides of 3 to 5 amino acids, resulting in poor versatility. Furthermore, it relies on manual experience in descriptor design, failing to achieve end-to-end feature self-learning. Additionally, it lacks multi-task capabilities and comprehensive evaluation functions such as classification and ranking.

[0006] The BERT-DPPIV (Bidirectional Encoder Representations from Transformers for Dipeptidyl Peptidase IV) model proposed by Guan et al. utilizes an attention-based peptide language model to perform binary classification of DPP-IV repressive peptides and incorporates amino acid composition. The limitation of this approach is that it can only perform binary classification and does not involve interferometry (IC). 50 The regression or ranking of activity values ​​cannot provide guidance for candidate priority in subsequent experiments; moreover, the geometric and physicochemical constraints of the DPP-IV activity pocket have not been considered, so the structural mechanism by which specific peptides are active cannot be explained, and it cannot guide rational modification.

[0007] In addition to the aforementioned shortcomings, this field also faces a common technical challenge: due to inconsistent measurement conditions in different research laboratories, the IC50 values ​​for the same peptide vary. 50 Experimental measurements exhibit significant systematic bias across different sources. Directly merging such heterogeneous data for training causes the model to learn from data source biases rather than the true structure-activity relationship, resulting in extremely low ranking reliability. The Spearman rank correlation coefficient of related methods is typically below 0.25.

[0008] In summary, the relevant technologies generally suffer from the following drawbacks: limited predictive features and weak characterization capabilities; inability to explicitly model substrate-enzyme binding modes; and inability to address cross-laboratory IC (integrated enzyme) problems. 50 The problems include systematic biases in data fusion and a lack of closed-loop capabilities from "qualitative prediction" to "quantitative ranking" and ultimately "guiding rational design". Summary of the Invention

[0009] This application provides a source-aware ranking learning-based method for screening DPP-IV repressive peptides, addressing the shortcomings of existing DPP-IV repressive peptide prediction methods, such as weak feature characterization capabilities, inability to explicitly model substrate-enzyme binding patterns, and cross-source IC (increase-increase ratio) issues. 50The data suffers from systematic biases that prevent reliable sorting, as well as a lack of closed-loop capabilities from activity prediction to rational design.

[0010] The first aspect of this application provides a method for screening DPP-IV repressive peptides based on source-aware ranking learning, including the following steps: S1. Construct a training dataset containing positive samples and at least three levels of negative samples. The three levels of negative samples include at least: a first negative sample randomly sampled from a public database, and a second negative sample generated by a generative model that is similar to the positive samples in physicochemical properties but does not possess DPP-IV inhibitory activity. Standardize the peptide samples, including at least: sequence redundancy removal, splitting the training and test sets by clusters, and adjusting the IC... 50 The data undergoes standardized preprocessing; S2. Extract multimodal features of peptide segments, including at least: sequence deep semantic embedding features extracted based on protein language models, and pocket structure features extracted from the active pocket region of DPP-IV protein through molecular docking as input to the model features. S3. Through the cross-attention mechanism, the sequence deep semantic embedding features are used as query vectors, and the pocket structure features are used as key vectors and value vectors, and explicit alignment is performed to generate fusion features that reflect the interaction mode between peptide sequences and active pockets. S4. Construct a decoupled prediction architecture consisting of a classification branch and a ranking branch, wherein the classification branch and the ranking branch each have independent parameter layers and output heads; the classification branch receives the fusion features and outputs the classification probability that the peptide has DPP-IV inhibitory activity; the ranking branch adopts an independent ranking model, which is executed in parallel with the classification branch. The ranking model is trained by constructing preference pairs based only on half-inhibition concentration data within the same assay source, and outputs the activity ranking score of the candidate peptide. S5. Virtual screening output: Based on the classification probability and the activity ranking score, the candidate peptides are comprehensively ranked, and the ranking results of the candidate peptides' activity strength are output.

[0011] Optionally, in some embodiments, the second negative sample is generated by an RBM (Restricted Boltzmann Machine); the RBM uses the amino acid composition, molecular weight, isoelectric point, and hydrophobicity of the positive sample as input to the visible layer, learns the joint probability distribution of the positive sample through a contrastive divergence algorithm, and then generates a synthetic sequence that is highly similar to the physicochemical properties of the positive sample but does not have DPP-IV inhibitory activity through Gibbs sampling.

[0012] Optionally, in some embodiments, the protein language model is an ESM-2 (Evolutionary Scale Modeling version 2) model; the extraction process of the sequence deep semantic embedding features is as follows: input the peptide sequence into the ESM-2 model, extract the average pooling representation of the last hidden state of the ESM-2 model, and obtain a fixed-dimensional embedding vector.

[0013] Optionally, in some embodiments, the cross-attention mechanism includes: using the residue-level embedding matrix of the peptide sequence as the query vector, and using the physicochemical feature matrix of the active pocket residues of the DPP-IV protein as the bond vector and value vector, wherein the physicochemical features include at least one of amino acid type, hydrophobicity scaling value, formal charge, number of hydrogen bond donors, number of hydrogen bond acceptors, van der Waals volume, and sub-pocket category number; calculating the attention weight matrix by scaling the dot product, and performing matrix multiplication of the attention weight matrix and the value vector to generate the fusion feature; The screening method further includes: generating an attention heatmap based on the attention weight matrix to identify the interaction strength between the position of each residue in the peptide sequence and each residue in the active pocket.

[0014] Optionally, in some embodiments, the training method of the ranking model in the ranking branch includes: based on IC from the same literature source or the same measurement batch. 50 Data constructs preference pairs; where, in any preference pair, if the IC of peptide i is... 50 IC50 values ​​less than peptide j 50 If the value is specified, the ordinal relationship between peptide i and peptide j is labeled as i being superior to j; with the normalized discounted cumulative gain as the optimization objective, the LambdaRank model is trained using the preference pair so that the ranking model can avoid systematic bias between different assay sources by learning the relative activity distribution within the same assay source.

[0015] Optionally, in some embodiments, the classification branch and the ranking branch are trained independently using different feature subsets; wherein, the classification branch takes full sequence statistical features as input; the ranking branch takes a feature subset filtered by the feature gain index as input, the feature subset being a dimensionality-reduced feature set obtained from the sequence statistical features and the sequence deep semantic embedding features; the classification branch and the ranking branch do not share a loss function and can be optimized independently.

[0016] Optionally, in some embodiments, the normalization process, when constructing the training dataset, further includes normalizing the IC. 50 The data preprocessing process includes: detecting ICs 50The order of magnitude of the raw data identifies the IC. 50 The initial unit of the raw data is then converted to pC. 50 (negative logarithm of the half maximal inhibitory concentration) value; filter out inaccurate activity records with range symbols and non-homologous data from non-enzymatic assay sources in the training dataset to construct a training set that satisfies ordinal constraints.

[0017] Optionally, in some embodiments, the training of the classification branch adopts a two-stage transfer learning strategy, specifically including: Stage 1: pre-training using activity data of non-peptide DPP-IV target small molecule inhibitors to enable the classification branch to learn the general chemical pattern of DPP-IV inhibitory activity; Stage 2: fine-tuning the model after Stage 1 pre-training using peptide data, wherein, during the fine-tuning process, the underlying feature extraction parameters are frozen, and a learning rate lower than that of Stage 1 is used to adapt to peptide sequence features.

[0018] Optionally, in some embodiments, the screening method further includes: based on the attention heatmap, identifying residue positions in the peptide sequence whose attention weight values ​​with key residues in the S1 or S2 sub-pockets of the active pocket reach or exceed a preset weight threshold; and outputting peptide site modification suggestions for enhancing hydrophobic interactions or enhancing charged interactions based on the physicochemical interaction type between the residue positions and the active pocket. The virtual filtering output also includes: The candidate peptides are subjected to multidimensional credibility assessment and safety assessment. The multidimensional credibility assessment includes: calculating the kNN (K-Nearest Neighbors) distance between the candidate peptide and the samples in the training dataset, using a preset quantile threshold of the training dataset as the boundary of the applicable domain to determine the applicable domain, and downweighting candidate peptides that exceed the boundary of the applicable domain. The safety assessment includes: using the ADMET (Aspect-Oriented Metabolism, Metabolism, Excretion and Toxicity) prediction model to quantitatively assess the biosafety of the candidate peptides, and screening out candidate peptides that do not meet the preset safety threshold.

[0019] A second aspect of this application provides a DPP-IV inhibitory peptide screening device based on source-aware ranking learning, comprising: The dataset construction and negative sample hierarchical module is used to construct the training dataset, which contains positive samples and at least three levels of negative samples. The three levels of negative samples include at least: a first negative sample randomly sampled from a public database, and a second negative sample generated by a generative model that is similar to the positive samples in physicochemical properties but does not possess DPP-IV inhibitory activity. The peptide samples are standardized, and the standardization process includes at least: sequence redundancy removal, splitting the training and test sets by clusters, and adjusting the IC... 50 The data undergoes standardized preprocessing; The multimodal feature extraction module is used to extract multimodal features of peptides, including at least: sequence deep semantic embedding features extracted based on a protein language model, and pocket structure features extracted from the active pocket region of the DPP-IV protein through molecular docking as input to the model features. The feature alignment module is used to explicitly align the sequence deep semantic embedding features as query vectors and the pocket structure features as key vectors and value vectors through a cross-attention mechanism, generating fusion features that reflect the interaction pattern between peptide sequences and active pockets. A dual-task prediction module is used to construct a decoupled prediction architecture consisting of a classification branch and a ranking branch. The classification branch and the ranking branch each have independent parameter layers and output heads. The classification branch receives the fused features generated by the feature alignment module and outputs the classification probability that the peptide has DPP-IV inhibitory activity. The ranking branch adopts an independent ranking model, which is executed in parallel with the classification branch. The ranking model is trained by constructing preference pairs based only on half-inhibition concentration data within the same assay source, and outputs the activity ranking score of the candidate peptide. The output module is used for virtual screening output. Based on the classification probability and the activity ranking score, it comprehensively ranks the candidate peptides and outputs the ranking results of the candidate peptides' activity strength. In the virtual screening output stage, it performs multi-dimensional credibility assessment and safety assessment on the candidate peptides. Specifically, it determines the applicable domain by calculating the k-nearest neighbor (kNN) distance between the candidate peptide and the training set samples, using a preset quantile threshold of the training set as the boundary of the applicable domain, and downweighting candidate peptides that exceed the boundary of the applicable domain. It also uses the ADMET prediction model to quantitatively assess the biosafety of the candidate peptide segments and screens out candidate peptides that do not meet the preset safety threshold.

[0020] The beneficial effects of the embodiments of this application are as follows: (1) By constructing a three-level negative sample hierarchical system including indistinguishable negative samples generated by restricted Boltzmann machine, the distribution distance between positive and negative samples in the feature space is reduced, forcing the model to learn more refined discrimination boundaries. The sequence deep semantic embedding features extracted by protein language model and the active pocket structure features based on molecular docking are integrated and explicitly aligned through cross attention mechanism, which effectively improves the model's discrimination ability and prediction accuracy for boundary samples. On the independent test set, our method achieved a classification accuracy of 0.8099, an ROC-AUC (Receiver Operating Characteristic Curve Area Under the Curve) of 0.8904, a PR-AUC (Precision-Recall Curve Area Under the Curve) of 0.9275, and an F1 score of 0.8301. Ablation experiments showed that removing difficult-to-distinguish negative samples reduced the ROC-AUC by approximately 0.74 percentage points, removing protein language model embedding features reduced the ROC-AUC by 3.81 percentage points, and removing transfer learning pre-training features reduced the ROC-AUC by 3.02 percentage points.

[0021] (2) By using a cross-attention mechanism with residue-level embedding of peptide sequences as query vectors and physicochemical characteristics of active pocket residues as bond and value vectors, an explicit and interpretable characterization of substrate-enzyme binding patterns was achieved. The intensity of attention interactions between each residue position of the output peptide and key residues in the active pocket can be quantified, thereby identifying key amino acid sites that play a decisive role in activity and their correspondence with specific pocket residues, providing a structurally operable basis for subsequent site-specific peptide modification.

[0022] (3) By constructing a decoupled prediction architecture with independent parameter layers and output heads for the classification branch and the ranking branch respectively, and using IC based only on the same determination source. 50 By constructing preference pairs from the data and training the LambdaRank ranking model with normalized discounted cumulative gain as the optimization objective, the cross-source IC is avoided at the training mechanism level. 50 The systematic bias in the data enabled a reliable ranking of candidate peptide activity. This included measured IC50 values. 50 On the test samples, the Spearman correlation coefficient was no less than 0.54, NDCG@20 no less than 0.55, and EF@5 no less than 2.4-fold. This provides an effective quantitative basis for prioritizing the experimental validation of candidate peptides.

[0023] (4) By generating an attention heatmap based on the cross-attention weight matrix, the residue positions in the peptide sequence that have a significant attention response to the key residues in the S1 or S2 sub-pocket can be identified. Based on the physicochemical interaction type, suggestions for site modification to enhance hydrophobic interaction or enhance charged interaction are output, thus opening up a complete closed loop from activity prediction and mechanism analysis to rational design, extending the computational tool from virtual screening function to peptide-directed modification guidance function.

[0024] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0025] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for screening DPP-IV repressive peptides based on source-sensing activity sequencing according to embodiments of this application; Figure 2 This is a schematic diagram of the sample category distribution according to a specific embodiment of this application; Figure 3 To measure the quantitative peptide sample according to a specific embodiment of this application in different pICs 50 Histogram of sample size distribution within the activity value range; Figure 4 A feature correlation heatmap according to a specific embodiment of this application; Figure 5 An attention heatmap according to a specific embodiment of this application; Figure 6 This is a schematic diagram of an integration model confusion matrix according to a specific embodiment of this application; Figure 7 This is a block diagram of a source-sensing activity-sequencing DPP-IV inhibitory peptide screening device provided according to an embodiment of this application. Detailed Implementation

[0026] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0027] This application first provides a method for screening DPP-IV repressive peptides based on source-sensory activity sequencing. This method uses IC50 to screen DPP-IV repressive peptides. 50The activity value was incorporated into the DPP-IV inhibitory peptide calculation and evaluation system, realizing the dual functions of "classification and qualitative judgment" and "ranking and quantitative evaluation".

[0028] Specifically, Figure 1 The flowchart of the source-aware activity sequencing-based DPP-IV inhibitory peptide screening method provided in the embodiments of this application is as follows: Figure 1 As shown, this source-sensing activity sequencing-based DPP-IV repressive peptide screening method includes the following steps: In step S1, a training dataset is constructed, containing positive samples and at least three levels of negative samples. The three levels of negative samples include at least: a first negative sample randomly sampled from a public database, and a second negative sample generated by a generative model that is similar to the positive samples in physicochemical properties but does not possess DPP-IV inhibitory activity. The peptide samples are standardized, and the standardization process includes at least: sequence redundancy removal, splitting the training and test sets by clusters, and adjusting the IC... 50 The data undergoes standardized preprocessing.

[0029] The purpose of step S1 is to construct a high-quality, highly discriminative training dataset to lay the data foundation for subsequent multimodal feature extraction and model training.

[0030] First, a high-confidence positive sample set is constructed. In this application embodiment, the positive samples refer to peptide sequences that have been verified and confirmed to have DPP-IV inhibitory activity through in vitro enzyme activity experiments.

[0031] The positive samples in this application's embodiments are integrated from the following six authoritative data sources: BIOPEP-UWM (Bioactive Peptide Database), PEPLab (Food-Derived Bioactive Peptide Database), BERT-DPPIV (BERT-based dipeptidyl peptidase IV inhibitory peptide benchmark dataset), and antidmpred (anti-diabetic peptide prediction dataset). The dataset includes a multi-feature training dataset of dipeptidyl peptidase IV repressor peptides and sequences validated by in vitro enzyme activity experiments from published literature. A total of 2639 original positive samples were obtained, which were then precisely deduplicated to 1012. These were further clustered using the CD-HIT (ClusterDatabase at High Identity with Tolerance, a sequence clustering and redundancy removal tool) tool at an 80% sequence consistency threshold, ultimately retaining 912 high-quality positive samples as the positive sample set for model training and evaluation.

[0032] To prevent test set leakage and overly optimistic evaluation results due to sequence similarity, this application employs a two-level CD-HIT redundancy removal scheme here: The first stage uses a sequence consistency threshold of no less than 80% to cluster all peptides and remove redundancy.

[0033] The second level ensures that the training and test sets are hard-split according to CD-HIT clusters, and sequences within the same cluster are prohibited from appearing across sets. This measure ensures that the test set is completely independent of the model and can more realistically reflect the model's generalization ability on unseen sequences. The final dataset is hierarchically divided in an 8:2 ratio into the iDPP-IV-CV (Dipeptidyl peptidase IV repressor peptide cross-validation training set) training set and the iDPP-IV-TS (Dipeptidyl peptidase IV repressor peptide independent test set) independent test set.

[0034] In this application, "negative sample" refers to a peptide sequence that does not possess DPP-IV inhibitory activity. To address the issue of a single negative sample construction strategy in the aforementioned background art, this application constructs a "three-level negative sample hierarchical system," which divides negative samples into three progressive levels based on the characteristic spatial distance between negative and positive samples.

[0035] Level 1 negative samples: Random negative samples: randomly sampled from the non-inhibitor partitions of public databases such as UniProt and PepBank. The sequence statistical characteristics (such as amino acid composition, hydrophobicity distribution, etc.) of this level of negative samples differ significantly from those of active peptides, providing a broad and clearly differentiated inactive reference baseline for the model. The number of samples included is no less than 800.

[0036] Level 2 Negative Samples: Difficult-to-Distinguish Negative Samples: These "difficult-to-distinguish negative samples" refer to synthetic peptide sequences generated by the generative model that are highly similar to positive samples in terms of amino acid composition, molecular weight, isoelectric point, hydrophobicity, and other physicochemical properties, but lack DPP-IV inhibitory activity due to differences in implicit structural features. The aim is to narrow the distance between positive and negative samples in the feature space, forcing the model to learn more refined discrimination boundaries, rather than relying solely on superficial differences in physicochemical distribution for coarse classification.

[0037] This application employs a Restricted Boltzmann Machine (RBM) as the generative model to generate second-level negative samples. A Restricted Boltzmann Machine is an undirected probabilistic graphical model consisting of a visible layer and hidden layers. The RBM energy function is defined as: ; (1) in, The visible layer state vector, This represents the state value of the i-th node in the visible layer. The hidden layer state vector, Let j be the state value of the j-th node in the hidden layer. For the visible layer bias term, For hidden layer bias terms, Visible layer node With hidden layer nodes The connection weights between them.

[0038] In the RBM construction of this application, the visible layer nodes correspond to the physicochemical characteristic parameters of the positive samples, including at least amino acid composition, molecular weight, isoelectric point, and hydrophobicity. After the RBM training is completed, the contrastive divergence algorithm is used to learn the joint probability distribution of the positive samples. After training converges, positive samples are used as seeds, and Gibbs sampling is used to iteratively sample in the probability space learned by RBM to generate synthetic sequences that are highly similar to positive samples in physicochemical properties but do not have DPP-IV inhibitory activity.

[0039] This application generated 293 poorly distinguishable negative samples, which were included in the training dataset after ADMET pharmacokinetic screening. After removing this module, the accuracy changed from 0.8099 to 0.8143 (+0.44%), but the ROC-AUC decreased from 0.8904 to 0.8830 (-0.74%). This indicates that the main value of this module lies in improving the ranking and resolution of boundary samples, rather than the accuracy under a single threshold.

[0040] Alternatively, in addition to using RBM generation, indistinguishable negative samples can be generated using generative adversarial networks or variational autoencoders, or they can be expanded at low cost by performing conserved amino acid substitutions on positive sample sequences.

[0041] Level 3 negative samples: High-confidence negative samples: Database non-inhibitory peptide sequences confirmed to contain no known DPP-IV activity inhibitory motifs, with low risk of false negatives, used to further enrich the diversity of negative samples and ensure sufficient coverage of training data.

[0042] All constructed sample sequences were screened using ADMET to remove sequences with potential toxicity or sensitization risks, ensuring that all sequences in the dataset have reasonable pharmacokinetic characteristics and safety.

[0043] For those with IC 50 The measured values ​​and peptide samples intended for subsequent ranking model training need to undergo the following preprocessing steps. 50 The half-inhibition concentration (WIC) is the concentration of an inhibitor required to reduce the activity of a target enzyme by 50%, and it is a core quantitative indicator for measuring inhibitor activity.

[0044] Specifically, the first step is to automatically identify and convert the units. This is because IC values ​​from different literature sources... 50 The original units of the data are inconsistent, including millimoles per liter (mM), micromoles per liter (μM), nanomoles per liter (nM), etc. This application uses IC testing... 50 The order of magnitude of the raw data is automatically identified by its initial unit, and then IC is... 50 Value to pIC50 Value. pIC 50 IC 50 The negative logarithmic form of molar concentration can effectively achieve linear comparison of activity values ​​and multidimensional distribution calibration. The specific conversion formula is as follows: If the initial unit identified is mM, then the conversion formula is: (2-1) If the initial unit identified is μM, then the conversion formula is: (2-2) If the initial unit identified is nM, then the conversion formula is: (2-3) After completing unit conversion and numerical linearization, this application further performs a screening process for inaccurate records and non-homologous data. For inaccurate activity records in the original data containing range symbols such as ">" or "<", since they can only indicate the activity range rather than the precise value, they cannot meet the precise "ordinal constraint" condition (i.e., it is impossible to clearly determine whether the activity of peptide i is superior to that of peptide j in subsequent steps). Therefore, this application automatically screens out such inaccurate records through rule matching and does not include them in the preference pair construction process of the ranking model.

[0045] Therefore, this application retains only enzymatic assay data of the same assay type (such as CHEMBL assay type=B), and removes data from non-enzymatic assay sources such as cell activity, in order to eliminate additional biases introduced by differences in assay conditions and ensure the consistency and comparability of training data sources.

[0046] Through the above preprocessing steps, this application constructs an IC that meets the "ordinal constraint consistency" requirement. 50 A dataset, meaning that there is a clear and consistent comparison of the activity levels of each data item, lays the data foundation for the subsequent ranking model's "same source preference pair" construction strategy.

[0047] Therefore, this application constructs a highly reliable dataset and a hierarchical system of negative samples through step S1. Figure 2 This is a schematic diagram illustrating the sample category distribution according to a specific embodiment of this application, such as... Figure 2 As shown in (a) of the paper, the original summary sample consisted of 2312 entries (1366 positive and 946 negative). After standard amino acid validity screening, 2311 entries were suitable for modeling, including 1848 entries in the training set and 463 entries in the test set. All subsequent classification metrics in this paper will be based on this modeling snapshot. The dataset described above meets the needs of model training and evaluation in terms of both scale and sample diversity. Figure 2 (b) shows the availability ratio structure of all 2312 data points when facing different decoupled branch tasks, including IC. 50The regression / ranking of the data yielded 20.60% of the groups (615 records in total), excluding IC. 50 The percentage of usable groups for classification only was 79.40% (1697 records in total). This asymmetric distribution strongly demonstrates the objective necessity of the dual-task decoupled prediction architecture of classification and ranking adopted in this application.

[0048] Figure 3 To measure the quantitative peptide sample according to a specific embodiment of this application in different pICs 50 Histogram of sample size distribution within the activity value range, such as Figure 3 As shown in the figure, this graph represents the quantitative subset of peptides at different pICs. 50 Histogram of sample number (frequency) distribution across the activity value range. Statistical results indicate that the pIC of the quantitative subset... 50 The activity values ​​exhibited excellent continuity, unimodality, and local symmetry. The overall mean of the peptide samples was 3.62, the median was 3.59, and the core activity range was smoothly and widely distributed between 2.0 and 5.0.

[0049] In step S2, multimodal features of peptides are extracted, including at least: sequence deep semantic embedding features extracted based on a protein language model, and pocket structure features extracted from the active pocket region of the DPP-IV protein through molecular docking as input to the model features.

[0050] The core objective of step S2 in this embodiment is to comprehensively characterize peptides from multiple information dimensions, overcoming the problem of insufficient capture of complex activity information caused by relying solely on single shallow sequence statistical features in the aforementioned background technologies. The "multimodal" nature of this embodiment refers to the integration of multiple feature types from different information sources, with different data formats and semantic levels. Step S2 constructs four feature extraction branches, extracting features from sequence statistics, deep semantics, molecular topology, and active pocket structure, respectively. It should be noted that the deep semantic embedding features (branch B) and the active pocket structure features (branch D) are the core feature sources for step S2.

[0051] Branch A is used to extract classic sequence statistical features of peptides, providing a basic feature space for the classification model. Specifically, the extracted sequence descriptors include: amino acid composition, i.e., the frequency of each of the 20 standard amino acids in the peptide sequence, generating a 20-dimensional feature vector; dipeptide composition, i.e., the frequency distribution of all adjacent amino acid pairs in the sequence, generating a 400-dimensional feature vector; composition-transformation-distribution, used to describe the distribution pattern of amino acids along the sequence; and pseudo-amino acid composition, incorporating physicochemical information such as the hydrophobicity, hydrophilicity, and side chain volume of amino acids to enhance the expressive power of sequence features. After feature screening and redundancy removal, 617 high-quality sequence statistical features are retained as the main input for the subsequent classification model.

[0052] like Figure 4 As shown, this application calculated the correlation heatmap between various encoding methods to assess the independence between different modal features. Specifically, AAC represents amino acid composition, DPC represents dipeptide composition, CTD represents composition-transition-distribution, and PAAC represents amphiphilic pseudo-amino acid composition. The results indicate that features of different dimensions capture different physicochemical properties and molecular structural information of peptides, exhibiting good feature complementarity and providing more comprehensive feature representation for subsequent classification and ranking models.

[0053] Branch B is used to extract deep semantic embedding features of peptide sequences.

[0054] Specifically, this application employs a protein language model as a deep semantic extractor. Protein language models can automatically learn contextual dependencies and evolutionary conservation information between amino acid residues. In this application, the protein language model used is specifically the ESM-2 model, a series of protein language models based on the Transformer architecture released by Meta AI.

[0055] Optionally, the ESM-2 model can be replaced with other protein-pretrained language models, or by concatenating or weighting the embedding vectors of multiple language models. This application loads ESM-2... Version. This model was pre-trained on the UniRef50 protein sequence database using a masked language modeling objective, enabling it to learn rich amino acid co-evolution patterns and physicochemical constraints in the hidden layers.

[0056] The extraction process of deep semantic embedding features is as follows. First, the peptide sequence is input into the ESM-2 model in the form of single-letter amino acid codes. The model uses alternating computations of multi-layer self-attention mechanism and multi-layer perceptron to output a hidden state matrix at each layer, which is the same length as the input sequence and has a fixed-dimensional hidden state vector corresponding to each residue position.

[0057] Next, the hidden state matrix of the last layer (i.e., the 12th layer) of the model is extracted. Average pooling is then performed on this hidden state matrix along the sequence length dimension. Average pooling is the element-wise arithmetic average of the hidden state vectors at all positions in the sequence, thus compressing the variable-length sequence representation into a fixed-length vector. This operation can be formally represented as: ; (3) in, This refers to the number of amino acid residues in the peptide sequence. For the 12th layer of the ESM-2 model in the... The hidden state vector output at each residue position This is the final sequence-level embedding vector.

[0058] The above average pooling operation yields a fixed-length sequence deep semantic embedding vector with a dimension of 480. This embedding vector captures the long-range contextual dependencies and evolutionary conservation information between amino acid residues in the peptide sequence, serving as an important semantic supplement for subsequent feature fusion.

[0059] Branch C is used to extract the molecular graph topological features of peptides. In this application, it is positioned as an auxiliary reference feature in the virtual screening stage and does not participate in the training of the main classification model and ranking model. Its purpose is to decouple the topological geometric information from the global semantic information and avoid noise in the molecular graph features from interfering with the generalization ability of the main feature space. In this embodiment, the auxiliary reference feature is activated only in the post-processing stage where fine-grained differentiation of highly active candidate peptides is required, so as to enhance the robustness of local screening through structural prior information while maintaining the generalization of the main model.

[0060] This application employs a 3-layer GAT (Graph Attention Network), with each layer configured with 4 attention heads and a hidden layer dimension of 128. Through message passing and node feature updates, the GAT ultimately outputs a global graph-level representation of the peptide molecular graph, serving as a quantitative expression of the molecular topological geometric features.

[0061] Branch D is used to extract the structural interaction features between the peptide and the active pocket of the DPP-IV target protein. In this application, the "active pocket" refers to the three-dimensional spatial region on the surface of the DPP-IV protein that contains and catalyzes the substrate, surrounded by a specific set of amino acid residues. Introducing the structural and physicochemical information of these "active pocket residues" into the model enables the prediction method of this application to have the explicit characterization capability of substrate-enzyme binding patterns. That is, it can not only determine whether a peptide is active, but also explain which residue on the peptide interacts with which specific residue in the active pocket and under what conditions.

[0062] First, structural preprocessing was performed. DPP-IV crystal structures 1NU8 and 5J3J were downloaded from the RCSB PDB (Collaborative Research Foundation for Structural Bioinformatics Protein Database). Non-functional chains, cocrystal ligands, and water of crystallization molecules were removed using PyMOL, retaining only the functional acceptor chain for subsequent docking calculations. Optionally, the DPP-IV crystal structure could be replaced with other DPP-IV inhibitor cocrystal structures (such as PDB ID: 1RWQ, 2P8S) or homology-modeled structures; docking software could be replaced with Glide, AutoDock4, or Rosetta FlexPepDock, still yielding usable structural features.

[0063] Furthermore, based on the catalytic triplet and the key substrate-binding residues reported in the literature, the active pocket range was determined. In this application, the "catalytic triplet" refers to the three core amino acid residues essential for DPP-IV to exert its catalytic function, specifically Ser630, Asp708, and His740.

[0064] The active pocket is further divided into the following sub-pocket regions: The S1 subpocket is responsible for containing the side chain of the amino acid at the P1 position of the substrate. Its key residues include Tyr631, Val656, Trp659, Tyr662, Tyr666, and Val711. The S2 subpocket is responsible for containing the side chain of the amino acid at the P2 position of the substrate. Its key residues include Arg125, Glu205, Glu206, Phe357, and Arg358. The S1' and S2' extension regions further expand the substrate binding range, and their key residues include Tyr547, Trp629, Ser209, Arg669, and Asn710.

[0065] In molecular docking and structural feature extraction, this application uses AutoDock Vina for peptide-enzyme molecular docking. The top 5 conformations with the best binding affinity are retained for each peptide docking. From the docking results, the following quantitative structural features are extracted as the output of branch D: optimal binding energy (Ebest), mean and standard deviation of the affinity scores of the top 5 conformations, number of hydrogen bond donor-acceptor pairs, number of hydrophobic contacts, contact indicator markers of each residue in the catalytic triplet, and contact residue counts between the peptide and each of the S1, S2, S1', and S2' sub-pocket regions.

[0066] The active pocket structure features extracted from branch D are used in the subsequent feature alignment step for cross-attention calculation with the sequence deep semantic embedding features, which is the key information basis for realizing explicit modeling of substrate-enzyme binding patterns.

[0067] After completing the feature extraction of the above four branches, feature organization and filtering are performed. Specifically, the features extracted by branch A (sequence statistical features) and branch B (deep semantic embedding features) are merged into the main feature space of the classification model and ranking model of this application, and the merged dimension is no less than 1000 dimensions.

[0068] Branch C (molecular graph topological features) and branch D (active pocket structural features) are only used as auxiliary references in the subsequent virtual screening stage and are not included in the main model training process. The purpose is to avoid the noise in the molecular graph topological features and the high-dimensional sparsity of the pocket structural features from interfering with the main feature space. Prior structural information is only introduced in the post-processing stage where it is necessary to finely distinguish the activity strength of candidate peptides.

[0069] For the merged branch A and branch B features (totaling 1097 dimensions), the XGBoost gain metric was used to rank the features by importance. The top 300 features were retained to form a dimensionality-reduced feature set for subsequent training of the ranking model. The selection results showed that in this dimensionality-reduced feature set, sequence statistical features retained 110 dimensions (contributing approximately 49.6% of the importance), and ESM-2 embedding features retained 190 dimensions (contributing approximately 50.4% of the importance), indicating that the two types of features are complementary in distinguishing DPP-IV inhibitory activity. The classification model was trained using all sequence statistical features from branch A, achieving optimal performance with the current data scale.

[0070] In step S3, the sequence deep semantic embedding features are used as query vectors and pocket structure features are used as key and value vectors through a cross-attention mechanism. These are then explicitly aligned to generate fusion features that reflect the interaction pattern between peptide sequences and active pockets.

[0071] The core objective of step S3 in this application embodiment is to establish an explicit association between the peptide sequence and the DPP-IV active pocket, so as to overcome the problem that "substrate-enzyme binding mode cannot be explicitly modeled and prediction results lack structural interpretability" caused by the simple splicing of multimodal features in the above-mentioned background art.

[0072] In this application, "explicit alignment" refers to the use of a cross-attention mechanism to calculate the interaction strength between each amino acid residue position in the peptide sequence and each key residue in the active pocket, generating a quantifiable, traceable, and visualizeable attention weight matrix. This mechanism enables the model to not only output a classification result of "whether the peptide has inhibitory activity," but also to clearly reveal "which residue position on the peptide contributes to the activity because of its strong interaction with which specific residue in the active pocket," providing structural guidance for subsequent rational design.

[0073] The cross-attention mechanism requires three types of input: query vector, key vector, and value vector. Their construction will be explained below.

[0074] Specifically, the query vector originates from the deep semantic embedding features of the peptide sequence, but its granularity differs from the full-sequence-level embedding used for classification in step S2. In this step S3, the residue-level embedding matrix output by the ESM-2 model is used as the query vector. "Residue-level embedding" refers to retaining the complete output, where each amino acid residue position corresponds to an independent embedding vector, without compression through average pooling. Let the peptide sequence length be... The hidden layer dimension of ESM-2 is Then the mathematical form of the query vector is a matrix. Each row of this matrix , corresponding to the first in the peptide sequence Deep semantic representation of amino acid residues under context-dependent and evolutionary conservation constraints.

[0075] The bond vector and value vector are derived from the physicochemical feature encoding of the DPP-IV active pocket residues. The term "pocket residue" refers to the key amino acid residues defined in step S2 that constitute the three-dimensional spatial region of the DPP-IV active pocket. This application extracts a physicochemical feature vector for each pocket residue that contains at least the following dimensional information.

[0076] Includes: Amino acid type: Residue identity represented by one-hot encoding or pre-trained embedding; Kyte-Doolittle hydrophobicity scale value: An empirically determined quantitative index of amino acid hydrophobicity, reflecting the hydrophilic or hydrophobic tendency of the residue side chain; Formal charge: Net charge state of the residue side chain under physiological pH conditions; Number of hydrogen bond donors: Number of hydrogen bond donor groups that the residue side chain can provide; Number of hydrogen bond acceptors: Number of hydrogen bond acceptor groups that the residue side chain can provide; Van der Waals volume: Spatial volume of the residue side chain (in ų); Sub-pocket category number: Identifies whether the residue belongs to the S1 sub-pocket, S2 sub-pocket, S1' extension region, S2' extension region, or catalytic triplet.

[0077] Let the number of encoded residues included in the active pocket be... The physicochemical feature vector dimension of each residue is Then the mathematical form of the key vector is a matrix. The mathematical form of a value vector is a matrix. In one embodiment of this application, The values ​​are 8-dimensional, specifically including: amino acid type, hydrophobicity scale value, formal charge, number of hydrogen bond donors, number of hydrogen bond acceptors, van der Waals volume, sub-pocket category number, and catalytic triplet label.

[0078] After constructing Q, K, and V, the scaled dot product attention calculation is performed: the dot product similarity matrix between Q and K is calculated and scaled by a scaling factor. After scaling, Softmax normalization is performed along the pocket residue dimension to obtain the attention weight matrix A, which is then multiplied by V to obtain the context vector matrix. The mathematical expression of this process is: ; (4) in, The function operates along the rows of the matrix. produce The similarity score matrix of dimension 1, after scaling and softmax, is obtained as follows: The attention weight matrix of dimension is denoted as . .

[0079] Extension of Multi-Head Attention: To further enhance the model's expressive power and its ability to capture different interaction patterns, this application employs a multi-head attention mechanism to compute multiple independent attention sets in parallel. In a multi-head configuration, [the following is implied:] ... Using different learnable projection matrices The projection is mapped to multiple subspaces, with each head having a projection dimension of d / h, for example, 480 / 8 = 60 dimensions. The scaling dot product attention calculation is performed independently in each subspace. Finally, the outputs of all heads are concatenated and processed through a linear projection matrix. Mapping back to the original dimension. Its mathematical expression is: ; (5) ; (6) in, In one embodiment of this application, the number of attention heads is considered. The value is 8.

[0080] Through the multi-head cross-attention calculation described above, the hidden state representation sequence of each residue position in the peptide sequence after "pocket information selective aggregation" is obtained. This sequence is then compared with the original query vector. (i.e., residue-level embedding of peptide segments) is performed to generate the final fusion feature.

[0081] The fusion method can employ a combination of residual connections and layer normalization. First, the output of multi-head cross-attention is combined with the original query vector. Element-wise addition (residual connection) followed by layer normalization stabilizes gradient propagation and preserves the semantic information of the original sequence. Then, it can be further processed through feedforward neural network layers and nonlinear activation functions to obtain the final fused feature matrix. Each row of this fused feature matrix simultaneously contains the intrinsic semantic information of the corresponding residue and the contextual structural information of its interaction with the entire active pocket.

[0082] This fusion feature will be used as input to the classification branch in the subsequent step S4 to determine whether the peptide has DPP-IV inhibitory activity.

[0083] Furthermore, while outputting prediction results, the method of this application can provide structural interpretability of the substrate-enzyme binding pattern. This effect is achieved by generating... Figure 5 The cross-attention weight heatmap of sequence residues and pocket residues is shown.

[0084] Generation of the attention heatmap: The attention weight matrix calculated in step S3 is used to generate the attention heatmap. Perform visualization processing. Specifically, visualize the matrix. Each element in This is mapped to the color depth or hue of a pixel block in a two-dimensional image, where the horizontal axis corresponds to the active pocket residues. The vertical axis corresponds to the position of peptide residues. The resulting heatmap can visually identify the interaction strength between the positions of each residue in the peptide sequence and the residues in the active pocket.

[0085] The attention heatmap analysis function identifies high-brightness areas on the attention heatmap.

[0086] In one specific embodiment, the cross-attention weight heatmap shows the attention weight distribution between 8 attention heads (h=8) and 12 key residues in the DPP-IV active pocket (including Glu205, Glu206, Tyr662, Ser630, Asp708, His740, Arg125, Asn710, Trp629, Phe357, Lys554, Tyr631). Each attention head focuses on a different binding subregion (such as the catalytic triplet, substrate binding site, S1 sub-pocket, S2 sub-pocket, etc.), revealing the position-specific interaction pattern between peptide sequences and DPP-IV pocket residues.

[0087] This attention heatmap, based on attention weights, guides targeted modifications to different sub-pockets. For key residue positions with high attention weights, if they exhibit strong interactions with hydrophobic pocket residues (such as Tyr662 and Val656 in the S1 sub-pocket), it is recommended to replace the amino acid at that position with a residue with a larger hydrophobic side chain to enhance the hydrophobic interaction; if they exhibit strong interactions with charged pocket residues (such as Glu205 / Glu206 and Arg125 in the S2 sub-pocket), it is recommended to replace them with amino acids with complementary charges to enhance the electrostatic interaction.

[0088] Through the above functions, this application establishes a complete closed loop from "prediction" to "interpretation" to "design". The classification branch answers "whether it is active", the ranking branch answers "how strong is the activity", and the cross-attention analysis answers "which site is active" and guides "how to modify it to make it more active".

[0089] In step S4, a decoupled prediction architecture consisting of a classification branch and a ranking branch is constructed. The classification and ranking branches each have independent parameter layers and output heads. The classification branch receives fusion features and outputs the classification probability of whether a peptide has DPP-IV inhibitory activity. The ranking branch uses an independent ranking model, executed in parallel with the classification branch, and uses IC50 based solely on the same assay source. 50 The data-constructed preference pairs are used to train the ranking model, which outputs the activity ranking scores of candidate peptides.

[0090] The core objective of step S4 in this embodiment is to construct a decoupled prediction architecture for the dual tasks of "classification + ranking". This is due to the cross-document source IC... 50 The data contains systematic biases. If classification and ranking tasks are coupled during training, the model will learn the biases from the data sources rather than the true structure-effect relationship, resulting in extremely poor ranking ability. Therefore, this application decouples classification and ranking into two independent model branches, each using the feature subset and training strategy best suited to its own task objective.

[0091] The decoupled prediction architecture consists of two parts: a classification branch and a ranking branch.

[0092] The classification branch is a heterogeneous ensemble classifier whose core function is to receive fused features and output a binary classification prediction result of whether a peptide has DPP-IV inhibitory activity. The classification branch has its own independent parameter layer, including feature weights of each base model, tree structure node splitting thresholds, leaf node scores, etc.; and it has its own independent output head, which merges the prediction probabilities of each base model through a soft voting mechanism to output the classification probability.

[0093] The ranking branch is a ranking model independent of the classification branch, with its own independent parameter layer and output header. It executes in parallel with the classification branch, and is based solely on IC (Integer Orientation) from the same determination source. 50 The preference pairs constructed from the data are used for training, and the output is a ranking result of the activity strength of candidate peptides. In this embodiment of the application, no less than 50 sample groups from the same source are selected and constructed from the original dataset to ensure that the subsequent LambdaRank model can learn the true cross-sample preference relationship, rather than the systematic bias between batches.

[0094] The "decoupling" of the classification branch and the ranking branch in this application embodiment is reflected in the following three levels: First, structural decoupling, with each branch having an independent set of parameter layers and output heads; second, input decoupling, with each branch using at least partially different feature subsets as inputs; and third, training decoupling, with each branch not sharing a loss function and not jointly optimizing in the same training process, and either branch can be independently upgraded or replaced without affecting the performance of the other branch.

[0095] The classification branch uses the full sequence statistical features of branch A as input to construct a heterogeneous ensemble classifier.

[0096] The heterogeneous ensemble classifier in this application consists of at least three heterogeneous base models. In one embodiment, a combination of four tree models—XGBoost, LightGBM, CatBoost, and Random Forest—is used. The consideration for choosing tree models and constructing heterogeneous ensembles is that tree models have a natural inductive bias advantage for tabular structured features (such as sequence statistical features like amino acid composition and dipeptide composition), while heterogeneous ensembles (i.e., combining base learners using different algorithmic principles) can reduce overall prediction variance and improve generalization robustness through model diversity.

[0097] This application employs a soft voting mechanism to fuse the predicted probabilities of various base models. Soft voting involves taking the arithmetic mean of the positive class probabilities output by each base model, using this mean as the final classification probability output. Compared to hard voting (which only takes the class labels of the majority of models), soft voting preserves prediction confidence information and exhibits smoother discriminative characteristics for boundary samples whose probabilities are close to the decision boundary.

[0098] This application employs the Optuna framework to search for hyperparameters of each basic model. The out-of-bag AUC with 5-fold hierarchical cross-validation is used as the optimization objective, and at least 30 trials are performed.

[0099] In a preferred embodiment, to further address the technical limitations of the limited size of the peptide dataset, the training of the classification branch can employ a two-stage transfer learning strategy. Transfer learning refers to a training paradigm that transfers knowledge learned from source domain data to a target domain task.

[0100] Phase 1: Pre-training is performed using activity data of non-peptide DPP-IV target small molecule inhibitors (such as 2879 DPP-IV target compound records in the ChEMBL database) to enable the classification branches to learn the general chemical patterns and structure-activity relationships of DPP-IV inhibitory activity. Optionally, the pre-training data can be replaced with data of other DPP-IV inhibitory compounds.

[0101] Phase Two: Fine-tuning. After pre-training, the model is transferred to peptide data for fine-tuning. During fine-tuning, the low-level feature extraction parameters are frozen, and only the high-level decision layer parameters are updated, with optimization performed at a learning rate lower than that used in the pre-training phase. This strategy of "freezing the low-level parameters + fine-tuning with a small learning rate" allows the model to gradually adapt to the specific feature distribution of peptide sequences while retaining the general DPP-IV activity patterns learned in the pre-training phase, effectively mitigating the risk of overfitting on small sample peptide data.

[0102] On the independent test set iDPP-IV-TS, the classification branch achieved classification performance of 0.8904 ROC-AUC, 0.9275 PR-AUC, 0.8301 F1 score, and 0.8099 precision. The PR-AUC of 0.9275 indicates that the model achieves a good balance between positive sample recall and precision, which is particularly important for applications where positive samples are relatively scarce, such as screening for DPP-IV inhibitory peptides.

[0103] Table 1 is a performance comparison table of the single-model iDPP-IV-TS test set.

[0104] Table 1

[0105] Table 1 shows the classification performance metrics of each base model. The ROC-AUC of each model exceeds 0.81. The ensemble model, by fusing the outputs of the four models through a soft voting mechanism, achieved stable performance of ROC-AUC = 0.8904 and PR-AUC = 0.9275.

[0106] Table 2 shows the overall classification performance of the final integrated model in this application embodiment on the iDPP-IV-TS independent test set.

[0107] Table 2

[0108] As shown in Table 2, the ensemble model achieves a good balance between positive sample recall and precision, making it suitable for screening DPP-IV inhibitory peptides. To further quantitatively and multidimensionally evaluate the boundary discrimination accuracy of the classification branches in the decoupled prediction architecture of this application, the embodiments of this application derive the confusion matrix of the ensemble model on the independent test set. Specific results are shown below. Figure 6 As shown.

[0109] Figure 6 The vertical axis represents the true labels, divided into true negative samples and true positive samples; the horizontal axis represents the model's predicted labels, divided into predicted negative samples and predicted positive samples. The intensity of the color bands on the right side of the figure represents the frequency density of samples falling into the corresponding intervals.

[0110] Combination Figure 6 At the top of the overall performance metrics, the ensemble model of this application ultimately achieved an accuracy of 0.810 and an F1 score of 0.830. These independent confusion matrix statistics demonstrate that this application, through its multi-model soft-voting ensemble architecture, can achieve refined modeling of the decision boundary of the peptide feature space. In particular, the number of false positives (FPs) was controlled to a low level of 30, demonstrating at the chemical effect level that this application possesses extremely strong background noise screening capabilities.

[0111] Furthermore, this application addresses "cross-laboratory IC" 50 The solution to the industry problem of "systematic bias" does not attempt to correct ICs through normalization or batch effect correction during the preprocessing stage. 50 Instead of "aligning" absolute values, this is avoided by using ordinal constraints at the training mechanism level.

[0112] The specific strategy is: to use only ICs from the same measurement source. 50 Data constructs preference pairs. A preference pair refers to a pair of data constructed from two ICs. 50 Sample pairs of comparable peptides are used to indicate the ordinal relationship in the model that "peptide A has a higher activity than peptide B".

[0113] The construction rules for preference pairs in this application embodiment are as follows: For any two peptides i and j from the same literature source or the same assay batch, obtain their pICs after preprocessing in step S1. 50 Value. If the IC50 of peptide i is... 50 IC50 values ​​less than peptide j 50 value (i.e. pIC) 50 i>pIC 50 If the order of peptides i and j is i, then the ordinal relationship between i and j is labeled as "i is superior to j", denoted as i>j. This preference does not care about how much stronger i is than j, but only about the relative ordinal relationship of "which is stronger".

[0114] The preference in this application is to construct only within the same source group and never across different source groups. For example, peptide a from literature A and peptide b from literature B, even though a has a numerical IC50 of 1,000, are different. 50 IC values ​​less than b 50 The values ​​do not constitute preference pairs. This is because the difference in IC50 values ​​between literature A and literature B may primarily stem from differences in experimental conditions, rather than differences in the activity of the peptides themselves. By constructing preference pairs within a group, the interference of cross-laboratory batch effects is completely avoided at the data organization level, ensuring the consistency of the model gradient source.

[0115] In one embodiment of this application, the term "same measurement source" not only refers to data strictly from the same literature, but can also be extended to data from the same measurement batch number, such as data records with the same assignment number in the CHEMBL database. This provides a more flexible and operable scope boundary for data grouping in actual engineering deployments.

[0116] In one embodiment, a preference pair set was constructed using 96 groups from the same source and 536 samples. It is used to train the ranking model.

[0117] In this application, the classification and ranking branches are trained independently using different feature subsets. Specifically, the classification branch uses the full sequence statistical features (617 dimensions) of branch A as input. The ranking branch uses a dimensionality-reduced feature set filtered by a feature gain metric as input. This dimensionality-reduced feature set is constructed from the 1097-dimensional complete feature space resulting from the merging of sequence statistical features and deep semantic embedding features, using XGBoost gain metrics to rank features by importance and retaining the top 300 features. In this dimensionality-reduced feature set, sequence statistical features retain 110 dimensions (contributing approximately 49.6% of importance), and ESM-2 embedding features retain 190 dimensions (contributing approximately 50.4% of importance). The contribution ratios of the two types of features are similar, indicating that sequence statistical features and deep semantic features are complementary in distinguishing activity ranking.

[0118] The ranking branch trains the LambdaRank model with Normalized Discounted Cumulative Gain (NDCG) as the direct optimization objective. NDCG is a widely used evaluation metric in information retrieval and ranking learning, used to measure whether highly active samples at the top of the ranking list can obtain higher model scores.

[0119] Define the set of preference pairs within the same measurement source s. For those belonging to For the preference pair (i, j) (i.e., i>j), the loss function of LambdaRank is: ; (7) in, and These are the model's prediction ranking scores for peptides i and j, respectively. In one embodiment of this application, hyperparameters for controlling the shape of the sigmoid function are described. The value is 1; The absolute value of the change in the NDCG index caused by swapping the positions of i and j in the sorted list is given by a base-2 logarithmic discount factor (log2(rank+1)).

[0120] The gradient properties of this loss function represent the mathematical implementation of the method used in this application to "avoid cross-source bias at the training mechanism level." Specifically, the summation range within the loss function is strictly limited to... In other words, "the set of preference pairs within the same measurement source" means that every gradient signal received by the model during training comes from the relative activity comparison results of two peptides within the same experimental batch; cross-source IC 50 The absolute difference never enters the gradient calculation chain, so the model does not learn any bias patterns related to cross-laboratory batch effects.

[0121] The ranking model is implemented using the LambdaRank algorithm within the LightGBM framework. LambdaRank is a list-level ranking learning algorithm based on gradient optimization. Its core idea is that when calculating the gradient, it not only considers the degree of matching between the model's predicted score and the true label, but also explicitly encodes the impact of swapping the ranking positions of any two samples on the final evaluation metric into the gradient direction. Optionally, the ranking model in this embodiment can use RankSVM, RankNet, or LambdaMART instead of LambdaRank, while still maintaining a certain level of ranking optimization capability.

[0122] The ranking branch trains the LambdaRank model with the normalized discounted cumulative gain as the direct optimization objective. In one specific embodiment, a set of preference pairs is constructed using 96 groups from the same source and 536 samples. The feature input is a Top300 dimension (sequence statistics + ESM-2 embedding), trained independently, and not jointly trained with the classification head.

[0123] In the test set IC 50 Subset (n=124, pIC) 50 The evaluation results of the ranking branch (range 3.30 to 8.33) are shown in Table 3. The Spearman rank correlation coefficient reached 0.542, indicating a significant positive correlation between the model's ranking of candidate peptide activity and the actual activity ranking; NDCG@20 reached 0.554, indicating that the quality of the active peptides in the top 20 of the ranking list is relatively high. In contrast, pure regression methods that do not use the same source preference strategy typically have Spearman rank correlation coefficients below 0.25 on the same data, and this application improves upon this by more than 100%.

[0124] Table 3

[0125] Experimental results show that the sorting head trained by the above ordinal constraints achieves an SCC of 0.542, which is significantly improved compared with the traditional regression method, and the NDCG@20 index reaches 0.554, verifying the superiority of the sorting branch in prioritizing candidate peptides.

[0126] Furthermore, the enrichment factor EF@5 reached 2.40-fold, indicating that the enrichment rate of bioactive peptides among the top 5 candidate peptides in the ranking list was 2.4 times that of random screening. This demonstrates that the ranking branch has practical value in determining the priority of candidates.

[0127] In addition to the two core tasks of classification and ranking, this application can also incorporate an auxiliary regression branch into the decoupled architecture to provide multi-task learning for systems with experimentally measured pICs. 50The value of the peptide provides a continuous reference value for its activity intensity.

[0128] This auxiliary regression branch uses Huber Loss as the loss function, which is robust to outliers and suitable for ICs with measurement noise. 50 Returning to the original scenario.

[0129] It's important to note that the classification and regression branches are trained decoupled and independently; that is, they are trained independently without sharing parameters or a joint loss function. The classification branch is optimized using cross-entropy loss, while the regression branch is optimized using Huber Loss. This decoupling design ensures that the prediction results of the regression branch do not negatively impact the discrimination boundary of the classification branch, thus guaranteeing that classification performance is not affected by the regression task.

[0130] In a multi-task learning framework, the model uses pIC 50 Regression is used as an auxiliary task. This application employs a dual-task approach of classification and LambdaRank ranking, with regression providing continuous reference values ​​for activity intensity. For applications with measured pIC... 50 The peptide samples with high values ​​were used to evaluate the fit quality of the regression branches. Table 4 provides a complete index of regression performance. This experiment used 615 IC50 values ​​(609 valid pairs after unit correction) to measure the regression performance. 50 Supervised regression training was performed using values ​​(μM), and 609 valid sequences were used in the model after being screened using standard amino acids. Because the training set came from literature data from more than 10 different laboratories (with varying substrates, buffers, temperatures, etc.), cross-laboratory IC50 was used. 50 There is a systematic bias, absolute value regression (R²) 2 However, the relative ranking ability reflected by Spearman Rank Correlation (SCC) has practical application value.

[0131] Table 4

[0132] The regression head provides a continuous reference for the activity intensity of candidate peptides; in one embodiment, its Spearman correlation coefficient is 0.599 and MAE is 0.342. This result indicates that the model has a moderate ability to fit activity trends; however, the upper limit of regression is still constrained by differences in cross-laboratory origins. The true ranking ability is assessed using the aforementioned LambdaRank head. The regression head is positioned as an auxiliary prediction tool; candidate selection is still determined jointly by classification probability and ranking score.

[0133] In step S5, the virtual screening output is generated. Based on the classification probability and activity ranking score, the candidate peptides are comprehensively ranked, and the ranking results of the candidate peptides' activity strength are output.

[0134] Step S5 combines the classification probabilities output from the classification branch and the activity ranking scores output from the ranking branch in step S4 to perform a comprehensive ranking of candidate peptides. In one embodiment, the comprehensive ranking score can be obtained by weighted fusion of the classification probabilities and ranking scores, ultimately outputting a ranking list of candidate peptides based on their activity levels.

[0135] Optionally, in some embodiments, the DPP-IV inhibitory peptide screening method based on source-aware activity ranking further includes: providing rational modification guidance for candidate peptides based on the attention heatmap generated in step S3.

[0136] Specifically, a preset weight threshold is first set to filter residue position pairs with significant interaction strength from the attention weight matrix. Each element A in the attention weight matrix A is then... ij Compared with a preset weight threshold, if A ij If the value is equal to or greater than a preset weight threshold, then the peptide residue position i is marked as a "high-response residue position". The value of this threshold can range from 20% to 50% of the maximum value of the attention weight matrix, and the specific value can be adjusted according to the actual application scenario and the desired analysis granularity.

[0137] During labeling, further attention is paid to whether the pocket residues corresponding to the highly responsive residues belong to the S1 or S2 sub-pocket. Key residues in the S1 sub-pocket include Tyr631, Val656, Trp659, Tyr662, Tyr666, and Val711, exhibiting strong hydrophobic properties; key residues in the S2 sub-pocket include Arg125, Glu205, Glu206, Phe357, and Arg358, which contain both charged and aromatic residues.

[0138] After identifying the locations of highly responsive residues and their corresponding pocket residues, a physicochemical interaction type determination is performed. The physicochemical interaction type refers to the category of primary non-covalent interaction patterns between peptide residues and active pocket residues based on their respective physicochemical properties. The specific determination rules are as follows: Hydrophobic interaction type: This type is determined when the pocket residues corresponding to the high-response residues have significant hydrophobic properties. For example, the pocket residues belong to the S1 sub-pockets dominated by hydrophobic residues (such as Tyr662 and Val656). Charged interaction type: This type is determined when the pocket residues corresponding to the high-response residues have a net charge under physiological pH conditions, such as Glu205 / Glu206 (negative charge) or Arg125 / Arg358 (positive charge) in the S2 sub-pocket. Mixed interaction type: This type is defined when the same peptide residues simultaneously exhibit high attention responses to multiple pocket residues (belonging to hydrophobic and charged types respectively).

[0139] Based on the determination of the physicochemical interaction type, site-specific modification suggestions for the peptide are provided. For hydrophobic interactions, it is recommended to replace the amino acid at that position with an amino acid with a larger hydrophobic side chain; for charged interactions, based on the charge properties of the pocket residues, it is recommended to replace them with amino acids with complementary charges.

[0140] In a preferred embodiment, the virtual screening output stage further includes: performing multidimensional reliability assessment and safety assessment on candidate peptides.

[0141] The multidimensional reliability assessment in this embodiment is performed as follows: Calculate the multidimensional features of the candidate peptide and the feature space distance (e.g., Euclidean distance) between each sample in the training set, and take the average k-nearest neighbor distance D. kNN Using training set D kNN A preset quantile threshold (e.g., the 95th quantile, approximately 2.8206 in one embodiment) is used as the boundary of the applicable domain for determining the applicable domain. When the D of the candidate peptide... kNN When the threshold is exceeded, it is determined to be an out-of-domain sample. A penalty factor γ (where γ∈[0.1, 0.5]) is introduced into its comprehensive ranking score to reduce its weight, thus lowering its ranking priority. However, because it may have a novel structure, it is given priority to enter the subsequent incremental verification queue.

[0142] The safety assessment in this application embodiment is performed as follows: ADMET prediction models are used to quantitatively assess the biocompatibility of candidate peptides across five dimensions: absorption, distribution, metabolism, excretion, and toxicity, eliminating candidate peptides that do not meet a preset safety threshold in any dimension. In one embodiment, the ADMET Lab 3.0 tool is used to perform the above assessment, supplemented by ToxinPred and AllerTOP tools for toxicity prediction and allergenicity assessment.

[0143] On the internal test set, approximately 94.6% of the samples fell within the applicable domain, indicating that the applicable domain determination method of this application has good predictive reliability for most candidate peptides. After temperature scaling calibration, the negative log-likelihood decreased from 1.415 to 0.735, and the Brier score decreased from 0.334 to 0.260, further validating the effectiveness of combining applicable domain determination with model calibration.

[0144] Through the above process, this application achieves a complete closed-loop technical route from "prediction" to "interpretation" and then to "design." Specifically, the classification branch answers "whether the peptide has DPP-IV inhibitory activity" (qualitative), the ranking branch answers "how the peptide ranks relatively in terms of activity in the candidate list" (quantitative); the cross-attention heatmap answers "which residue on the peptide contributes to its activity due to a strong interaction with which specific residue in the active pocket," i.e., mechanism analysis; based on the structured interaction information output by the interpretation layer, the above process answers "what amino acid substitutions at specific residue positions of the peptide might enhance its activity," i.e., rational modification guidance, and outputs specific and actionable site-specific modification suggestions. Thus, this application achieves a technological leap from automated screening to intelligent design by transforming the attention weight matrix output by the model into executable molecular design instructions.

[0145] To verify the effectiveness of the multimodal feature fusion framework and ordinal constraint learning strategy proposed in this application, this embodiment comprehensively evaluates the model performance and sets up multiple ablation experiments to demonstrate the contribution of each technical module.

[0146] The experimental method involved removing one component at a time while keeping the rest of the configuration unchanged, and observing the changes in model performance. Using the complete ensemble model (ROC-AUC=0.8904) as a baseline, the experimental results for each ablation variant are shown in Table 5. Table 5 presents the ablation experimental results for the embodiments of this application. After removing the difficult-to-distinguish negative samples, the accuracy changed from 0.8099 to 0.8143 (+0.44%), but the ROC-AUC decreased from 0.8904 to 0.8830 (-0.74%), indicating that the main value of this module lies in improving the sorting and discrimination ability of boundary samples, rather than the accuracy under a single threshold. After removing the ESM-2 embedding, the ROC-AUC decreased to 0.8523 (-3.81%), confirming the important role of the protein language model in sequence representation. After removing the transfer learning strategy, the ROC-AUC decreased to 0.8602 (-3.02%), verifying the gain effect of ChEMBL pre-training on peptide small-sample learning.

[0147] Table 5

[0148] It should be noted that the ablation rows marked with * in Table 5 are from the estimated values ​​of the full deep learning model v1, and are marked with an asterisk to distinguish them.

[0149] This embodiment also includes a control group for head ablation, with SCC as the primary evaluation metric and EF@5 as a secondary metric. SCC = 0.542, EF@5 = 2.40×, balancing ordination stability and candidate screening interpretability.

[0150] This embodiment also conducted a feature source ablation experiment (PAAC vs ESM-2 vs fusion), that is, an independent reproduction experiment of "using only pseudo-amino acid composition vs using only ESM-2 embedding vs fusing both". Under the same training / test division, the three sets of features were compared using a unified XGBoost classifier. The results are shown in Table 6, which is a comparison table of feature source ablation results of this embodiment.

[0151] Table 6

[0152] Based on ROC-AUC, the optimal solution is PAAC+ESM2_fusion (530D) (0.8721). The results show that ESM-2 alone provides sufficient information but has high noise; PAAC is more robust with small sample sizes; fusion of the two can improve AUC but not necessarily simultaneously improve MCC. Therefore, the main workflow still prioritizes classification head stability while optimizing the ranking head independently as an engineering compromise.

[0153] On the internal test set, approximately 94.6% of the samples fell within the applicable domain, indicating that the applicable domain determination method of this application has good predictive reliability for most candidate peptides. After temperature scaling calibration, the negative log-likelihood decreased from 1.415 to 0.735, and the Brier score decreased from 0.334 to 0.260, verifying the effectiveness of combining applicable domain determination with model calibration.

[0154] Finally, this application proposes a DPP-IV inhibitory peptide screening device based on source-aware activity sequencing.

[0155] Figure 7 This is a block diagram of a source-aware activity-sequencing DPP-IV inhibitory peptide screening device according to an embodiment of this application.

[0156] like Figure 7 As shown, the source-aware activity sorting-based DPP-IV inhibitory peptide screening device 10 includes: a dataset construction and negative sample stratification module 100, a multimodal feature extraction module 200, a feature alignment module 300, a dual-task prediction module 400, and an output module 500.

[0157] Specifically, the dataset construction and negative sample hierarchical module 100 is used to construct the training dataset, which contains positive samples and at least three levels of negative samples. The three levels of negative samples include at least: a first negative sample randomly sampled from a public database, and a second negative sample generated by a generative model that is similar to the positive sample in physicochemical properties but does not possess DPP-IV inhibitory activity. The peptide samples are standardized, and the standardization process includes at least: sequence redundancy removal, splitting the training and test sets according to clusters, and IC... 50 The data undergoes standardized preprocessing; The multimodal feature extraction module 200 is used to extract multimodal features of peptides, including at least: sequence deep semantic embedding features extracted based on a protein language model, and pocket structure features extracted from the active pocket region of DPP-IV protein through molecular docking as input to the model features. The feature alignment module 300 is used to explicitly align sequence deep semantic embedding features as query vectors and pocket structure features as key vectors and value vectors through a cross-attention mechanism, generating fusion features that reflect the interaction pattern between peptide sequences and active pockets. The dual-task prediction module 400 is used to construct a decoupled prediction architecture consisting of a classification branch and a ranking branch. The classification branch and the ranking branch each have independent parameter layers and output heads. The classification branch receives fused features generated by the feature alignment module and outputs the classification probability that the peptide has DPP-IV inhibitory activity. The ranking branch adopts an independent ranking model, which is executed in parallel with the classification branch. It trains the ranking model by constructing preference pairs based only on half-inhibition concentration data within the same assay source and outputs the activity ranking score of the candidate peptide. The output module 500 is used for virtual screening output. Based on classification probability and activity ranking score, it comprehensively ranks candidate peptides and outputs the ranking results of the activity strength of candidate peptides. In the virtual screening output stage, it performs multi-dimensional credibility assessment and safety assessment of candidate peptides. Specifically, it determines the applicable domain by calculating the kNN distance between the candidate peptide and the training set samples, using the preset quantile threshold of the training set as the boundary of the applicable domain, and downweighting candidate peptides that exceed the boundary of the applicable domain. In addition, it uses the ADMET prediction model to assess the biosafety of candidate peptide segments and screens out candidate peptides that do not meet the preset safety threshold.

[0158] It should be noted that the foregoing explanation of the DPP-IV inhibitory peptide screening method based on source-sensing activity sequencing also applies to the DPP-IV inhibitory peptide screening device based on source-sensing activity sequencing in this embodiment, and will not be repeated here.

[0159] The source-aware activity ranking-based DPP-IV inhibitory peptide screening method and apparatus proposed in this application achieves a classification ROC-AUC of 0.8904 and a ranking Spearman rank correlation coefficient of 0.542 on an independent test set through three-level negative sample stratification, multimodal feature extraction and explicit cross-attention alignment, decoupled prediction architecture, and same-source preference-based ordinal constraint training strategy. This fills the technical gap in the qualitative classification and quantitative ranking of DPP-IV inhibitory peptides and establishes a complete closed loop from activity prediction and mechanism analysis to rational design.

Claims

1. A method for screening DPP-IV repressive peptides based on source-sensing activity sequencing, characterized in that, Includes the following steps: S1. Construct a training dataset containing positive samples and at least three levels of negative samples. The three levels of negative samples include at least: a first negative sample randomly sampled from a public database, and a second negative sample generated by a generative model that is similar to the positive samples in physicochemical properties but does not possess dipeptidyl peptidase DPP-IV inhibitory activity. Standardize the peptide samples, including at least: sequence redundancy removal, splitting the training and test sets by clusters, and halving the inhibitory concentration IC50. 50 The data undergoes standardized preprocessing; S2. Extract multimodal features of peptide segments, including at least: sequence deep semantic embedding features extracted based on protein language models, and pocket structure features extracted from the active pocket region of DPP-IV protein through molecular docking as input to the model features. S3. Through the cross-attention mechanism, the sequence deep semantic embedding features are used as query vectors, and the pocket structure features are used as key vectors and value vectors, and explicit alignment is performed to generate fusion features that reflect the interaction mode between peptide sequences and active pockets. S4. Construct a decoupled prediction architecture consisting of a classification branch and a ranking branch, wherein the classification branch and the ranking branch each have independent parameter layers and output heads; the classification branch receives the fusion features and outputs the classification probability that the peptide has DPP-IV inhibitory activity; the ranking branch adopts an independent ranking model, which is executed in parallel with the classification branch. The ranking model is trained by constructing preference pairs based only on half-inhibition concentration data within the same assay source, and outputs the activity ranking score of the candidate peptide. S5. Virtual screening output: Based on the classification probability and the activity ranking score, the candidate peptides are comprehensively ranked, and the ranking results of the candidate peptides' activity strength are output.

2. The screening method according to claim 1, characterized in that, The second negative sample was generated by a Restricted Boltzmann Machine (RBM). The RBM uses the amino acid composition, molecular weight, isoelectric point, and hydrophobicity of the positive sample as the input to the visible layer. It learns the joint probability distribution of the positive sample through a contrastive divergence algorithm, and then generates a synthetic sequence that is highly similar to the physicochemical properties of the positive sample but does not have DPP-IV inhibitory activity through Gibbs sampling.

3. The screening method according to claim 1, characterized in that, The protein language model is the ESM-2 model, which is used for evolutionary scale modeling. The extraction process of the sequence deep semantic embedding features is as follows: input the peptide sequence into the ESM-2 model, extract the average pooling representation of the last hidden state of the ESM-2 model, and obtain a fixed-dimensional embedding vector.

4. The screening method according to claim 1, characterized in that, The cross-attention mechanism includes: The residue-level embedding matrix of the peptide sequence is used as the query vector, and the physicochemical feature matrix of the active pocket residues of the DPP-IV protein is used as the bond vector and value vector. The physicochemical features include at least one of the following: amino acid type, hydrophobic scaling value, formal charge, number of hydrogen bond donors, number of hydrogen bond acceptors, van der Waals volume, and sub-pocket category number. The attention weight matrix is ​​calculated by scaling the dot product, and the attention weight matrix is ​​multiplied by the value vector to generate the fused feature. The screening method further includes: An attention heatmap is generated based on the attention weight matrix to identify the interaction strength between the position of each residue in the peptide sequence and the residues in the active pocket.

5. The screening method according to claim 1, characterized in that, The training methods for the sorting model in the sorting branch include: IC based on the same literature source or the same batch of tests 50 Data constructs preference pairs; where, in any preference pair, if the IC of peptide i is... 50 IC50 values ​​less than peptide j 50 If the value is specified, then the ordinal relationship between peptide i and peptide j is marked as i being superior to j; Using the normalized discounted cumulative gain as the optimization objective, the LambdaRank model is trained using the aforementioned preference pair, so that the ranking model can avoid systematic bias between different measurement sources by learning the relative activity distribution within the same measurement source.

6. The screening method according to claim 1, characterized in that, The classification branch and the ranking branch are trained independently using different feature subsets; The classification branch takes the statistical features of the entire sequence as input; The sorting branch takes a subset of features filtered by the feature gain index as input. The subset of features is a dimensionality-reduced feature set obtained from the sequence statistical features and the sequence deep semantic embedding features. The classification branch and the ranking branch do not share the loss function and can be optimized independently.

7. The screening method according to claim 1, characterized in that, When constructing the training dataset, the normalization process also includes applying the IC... 50 The data preprocessing process includes: By detecting IC 50 The order of magnitude of the raw data identifies the IC. 50 The initial units of the raw data were then uniformly converted to the negative logarithmic half-inhibition concentration (pIC). 50 value; Inaccurate activity records with range symbols and non-homologous data from non-enzymatic assay sources are filtered out from the training dataset to construct a training set that satisfies the ordinal constraint consistency.

8. The screening method according to claim 1, characterized in that, The training of the classification branch employs a two-stage transfer learning strategy, specifically including: Phase 1: Pre-training is performed using activity data of non-peptide DPP-IV target small molecule inhibitors to enable the classification branch to learn the general chemical pattern of DPP-IV inhibitory activity. Phase Two: Fine-tuning the model pre-trained in Phase One using peptide data. During fine-tuning, the underlying feature extraction parameters are frozen, and a learning rate lower than that in Phase One is used to adapt to peptide sequence features.

9. The screening method according to claim 4, characterized in that, Also includes: Based on the attention heatmap, identify residue positions in the peptide sequence whose attention weight values ​​for key residues in the S1 or S2 sub-pockets of the active pocket reach or exceed a preset weight threshold. Based on the type of physicochemical interaction between the residue position and the active pocket, suggestions for site-specific peptide modification to enhance hydrophobic or charged interactions are output. The virtual filtering output also includes: The candidate peptides are subjected to multidimensional credibility assessment and safety assessment. The multidimensional credibility assessment includes: calculating the k-nearest neighbor (kNN) distance between the candidate peptide and the samples in the training dataset; using a preset quantile threshold of the training dataset as the boundary of the applicable domain to determine the applicable domain; and reducing the weight of candidate peptides that exceed the boundary of the applicable domain. The safety assessment includes: using the ADMET pharmacokinetic and toxicological property prediction model to quantitatively assess the biosafety of the candidate peptides and screening out candidate peptides that do not meet the preset safety threshold.

10. A DPP-IV inhibitory peptide screening device based on source-sensing activity sequencing, characterized in that, include: The dataset construction and negative sample hierarchical module is used to construct the training dataset, which contains positive samples and at least three levels of negative samples. The three levels of negative samples include at least: a first negative sample randomly sampled from a public database, and a second negative sample generated by a generative model that is similar to the positive sample in physicochemical properties but does not possess dipeptidyl peptidase DPP-IV inhibitory activity. The peptide samples are standardized, and the standardization process includes at least: sequence redundancy removal, splitting the training and test sets according to clusters, and halving the IC50 inhibition concentration. 50 The data undergoes standardized preprocessing; The multimodal feature extraction module is used to extract multimodal features of peptides, including at least: sequence deep semantic embedding features extracted based on a protein language model, and pocket structure features extracted from the active pocket region of the DPP-IV protein through molecular docking as input to the model features. The feature alignment module is used to explicitly align the sequence deep semantic embedding features as query vectors and the pocket structure features as key vectors and value vectors through a cross-attention mechanism, generating fusion features that reflect the interaction pattern between peptide sequences and active pockets. A dual-task prediction module is used to construct a decoupled prediction architecture consisting of a classification branch and a ranking branch. The classification branch and the ranking branch each have independent parameter layers and output heads. The classification branch receives the fused features generated by the feature alignment module and outputs the classification probability that the peptide has DPP-IV inhibitory activity. The ranking branch adopts an independent ranking model, which is executed in parallel with the classification branch. The ranking model is trained by constructing preference pairs based only on half-inhibition concentration data within the same assay source, and outputs the activity ranking score of the candidate peptide. The output module is used for virtual screening output. Based on the classification probability and the activity ranking score, candidate peptides are comprehensively ranked, and the ranking results of the activity strength of the candidate peptides are output. During the virtual screening output stage, multi-dimensional credibility and safety assessments are performed on the candidate peptides. Specifically, the k-nearest neighbor (kNN) distance between the candidate peptide and the training set samples is calculated, and a preset quantile threshold of the training set is used as the boundary of the applicable domain for domain determination. Candidate peptides outside the applicable domain boundary are downweighted. Furthermore, the biosafety of the candidate peptides is quantitatively assessed using the ADMET pharmacokinetic and toxicological property prediction model, and candidate peptides that do not meet the preset safety threshold are screened out.

Images