A method for predicting a conformational ensemble of a promiscuous protein based on a generative deep learning model

Abstract

The application relates to a kind of allosteric protein conformation ensemble prediction methods based on generative deep learning model, comprising the following steps: using quaternion to characterize protein skeleton;The probability distribution information implied in conformation data is learned using the generative modeling based on score matching, a protein structure diffusion model is constructed, the diffusion model takes protein sequence as input, and adopts the dynamic conformation of allosteric protein structure representation as output, point invariant attention mechanism and Transformer are used in the score matching network in the model;The loss function of the model is constructed based on DSM loss, the position of skeleton atom and the loss on distance matrix, and the protein structure diffusion model is trained;Protein sequence is obtained, input into the trained protein structure diffusion model, and the predicted allosteric protein conformation ensemble is obtained.Compared with the prior art, the application realizes sampling conformation at the skeleton level, and can provide more detailed structural features and more accurate property estimation.

Description

Technical Field

[0001] This invention relates to the field of allosteric protein conformation ensemble prediction, and in particular to a method for allosteric protein conformation ensemble prediction based on a generative deep learning model. Background Technology

[0002] Allosteric proteins are important regulatory elements of physiological functions and have attracted widespread attention from academia and industry as drug targets with high selectivity and safety potential. However, current experimental methods have limitations in resolving allosteric protein structures and identifying allosteric sites, failing to capture the conformational ensemble of allosteric proteins, and also suffer from drawbacks such as excessive cost and the ability to resolve only average structural properties. Therefore, there is an urgent need to develop accurate and efficient artificial intelligence methods for the study of the conformational ensemble of allosteric proteins.

[0003] The development of artificial intelligence methods, particularly generative models such as diffusion models, has become an important research tool, including in the biomedical field. Generative deep learning learns patterns from data, modeling data from a top-down perspective. A common paradigm for this approach is to design a simple distribution (such as a Gaussian distribution) and then use a deep learning network combined with a specific algorithm to map the data from the simple distribution to the target data space. Converged generative models can sample from a simple distribution and, through network transformation, obtain samples that approximate the target data distribution. Its core advantage lies in simplifying the iterative sampling process of non-independent and identically distributed (IOD) data into IOD sampling, significantly improving sampling efficiency and ensuring that the sampling range is not affected by the imbalance of the target data's probability distribution. This paradigm has extremely high potential value in studying the structural ensembles of biological macromolecules, including allosteric proteins. However, currently, no research has applied generative deep learning to the conformational ensemble study of allosteric proteins, making it difficult to accurately express the relevant characteristics of allosteric proteins. Summary of the Invention

[0004] The purpose of this invention is to solve the problem of predicting the conformational ensemble of allosteric proteins, and to provide a method for predicting the conformational ensemble of allosteric proteins based on a generative deep learning model. By using generative deep learning methods and fine-tuning a diffusion model, the conformational ensemble of allosteric proteins can be predicted directly from the sequence, which helps to elucidate their structure-function mechanism and develop targeted drugs.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A generative deep learning model-based method for predicting the conformational ensemble of allosteric proteins directly predicts the dynamic conformational set of allosteric proteins from the sequence through a fine-tuned protein structure diffusion model. The method includes the following steps:

[0007] Protein structure characterization: using N, C α A quaternion consisting of four atoms (C, O) represents the protein backbone. A protein backbone segment formed by a single residue is considered a triangular rigid body, with C... α Centered on;

[0008] Protein structure diffusion model construction: The protein structure diffusion model is constructed by using fractional matching-based generative modeling to learn the probability distribution information hidden in the conformational data. The protein structure diffusion model takes the protein sequence as input and the dynamic conformation of the allosteric protein represented by the protein structure as output. The fractional matching network in the model uses a point-invariant attention mechanism to capture the interaction and relationship between nearby residues in the coordinate space. After the point-invariant attention mechanism, a Transformer is connected to learn the overall characteristics and long-range interactions of the chain.

[0009] Protein structure diffusion model training: A loss function is constructed based on DSM loss, the position of backbone atoms, and the loss on the distance matrix to train the protein structure diffusion model;

[0010] Allosteric protein conformation ensemble prediction: Obtain the protein sequence, input it into the trained protein structure diffusion model, and obtain the predicted allosteric protein conformation ensemble.

[0011] In the protein structure diffusion model, generative modeling is represented by a diffusion process defined by a stochastic differential equation. This diffusion process is divided into forward diffusion and backward diffusion, wherein the forward diffusion process is characterized by the following equation:

[0012] dx=f(x,t)dt+g(t)dw

[0013] Where t∈[0,T] is a continuous variable. This represents a standard Wiener procedure. The drift factor is a parameter in vector form. The diffusion factor is a scalar parameter, where T represents the endpoint of the forward diffusion process and n represents the data dimension in which the diffusion process takes place. The forward diffusion process defines the process by which protein structure data x is noisy by the drift factor as the time factor t continuously increases.

[0014] The reverse diffusion process is also defined in the form of stochastic differential equations:

[0015]

[0016] in, It is the partial derivative of the target data distribution in each dimension of the data, that is, the score of the data; It is a standard Wiener process that propagates back from T to 0 with time factor t, where dt is an infinitesimally small time step with a negative value, and x t＝0 Indicates protein conformation, x t＝T The endpoint of the noise addition process is represented by , which represents the pure noise sampled from the Gaussian distribution. By solving the stochastic differential equation of the back diffusion process at each t, ​​data sampled from the Gaussian distribution is transformed to obtain a protein conformation ensemble that conforms to the Boltzmann distribution.

[0017] The protein structure diffusion process is rotationally and translationally isotropic, meaning that the overall rotation and translation should have the same effect on the input data and the output data after the diffusion process or network transformation, as shown in the following equation:

[0018]

[0019] in, ρ represents the diffusion process or network transformation, and ρ represents the rotation and translation operation.

[0020] To satisfy the conditions for rotation and translation equivariance, for each frame T i :=[R i ,v i ], the rotation matrix R under the SO(3) group i and Translation vector v in space i The diffusion process is handled independently, as follows:

[0021]

[0022] in Controlling the noise level during the diffusion process In the group Brownian motion as defined in P: Used to remove the center of mass;

[0023] During forward diffusion or noise addition, the addition of noise to the rotation matrix is ​​determined by the noise kernel p. t|0 (R t |R0) is determined, and the noise kernel is obtained from an isotropic Gaussian distribution under the SO(3) group, which has the following form:

[0024]

[0025] in It is a combined rotation matrix The representation after the axis-angle transformation;

[0026] The addition of noise to the translation vector satisfies the Ornstein-Uhlenbeck process. The noise kernel of the translation vector is shown as follows and will eventually converge to

[0027]

[0028] where means that the noise is sampled from a Gaussian distribution.

[0029] The fractional matching network is used to learn the fraction of data x in the reverse diffusion process to make it as close as possible to the true fraction of x, achieving the purpose of generating conformations;

[0030] The fractional matching network requires three inputs in each layer: a one-dimensional vector representation s l , a pairwise feature representation z l and the rotation and translation set T l to be updated; where, the protein sequence features are extracted by ESM2-650M and spliced with the residue position encoding and the time encoding represented by trigonometric functions as the initial one-dimensional vector representation s0 of the model, and the pairwise feature representation z0 is obtained by transforming s0 according to the relative position encoding; after the pointwise invariant attention mechanism-Transformer transformation in each layer, the one-dimensional vector representation is updated through a fully connected network, and the updated vector is cross-multiplied to update the pairwise features.

[0031] The DSM loss is expressed as:

[0032]

[0033] where s θ (T t ,t) represents the data fraction predicted by the network, refers to the actual fraction of the data; λ(t) is the weight, which is set as:

[0034]

[0035] For the samples with fewer forward diffusion steps, the mean square error is used to supervise the positions of the backbone atoms and the differences on the distance matrix. Among them, the samples with fewer forward diffusion steps are the samples that satisfy t < T / 4.

[0036] The total loss function of the protein structure diffusion model is expressed as:

[0037]

[0038] Where ω1=ω2=0.25 controls the weights of conformational quality loss. For DSM loss, This represents the mean square error in the skeleton atomic coordinates. The mean square error between the distance matrix of the generated conformation and the actual structure.

[0039] The protein structure diffusion model was trained with a learning rate of 10. -4 The Adam optimizer optimizes the score matching network.

[0040] For model training of the translation vector part, a linear noise-adding strategy in VP-SDE is used, and for model training of the rotation matrix part, a logarithmic noise-adding strategy in VE-SDE is used, as shown in the following equation:

[0041]

[0042] Where, β min =0.1,β max =20,σ min =0.1,σ max =1.5, both constants, used to adjust the noise scales β(t) and σ(t) at each time step t, where t∈[0,T] is a continuous variable and T represents the endpoint of the forward diffusion process.

[0043] The training strategy for the protein structure diffusion model consists of two parts: the first part is to pre-train the model on the experimental structure dataset, and the second part is to fine-tune the model on the IDRome dataset. The conformational data in the training set for the second part must simultaneously meet the conditions of high structural diversity and high sequence specificity.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This invention proposes a novel generative deep learning model for predicting the conformational ensembles of allosteric proteins. This method, through a fine-tuned diffusion model, can directly predict conformational sets from protein sequences without relying on expensive multiple sequence alignment (MSA) or experimental data. By starting from a single protein sequence, it can accurately predict the distribution characteristics of allosteric protein multiconformation ensembles, contributing to the revelation of these ensembles and their transformational functional patterns. This method not only provides an important tool for allosteric protein ensemble prediction and dynamic conformation research but will also significantly promote the design of allosteric protein-targeted drugs and the development of related therapies. Attached Figure Description

[0046] Figure 1 This is a flowchart of the method of the present invention;

[0047] Figure 2 This is a schematic diagram of the protein backbone structure;

[0048] Figure 3 This is a schematic diagram of the protein structure diffusion model of the present invention;

[0049] Figure 4 This is a conformational ensemble of the allosteric protein KRAS predicted using the method of the present invention in one embodiment. Detailed Implementation

[0050] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0051] This embodiment provides a method for predicting the conformational ensemble of allosteric proteins based on a generative deep learning model. It directly predicts the dynamic conformational set of allosteric proteins from the sequence using a fine-tuned protein structure diffusion model. This method can directly predict the protein conformational set from the sequence using a fine-tuned diffusion model. The proposed method does not rely on costly multiple sequence alignment or crystal structures, achieving accurate prediction of the conformational ensemble of allosteric proteins starting from a single protein sequence. Compared with state-of-the-art methods, this invention achieves conformational sampling at the scaffold level, providing more detailed structural features and more accurate property estimates, and will help reveal the conformational ensemble and structure-function paradigm of allosteric proteins. This invention provides an important tool for predicting the conformational ensemble and studying the dynamic conformation of allosteric proteins, and can also provide strong support for drug discovery targeting allosteric proteins and the development of related therapies.

[0052] Specifically, such as Figure 1 As shown, the method includes the following steps:

[0053] Step 1) Protein structure characterization

[0054] Previous methods typically used Cartesian coordinates to represent the positions of atoms in proteins, where Cartesian coordinates represent the number of atoms that make up the protein. However, due to the redundancy of information in coordinate representation, we must extract necessary protein features for network training, using as few features as possible to characterize the complete internal structure of the protein. This embodiment uses quaternions to characterize the protein backbone. The protein backbone is mainly composed of N, C α It is composed of four atoms: C, O, and O. O and C are connected by double bonds, which gives O very little freedom in the chain-like structure of the skeleton. Figure 2As shown, the main backbone feature related to the position of O is the dihedral angle ω, which is distributed around 180° in almost all proteins, indicating that the relative position of O is very fixed. Therefore, to characterize the protein backbone, the main focus should be on N and C. α The protein backbone consists of three atoms, C, and T. Based on the characteristics of crystal structure, these three atoms are relatively fixed internally; that is, their bond lengths and internal bond angles are essentially constant. Therefore, in this embodiment, a protein backbone fragment formed by a single residue is considered as a triangular rigid body, with C as the bond length. α Centered on.

[0055] Step 2) Construction of a protein structure diffusion model

[0056] To enable the model to capture the complex Boltzmann distribution of allosteric proteins under physiological conditions, this embodiment uses score-based generative modeling (SGM) to learn the implicit probability distribution information in conformational data and construct a protein structure diffusion model. This model takes the protein sequence as input and outputs the dynamic conformation of the allosteric protein represented by its protein structure. The score-matching network in the model employs invariant point attention (IPA) to capture the interactions and relationships between nearby residues in coordinate space. A Transformer is then connected after the IPA to learn the overall chain characteristics and long-range interactions.

[0057] SGM can be represented by a diffusion process defined by a stochastic differential equation. This diffusion process is divided into a forward diffusion process and a backward diffusion process, also known as the noise addition process and the noise reduction process, respectively. The forward diffusion process is characterized by the following equation:

[0058] dx=f(x,t)dt+g(t)dw (1)

[0059] Where t∈[0,T] is a continuous variable. This represents the standard Wiener process (also known as Brownian motion). The drift factor is a parameter in vector form. Let the diffusion factor be a scalar parameter, T represent the endpoint of the forward diffusion process, and n represent the data dimension during the diffusion process. It can be seen that the forward diffusion process defines the process by which protein structure data x is noisily added by the drift factor as the time factor t continuously increases. Correspondingly, the backward diffusion process has also been shown to be defined in the form of a stochastic differential equation:

[0060]

[0061] in, It is the partial derivative of the target data distribution in each dimension of the data, that is, the score of the data; It is a standard Wiener process that propagates back from T to 0 with time factor t, where dt is an infinitesimally small time step with a negative value, and x t＝0 This represents the protein conformation, which is the real data that the network wants to learn from. t＝T The endpoint of the noise addition process is represented by , which represents the pure noise sampled from the Gaussian distribution. By solving the stochastic differential equation of the back-diffusion process at each t, ​​data sampled from the Gaussian distribution is transformed to obtain a protein conformation ensemble conforming to the Boltzmann distribution. In this equation, besides the fraction of data x... All other parts have fixed solutions. Therefore, we only need one neural network for score matching to learn. The goal of generating a conformation is to make it conform as closely as possible to the true fraction of x.

[0062] This embodiment has thus explained the basic principle of SGM. However, protein structure cannot be represented by a simple x; it is necessary to consider each frame T. i :=[R i ,v i Both parts of the [structure] are designed with reasonable diffusion processes. It is important to note that this embodiment requires the diffusion process to be rotationally and translationally equivalent; that is, the overall rotation and translation should have the same effect on the input data and the output data after the diffusion process or network transformation, as shown in the following equation:

[0063]

[0064] in, ρ represents the diffusion process or network transformation, and ρ represents the rotation and translation operation.

[0065] To satisfy the conditions for rotation and translation equivariance, for each frame T i :=[R i ,v i ], the rotation matrix R under the SO(3) group i and Translation vector v in space i The diffusion process is handled independently, as follows:

[0066]

[0067] in Controlling the noise level during the diffusion process In the group Brownian motion as defined in P: Used to remove the center of mass.

[0068] During forward diffusion or noise addition, the addition of noise to the rotation matrix is ​​determined by the noise kernel p. t|0 (R t |R0) is determined, and the noise kernel is obtained from an isotropic Gaussian distribution under the SO(3) group, which has the following form:

[0069]

[0070] in It is a combined rotation matrix The representation after axis-angle transformation.

[0071] The addition of noise to the translation vector follows an Ornstein-Uhlenbeck process, also known as VP-SDE. The noise kernel of the translation vector is shown in the following equation, and it will eventually converge to...

[0072]

[0073] in, This indicates that the noise is sampled from a Gaussian distribution.

[0074] Equation (6) describes the diffusion process on the protein structure represented by the rotation and translation matrix. In this process, isotropic Gaussian noise is added to the group containing the rotation matrix and the translation vector to achieve the purpose of the diffusion process being equivalent to the rotation and translation.

[0075] The network architecture used for fractional matching on rotation and translation matrices is described below: Since the diffusion process is designed to be strictly rotation-translation equivariant, network transformations should obviously not violate this. Therefore, this embodiment adopts a variant of the structural module DenoisingIPA from AlphaFold2, and the overall network architecture design is as follows. Figure 3 As shown, the network employs Invariant Point Attention (IPA) to capture the interactions and relationships between nearby residues in the coordinate space. A Transformer is then connected after IPA to learn the overall chain features and long-range interactions. This network design requires three inputs in each layer: a one-dimensional vector representation s... l Pairwise feature representation z l And the set of rotations and translations T that needs to be updated lIn this embodiment, ESM2-650M is used to extract protein sequence features, which are concatenated with residue position encoding and time encoding represented by trigonometric functions as the initial one-dimensional vector representation s0 of the model. And pairwise feature representations z0 are obtained by transforming s0 according to relative position encoding. After each layer of IPA-Transformer transformation, the one-dimensional vector representation is updated through a fully connected network, and the updated vectors are cross-multiplied to update the pairwise features.

[0076] Step 3) Training of the protein structure diffusion model

[0077] Based on the DSM loss, the positions of backbone atoms, and the loss on the distance matrix, the loss function of the model is constructed to train the protein structure diffusion model. The network training strategy includes two steps: one is pre-training on the experimental dataset; the other is fine-tuning on the IDRome molecular dynamics simulation dataset to obtain the final network model.

[0078] Specifically, the score matching network has a different training objective from conventional neural networks. It does not fit the protein conformation but the scores of the data after a certain degree of noise addition, that is, it fits the distribution of the noisy data. To measure the fitting degree of the predicted scores to the actual distribution, this embodiment calculates the DSM loss as shown in the following formula:

[0079]

[0080] where s θ (T t , t) represents the data score predicted by the network, denotes the actual score of the data; λ(t) is the weight. To ensure that the DSM loss scores 1 for perfect fitting at all time steps t, making the contributions of each time step to the loss function the same, it is set as:

[0081]

[0082] In addition, to ensure that the model learns the detailed features in the protein structure, in addition to the DSM loss on the rotation and translation matrices, this embodiment uses the mean square error to supervise the differences in the positions of backbone atoms and the distance matrix for samples with fewer forward diffusion steps (t < T / 4).

[0083] Therefore, the total loss function of the protein structure diffusion model is expressed as:

[0084]

[0085] where ω1 = ω2 = 0.25 controls the weights of the conformational quality loss, is the DSM loss, is the mean square error on the backbone atom coordinates, The mean square error between the distance matrix of the generated conformation and the actual structure.

[0086] During the optimization process, the maximum time step t = 1.0, and the learning rate used for model training is 10. -4 The Adam optimizer optimizes DenoisingIPA. For the translation vector part of the network training, a linear noise-adding strategy from VP-SDE is used, while for the rotation matrix part, a logarithmic noise-adding strategy from VE-SDE is used, as shown in the following equation:

[0087]

[0088] β min =0.1,β max =20 (10)

[0089]

[0090] σ min =0.1,σ max =1.5 (11) where β min ,β max ,σ min ,σ max , are both constants, used to adjust the noise scales β(t) and σ(t) at each time step t, where t∈[0,T] is a continuous variable, and T represents the endpoint of the forward diffusion process.

[0091] The network training strategy is described below: It consists of two parts. First, pre-training on experimental structure data: This invention divides the experimental structure dataset into two parts: crystal structures and NMR-resolved conformations. The crystal structure part references the trRosetta protein structure prediction work published by Jianyi Yang et al. in 2020, selecting a resolution lower than […]. The protein sequences were deredundant using the MMseq2 tool, resulting in 15,051 protein systems with a sequence similarity of no more than 30%. To increase the conformational diversity of individual protein systems in the training set, all NMR-resolved systems from the PDB were downloaded and screened using the same 30% sequence similarity threshold, resulting in 539 systems with an average of 20 conformations per system, totaling 10,454 structures. Pre-training was performed on these structure datasets. Secondly, fine-tuning was performed on the IDRome dataset: For the model's training set, this embodiment believes that the conformational data should not only have high structural diversity but also high sequence specificity. Since the direct input to the model is only protein sequences, if the model does not encounter sufficiently diverse sequence inputs during training, it is likely that different sequences of the same length will generate similar or even identical conformations, i.e., the model's sequence specificity is too low. To solve this problem, this embodiment references the coarse-grained protein database IDRome published in Nature in January 2024 by Giulio Tesei et al. This database contains 28,058 deduplicated natural random protein systems from the human proteome. These systems account for 35% of the human proteome, demonstrating the wide distribution and important functions of natural random proteins. The simulation duration for each system in this dataset is no less than 68 ns. 1000 frames of low-correlation conformations were extracted at equal intervals from the simulation trajectory. The pre-trained model was then fine-tuned on this conformation dataset to obtain the final model.

[0092] Step 4) Prediction of allosteric protein conformation ensemble

[0093] Specifically: obtain the protein sequence, input it into the trained protein structure diffusion model, and obtain the predicted allosteric protein conformation ensemble.

[0094] The accuracy and efficiency of the method of this invention were verified by predicting and evaluating the classic allosteric protein KRAS. Figure 4 As shown, the method of the present invention can capture different experimental conformations (B and C) of allosteric proteins. This result demonstrates the accuracy of the method of the present invention. At the same time, since the method does not rely on multiple sequence alignment and the sampling process is not limited by energy barriers, it can quickly capture multiple conformations including KRAS, thus verifying its high efficiency.

[0095] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for predicting allosteric protein conformation ensembles based on generative deep learning models, characterized in that, The method for directly predicting the dynamic conformation set of allosteric proteins from a sequence using a fine-tuned protein structure diffusion model includes the following steps: Protein structure characterization: using N, C α A quaternion consisting of four atoms (C, O) represents the protein backbone. A protein backbone segment formed by a single residue is considered a triangular rigid body, with C... α Centered on; Protein structure diffusion model construction: The protein structure diffusion model is constructed by using fractional matching-based generative modeling to learn the probability distribution information hidden in the conformational data. The protein structure diffusion model takes the protein sequence as input and the dynamic conformation of the allosteric protein represented by the protein structure as output. The fractional matching network in the model uses a point-invariant attention mechanism to capture the interaction and relationship between nearby residues in the coordinate space. After the point-invariant attention mechanism, a Transformer is connected to learn the overall characteristics and long-range interactions of the chain. Protein structure diffusion model training: A loss function is constructed based on DSM loss, the position of backbone atoms, and the loss on the distance matrix to train the protein structure diffusion model; Allosteric protein conformation ensemble prediction: Obtain the protein sequence, input it into the trained protein structure diffusion model, and obtain the predicted allosteric protein conformation ensemble.

2. The method for predicting allosteric protein conformation ensembles based on a generative deep learning model according to claim 1, characterized in that, In the protein structure diffusion model, generative modeling is represented by a diffusion process defined by a stochastic differential equation. This diffusion process is divided into forward diffusion and backward diffusion, wherein the forward diffusion process is characterized by the following equation: dx=f(x,t)dt+g(t)dw Where t∈[0,T] is a continuous variable. This represents a standard Wiener procedure. The drift factor is a parameter in vector form. The diffusion factor is a scalar parameter, where T represents the endpoint of the forward diffusion process and n represents the data dimension in which the diffusion process takes place. The forward diffusion process defines the process by which protein structure data x is noisy by the drift factor as the time factor t continuously increases. The reverse diffusion process is also defined in the form of stochastic differential equations: in, It is the partial derivative of the target data distribution in each dimension of the data, that is, the score of the data; It is a standard Wiener process that propagates back from T to 0 with time factor t, where dt is an infinitesimally small time step with a negative value, and x t＝0 Indicates protein conformation, x t＝T The endpoint of the noise addition process is represented by , which represents the pure noise sampled from the Gaussian distribution. By solving the stochastic differential equation of the back diffusion process at each t, ​​data sampled from the Gaussian distribution is transformed to obtain a protein conformation ensemble that conforms to the Boltzmann distribution.

3. The method for predicting allosteric protein conformation ensembles based on a generative deep learning model according to claim 1, characterized in that, The protein structure diffusion process is rotationally and translationally isotropic, meaning that the overall rotation and translation should have the same effect on the input data and the output data after the diffusion process or network transformation, as shown in the following equation: in, ρ represents the diffusion process or network transformation, and ρ represents the rotation and translation operation. To satisfy the conditions for rotation and translation equivariance, for each frame T i :=[R i ,v i ], the rotation matrix R under the SO(3) group i and Translation vector v in space i The diffusion process is handled independently, as follows: in Controlling the noise level during the diffusion process In the group Brownian motion as defined in [the text] Used to remove the center of mass; During forward diffusion or noise addition, the addition of noise to the rotation matrix is ​​determined by the noise kernel p. t|0 (R t |R0) is determined, and the noise kernel is obtained from an isotropic Gaussian distribution under the SO(3) group, which has the following form: in It is a combined rotation matrix Representation after axis-angle transformation; The addition of noise to the translation vector follows an Ornstein-Uhlenbeck process, and the noise kernel of the translation vector is shown in the following equation, which will eventually converge to... in, This indicates that the noise is sampled from a Gaussian distribution.

4. The method for predicting allosteric protein conformation ensembles based on a generative deep learning model according to claim 2, characterized in that, The score matching network is used to learn the score of data x during the backdiffusion process. To make it conform as closely as possible to the true fraction of x, thereby achieving the purpose of generating a conformation; The score matching network requires three inputs in each layer: a one-dimensional vector representation s l Pairwise feature representation z l And the set of rotations and translations T that needs to be updated l In this model, ESM2-650M is used to extract protein sequence features, which are then concatenated with the time codes of residue position encoding and trigonometric function representation as the initial one-dimensional vector representation s0. Paired feature representations z0 are obtained by transforming s0 according to the relative position encoding. After the point-invariant attention mechanism-Transformer transformation in each layer, the one-dimensional vector representation is updated through a fully connected network, and the updated vector is then cross-multiplied to update the paired features.

5. The method for predicting allosteric protein conformation ensembles based on a generative deep learning model according to claim 1, characterized in that, The DSM loss is expressed as: Among them, s θ (T t ,t) represents the data score predicted by the network. Then it refers to the actual score of the data; λ(t) is the weight, which is set as:

6. The method for predicting allosteric protein conformation ensembles based on a generative deep learning model according to claim 1, characterized in that, For samples with fewer forward diffusion steps, mean squared error is used to monitor the positions of skeleton atoms and the differences in the distance matrix, wherein the samples with fewer forward diffusion steps are those that satisfy t < T / 4.

7. The method for predicting allosteric protein conformation ensembles based on a generative deep learning model according to claim 1, characterized in that, The total loss function of the protein structure diffusion model is expressed as: Where ω1=ω2=0.25 controls the weights of conformational quality loss. For DSM loss, This represents the mean square error in the skeleton atomic coordinates. The mean square error between the distance matrix of the generated conformation and the actual structure.

8. The method for predicting allosteric protein conformation ensembles based on generative deep learning models according to claim 1, characterized in that, The protein structure diffusion model was trained with a learning rate of 10. -4 The Adam optimizer optimizes the score matching network.

9. The method for predicting allosteric protein conformation ensembles based on a generative deep learning model according to claim 8, characterized in that, For model training of the translation vector part, a linear noise-adding strategy in VP-SDE is used, and for model training of the rotation matrix part, a logarithmic noise-adding strategy in VE-SDE is used, as shown in the following equation: Where, β min =0.1,β max =20,σ min =0.1,σ max =1.5, both constants, used to adjust the noise scales β(t) and σ(t) at each time step t, where t∈[0,T] is a continuous variable and T represents the endpoint of the forward diffusion process.

10. The method for predicting allosteric protein conformation ensembles based on a generative deep learning model according to claim 1, characterized in that, The training strategy for the protein structure diffusion model consists of two parts: the first part is to pre-train the model on the experimental structure dataset, and the second part is to fine-tune the model on the IDRome dataset. The conformational data in the training set for the second part must simultaneously meet the conditions of high structural diversity and high sequence specificity.

Images