Calculation design method and system of antioxidant peptide

Abstract

The invention discloses a calculation design method and system of antioxidant peptide, and belongs to the technical field of bioinformatics and polypeptide design. The method comprises the following steps: constructing an antioxidant peptide activity prediction model and an antioxidant peptide sequence generation model based on an adapted language model; and guiding a sequence generation model to generate a candidate antioxidant peptide sequence through a target-oriented optimization process including reinforcement learning by utilizing the output of the activity prediction model. By constructing the prediction and design integrated calculation framework, the novel antioxidant peptide can be efficiently and reliably found and designed, and the method can be applied to the technical fields of food preservation, health care product development and medical biology.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and peptide design technology, specifically to an antioxidant peptide prediction and design method and its integrated framework based on deep learning and reinforcement learning. Background Technology

[0002] Oxidation is a fundamental chemical process in biological systems. Excessive accumulation of reactive oxygen species (ROS), a byproduct of oxidation, can trigger oxidative stress, which is closely related to cardiovascular disease, neurodegenerative diseases, cancer, and the aging process. Antioxidant peptides (AOPs) are a class of promising bioactive molecules, typically composed of 2-20 amino acids, capable of combating oxidative damage through mechanisms such as free radical scavenging. As natural antioxidants, AOPs are highly favored in the fields of food preservation and functional health products due to their high safety and efficacy, and are expected to replace synthetic antioxidants.

[0003] However, traditional experiment-based AOPs screening methods are cumbersome, time-consuming, and costly, severely hindering the rapid discovery of novel AOPs. To overcome this bottleneck, computational methods based on machine learning and deep learning have become powerful tools for predicting and designing AOPs.

[0004] Despite significant progress in existing computational methods, key challenges remain. First, the accuracy and generalization ability of current predictive models still have room for improvement, which is fundamental for reliably identifying candidate peptides. Second, most studies focus on determining the activity of known peptides, while the ability to design novel peptide sequences with high antioxidant potential de novo remains limited. Finally, a gap often exists between computational predictions and actual biological activity, necessitating an integrated framework that not only enables computational design but also allows for experimental verification of efficacy. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an antioxidant peptide calculation method that integrates efficient prediction and innovative design, aiming to solve the technical problems of insufficient prediction accuracy, limited innovative design capabilities, and disconnect between calculation and experiment in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a computational design method and system for antioxidant peptides, which constitutes an integrated framework for efficient prediction and de novo design. The method includes: Step S1, using a protein sequence knowledge transfer strategy for language model adaptation: on an antioxidant-related protein sequence dataset, a language model based on the Transformer architecture is fine-tuned for domain adaptation to obtain an adapted language model backbone; Step S2, constructing an activity prediction model optimized by reinforcement learning: based on the adapted language model backbone, an antioxidant peptide activity prediction model is constructed, and its decision boundary is optimized using a first reinforcement learning stage to improve prediction accuracy; Step S3, constructing a sequence generation model guided by a predictor and performing goal-oriented sequence generation: based on the adapted language model backbone, an antioxidant peptide sequence generation model is constructed, the training process of which includes a supervised learning stage and a second reinforcement learning stage; and using the optimized activity prediction model as a fixed evaluator, the sequence generation model is guided to generate candidate antioxidant peptide sequences with high predictive activity, novelty, and diversity through a multi-objective reward function driven by the second reinforcement learning stage.

[0007] In a preferred embodiment of the present invention, the language model is a language model based on the Transformer architecture, and the adaptation process is performed on an antioxidant-related protein sequence dataset.

[0008] In another preferred embodiment of the present invention, the goal-oriented optimization process is driven by a multi-objective reward function that comprehensively evaluates the predicted antioxidant activity, sequence novelty, and sequence diversity of the generated sequence.

[0009] The beneficial effects of this invention are as follows: 1. Excellent predictive performance: The predictor RLP-T5Pred of this invention shows performance far exceeding that of existing methods on independent test sets, with an AUC-ROC of 0.9589 and an F1 score of 0.9236, providing a highly reliable tool for screening and evaluating peptides.

[0010] 2. Strong innovative design capabilities: The generator RLP-T5Gen of this invention successfully generated a large number of peptide sequences with high predictive activity and novel structures. UMAP visualization analysis shows that the generated sequences not only possess the characteristics of known AOPs, but also explore new chemical spaces.

[0011] 3. Experimental Verification: This invention, through computational design, chemical synthesis, and in vitro experiments, verified 17 novel peptides. The results confirmed that most peptides possess antioxidant activity. In the ABTS free radical scavenging assay, the activity of many peptides was equal to or even superior to the industry benchmark glutathione (GSH), particularly peptide pep16, which performed exceptionally well in both free radical scavenging assays. This demonstrates the significant practical value of this invention's framework in accelerating the discovery of functional peptides. Attached Figure Description

[0012] Figure 1 This is a schematic diagram illustrating the overall architecture and workflow of the RLAnOxPeptide framework of this invention.

[0013] Figure 2 This is a performance comparison chart of the predictor model components of this invention in ablation studies, wherein... Figure 2 (a) shows the validation set results. Figure 2 (b) shows the results for the independent test set.

[0014] Figure 3 This is a characteristic analysis diagram of the sequence generated by the generator RLP-T5Gen of this invention, wherein... Figure 3 (a) Visualization of the UMAP distribution of the generated sequence and the training set sequence in the feature space. Figure 3 (b) Comparison of sequence length distributions Figure 3 (c) represents the key performance indicators of the generated sequence.

[0015] Figure 4 The figure shows the experimental verification results of the 17 peptides designed in this invention. Figure 4 (a) shows the peptide sequence, identifier, and prediction score. Figure 4 (b) Results of DPPH free radical scavenging activity assay. Figure 4 (c) shows the results of the ABTS free radical cation scavenging activity assay. Detailed Implementation

[0016] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0017] like Figure 1 As shown, the implementation process of the RLAnOxPeptide framework of the present invention includes steps such as data construction, model training, sequence generation, and experimental verification.

[0018] Step 1: Dataset Construction and Preprocessing. First, positive AOPs sequences were collected from multiple public databases and literature, including DFBP and BIOPEP-UWM. Sequences containing antioxidant-related keywords were screened and excluded from the UniProt database to form the negative non-AOPs dataset. The original data underwent preprocessing including deduplication, length filtering (2-20 amino acids), and removal of sequences containing non-standard amino acids. Subsequently, the CleanLab confidence learning algorithm was used for automated cleaning of the dataset, identifying and removing samples with potential labeling errors, ultimately resulting in a high-purity dataset containing 1412 AOPs and 1442 non-AOPs.

[0019] Step Two: Domain-Adaptive Fine-Tuning of the Model Backbone To enable the model to better understand the sequence features of antioxidant peptides, this invention selects the ProtT5 large language model as its basic architecture. On a self-built antioxidant protein dataset, ProtT5 is fine-tuned for domain adaptation using the Masked Language Modeling (MLM) task. After fine-tuning, the model can generate more informative peptide sequence feature representations, serving as a shared backbone for subsequent predictors and generators.

[0020] Step 3: Construction and training of the high-precision predictor RLP-T5Pred, such as... Figure 1 As shown in the “RLP-T5Pred” module, the predictor is built based on the aforementioned fine-tuned ProtT5 backbone. Its training is divided into two stages: first, supervised learning (SL) is performed, using techniques such as label smoothing and logit penalty to prevent model overfitting; then, reinforcement learning (RL) is introduced for joint training to further optimize the decision boundary of the classifier, enabling it to more accurately distinguish between highly active and inactive molecules.

[0021] like Figure 2 The ablation experiments shown demonstrate the effectiveness of each technical component in this invention. Compared to the baseline model (untuned, no reinforcement learning), introducing either domain-adaptive fine-tuning or reinforcement learning (RL) alone improves performance, while the model combining both (fine-tuned and reinforcement learning) achieves the best performance across all metrics, including AUC-ROC, F1 score, precision, recall, and accuracy, on the independent test set ( Figure 2 The AUC-ROC on (b) reached 0.9589.

[0022] Step 4: Construction and training of the innovative sequence generator RLP-T5Gen, such as... Figure 1As shown in the "RLP-T5Gen" module, the generator is also built on a finely tuned ProtT5 backbone. Its training employs an innovative iterative loop strategy: in the supervised learning phase, the model learns sequence patterns of known AOPs to ensure the "syntactic" correctness of the generated sequences; in the reinforcement learning phase, the model generates new sequences and updates them using the reinforcement learning policy gradient algorithm. This update is driven by a multi-objective reward function that combines the activity score given by the fixed predictor RLP-T5Pred with evaluations of sequence length, diversity, and novelty.

[0023] like Figure 3 As shown, the generator exhibits excellent performance. Figure 3 The UMAP visualization results in (a) show that the generated sequence (orange dots) closely overlaps with the known AOPs (blue dots) in the training set in the feature space, indicating that they have similar high-activity features. At the same time, the generated sequence also explores new regions not covered by the training set, demonstrating its innovation. Figure 3 (b) shows that the length distribution of the generated sequences is highly consistent with that of the training set, with average lengths of 5.5 and 6.3, respectively. Figure 3 The performance table in (c) shows that the average predicted activity probability of the generated sequences is as high as 0.942, and it has high character-level diversity (0.891) and binary diversity (0.989).

[0024] Step 5: Peptide sequence design and experimental verification, such as... Figure 1 As shown in the flowchart below, a candidate library containing 1000 new sequences was generated using the trained RLP-T5Gen. Seventeen candidate peptides with high prediction scores and sequence diversity were selected for chemical synthesis, such as... Figure 4 As shown in (a), the antioxidant activity of these 17 peptides was determined using two widely accepted in vitro assays: the DPPH radical scavenging assay and the ABTS radical cation scavenging assay.

[0025] like Figure 4As shown in (b) and 4(c), the experimental results strongly confirm the effectiveness of the framework of this invention. In the DPPH experiment, 10 out of 17 peptides exhibited antioxidant activity exceeding that of the positive control GSH by more than 50%, with pep16 showing no significant difference in antioxidant activity compared to GSH. In the ABTS experiment, 11 out of 17 peptides showed antioxidant activity comparable to or even higher than that of GSH. The letters at the top of the data bars in the figures (e.g., a, b, c, etc.) are used to indicate statistically significant differences between different experimental groups. According to scientific convention, data bars labeled with the same letter indicate no statistically significant difference (p>0.05), while data bars labeled with different letters indicate statistically significant differences. These labels are an important part of the experimental data analysis, used to rigorously compare the activity of each designed peptide with that of the control group (GSH). These results fully demonstrate that the framework proposed in this invention can successfully translate computational design into molecules with practical biological functions, verifying its reliability and effectiveness in practical applications.

[0026] It should be noted that the above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. A computational design method for antioxidant peptides, characterized in that, The process includes the following steps: S1: Constructing an antioxidant peptide activity prediction model based on an adapted language model; S2: Constructing an antioxidant peptide sequence generation model based on the adapted language model; S3: Using the output of the activity prediction model, guiding the sequence generation model to generate candidate antioxidant peptide sequences through a goal-oriented optimization process involving reinforcement learning.

2. The method according to claim 1, characterized in that, The language model is a language model based on the Transformer architecture.

3. The method according to claim 1, characterized in that, The language model was adapted on a dataset of antioxidant-related protein sequences.

4. The method according to claim 3, characterized in that, The adaptation is achieved by performing a masked language model (MLM) task.

5. The method according to claim 1, characterized in that, The process of constructing the antioxidant peptide activity prediction model in step S1 includes a reinforcement learning phase for optimizing its decision boundary.

6. The method according to claim 1, characterized in that, The process of constructing the antioxidant peptide sequence generation model in step S2 includes a supervised learning phase of training on known antioxidant peptide sequences.

7. The method according to claim 1, characterized in that, The goal-oriented optimization process is driven by a multi-objective reward function, which comprehensively evaluates the predicted antioxidant activity, sequence novelty, and sequence diversity of the generated sequences.

8. The method according to claim 1, characterized in that, The method also includes cleaning the training data using a confidence learning algorithm before building the model.

9. A computational design system for performing the method as described in any one of claims 1 to 8, characterized in that, It includes a language model-based peptide activity prediction module and a peptide sequence generation module; The generation module can receive the output of the prediction module and use reinforcement learning to optimize the generation of candidate antioxidant peptide sequences.

10. The system according to claim 9, characterized in that, The training process of the peptide sequence generation module includes a supervised learning phase and a reinforcement learning phase, and is driven by a multi-objective reward function that includes activity, novelty, diversity, and length.

Images