Multi-objective antimicrobial peptide generation method based on constrained monte carlo tree search and diffusion model

Abstract

This invention discloses a multi-objective antimicrobial peptide generation method based on a constrained Monte Carlo tree search and diffusion model, belonging to the field of bioengineering. By constructing and fine-tuning a new conditional mask diffusion language model based on a protein language model, a deeper understanding of the complex evolutionary information and structural patterns in protein sequences is obtained, improving the prediction accuracy of the minimum inhibitory concentration (MIC) of antimicrobial peptides. Addressing the challenge of simultaneously considering multiple conflicting attributes such as activity, toxicity, and hemolytic activity in antimicrobial peptide development, a Pareto reward-based constrained Monte Carlo tree search is further introduced to explicitly model the competitive relationships between different objectives, achieving high-quality, multi-objective balanced antimicrobial peptide design. This invention not only significantly improves the efficiency of producing high-quality candidate drugs but also effectively reduces the burden of subsequent biological experiments for verification, accelerating the development of novel anti-infective drugs.

Description

Technical Field

[0001] This invention relates to the field of bioengineering, specifically to a method for generating multi-objective antimicrobial peptides based on a constrained Monte Carlo tree search and diffusion model. Background Technology

[0002] Antimicrobial peptides are widely distributed in organisms and are an important component of the innate immune system. Due to their broad-spectrum antibacterial properties, they are considered highly promising candidates for next-generation antibiotics. These small molecule peptides are generally less than 50 amino acids in length and have a unique bactericidal mechanism: they generate strong electrostatic interactions between positively charged amino acid residues and negatively charged lipopolysaccharides (in Gram-negative bacteria) or lipoteichoic acid (in Gram-positive bacteria) on the bacterial cell membrane. Subsequently, the antimicrobial peptides adsorb and cross the cell membrane via Coulomb forces, inducing membrane perforation, cytoplasmic leakage, and even physical rupture, thereby completely killing the pathogens.

[0003] Because this mechanism directly targets the physical boundary of bacteria (cell membrane) rather than a single protein receptor, pathogenic bacteria are extremely difficult to develop drug resistance through gene mutation. At a time when global antibiotic overuse has triggered a superbug crisis and traditional chemical antibiotic development has reached a plateau, the development of antimicrobial peptides has extremely important strategic significance in clinical medicine, food preservation, and livestock farming.

[0004] Despite its promising prospects, the development of antimicrobial peptides still faces multiple obstacles. First, the yield of natural antimicrobial peptides extracted from organisms in nature is extremely low, and large-scale industrial production is expensive. Second, traditional rational design based on natural templates (such as point mutations and sequence heterozygosity) relies heavily on prior knowledge from experts, resulting in lengthy research and development cycles and extremely high trial-and-error costs.

[0005] Even more challenging is the fact that the development of antimicrobial peptides is essentially a multi-objective optimization problem. An ideal antimicrobial peptide must enhance antibacterial activity while simultaneously exhibiting low hemolytic activity (cytotoxicity), high metabolic stability, and low synthesis cost. However, these properties are often strongly coupled or even conflicting: for example, increasing the hydrophobicity of the peptide chain usually improves bactericidal efficacy, but this is often accompanied by a significant increase in cytotoxicity to mammalian cells. Traditional methods struggle to find Pareto optimal solutions for each performance metric within such a complex global search space, leading to the elimination of numerous candidate sequences in preclinical trials due to toxicity or stability issues.

[0006] In recent years, breakthroughs in deep learning technology have opened up new avenues for protein sequence design. Given the high similarity between protein amino acid sequences and natural language logic, computationally assisted design based on generative AI (such as generative adversarial networks, variational autoencoders, and autoregressive models) has shown great potential. Among these, diffusion models, by simulating the process of adding noise to ordered data and then learning the reverse denoising process to generate new data, demonstrate superior robustness and diversity compared to traditional models in capturing the high-dimensional distribution characteristics of proteins. For example, the invention patent CN121789781A discloses an antimicrobial peptide generation and screening method based on a conditional latent diffusion model. This method models the distribution of antimicrobial peptide sequences in a continuous latent space, achieves controllable generation of antimicrobial peptide sequences through conditional constraints, and combines multi-level computational screening with experimental verification strategies to improve the efficiency and reliability of the antimicrobial peptide discovery process. This addresses the problem of limited antimicrobial peptide discovery efficiency in existing antimicrobial peptide design methods due to factors such as the large sequence space, difficulty in controlling the generation process, high candidate sequence redundancy, and high screening costs.

[0007] However, diffusion models have significant limitations when applied to the design of specific functional peptides: their generation process is essentially random sampling, lacking explicit control over the generation path. For antimicrobial peptide design tasks that require simultaneous satisfaction of multiple constraints such as antibacterial properties, safety, and physicochemical properties, pure diffusion models often generate a large number of invalid sequences, resulting in low search efficiency and difficulty in accurately hitting rare spaces with specific biological activities.

[0008] To address the issue of uncontrolled growth during the generation process, the introduction of decision-making algorithms for guidance has become a cutting-edge technological approach. Monte Carlo Tree Search (MCTS), an algorithm that finds the optimal path in a discrete decision space by balancing exploration and utilization, has demonstrated its powerful search capabilities in the field of complex game theory. Introducing Constrained Monte Carlo Tree Search (CMCTS) into the inverse denoising process of a diffusion model can be seen as adding an intelligent navigator at each step of the generation process. By defining complex constraint reward functions (such as deep learning-based toxicity predictors and activity scores), CMCTS can dynamically evaluate and prune potential sampling paths in the diffusion model. This approach not only restricts the generation process to within biophysical constraints but also guides the model towards collaborative optimization towards multiple objectives—high antibacterial activity, high activity, low toxicity, and low minimum inhibitory concentration—through a multi-objective reward mechanism.

[0009] In summary, while existing technologies have attempted to extract high-dimensional semantic features using protein language models, mature solutions are still lacking in how to efficiently and purposefully search for optimal multi-target sequences. Current diffusion model designs often focus on the generation speed or similarity of single sequences, neglecting the core functional constraint coordination problem in biomolecular design. Therefore, there is an urgent need to develop a novel antimicrobial peptide design method that integrates the underlying features of protein language models, the diffusion model generation framework, and utilizes constrained Monte Carlo tree search for multi-target path guidance. This would not only significantly improve the efficiency of producing high-quality candidate drugs but also effectively reduce the burden of subsequent biological experiments for validation, accelerating the development of novel anti-infective drugs. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this invention provides a multi-objective antimicrobial peptide generation method based on constrained Monte Carlo tree search and diffusion model. A new conditional mask diffusion language model is constructed and fine-tuned using a protein language model as the base model. This model is used to deeply capture complex evolutionary information and structural patterns in protein sequences, improving the prediction accuracy of the minimum inhibitory concentration (MIC) of antimicrobial peptides. Furthermore, addressing the challenge of simultaneously considering multiple conflicting attributes such as activity, toxicity, and hemolytic activity in antimicrobial peptide development, a Pareto reward-based constrained Monte Carlo tree search is introduced. This explicitly models the competitive relationships between different objectives, achieving high-quality, multi-objective balanced antimicrobial peptide design.

[0011] To achieve the above-mentioned objectives, an embodiment provides a method for generating multi-objective antimicrobial peptides based on a constrained Monte Carlo tree search and diffusion model, comprising the following steps: Antimicrobial peptide sequences and non-antimicrobial peptide sequences were obtained and preprocessed to obtain a dataset. A conditional mask diffusion language model is constructed, and a generative model is obtained by training and fine-tuning it using a dataset. An initial seed sequence is generated through conditional diffusion, and an optimization direction vector containing multiple optimization objectives is set. The initial seed sequence is masked and noise is added to obtain a noisy state. Using the noisy state as the root node, combined with the conditional probability distribution and optimization direction vector output by the generative model, a constrained Monte Carlo tree search including selection, expansion, simulation evaluation and backpropagation stages is performed to obtain the updated Pareto front non-dominated solution set and the corresponding candidate sequence. The external multi-attribute evaluation module is called to obtain the attribute values ​​of the candidate sequence on each optimization objective, calculate the multi-attribute improvement score, and update the Pareto front non-dominated solution set accordingly. Some sequences are selected from the current Pareto front and then masked and noise-added as seed sequences for the next iteration. Based on the current iteration round and the quality index of the Pareto front, determine whether the termination condition has been met. If so, output all antimicrobial peptide sequences in the final saved Pareto front and their corresponding multi-attribute prediction values.

[0012] Preferably, the open-source antimicrobial peptide sequences include antimicrobial peptide sequences from three databases: dbAMP, AMP Scanner, and DRAMP.

[0013] Preferably, the pretreatment includes: Remove repetitive antimicrobial peptide sequences; The filtration process removes antimicrobial peptide sequences and non-antimicrobial peptide sequences with lengths greater than 50 amino acids and less than 11 amino acids, as well as those containing uncommon amino acids.

[0014] Optionally, the dataset can be divided into a training set and a validation set in a 9:1 ratio.

[0015] Preferably, the conditional mask diffusion language model is based on the protein language model ESM, and introduces a conditional embedding layer. The conditional embedding layer is used to map the preset conditional labels into vectors of the same dimension as the word embeddings, and add the vectors to the word embedding vectors.

[0016] Preferably, an implicit classifier guidance strategy is used when training and fine-tuning using the dataset, and conditional labels are randomly discarded during training and fine-tuning; a composite strategy is used during the training and fine-tuning process, including: Full-supervised fine-tuning is performed on the word embedding layer, conditional embedding layer, and output head in the conditional masked diffusion language model to update the parameters; LoRA low-rank adaptation is introduced for the high-level modules of the Transformer encoder. Specifically, LoRA low-rank adaptation includes adding a branch consisting of two low-rank matrices next to the original pre-trained weight matrix for the high-level modules of the Transformer encoder. The first low-rank matrix is ​​initialized with a Gaussian distribution, and the second low-rank matrix is ​​initialized to zero. During training fine-tuning, only the parameters of these two low-rank matrices are updated.

[0017] More preferably, the fine-tuning of the conditional masked diffusion language model employs the AdamW optimizer, with a learning rate of 1e. -4 up to 1e -3 The linear scheduling is performed with a weight decay coefficient of 0.01 and a beta value of (0.9, 0.98); the loss function is the reweighted cross-entropy loss function.

[0018] Preferably, the multiple optimization objectives include: antibacterial activity, minimum inhibitory concentration, toxicity, and hemolysis.

[0019] More preferably, the antimicrobial activity is obtained through an antimicrobial prediction unit in an external multi-attribute evaluation module. The antimicrobial prediction unit is used to predict the antimicrobial probability of each antimicrobial peptide sequence by taking the generated antimicrobial peptide sequence as input and using the AMP Scanner-v2 model. Toxicity is obtained through a toxicity prediction unit in an external multi-attribute assessment module. This toxicity prediction unit is used to predict the toxicity probability of each antimicrobial peptide sequence by taking the generated antimicrobial peptide sequence as input and using the ToxinPred2 model. Hemolytic activity is obtained from the hemolytic prediction unit in the external multi-attribute evaluation module. The hemolytic prediction unit is used to predict the hemolytic probability of each antimicrobial peptide sequence by taking the generated antimicrobial peptide sequence as input and using the HemoPI model.

[0020] Preferably, the selection phase in the constrained Monte Carlo tree search includes: based on the root node, traversing all expanded child nodes under the root node, calculating the improved upper confidence bound score of the child nodes, comparing the improved upper confidence bound scores of all child nodes, selecting the node corresponding to the maximum value, and if the currently selected child node still has unexpanded child nodes, the selection phase ends and the currently selected child node is returned to enter the expansion phase; otherwise, the above steps are repeated with the child node as the new current node, and the above traversal, calculation, comparison and judgment steps are repeated until a node that has not been fully expanded is reached. The formula for calculating the improved upper confidence boundary score of child nodes is as follows: , In the formula, For the improved upper confidence bound score, As the root node, for The first node corresponding to the time Number of child nodes To accumulate rewards, For the number of visits, The noise probability is the output of the conditional masked diffusion language model. To explore coefficients.

[0021] Preferably, the expansion phase in the constrained Monte Carlo tree search includes: based on the currently expanded leaf nodes output by the selection phase, when generating child nodes, independently injecting Gumbel noise into the unnormalized output value of the conditional mask diffusion language model, and generating child nodes according to the conditional probability distribution after adding noise.

[0022] Preferably, the simulation evaluation phase in constrained Monte Carlo tree search includes: The partial mask sequence corresponding to the child nodes generated in the expansion stage is completely denoised to obtain candidate sequences; the candidate sequences are then sent to an external multi-attribute evaluation module to obtain the antibacterial probability, the logarithm of the minimum inhibitory concentration, the toxicity probability, and the hemolysis probability, respectively. Calculate the attribute improvement vector of the candidate sequence relative to the seed sequence in the previous round for each optimization objective; Calculate the cosine of the angle between the attribute boosting vector and the preset optimization direction vector to obtain the angle value; if the angle value is greater than the current dynamic angle threshold, discard this candidate sequence; if the angle value is less than or equal to the dynamic angle threshold, compare the multi-attribute score of this candidate sequence with the sequences in the current Pareto front non-dominated solution set; when this candidate sequence is better than all existing sequences in the non-dominated solution set on at least one optimization objective and not worse than existing sequences on other optimization objectives, add this candidate sequence to the non-dominated solution set and remove the old sequences dominated by this candidate sequence; calculate the reward value based on the position of this candidate sequence in the updated Pareto front for backpropagation.

[0023] More preferably, the dynamic angle threshold is initially set to 45°, and is dynamically adjusted using an exponential moving average method based on the proportion of candidate sequences rejected in the current round, with the adjustment range limited to between 15° and 75°.

[0024] Preferably, the minimum inhibitory concentration prediction in the external multi-attribute evaluation module adopts a regression method based on a protein language model, including: Obtain the antimicrobial peptide sequences and the corresponding experimental measurements of minimum inhibitory concentrations, construct a training dataset, and preprocess it. The preprocessed antimicrobial peptide sequence is input into the protein language model encoder to obtain the sequence embedding tensor, and then average pooling is performed. The average pooled sequence embedding tensor is input into the multilayer perceptron regression model to output the predicted log minimum inhibitory concentration value. The loss function is calculated based on the mean squared error between the predicted and actual values. The parameters of the multilayer perceptron regression model are updated through backpropagation to complete the training. During prediction, the protein language model encoder is frozen, and only the multilayer perceptron regression model is run to output the predicted logarithmic minimum inhibitory concentration value.

[0025] Preferably, calculating the multi-attribute improvement score includes: for each optimization objective, calculating the improvement of the candidate sequence on this optimization objective, and determining the ranking of this improvement among all current candidate sequences; based on the ranking of each optimization objective, the following calculation formula is used: , in, Candidate sequences to be evaluated The corresponding multiple attributes increase the score. For the standardized function, To optimize the total number of targets, For the first The weighting coefficients of each optimization objective. Candidate sequences In the The magnitude of attribute improvement for each optimization objective. For ranking functions, For the current iteration round The corresponding noise level, These are the preset noise scheduling parameters. The projection length of the attribute boost vector onto the optimization direction. The weighting coefficients for the alignment terms.

[0026] Preferably, the termination condition is reaching a preset number of iterations.

[0027] Preferably, the multi-target antimicrobial peptide generation method further includes: before outputting the antimicrobial peptide sequence in the final Pareto front, sending the sequence to a physicochemical property prediction unit to calculate physicochemical indices including sequence length, molecular weight, aromaticity, instability index, isoelectric point, charge, secondary structure fraction, and hydrophilicity; sending the sequence to a structure prediction unit to predict the three-dimensional structure using AlphaFold Server to obtain the pLDDT index of structural foldability; and filtering the sequences according to a preset threshold to remove sequences whose physicochemical properties or structural foldability do not meet the requirements.

[0028] The present invention also provides a multi-objective antimicrobial peptide generation device based on a constrained Monte Carlo tree search and diffusion model, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the multi-objective antimicrobial peptide generation method based on the constrained Monte Carlo tree search and diffusion model.

[0029] Compared with the prior art, the beneficial effects of the present invention include at least the following: 1) This invention introduces a high-parameter protein language model, which can deeply capture the complex evolutionary information and structural patterns in protein sequences and extract high-dimensional semantic features. Compared with traditional prediction models based on physicochemical properties or small convolutional networks, this deep semantic embedding greatly improves the prediction accuracy of the minimum inhibitory concentration of antimicrobial peptides, providing a reliable performance feedback basis for subsequent generation tasks.

[0030] 2) This invention addresses the challenge of simultaneously considering multiple conflicting attributes such as activity, toxicity, and hemolytic activity in antimicrobial peptide development by introducing a Pareto reward-based constrained Monte Carlo tree search. Compared to traditional static weighting methods, this approach explicitly models the competitive relationships between different objectives, preventing the model from collapsing to the optimum of a single attribute, thereby producing high-quality candidate sequences that achieve a Pareto optimal balance between antimicrobial efficacy and safety.

[0031] 3) This invention employs a generative paradigm that does not require retraining. Once the conditional mask diffusion language model is trained, different attributes can be optimized by changing the external evaluation operator. Simultaneously, a strategy of freezing encoder parameters is used in the training of the prediction model, with only the lightweight multilayer perceptron regression head being fine-tuned. This not only significantly reduces the consumption of computational resources but also effectively avoids the risk of overfitting on small-sample biological data.

[0032] 4) By combining global iterative optimization with constrained Monte Carlo tree search, this invention breaks through the limitation of traditional diffusion models' one-stop generation. The model can dynamically correct suboptimal decisions in the early denoising process through multiple local masking-re-optimization loops, effectively reducing search bias caused by global evaluation noise and significantly improving the structural foldability and realism of the generated sequence.

[0033] 5) This invention innovatively introduces a dynamic Pareto constraint filtering mechanism, utilizing cosine similarity to calculate the alignment between the improved vector and the preset optimization direction. This strategy, while ensuring sampling diversity, forcibly filters high-quality candidate branches through pruning operations, effectively solving the common Pareto front explosion and candidate sequence redundancy problems in multi-objective search, and greatly improving the efficiency of searching the target chemical space. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0035] Figure 1 This is a schematic flowchart of a multi-objective antimicrobial peptide generation method based on a constrained Monte Carlo tree search and diffusion model, provided in an embodiment of the present invention.

[0036] Figure 2 This is a schematic diagram of the conditional mask diffusion language model architecture and conditional guidance provided in an embodiment of the present invention.

[0037] Figure 3 This is a schematic diagram of the minimum inhibitory concentration prediction process provided in an embodiment of the present invention.

[0038] Figure 4 A violin diagram showing the performance comparison of common peptides, general antimicrobial peptides, and antimicrobial peptides generated in the embodiments of the present invention in terms of multidimensional properties. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0040] The concept of this invention lies in the fact that existing methods for generating antimicrobial peptide sequences mostly focus on the generation speed or similarity of a single sequence, neglecting the core functional constraint coordination problem in biomolecular design.

[0041] To address the aforementioned issues, this paper presents a multi-objective antimicrobial peptide generation method based on constrained Monte Carlo tree search and a diffusion model. A new conditional mask diffusion language model is constructed and fine-tuned using a protein language model as the base model. This model is used to deeply capture complex evolutionary information and structural patterns in protein sequences, improving the prediction accuracy of the minimum inhibitory concentration (MIC) of antimicrobial peptides. Furthermore, to address the challenge of simultaneously considering multiple conflicting attributes such as activity, toxicity, and hemolytic activity in antimicrobial peptide development, a Pareto reward-based constrained Monte Carlo tree search is introduced. This explicitly models the competitive relationships between different objectives, achieving high-quality, multi-objective balanced antimicrobial peptide design.

[0042] like Figure 1 As shown in the embodiment, a multi-objective antimicrobial peptide generation method based on a constrained Monte Carlo tree search and diffusion model is provided, comprising the following steps: S1. Obtain open-source antimicrobial peptide sequences and non-antimicrobial peptide sequences and preprocess them to obtain a dataset.

[0043] In this embodiment, the antimicrobial peptide sequence data includes antimicrobial peptide sequence data from three databases: dbAMP, AMP Scanner, and DRAMP.

[0044] Preprocessing of antimicrobial peptide sequences and non-antimicrobial peptide sequences includes: First, duplicate antimicrobial peptide sequence data were removed from different databases.

[0045] Then, antimicrobial peptide sequences and non-antimicrobial peptide sequences with a length greater than 50 amino acids and less than 11 amino acids were removed from the antimicrobial peptide sequence data, as well as antimicrobial peptide sequences and non-antimicrobial peptide sequences containing single-letter amino acids represented by X, Z, B, and U were eliminated.

[0046] Finally, the dataset was divided into a training set and a validation set in a ratio of 9:1.

[0047] S2. Construct a conditional mask diffusion language model, train and fine-tune it using a dataset to obtain a generative model, generate an initial seed sequence through conditional diffusion, and set an optimization direction vector containing multiple optimization objectives.

[0048] Figure 2This is a schematic diagram of the Conditional Masked Diffusion Language Model (CMDLM) architecture and conditional guidance provided in an embodiment of the present invention. In this embodiment, the CMDLM model uses the ESM-2 150M model as its basic framework. The choice of a 150M model aims to achieve faster diffusion denoising sampling speed while maintaining semantic expressiveness during subsequent complex Monte Carlo tree search iterations. The CMDLM includes 30 Transformer blocks, 20 attention heads, a hidden layer dimension of 640, and an intermediate layer size of 2560. A conditional embedding layer is introduced on top of this, which maps preset conditional labels to vectors of the same dimension as word embeddings and adds these vectors to the word embedding vectors.

[0049] During training, this embodiment uses approximately 2.6M short peptide datasets (10-50 amino acids in length) collected from the UniProt database to perform a first round of supervised full-scale fine-tuning (SFT) on the model. In the second step, building upon general grammar learning, this embodiment uses a labeled dataset containing both antimicrobial peptides (AMP) and non-antimicrobial peptides (non-AMP) to perform a second round of domain-adaptive fine-tuning on the conditional masked diffusion language model. To achieve controlled generation, this embodiment introduces a Classifier-Free Guidance (CFG) technique, randomly discarding conditional labels with a probability of 0.1 during training. In this training phase, the conditional masked diffusion language model needs to simultaneously learn both the label-guided probability distribution and the unconditional baseline distribution. The fine-tuning process employs a composite strategy combining SFT and LoRA lightweight techniques: updating parameters for key components in the conditional masked diffusion language model that directly control semantic mapping, using the AdamW optimizer, and setting the learning rate to 1e. -4 up to 1e -3 The linear scheduling is used, with a weight decay coefficient of 0.01 and a beta value of (0.9, 0.98). The loss function adopted is the reweighted cross-entropy loss function.

[0050] Specifically, this embodiment performs full parameter fine-tuning on the newly introduced conditional embedding layer, the original word embedding layer, and the output head. By updating the word embeddings and the output head, the conditional masked diffusion language model can spatially reconstruct the preferred amino acid frequencies specific to antimicrobial peptides, thereby breaking the over-reliance of the general model on high-frequency amino acids such as L (leucine) and S (serine). For the encoder modules (layers 20 to 29) of the conditional masked diffusion language model (i.e., the high-level semantic layer of the conditional masked diffusion language model), this embodiment introduces LoRA technology to capture the high-dimensional functional features of antimicrobial peptides. In its specific mathematical implementation, LoRA technology uses the original pre-trained weight matrix... This is achieved by adding a branch path as a bypass. This branch path consists of two low-rank matrices. and Composition by multiplication, where the rank , and These are the original pre-trained weight matrices. The number of rows and columns. During training, the original weights... Keep frozen, only for lightweight matrices and Perform gradient updates. The forward computation process is transformed into: , matrix Initialization is performed using a Gaussian distribution, matrix Initializing to zero ensures that no interference noise is introduced during the initial fine-tuning phase. In this way, this embodiment only needs to update a very small percentage of parameters (approximately 0.1%-1% of the total parameters) to accurately induce the model to capture the complex nonlinear features of antimicrobial peptides in the 20-29 layer encoder without disrupting the underlying protein grammar.

[0051] S3. Add noise to the initial seed sequence mask to obtain the noisy state. Using the noisy state as the root node, combine the conditional probability distribution and optimization direction vector output by the generative model, and perform a constrained Monte Carlo tree search that includes selection, expansion, simulation evaluation and backpropagation stages to obtain the updated Pareto front non-dominated solution set and the corresponding candidate sequence.

[0052] In this embodiment, the Constrained Monte Carlo Tree Search (CMCTS) process is first initialized. This embodiment treats the protein sequence generation process as a sequential decision-making task in a discrete space. First, a sequence completely covered by a mask is initialized as the root node of the search tree; simultaneously, a set of multi-objective optimization direction vectors is set to guide the search. These direction vectors represent the desired improvement targets in multiple dimensions such as antibacterial activity, hemolytic activity, toxicity, and minimum inhibitory concentration (MIC).

[0053] Next, constrained MCTS diffusion sampling (CMCTD) at the local level is performed. This is the core search phase of the invention, which explores the optimal denoising path through the following four iterative steps: First, in the selection phase, starting from the current node, the potential of each path is evaluated using an improved UCB (upper confidence interval) score. This score not only considers the Pareto reward already obtained by the path (representing its multi-objective performance) but also incorporates the prior probability of the diffusion model, thus ensuring that the search process both approaches high-performance regions and does not deviate from reasonable protein syntax. Specifically: based on the root node, all expanded child nodes under the root node are traversed, the improved upper confidence interval score of the child nodes is calculated, the improved upper confidence interval scores of all child nodes are compared, and the node corresponding to the maximum value is selected. If the currently selected child node still has unexpanded child nodes, the selection phase ends, and the currently selected child node is returned to enter the expansion phase; otherwise, the above steps are repeated with the aforementioned child node as the new current node, repeating the traversal, calculation, comparison, and judgment steps until a node that has not yet been fully expanded is reached. The formula for calculating the improved upper confidence boundary score of child nodes is as follows: , In the formula, For the improved upper confidence bound score, As the root node, for The first node corresponding to the time Number of child nodes To accumulate rewards, For the number of visits, The noise probability is the output of the conditional masked diffusion language model. To explore coefficients.

[0054] Then, in the expansion phase, for the selected branch, the CMDLM model is used to predict the possible next token state, and Gumbel noise is injected to increase the diversity of sampling. Child nodes are generated based on the noisy conditional probability distribution to prevent the model from getting trapped in local optima. The Gumbel noise formula is: , In the formula, The generated Gumbel noise value, To conform to a uniform distribution A random number, taking values ​​in the range (0, 1). These are the parameter tuning coefficients.

[0055] Then, in the simulation evaluation and dynamic filtering stage, for the newly generated child nodes generated in the expansion stage, the corresponding partial mask sequence is subjected to fast and complete denoising to obtain candidate sequences. The candidate sequences are sent to the external multi-attribute evaluation module, and multiple attribute prediction operators in the external multi-attribute evaluation module are called to calculate the antibacterial probability, the logarithm of the minimum inhibitory concentration, the toxicity probability, and the hemolysis probability, respectively.

[0056] Subsequently, the attribute boost vector of the candidate sequence relative to the seed sequence of the previous round is calculated for each optimization objective, and its alignment with the preset optimization direction is measured. If the optimization direction of a certain path deviates significantly from the preset trajectory, automatic pruning filtering will be performed to effectively suppress the ineffective expansion of the search space. Specifically, the attribute boost vector is calculated. With preset optimized direction vector The cosine of the included angle between them is used to obtain the angle measure. If the angle is greater than the current dynamic angle threshold. If the angle is less than or equal to the dynamic angle threshold, then discard this candidate sequence; if the angle is less than or equal to the dynamic angle threshold, then compare the multi-attribute score of this candidate sequence with the sequence in the current Pareto front non-dominated solution set; wherein, the angle threshold... Initial settings The value is dynamically adjusted using an exponential moving average based on the rejection rate, with an adjustment range of [value missing]. .

[0057] When this candidate sequence is superior to the nondominated solution set on at least one optimization objective. If a candidate sequence is found to be no worse than existing sequences in other optimization objectives, it is added to the nondominated solution set. Remove old sequences dominated by this candidate sequence; calculate the reward value based on the position of this candidate sequence in the updated Pareto front, and use it for backpropagation; Finally, during the backpropagation phase, the final reward is calculated based on the sequence's position in the current Pareto front, and the value weight of the entire search path is updated upwards to guide subsequent search actions.

[0058] S4. Call the external multi-attribute evaluation module to obtain the attribute values ​​of the candidate sequence on each optimization objective, calculate the multi-attribute improvement score, and update the Pareto front non-dominated solution set accordingly. Select some sequences from the current Pareto front, mask and add noise, and use them as seed sequences for the next iteration.

[0059] In this embodiment, the external multi-attribute evaluation module includes an antimicrobial activity prediction unit, a toxicity prediction unit, a hemolytic activity prediction unit, and a minimum inhibitory concentration prediction unit. Antimicrobial prediction unit: The generated antimicrobial peptide sequence is input into the AMP Scanner-v2 model. The built-in prediction operator in the AMP Scanner-v2 model predicts the antimicrobial probability Pamp of each antimicrobial peptide.

[0060] Toxicity prediction unit: The generated antimicrobial peptide sequence is input into the open-source ToxinPred2 model. The ToxinPred2 model has a built-in prediction operator that predicts the toxicity probability Ptoxin of each antimicrobial peptide.

[0061] Hemolysis prediction unit: The generated antimicrobial peptide sequence is input into the open-source HemoPI model, and the HemoPI model has built-in prediction operators that predict the hemolysis probability of each antimicrobial peptide.

[0062] Minimum inhibitory concentration (MIC) prediction unit: A regression model based on the protein language model ESM-2 650M and superimposed with a three-layer multilayer perceptron (MLP) regression head is used to input the generated antimicrobial peptide sequence into this regression model to obtain the regression score, i.e. the logarithm of the MIC, which is used to predict the MIC of each antimicrobial peptide.

[0063] To further eliminate noise in biological attribute prediction and correct early denoising errors, this embodiment introduces a global iterative refinement mechanism. After one round of search, all selected sequences are ranked according to their Multi-Attribute Improvement (MPI) scores. The MPI score comprehensively weighs the improvement of each objective and its alignment with the optimization objective. The best-performing set of sequences is selected, partially masked, and then reintroduced as the starting point for the MCTS search. The multi-attribute improvement score is calculated based on the attribute values ​​of the candidate sequences for each optimization objective, expressed as: , in, Candidate sequences to be evaluated The corresponding multiple attributes increase the score. For the standardized function, To optimize the total number of targets, For the first The weighting coefficients of each optimization objective. Candidate sequences In the The magnitude of attribute improvement for each optimization objective. For ranking functions, For the current iteration round The corresponding noise level, These are the preset noise scheduling parameters. The projection length of the attribute boost vector onto the optimization direction. The weighting coefficients for the alignment terms.

[0064] In this embodiment, a physicochemical property prediction unit and a structure prediction unit are also included. Before outputting the antimicrobial peptide sequence in the final Pareto front, the sequence is sent to the Biopython ProteinAnalysis tool in the physicochemical property prediction unit. This tool can calculate 15 types of physicochemical indicators such as sequence length, molecular weight, aromaticity, instability index, isoelectric point, gravity, charge at pH 6, charge at pH 7, helix fraction, turn structure fraction, plate structure fraction, molar extinction coefficient reduction, molar extinction coefficient value oxidation, average hydrophilicity, and average surface accessibility. Structure prediction unit: The generated antimicrobial peptide sequence is input into the Alphafold Server tool, which can predict the three-dimensional folding structure of the protein sequence and output pdb and cif files for the pLDDT index, which reflects the structural foldability of the protein sequence.

[0065] Finally, the sequences generated by the MP2D framework will be evaluated in a comprehensive evaluation system. The evaluation covers 15 categories of indicators, including: sequence reliability metrics based on language models, pLDDT (particle-dependent foldability) metrics based on AlphaFold Server, and solubility and charge distribution metrics based on physicochemical analysis tools. Experimental data show that the method generated in this embodiment can produce high-quality antimicrobial peptides that satisfy four conflicting attribute constraints in a single round of inference, without requiring repeated training of the generation model for different targets, demonstrating high flexibility and practical value.

[0066] like Figure 3 As shown, in this embodiment, a high-precision attribute prediction system is first constructed to provide quantitative feedback signals for subsequent sequence generation. This system focuses on predicting the minimum inhibitory concentration (MIC), a core activity indicator of antimicrobial peptides.

[0067] Step 1, Data Acquisition and Preliminary Processing: This embodiment first collects raw antimicrobial peptide sequences and their corresponding MIC experimental measurements from multiple authoritative public databases such as dbAMP, AMP Scanner, and DRAMP. After acquiring the data, sequences from different sources are first merged and deduplicated to ensure that each sequence entering the training set has a unique experimental label. Subsequently, based on the actual needs of drug development, excessively long sequences exceeding 40 amino acid residues are removed, and invalid entries containing illegal or uncertain amino acid characters such as X, B, U, and Z are strictly filtered out. To make the model more conducive to capturing activity patterns, this embodiment performs a base-10 logarithmic transformation on the raw MIC values, converting them into logMIC form, thereby effectively reducing the common skewness in the distribution of biological experimental data.

[0068] Step 2, Deep Semantic Feature Extraction: This embodiment uses the ESM-2650M, a large language model with a parameter scale of 650M, as the feature extraction engine. This model is pre-trained on massive amounts of natural protein sequences and possesses extremely strong biological semantic perception capabilities. Specifically, the preprocessed amino acid sequence is input into the ESM-2 encoder. The model generates a 1280-dimensional high-dimensional embedding vector for each amino acid, forming an embedding tensor that reflects the local features of the sequence. To obtain global features that represent the entire sequence, this embodiment performs element-wise average pooling on this tensor, calculating the average of all position vectors in the sequence, ultimately obtaining a fixed-dimensional sequence-level global embedding vector. The calculation formula is as follows: , in, It is the first An embedding vector of an amino acid. A sequence is represented as a vector. This represents the total number of amino acids.

[0069] Step 3, Regression Model Training and Parameter Freezing: The 1280-dimensional sequence representation vector is input into a carefully designed MLP regression model. This model consists of an input layer, multiple fully connected modules, and an output linear layer. The fully connected modules integrate linear mapping, batch normalization, the GELU Gaussian error linear unit activation function, and a Dropout layer to prevent overfitting. During the training phase, this embodiment employs a parameter freezing strategy, keeping the large ESM-2 encoder parameters unchanged and updating the lightweight MLP regression model only through the AdamW optimizer. The learning rate is set to 1×10⁻⁶. -4 The weight decay coefficient is 0.01; the beta parameter is (0.9, 0.98). This approach significantly reduces computational costs and increases training speed while effectively avoiding the risk of overfitting due to the limited scale of biological data. After training, the model can directly output accurate predicted logMIC values ​​based on the input antimicrobial peptide sequence.

[0070] Based on the same technical concept, this application also provides a multi-target antimicrobial peptide generation device based on a constrained Monte Carlo tree search and diffusion model, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the multi-target antimicrobial peptide generation method based on the constrained Monte Carlo tree search and diffusion model.

[0071] To further illustrate the effectiveness of the multi-objective antimicrobial peptide generation method based on a constrained Monte Carlo tree search and diffusion model provided by this invention, experimental verification was conducted. For example... Figure 4The diagram shows a violin illustration comparing the performance of ordinary peptides, general antimicrobial peptides, and antimicrobial peptides generated in this invention across multiple dimensions (including antibacterial activity, non-toxicity, non-hemolytic activity, and minimum inhibitory concentration). The horizontal axis represents the performance of ordinary peptides (peptides), general antimicrobial peptides (AMP), and the multi-target antimicrobial peptides generated in this invention (optimized). The vertical axis represents their scores on the corresponding classifiers (top left and top right for antibacterial activity (AMP-V2) and non-toxicity (NONTOXICITY), and bottom left and bottom right for non-hemolytic activity (NONHEMOLYSIS) and minimum inhibitory concentration (MIC)). Higher scores for antibacterial activity, non-toxicity, and non-hemolytic activity are better, while lower scores for minimum inhibitory concentration are better.

[0072] In summary, this invention provides a multi-objective antimicrobial peptide generation method based on constrained Monte Carlo tree search and diffusion model. By constructing and fine-tuning a new conditional mask diffusion language model based on a protein language model, a deeper understanding of the complex evolutionary information and structural patterns in protein sequences is obtained, improving the prediction accuracy of the minimum inhibitory concentration (MIC) of antimicrobial peptides. Addressing the challenge of simultaneously considering multiple conflicting attributes such as activity, toxicity, and hemolytic activity in antimicrobial peptide development, a Pareto reward-based constrained Monte Carlo tree search is further introduced to explicitly model the competitive relationships between different objectives, thereby achieving high-quality, multi-objective balanced antimicrobial peptide design.

[0073] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for generating multi-objective antimicrobial peptides based on constrained Monte Carlo tree search and diffusion model, characterized in that, Includes the following steps: Antimicrobial peptide sequences and non-antimicrobial peptide sequences were obtained and preprocessed to obtain a dataset. A conditional mask diffusion language model is constructed, and a generative model is obtained by training and fine-tuning it using a dataset. An initial seed sequence is generated through conditional diffusion, and an optimization direction vector containing multiple optimization objectives is set. The initial seed sequence is masked and noise is added to obtain a noisy state. Using the noisy state as the root node, combined with the conditional probability distribution and optimization direction vector output by the generative model, a constrained Monte Carlo tree search including selection, expansion, simulation evaluation and backpropagation stages is performed to obtain the updated Pareto front non-dominated solution set and the corresponding candidate sequence. The external multi-attribute evaluation module is called to obtain the attribute values ​​of the candidate sequence on each optimization objective, calculate the multi-attribute improvement score, and update the Pareto front non-dominated solution set accordingly. Some sequences are selected from the current Pareto front and then masked and noise-added as seed sequences for the next iteration. Based on the current iteration round and the quality index of the Pareto front, determine whether the termination condition has been met. If so, output all antimicrobial peptide sequences in the final saved Pareto front and their corresponding multi-attribute prediction values.

2. The method for generating multi-target antimicrobial peptides according to claim 1, characterized in that, The conditional mask diffusion language model is based on the protein language model ESM and introduces a conditional embedding layer. The conditional embedding layer is used to map the preset conditional labels into vectors of the same dimension as the word embeddings and add the vectors to the word embedding vectors.

3. The method for generating multi-target antimicrobial peptides according to claim 2, characterized in that, When using the dataset for training and fine-tuning, an implicit classifier guidance strategy is adopted, and conditional labels are randomly discarded during training and fine-tuning. A composite strategy is employed during training and fine-tuning, including: Full-supervised fine-tuning is performed on the word embedding layer, conditional embedding layer, and output head in the conditional masked diffusion language model to update the parameters; LoRA low-rank adaptation is introduced for the high-level modules of the Transformer encoder. Specifically, LoRA low-rank adaptation includes adding a branch consisting of two low-rank matrices next to the original pre-trained weight matrix for the high-level modules of the Transformer encoder. The first low-rank matrix is ​​initialized with a Gaussian distribution, and the second low-rank matrix is ​​initialized to zero. During training fine-tuning, only the parameters of these two low-rank matrices are updated.

4. The method for generating multi-target antimicrobial peptides according to claim 1, characterized in that, The selection phase in constrained Monte Carlo tree search includes: based on the root node, traversing all expanded child nodes under the root node, calculating the improved upper confidence bound score of the child nodes, comparing the improved upper confidence bound scores of all child nodes, selecting the node corresponding to the maximum value, and if the currently selected child node still has unexpanded child nodes, the selection phase ends and the currently selected child node is returned to enter the expansion phase; otherwise, the above steps are repeated with the child node as the new current node, repeating the above traversal, calculation, comparison and judgment steps until a node that has not been fully expanded is reached. The formula for calculating the improved upper confidence boundary score of child nodes is as follows: ， In the formula, For the improved upper confidence bound score, As the root node, for The first node corresponding to the time Number of child nodes To accumulate rewards, For the number of visits, The noise probability is the output of the conditional masked diffusion language model. To explore coefficients.

5. The method for generating multi-target antimicrobial peptides according to claim 4, characterized in that, The expansion phase in constrained Monte Carlo tree search includes: based on the currently expanded leaf nodes from the selection phase output, when generating child nodes, independently injecting Gumbel noise into the unnormalized output value of the conditional mask diffusion language model output, and generating child nodes according to the noisy conditional probability distribution.

6. The method for generating multi-target antimicrobial peptides according to claim 5, characterized in that, The simulation evaluation phase in constrained Monte Carlo tree search includes: The partial mask sequence corresponding to the child nodes generated in the expansion stage is completely denoised to obtain candidate sequences; the candidate sequences are then sent to an external multi-attribute evaluation module to obtain the antibacterial probability, the logarithm of the minimum inhibitory concentration, the toxicity probability, and the hemolysis probability, respectively. Calculate the attribute improvement vector of the candidate sequence relative to the seed sequence in the previous round for each optimization objective; Calculate the cosine of the angle between the attribute boosting vector and the preset optimization direction vector to obtain the angle value; if the angle value is greater than the current dynamic angle threshold, discard this candidate sequence; if the angle value is less than or equal to the dynamic angle threshold, compare the multi-attribute score of this candidate sequence with the sequences in the current Pareto front non-dominated solution set; when this candidate sequence is better than all existing sequences in the non-dominated solution set on at least one optimization objective and not worse than existing sequences on other optimization objectives, add this candidate sequence to the non-dominated solution set and remove the old sequences dominated by this candidate sequence; calculate the reward value based on the position of this candidate sequence in the updated Pareto front for backpropagation.

7. The method for generating multi-target antimicrobial peptides according to claim 1, characterized in that, The minimum inhibitory concentration (MIC) prediction in the external multi-attribute evaluation module employs a regression method based on a protein language model, including: Obtain the antimicrobial peptide sequences and the corresponding experimental measurements of minimum inhibitory concentrations, construct a training dataset, and preprocess it. The preprocessed antimicrobial peptide sequence is input into the protein language model encoder to obtain the sequence embedding tensor, and then average pooling is performed. The average pooled sequence embedding tensor is input into the multilayer perceptron regression model to output the predicted log minimum inhibitory concentration value. The loss function is calculated based on the mean squared error between the predicted and actual values. The parameters of the multilayer perceptron regression model are updated through backpropagation to complete the training. During prediction, the protein language model encoder is frozen, and only the multilayer perceptron regression model is run to output the predicted logarithmic minimum inhibitory concentration value.

8. The method for generating multi-target antimicrobial peptides according to claim 7, characterized in that, Calculating the multi-attribute improvement score includes: for each optimization objective, calculating the improvement of the candidate sequence on that objective, and determining the ranking of this improvement among all current candidate sequences; based on the ranking of each optimization objective, the score is obtained using the following formula: ， in, Candidate sequences to be evaluated The corresponding multiple attributes increase the score. For the standardized function, To optimize the total number of targets, For the first The weighting coefficients of each optimization objective. Candidate sequences In the The magnitude of attribute improvement for each optimization objective. For ranking functions, For the current iteration round The corresponding noise level, These are the preset noise scheduling parameters. The projection length of the attribute boost vector onto the optimization direction. The weighting coefficients for the alignment terms.

9. The method for generating multi-target antimicrobial peptides according to claim 1, characterized in that, The multi-target antimicrobial peptide generation method further includes: before outputting the sequence in the final Pareto front, sending the sequence into a physicochemical property prediction unit to calculate physicochemical indices including sequence length, molecular weight, aromaticity, instability index, isoelectric point, charge, secondary structure fraction, and hydrophilicity; and sending the sequence into a structure prediction unit to predict the three-dimensional structure and obtain the structural foldability index. The sequences are filtered according to a preset threshold, and sequences that do not meet the requirements for physicochemical properties or structural foldability are removed.

10. A multi-objective antimicrobial peptide generation device based on a constrained Monte Carlo tree search and diffusion model, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the multi-objective antimicrobial peptide generation method based on the constrained Monte Carlo tree search and diffusion model as described in any one of claims 1-9.

Images