Protein mutation effect prediction method based on latent diffusion and structure guided denoising

Abstract

The application relates to the field of artificial intelligence and bioinformatics, and particularly relates to a protein mutation effect prediction method based on latent diffusion and structure-guided denoising, which comprises the following steps: extracting one-dimensional sequence features and two-dimensional residue pair features of a protein sequence; mapping the two-dimensional residue pair features into bias information with the same scale as the one-dimensional sequence features, and fusing the one-dimensional sequence features and the bias information to obtain fused protein features; mapping the fused protein features to a continuous latent space, performing forward diffusion and adding noise in the continuous latent space, and obtaining a noise latent representation; performing reverse denoising on the noise latent representation in the continuous latent space with a wild-type protein structure embedding as a condition; decoding the obtained denoised latent representation to obtain a reconstructed sequence distribution; and determining a protein mutation effect prediction value according to a log probability difference of a mutant amino acid and a wild-type amino acid at a mutation site of the protein sequence in the reconstructed sequence distribution, so that the accuracy of protein mutation effect prediction is significantly improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence and bioinformatics, and more specifically, to a method for predicting protein mutation effects based on potential diffusion and structure-guided denoising. Background Technology

[0002] Predicting protein mutation effects is a crucial fundamental issue in protein design, enzyme engineering optimization, disease-related variant analysis, and drug development. Due to the extremely vast space of possible protein mutation combinations, traditional wet-lab screening methods are costly and time-consuming, failing to meet the demands of high-throughput protein optimization and functional analysis. Therefore, using computational models to predict mutation effects has become an important technological direction in this field. With the development of large-scale protein sequence databases and deep learning technologies, mutation effect prediction methods based on protein language models have gradually become mainstream. Early representative methods, such as ESM2, Tranception, ProteinNPT, and ESM-MSA, primarily learn evolutionary constraints and contextual dependencies in protein sequences through self-supervised pre-training on massive sequences. To further enhance the model's ability to model protein spatial dependencies, existing techniques have introduced sequence-structure joint modeling schemes, such as ESM-IF, SaProt, and ProtSSN, which utilize structural conditional sequence modeling to enhance the characterization of local conformational changes and residue spatial interactions. In recent years, diffusion models have also begun to be applied in this field. For example, DePLM and other schemes restore and optimize protein representation by progressively adding noise forward and progressively removing noise backward, providing new technical ideas for protein characterization and refinement.

[0003] However, the aforementioned existing technologies still have many shortcomings. First, pure sequence mutation effect prediction methods are mainly based on linear sequences. Although they can learn contextual dependencies, they lack explicit modeling of two-dimensional relationships and spatial constraints of residues, making it difficult to fully express the interactions between distant residues. This results in insufficient characterization of long-range synergistic effects in complex functional properties, and the prediction accuracy in tasks such as stability and binding capacity is easily limited. Second, sequence-structure joint modeling methods usually focus on "adding structural information" rather than "purifying evolutionary representations." These schemes only use structure as an additional input or auxiliary representation source, lacking dedicated feature denoising and task-related signal purification processes. As a result, irrelevant constraints and noise information that are not directly related to the specific target properties are still retained in the original protein language model representation, and the model is still susceptible to interference from weakly relevant information.

[0004] Furthermore, existing diffusion-based schemes have significant limitations in utilizing the representation space and fusing multi-source information. Most existing diffusion-based methods model within discrete sequence spaces or constrained representation spaces. Since discrete spaces are more suitable for symbol-level generation or replacement operations, they are easily limited by the discrete state transition methods during noise injection, state recovery, and multi-step iterations. This hinders the use of standard Gaussian processes to smoothly perturb and recover deep continuous features, resulting in insufficient smoothness of representations and inadequate multimodal information fusion capabilities. It is also difficult to simultaneously and uniformly handle one-dimensional sequence features, two-dimensional residue pair features, and structural conditional information. Finally, existing methods typically emphasize only a single task objective in their training objective design, such as focusing solely on sequence reconstruction capability or mutation effect regression and ranking. They lack a unified multi-objective joint optimization mechanism and a training strategy that simultaneously constrains sequence reconstruction, latent space denoising, and mutation effect prediction, making it difficult for the model to learn intermediate representations that are both biologically plausible and highly sensitive to mutation effects.

[0005] In summary, existing technologies still struggle to accurately target the key signals that determine mutation effects. It is necessary to propose a new technical solution to further improve the accuracy and robustness of protein mutation effect prediction. Summary of the Invention

[0006] To address the aforementioned technical problems, this application provides a protein mutation effect prediction method based on latent diffusion and structure-guided denoising. This method can denoise and refine protein evolutionary characterization in a continuous latent feature space, and simultaneously integrate sequence information, residue pair relationship information, and structural condition information to predict protein mutation effects, thereby improving prediction accuracy, robustness, and generalization ability.

[0007] The technical solution provided in this application is as follows: A method for predicting protein mutation effects based on potential diffusion and structure-guided denoising includes: Extracting one-dimensional sequence features and two-dimensional residue pair features from protein sequences; The two-dimensional residue pair features are mapped to bias information of the same scale as the one-dimensional sequence features, and then fused with the one-dimensional sequence features to obtain fused protein features; The fusion protein features are mapped to a continuous latent space, and forward diffusion noise is added in the continuous latent space to obtain a noisy latent representation; With wild-type protein structure embedding as a condition, inverse denoising is performed on the noisy latent representation in the continuous latent space to obtain a denoised latent representation; The denoised latent representation is decoded to obtain the reconstructed sequence distribution; The predicted value of protein mutation effect is determined based on the difference in log probability between mutant and wild-type amino acids at the mutation site in the reconstructed sequence.

[0008] One possible implementation involves extracting one-dimensional sequence features and two-dimensional residue pair features from the protein sequence, including: The protein sequence is input into a pre-trained protein language model, and the contextual representation of each residue site in the protein sequence is extracted as the one-dimensional sequence feature. The attention weights of the multi-layer attention mechanism in the protein language model are extracted to construct the two-dimensional residue pair features that characterize the implicit interaction relationship between any two residues.

[0009] In one possible implementation, the fusion protein features are constructed. , represented as:

[0010] In the formula, Representing one-dimensional sequence features, Representing two-dimensional residue pairs as features, This indicates the mapping of residue pairs to sequences and the enhancement of self-attention.

[0011] In one possible implementation, the fused protein features are mapped to a continuous latent space, and forward diffusion noise is performed in the continuous latent space to obtain a noisy latent representation, including: The fusion protein features are input into the encoder and mapped to a continuous latent space to obtain an initial latent representation. In the continuous latent space, a forward diffusion process is constructed based on a Markov chain, and Gaussian noise is gradually injected into the initial latent representation according to a preset number of diffusion time steps to generate the noisy latent representation.

[0012] One possible implementation involves performing inverse denoising on the noisy latent representation in the contiguous latent space, conditioned on wild-type protein structure embedding, including: The wild-type protein structure embedding is used as a conditional input and is fed into a pre-defined denoising network along with the noise latent representation of the current time step. The mean term corresponding to the reverse denoising process is determined based on the output of the denoising network with structured conditions. Combined with the preset variance term corresponding to the reverse denoising process, the latent representation of the previous step is recovered from the latent noise representation of the current step. Using the wild-type protein structure embedding as a prior constraint for the direction of potential representation recovery, the denoised potential representation is obtained through iterative recovery.

[0013] In one possible implementation, decoding the denoised latent representation to obtain the reconstructed sequence distribution includes: The denoised latent representation is input into the decoder, and the output is a reconstructed sequence; wherein the decoding process is constrained by a sequence reconstruction loss, which is used to preserve protein sequence information.

[0014] In one possible implementation, the predicted value of the protein mutation effect is determined based on the logarithmic probability difference between the mutant amino acid and the wild-type amino acid at the protein sequence mutation site under the reconstructed sequence distribution, including: For a given mutant sample, at each mutation site, calculate the conditional log probability of the mutant amino acid and the wild-type amino acid under the reconstructed distribution; The difference between the conditional log probability of the mutant amino acid and the conditional log probability of the wild-type amino acid is calculated as the mutation contribution at that site. The mutational contributions from all mutation sites are summed to obtain the predicted value of the protein mutation effect.

[0015] In one possible implementation, the method is based on a pre-trained protein mutation effect prediction model, wherein the total loss function of the protein mutation effect prediction model is the sum of sequence reconstruction loss, diffusion loss and mutation effect loss; The sequence reconstruction loss is expressed as:

[0016] In the formula, Represents the actual protein sequence. This represents the sequence distribution obtained from model reconstruction; The diffusion loss is expressed as:

[0017] In the formula, This indicates the actual protein sequence. Condition Information It follows a standard Gaussian distribution. noise variables and diffusion time step After joint sampling, the expected value of the corresponding objective function is taken; This represents the actual Gaussian noise injected during the forward diffusion process. This indicates that the denoising network with structured conditions is at the diffusion time step Given a latent noise representation and wild-type protein structure embedding The noise predicted under the given conditions Indicates the first The potential noise representation corresponding to each diffusion time step Indicates the diffusion time step. This indicates the embedding of wild-type protein structures; The loss due to the mutation effect is expressed as:

[0018] In the formula, This represents the true mutation effect value; Indicates removing the first... Sequence context after each site and condition information In this case, the model predicts the first The conditional probability that each site contains a mutant amino acid. Indicates the first The reconstructed amino acid random variable output at each site, , ; This means removing the first given number. Sequence context after each site and condition information In this case, the model predicts the first The conditional probability that each site contains a wild-type amino acid. .

[0019] Compared with the prior art, the technical solution provided in this application has the following beneficial effects: This application performs diffusion-based denoising in a continuous latent feature space, explicitly removing evolutionary noise unrelated to the target property through forward noise addition and backward recovery mechanisms. Compared with most existing schemes that directly use pre-trained representations or only perform simple structural enhancements and lack explicit noise stripping processes, this application can obtain a latent representation optimized for mutation effect prediction tasks, significantly improving the accuracy of protein mutation effect prediction.

[0020] Meanwhile, this application extracts both one-dimensional site features and two-dimensional residue pair features from the pre-trained protein language model and performs unified modeling through a feature fusion module. This overcomes the shortcomings of existing pure sequence methods that typically only utilize one-dimensional sequence context, and can better model long-distance residue coupling relationships and implicit structural constraints, demonstrating stronger characterization capabilities in tasks involving higher-order spatial interactions.

[0021] By embedding wild-type structures as conditional inputs to the inverse denoising network, the direction of latent representation recovery is dynamically guided during the denoising recovery process. Compared with existing sequence-structure joint schemes that often treat structure as a static additional input, this approach achieves more full utilization of structural information.

[0022] By using sequence reconstruction loss to ensure that the latent representation retains sequence information, using diffusion loss to learn the noise recovery process, and using mutation effect loss to directly connect to the target task, a unified multi-objective training mechanism is adopted, which breaks the limitation of existing schemes that usually only optimize a single objective.

[0023] By employing a unified framework of potential diffusion denoising, sequence-residue pair fusion, and structure condition guidance, sequence, structure, and task supervision information are collaboratively modeled. Compared to existing methods that are typically only suitable for a specific modeling paradigm, this approach is widely applicable to various mutation effect prediction tasks, such as protein stability, expression, binding, and activity. It is suitable for scenarios such as enzyme engineering optimization, protein design, and disease variant screening. Attached Figure Description

[0024] Figure 1 This is a flowchart of a protein mutation effect prediction method based on potential diffusion and structure-guided denoising, provided in Embodiment 1 of this application.

[0025] Figure 2 This is a schematic diagram of the structure of Mut-LDM, a protein mutation effect prediction model based on potential diffusion and structure-guided denoising, provided in Embodiment 1 of this application.

[0026] Figure 3 This is a schematic diagram of the ablation experiment results of the Mut-LDM protein mutation effect prediction model provided in Embodiment 3 of this application.

[0027] Figure 4 This is a schematic diagram showing the results of a comparative experiment under different diffusion steps provided in Embodiment 3 of this application. Detailed Implementation

[0028] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0029] Example 1 See Figure 1 This is a flowchart of a protein mutation effect prediction method based on latent diffusion and structure-guided denoising, provided in Embodiment 1 of this application. Figure 1 As shown, the specific implementation steps of the above method include: Step 101: Extract one-dimensional sequence features and two-dimensional residue pair features from the protein sequence.

[0030] like Figure 2As shown in the embodiments of this application, the pre-trained protein language model ESM2 is used to process the input protein sequence. x Feature extraction is performed to obtain one-dimensional sequence features. This corresponds to the contextual representation of each residue site. Simultaneously, the multi-layer attention weights of the aforementioned protein language model are extracted to form two-dimensional residue pair features. , is used to characterize the implicit interaction between any two residues.

[0031] Specifically, for a length of L The protein sequence, the above one-dimensional sequence characteristics The corresponding one-dimensional feature dimension is Two-dimensional residue pairs Two-dimensional feature dimension .in, For the hidden layer dimension, For the number of network layers, For the number of attention heads.

[0032] Compared to existing technologies that rely on one-dimensional sequence context to directly utilize sequence representation or sequence likelihood difference of protein language models for prediction, this application further extracts two-dimensional residue pair features, which can not only utilize sequence context, but also explicitly enhance the modeling of residue relationships and structural constraints, and remove evolutionary noise that is irrelevant to the target properties.

[0033] Step 102: Map the above two-dimensional residue pair features to bias information of the same scale as the above one-dimensional sequence features, and fuse them with the above one-dimensional sequence features to obtain fusion protein features.

[0034] like Figure 2 As shown, this application constructs a feature fusion module that maps the aforementioned two-dimensional residue pair features to bias information at the same scale as the one-dimensional sequence features via Pair-To-Sequence mapping, and injects it into the self-attention update process of the sequence pathway to obtain fused protein features. This process can be represented as:

[0035] in, This represents the mapping of residue pairs to sequences and the self-attention enhancement process. Through the aforementioned feature fusion module, both sequence context information and residue pair relationship information can be utilized simultaneously during site characterization updates, thereby enhancing the modeling capability for higher-order couplings and implicit structural constraints.

[0036] Step 103: Map the above-mentioned fusion protein features to a continuous latent space, and perform forward diffusion noise addition in the above-mentioned continuous latent space to obtain a noisy latent representation.

[0037] Specifically, the fusion protein features The input encoder is mapped to the latent space to obtain the initial latent representation. , represented as:

[0038] In the aforementioned continuous latent space, a forward diffusion process is constructed based on a Markov chain, proceeding to the aforementioned initial latent representation according to a preset number of diffusion time steps. Gaussian noise is gradually injected to generate a latent noise representation.

[0039] Specifically, this application models the forward diffusion process as a Markov chain, represented as:

[0040] In the formula, Indicates from the first The diffusion time step to the 1st The sequence consisting of all potential representations at each diffusion time step Indicates the first The potential representation corresponding to each diffusion time step This indicates that, given an initial latent representation Under the condition, from the first Step to the first Joint conditional probability distribution of forward diffusion step; From the The diffusion time step to the 1st Multiply the single-step transition probabilities of each diffusion time step together. Indicates by the first The potential representation of each diffusion time step Transfer to the The potential representation of each diffusion time step The single-step conditional probability distribution.

[0041] The single-step transfer satisfies:

[0042] In the formula, This represents the noise figure, used to control the intensity of the Gaussian noise injected at this time step; This represents the identity matrix that matches the dimensions of the latent representation. The mean is The covariance matrix is The Gaussian distribution.

[0043] remember , Then it can be further written as:

[0044] In the formula, Indicates from the first The diffusion time step to the 1st The cumulative signal retention coefficient at each diffusion time step. Indicates the first Signal retention coefficients for each diffusion time step This indicates that the mean is 0 and the covariance matrix is ​​the identity matrix. The standard Gaussian distribution. Therefore, it can be directly derived from the standard Gaussian distribution at any diffusion step. Sampling .

[0045] Compared to existing technologies that directly model in discrete protein sequence space, this application first maps the fusion protein representation to a continuous latent space, then performs Gaussian noise addition and reverse denoising, and unifies the fusion sequence features, residue pair features and structural condition information. The denoising trajectory is smoother, the feature representation is more stable, and the multimodal information fusion is more complete, thereby achieving smooth recovery of deep evolutionary representation and purification of task-related signals.

[0046] Step 104: Using wild-type protein structure embedding as a condition, perform inverse denoising on the above-mentioned noisy latent representation in the above-mentioned continuous latent space to obtain a denoised latent representation.

[0047] In the backdiffusion stage, this application employs a denoising network to progressively recover the latent representation and embeds the wild-type structure. s As a conditional input to the denoising network, structural information is not just a static additional feature, but dynamically participates in the recovery process of the latent representation. Thus, even when the real structure of the mutant is lacking, structural priors can still be used to improve prediction accuracy, making the recovered latent representation more in line with the requirements of the target task.

[0048] Specifically, the reverse denoising process is expressed as follows:

[0049] In the formula, The parameter is The reverse denoising process from the first The diffusion time step to the 1st All potential representations of each diffusion time step The joint probability distribution constituting the sequence, Indicates the first Each diffusion time step corresponds to a potential representation The prior probability distribution, The parameter is The denoising network, given the current diffusion time step, has a latent representation Under the condition of restoring the potential representation of the previous diffusion time step The conditional probability distribution satisfies .in, For the mean term, This is the variance term.

[0050] In this embodiment of the application, the variance term is set as... .in, Indicates the first The scalar variance coefficients corresponding to each inverse denoising time step are used to characterize the intensity of randomness introduced during the inverse sampling process of that time step.

[0051] The mean term is represented as .in, This represents the output of a denoising network with structure conditions. Based on the output of the denoising network with structure conditions, this application determines the mean term corresponding to the reverse denoising process, and combines it with the preset variance term corresponding to the reverse denoising process to recover the previous step's latent representation from the current step's noisy latent representation. Furthermore, by introducing structure conditions, this application can utilize the prior constraint of wild-type structure to constrain the direction of latent representation recovery when the true structure of the mutant is lacking.

[0052] Step 105: Decode the above denoised latent representation to obtain the reconstructed sequence distribution.

[0053] Specifically, the denoised latent representation is input into the decoder to obtain the reconstructed sequence, which is represented as:

[0054] Step 106: Determine the predicted value of the protein mutation effect based on the log probability difference between the mutant amino acid and the wild-type amino acid at the mutation site in the above reconstructed sequence distribution.

[0055] In the embodiments of this application, for a given mutated sample, at each mutation site At each site, the conditional logarithmic probabilities of the mutant and wild-type amino acids under the reconstructed distribution are calculated, and the difference is taken as the mutation contribution at that site. The overall mutational effect of multi-site mutations can be expressed as:

[0056] in, Indicates a mutated amino acid. This indicates wild-type amino acids. This represents the sequence context after removing the current site. Indicates conditional information.

[0057] Furthermore, the method described above in this application is based on a pre-trained protein mutation effect prediction model, the structure of which is as follows: Figure 2As shown in the diagram, it consists of a one-dimensional / two-dimensional feature extraction module, a feature fusion module, a latent space coding module, a diffusion denoising module, a structural condition module, a decoding module, and a multi-objective training module. It is used to execute the protein mutation effect prediction method described in Embodiment 1 of this application, which will not be elaborated here.

[0058] In this embodiment, the total loss function of the protein mutation effect prediction model is the sum of sequence reconstruction loss, diffusion loss and mutation effect loss. A multi-objective training method is adopted to simultaneously optimize sequence reconstruction, diffusion denoising and mutation effect prediction.

[0059] The sequence reconstruction loss is expressed as:

[0060] In the formula, Represents the actual protein sequence. This represents the sequence distribution obtained from model reconstruction; The diffusion loss is expressed as:

[0061] In the formula, This indicates the actual protein sequence. In condition information It follows a standard Gaussian distribution. noise variables and diffusion time step After joint sampling, the expected value of the corresponding objective function is taken. This represents the actual Gaussian noise injected during the forward diffusion process. This indicates that the denoising network with structured conditions is at the diffusion time step Given a latent noise representation and wild-type protein structure embedding The noise predicted under the given conditions Indicates the first The potential noise representation corresponding to each diffusion time step Indicates the diffusion time step. This indicates the embedding of wild-type protein structures; The loss due to the mutation effect is expressed as:

[0062] In the formula, This represents the true mutation effect value; Indicates removing the first... Sequence context after each site and condition information In this case, the model predicts the first The conditional probability that each site contains a mutant amino acid. Indicates the first The reconstructed amino acid random variable output at each site, , ; This means removing the first given number. Sequence context after each site and condition information In this case, the model predicts the first The conditional probability that each site contains a wild-type amino acid. .

[0063] The total loss function described above is expressed as follows:

[0064] Through the above-mentioned joint optimization, this application is able to gradually remove noise characterization that is irrelevant to the target properties while maintaining protein sequence information, and strengthen potential signals related to mutation effect prediction.

[0065] Example 2 Building upon Embodiment 1, Embodiment 2 further clarifies the specific configuration and implementation method of the technical solution proposed in this application. Specifically, the method proposed in this application is based on a pre-trained protein language model and a latent diffusion denoising framework. In a preferred embodiment, the backbone sequence encoder uses the 650M parameter version of ESM2, which contains 33 TransformerBlock layers with a hidden layer dimension of 1280, used to extract a one-dimensional contextual representation of the input protein sequence. Simultaneously, the attention weights of all hidden layers are extracted and concatenated to construct two-dimensional residue pair features, thereby characterizing the implicit structural relationships between any residue pairs.

[0066] In the feature fusion stage, two-dimensional residue pairs are mapped to bias information at the same scale as one-dimensional sequence features via a Pair-To-Sequence module and injected into the self-attention update process of the sequence pathway to obtain the fused protein characterization. This fused characterization is then mapped to a continuous latent space, where Gaussian diffusion is performed. Preferably, the diffusion time step number is set to 300, and the initial noise coefficient is... Set as End noise figure Set as .

[0067] In terms of structural condition modeling, the structural information of wild-type proteins is preferably used as the conditional input for the reverse denoising stage to impose geometric constraints on the latent representation recovery process. Preferably, the denoising network adopts a U-Net architecture to balance local pattern and global context modeling capabilities. The contributions of the structural condition module, feature fusion module, and diffusion module can be verified by the ablation results.

[0068] Regarding the training strategy, this invention employs a multi-objective joint optimization approach to simultaneously optimize sequence reconstruction loss, potential diffusion loss, and mutation effect loss. The optimizer preferably uses AdamW, with a learning rate set to [value missing]. The weight decay coefficient is set to The optimal number of training epochs is 50.

[0069] In terms of experimental data setup, the overall performance experiments used four benchmark datasets: ProteinGym, β-lactamase, GB1, and Fluorescence. For ProteinGym, 201 DMS datasets were selected and further divided into 66 stability datasets, 69 fitness datasets, 16 expression datasets, 12 binding datasets, and 38 activity datasets. Randomized five-fold cross-validation was used for each dataset, and the final results were reported as the five-fold mean. The Spearman rank correlation coefficient was used as the evaluation metric.

[0070] In terms of hardware implementation environment, training is preferably performed on a multi-GPU platform. The training environment includes two NVIDIA RTX-4090 processors, each with 24GB of video memory.

[0071] Example 3 To verify the effectiveness and advancement of the method proposed in this application, Example 3 of this application conducted control experiments on several publicly available benchmark datasets for predicting protein mutation effects. The datasets used included ProteinGym, β-lactamase, GB1, and Fluorescence. ProteinGym includes five tasks: stability, adaptation, expression, binding, and activity; β-lactamase is used to characterize mutation effects related to antibiotic resistance; GB1 is used to assess the impact of mutations on IgG binding adaptation; and Fluorescence is used to characterize the impact of mutations on fluorescence intensity. To ensure comparability with existing methods, 201 DMS datasets were preferably selected for evaluation on ProteinGym. All tasks employed randomized five-fold cross-validation, and Spearman's rank correlation coefficient was used as the evaluation metric.

[0072] As shown in Table 1, the proposed method achieves state-of-the-art results on ProteinGym stability, adaptability, expression, and activity tasks, as well as on the β-lactamase, GB1, and Fluorescence datasets. It also achieves near-optimal performance on the ProteinGym binding task, with overall performance superior to or no less than existing mainstream methods. Comparisons include pure sequence modeling methods such as ESM2, ESM-MSA, and Tranception; sequence-structure joint modeling methods such as ProtSSN and SaProt; and methods incorporating diffusion mechanisms such as ProteinNPT and DePLM. These results demonstrate that the unified framework proposed in this application—combining sequence features, residue pair features, structural conditions, and potential diffusion denoising—maintains high accuracy and robustness across various types of mutation effect prediction tasks.

[0073] Table 1: Comparison on protein mutation benchmark datasets (bold: best, underline: second best)

[0074] Further analysis reveals that, compared to pure sequence methods, the proposed method exhibits a more significant advantage in tasks requiring stronger spatial awareness. Compared to sequence-structure joint modeling methods, this application still demonstrates superior overall results, indicating that simply adding structural information is insufficient to significantly improve prediction accuracy; explicit purification of irrelevant information in the original evolutionary representation and purification of task-related signals are also necessary. For the diffusion-based scheme DePLM, which is closest to this invention, this application still shows superior results on most evaluation tasks, demonstrating that performing diffusion-based denoising in a continuous latent feature space and introducing wild-type structural conditions for guidance in the reverse denoising stage is more conducive to learning stable feature recovery trajectories and fusing multi-source information.

[0075] In addition to overall performance comparison, further ablation studies were conducted on this application to verify the contribution of each component module to the final performance. The ablation settings included: a structural information removal module (w / o ST), a feature fusion removal module (w / o FF), a diffusion removal module (w / o DB), and a module that simultaneously removes structural information, feature fusion, and diffusion (w / o ST+FF+DB). Figure 3 As shown, removing any module significantly reduces model performance; when the structural information, feature fusion, and diffusion modules are removed simultaneously, the model degenerates into an implementation relying solely on the basic sequence backbone, exhibiting the lowest performance. This indicates that the structural condition module, feature fusion module, and latent diffusion denoising module are all irreplaceable key components of the technical solution presented in this application.

[0076] Furthermore, Embodiment 3 of this application also conducted experimental analysis on the key parameter of diffusion steps. For example... Figure 4 As shown, when When no diffusion denoising is performed and the latent features are decoded directly, the model performance is significantly lower. As the number of diffusion steps increases, the model performance gradually improves and tends to stabilize. Considering both performance and training efficiency, 300 diffusion time steps are preferred. These results further demonstrate that the design of feature refinement through latent space diffusion denoising in this invention is effective.

[0077] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting protein mutation effects based on potential diffusion and structure-guided denoising, characterized in that, include: Extracting one-dimensional sequence features and two-dimensional residue pair features from protein sequences; The two-dimensional residue pair features are mapped to bias information of the same scale as the one-dimensional sequence features, and then fused with the one-dimensional sequence features to obtain fused protein features; The fusion protein features are mapped to a continuous latent space, and forward diffusion noise is added in the continuous latent space to obtain a noisy latent representation; With wild-type protein structure embedding as a condition, inverse denoising is performed on the noisy latent representation in the continuous latent space to obtain a denoised latent representation; The denoised latent representation is decoded to obtain the reconstructed sequence distribution; The predicted value of protein mutation effect is determined based on the difference in log probability between mutant and wild-type amino acids at the mutation site in the reconstructed sequence.

2. The protein mutation effect prediction method based on potential diffusion and structure-guided denoising according to claim 1, characterized in that, Extracting one-dimensional sequence features and two-dimensional residue pair features from protein sequences, including: The protein sequence is input into a pre-trained protein language model, and the contextual representation of each residue site in the protein sequence is extracted as the one-dimensional sequence feature. The attention weights of the multi-layer attention mechanism in the protein language model are extracted to construct the two-dimensional residue pair features that characterize the implicit interaction relationship between any two residues.

3. The protein mutation effect prediction method based on potential diffusion and structure-guided denoising according to claim 1, characterized in that, Constructing the fusion protein features , represented as: In the formula, Representing one-dimensional sequence features, Representing two-dimensional residue pairs as features, This indicates the mapping of residue pairs to sequences and the enhancement of self-attention.

4. The protein mutation effect prediction method based on potential diffusion and structure-guided denoising according to claim 1, characterized in that, The fusion protein features are mapped to a continuous latent space, and forward diffusion noise is added in the continuous latent space to obtain a noisy latent representation, including: The fusion protein features are input into the encoder and mapped to a continuous latent space to obtain an initial latent representation. In the continuous latent space, a forward diffusion process is constructed based on a Markov chain, and Gaussian noise is gradually injected into the initial latent representation according to a preset number of diffusion time steps to generate the noisy latent representation.

5. The method for predicting protein mutation effects based on potential diffusion and structure-guided methods according to claim 1, characterized in that, Conditioned by wild-type protein structure embedding, inverse denoising is performed on the noisy latent representation in the contiguous latent space, including: The wild-type protein structure embedding is used as a conditional input and is fed into a pre-defined denoising network along with the noise latent representation of the current time step. The mean term corresponding to the reverse denoising process is determined based on the output of the denoising network with structured conditions. Combined with the preset variance term corresponding to the reverse denoising process, the latent representation of the previous step is recovered from the latent noise representation of the current step. Using the wild-type protein structure embedding as a prior constraint for the direction of potential representation recovery, the denoised potential representation is obtained through iterative recovery.

6. The method for predicting protein mutation effects based on potential diffusion and structure-guided methods according to claim 1, characterized in that, Decoding the denoised latent representation to obtain the reconstructed sequence distribution includes: The denoised latent representation is input into the decoder, and the output is a reconstructed sequence; wherein the decoding process is constrained by the sequence reconstruction loss to preserve protein sequence information.

7. The method for predicting protein mutation effects based on potential diffusion and structure-guided methods according to claim 1, characterized in that, Based on the difference in log probabilities between mutant and wild-type amino acids at the protein sequence mutation site under the reconstructed sequence distribution, the predicted value of the protein mutation effect is determined, including: For a given mutant sample, at each mutation site, calculate the conditional log probability of the mutant amino acid and the wild-type amino acid under the reconstructed distribution; The difference between the conditional log probability of the mutant amino acid and the conditional log probability of the wild-type amino acid is calculated as the mutation contribution at that site. The mutational contributions from all mutation sites are summed to obtain the predicted value of the protein mutation effect.

8. The method for predicting protein mutation effects based on potential diffusion and structure-guided methods according to claim 1, characterized in that, The method is based on a pre-trained protein mutation effect prediction model, and the total loss function of the protein mutation effect prediction model is the sum of sequence reconstruction loss, diffusion loss and mutation effect loss. The sequence reconstruction loss is expressed as: In the formula, Represents the actual protein sequence. This represents the sequence distribution obtained from model reconstruction; The diffusion loss is expressed as: In the formula, This indicates the actual protein sequence. In condition information It follows a standard Gaussian distribution. noise variables and diffusion time step After joint sampling, the expected value of the corresponding objective function is taken; This represents the actual Gaussian noise injected during the forward diffusion process. This indicates that the denoising network with structured conditions is at the diffusion time step Given a latent noise representation and wild-type protein structure embedding The noise predicted under the given conditions Indicates the first The potential noise representation corresponding to each diffusion time step Indicates the diffusion time step. This indicates the embedding of wild-type protein structures; The loss due to the mutation effect is expressed as: In the formula, This represents the true mutation effect value; Indicates removing the first... Sequence context after each site and condition information In this case, the model predicts the first The conditional probability that each site contains a mutant amino acid. Indicates the first The reconstructed amino acid random variable output at each site, , ; This means removing the first given number. Sequence context after each site and condition information In this case, the model predicts the first The conditional probability that each site contains a wild-type amino acid. .

Images