Method, apparatus and computer program product for protein structure prediction

Abstract

A method for protein structure prediction is presented. In the method, an initial structure of a target protein is obtained based on composition information related to amino acid of the target protein for a diffusion process. A diffusion process is performed on the initial structure based on the composition information and the force field prediction for the target protein. And determining a target structure of the target protein based on a result of the diffusion process.

Description

Technical Field

[0001] The present disclosure relates generally to the field of computers, and more particularly, to methods, apparatus, and computer program products for protein structure prediction. Background Art

[0002] Proteins are dynamic macromolecules that play important roles in a variety of biological processes. Their functions are primarily achieved through conformational changes—structural modifications—that enable proteins to interact with other molecules. Mapping the protein conformational landscape provides crucial insights for identifying potentially druggable sites hidden beneath the protein surface and revealing transition pathways between multiple metastable states. A comprehensive understanding of protein conformation facilitates the elucidation of biological response mechanisms, enabling researchers to design targeted inhibitors and therapeutics with improved specificity and efficacy. Summary of the Invention

[0003] In a first aspect of the present disclosure, a method for protein structure prediction is provided. The method comprises: obtaining an initial structure of a target protein based on compositional information related to amino acids of the target protein through a diffusion process; performing a diffusion process on the initial structure based on the compositional information and a force field prediction for the target protein; and determining a target structure of the target protein based on the results of the diffusion process.

[0004] In a second aspect of the present disclosure, an electronic device is provided, comprising: a computer processor coupled to a computer-readable memory unit, wherein the memory unit comprises instructions, which, when executed by the computer processor, implement the method according to the first aspect of the present disclosure.

[0005] In a third aspect of the present disclosure, a computer program product is provided, which includes a computer-readable storage medium having program instructions implemented therein, and the program instructions can be executed by an electronic device to enable the electronic device to perform the method according to the first aspect of the present disclosure.

[0006] This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] The above and other objects, features and advantages of the present disclosure will become more apparent by describing some embodiments of the present disclosure in more detail in the accompanying drawings, in which like reference numerals generally refer to like parts in the embodiments of the present disclosure.

[0008] Figure 1shows an example environment in which example embodiments of the present disclosure can be implemented;

[0009] Figure 2 A schematic diagram illustrating an example diffusion process for protein structure prediction according to some embodiments of the present disclosure is shown;

[0010] Figure 3 A schematic diagram illustrating an example generation process of a protein structure according to some embodiments of the present disclosure;

[0011] Figure 4 An example flowchart showing a method for protein structure prediction according to some embodiments of the present disclosure; and

[0012] Figure 5 A block diagram of an electronic device is shown in which various embodiments of the present disclosure can be implemented. DETAILED DESCRIPTION

[0013] The principles of the present disclosure will now be described with reference to some embodiments. It should be understood that these embodiments are described only to illustrate and help those skilled in the art understand and implement the present disclosure, and do not imply any limitation on the scope of the present disclosure. The embodiments described herein can be implemented in various ways except for the manner described below.

[0014] In the following description and claims, unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs.

[0015] References in this disclosure to "one embodiment," "an embodiment," "an example embodiment," etc., indicate that the described embodiment may include a particular feature, structure, or characteristic, but not every embodiment necessarily includes the particular feature, structure, or characteristic. Furthermore, such phrases do not necessarily refer to the same embodiment. Furthermore, when a particular feature, structure, or characteristic is described in conjunction with an embodiment, it is understood that it is within the knowledge of those skilled in the art that other embodiments may be combined to effect such feature, structure, or characteristic, whether or not explicitly described.

[0016] It should be understood that although the terms "first," "second," etc. may be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, a first element can be referred to as a second element, and similarly, a second element can be referred to as a first element without departing from the scope of the example embodiments. As used herein, the term "and / or" includes any and all combinations of one or more of the listed terms.

[0017] The terms used herein are used only to describe particular embodiments and are not intended to limit the example embodiments. As used herein, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprise," "including," "having," "include," and / or "comprising," when used herein, specify the presence of the recited features, elements, and / or components, but do not preclude the presence or addition of one or more other features, elements, components, and / or combinations thereof.

[0018] The principles of the present disclosure will now be described with reference to some exemplary embodiments. It should be understood that these embodiments are described only to illustrate and help those skilled in the art understand and implement the present disclosure, and do not imply any limitation on the scope of the present disclosure. The embodiments described herein can be implemented in various ways other than those described below. In the following description and claims, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the art to which the present disclosure belongs.

[0019] It is understandable that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) must comply with the requirements of relevant laws, regulations and relevant provisions.

[0020] It is understandable that before using the technical solutions disclosed in the various embodiments of this disclosure, the type, scope of use, usage scenarios, etc. of the personal information involved in this disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0021] For example, in response to a user's active request, a prompt message is sent to the user to clearly inform the user that the operation requested will require the acquisition and use of the user's personal information. This allows the user to independently choose whether to provide personal information to the electronic device, application, server, storage medium, or other software or hardware that performs the operations of the disclosed technical solution based on the prompt message.

[0022] As an optional but non-limiting implementation, in response to receiving a user's active request, a prompt message may be sent to the user, for example, in the form of a pop-up window, in which the prompt message may be presented in text form. Furthermore, the pop-up window may also include a selection control for the user to select "agree" or "disagree" to provide personal information to the electronic device.

[0023] It is understandable that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of the present disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of the present disclosure.

[0024] As used herein, the term "model" is referred to as the association between inputs and outputs learned from training data, and thus a corresponding output can be generated for a given input after training. The generation of the model can be based on machine learning techniques. Typically, a machine learning model can be constructed that receives input information and makes predictions based on the input information. For example, a classification model can predict the classification of input information within a set of predetermined classifications. As used herein, a "model" may also be referred to as a "machine learning model," "learning model," "machine learning network," or "learning network," which are used interchangeably herein.

[0025] Traditional physics-based simulation methods, such as molecular dynamics (MD) simulations, have been widely studied for protein conformational sampling. Utilizing well-designed empirical force fields and numerical integrators, the model propagates the three-dimensional (3D) structure of the protein system over time using Newtonian mechanics. MD simulations converge toward an equilibrium distribution (i.e., a Boltzmann distribution) given sufficient time, which helps estimate significant thermodynamic properties, such as changes in bonding free energy.

[0026] However, to adhere to energy conservation and ensure numerical stability, MD simulations typically use time steps of only a few femtoseconds. This presents a challenge, as some biological processes of interest, such as protein folding, span much longer timescales, ranging from microseconds to seconds. This results in limited sampling efficiency within traditional MD simulations, further compounded by the rare event sampling problem, hindering widespread adoption of MD for high-throughput studies.

[0027] With the development of machine learning, machine learning models such as deep neural networks have been used for protein structure prediction. Built on the foundation of powerful folding models, several attempts have been made to customize these deep neural networks for protein conformation sampling. By perturbing the model input, such as multiple sequence alignment (MSA) masking or clustering, the folding model provides a more diverse set of possible folded structures, i.e., alternative conformations. However, this heuristic approach cannot guarantee that the predicted structure is a low-energy state of the target sequence.

[0028] Several solutions have incorporated diffusion models for protein conformation generation. By pre-training on a large number of known protein structures and effectively sampling via a predefined stochastic process, these models show promise in exploring various protein conformational states. However, existing diffusion models lack important physical prior information (such as MD force fields) to guide the diffusion process, hindering the ability to reliably sample different protein conformations that conform to the Boltzmann distribution.

[0029] Embodiments of the present disclosure provide a method for protein structure prediction. According to the embodiments of the present disclosure, an initial structure of a target protein is obtained based on compositional information related to the amino acids of the target protein during a diffusion process. Based on this compositional information and a force field prediction of the target protein, a diffusion process is performed on the initial structure. Based on the results of the diffusion process, a target structure of the target protein is determined.

[0030] In the disclosed embodiments, physics-based force guidance is used for protein structure prediction. This novel, theoretically guaranteed, physics-based guidance strategy effectively guides the diffusion sampler to generate low-energy conformations that better conform to the underlying Boltzmann distribution. In this way, prediction quality and fidelity can be ensured.

[0031] The theoretical basis and embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0032] Figure 1 1 shows a block diagram of an example environment 100 in which various embodiments of the present disclosure can be implemented. Figure 1 In an environment 100, an electronic device 120 receives composition information 101 and outputs a target structure 102. Composition information 101 is related to the amino acids of a target protein, such as the amino acid sequence of the target protein. Target structure 102 is the structure of the target protein. In this article, this structure is also referred to as a conformation.

[0033] The electronic device 120 includes or is equipped with a structure prediction model 110. To predict the structure of a target protein, the structure prediction model 110 generates an initial structure of the target protein based on composition information. The initial structure can be a noisy structure, for example, sampled from a Gaussian distribution.

[0034] Based on the composition information and the force field prediction for the target protein, a diffusion process is performed on the initial structure 101. The diffusion process and force field prediction are described below. Then, based on the results of the diffusion process, a target structure of the target protein can be determined 102. For example, the output of the diffusion process can be used as a predicted structure for the target protein.

[0035] To better understand the embodiments of the present disclosure, some principles are first described. The protein backbone can be represented as follows: for a protein with N amino acid residues, its backbone atomic coordinates can be parameterized by a series of N orientation-preserving rigid transformations (i.e., coordinate systems) for the local [N, C, C, O] backbone atoms in each residue. By x0 = [T0, R0] ∈ SE (3) N Unified representation of the positions of all N coordinate systems, where and R0∈SO(3) NDenote the corresponding translation and rotation operations respectively. Use the description around the oxygen atom in each residue around CC α The additional backbone dihedral angle ψ of the bond rotation allows the reconstruction of the protein backbone structure from the coordinate system representation.

[0036] On the manifold SE(3) N Diffusion modeling is used for protein backbone generation. In some embodiments, two independent diffusion processes can be defined for the translational and rotational subspaces, respectively: The subscript t represents the diffusion time variable, which ranges from [0, 1], and β t and σ t is a predefined time-dependent noise plan, P is a projection operator to remove the centroid, Indicates The standard Wiener process in T satisfies in The rotation transformation kernel satisfies p t (R t |R0)=IGSO3(R t ; R0, t), where IGSO3 is the isotropic Gaussian distribution on SO(3).

[0037] The associated inverse time stochastic differential equation (SDE) is as follows: in represents another standard Wiener process in reverse time.

[0038] Guided sampling has become a key strategy for developing diffusion models that can generate samples that conform to human instructions. Considering the existing paired data x0~p0(x0|c) and a conditional variable c (the subscript represents the diffusion time, where t=() corresponds to the original data), the conditional probability density is usually regarded as p at time t. t (x t |c). Applying Bayes’ theorem, we can obtain Therefore, a classifier can be trained to predict given noisy data x t The conditional probability p t (c|x t ), and use the scores output by the classifier as a guide. Instead of training a separate classifier to estimate Instead, we propose to use an implicit classifier In view of this, in some embodiments, an unconditional score estimator s can be used θ (x t ) and a conditional score estimator s θ(x t , c) to jointly estimate the target scoring function: where γ is a hyperparameter that controls the strength of the guidance. When γ = 0, it reduces to an unconditional model. When γ = 1, it becomes a purely conditional model. Both models can be trained simultaneously within the same framework, where the unconditional model receives the masked conditioning variable c during training.

[0039] Some example embodiments and corresponding theoretical foundations are now described to better understand the benefits of these embodiments. In some embodiments, the diffusion process for protecting structure prediction may include multiple denoising steps. In any given denoising step in the multiple denoising steps, the target scoring function can be determined based on the first intermediate structure (which is the input structure of the given denoising step) and the composition information. The first intermediate structure may include an initial structure or an output structure from a denoising step before the given denoising step. Specifically, for the first denoising step, the first intermediate structure is the initial structure. For any denoising step after the first denoising step, the first intermediate structure is the output structure from the previous denoising step.

[0040] In some embodiments, the target scoring function may include a translation component and a rotation component. In other words, the structure of the target protein can be represented by a translation component and a rotation component.

[0041] For example, the example structure is represented as x N , which has two components, such as the translation component T N and the rotation component R N Then, x N =[T N , R N As mentioned above, the initial structure is a structure that can be expressed as The structure prediction model 110 generates such an initial structure and performs a diffusion process starting from the initial structure. In a time step (i.e., a given denoising step) i=N, ..., 1, the input structure can be represented by the structure x i It can be expressed as Finally, the structure prediction model 110 determines the target structure through multiple denoising steps.

[0042] In some embodiments, the composition information may include the amino acid sequence of the target protein. A scoring-based model conditioned on the amino acid sequence and a scoring-based model not conditioned on the amino acid sequence may be combined. For example, a target scoring function may include a first scoring function and a second scoring function. A first scoring function may be generated based on a first intermediate structure using a first diffusion model conditioned on the amino acid sequence. A second scoring function may be generated based on the first intermediate structure using a second diffusion model not conditioned on the amino acid sequence. A target scoring function may then be derived based on the first scoring function, the second scoring function, and a first predetermined weight. In this manner, a diffusion model may be introduced for protein structure generation.

[0043] Now refer to Figure 2 , which shows an example diffusion process 200 of a protein structure according to some embodiments of the present disclosure. The first intermediate structure is represented as x i The diffusion model 210 may be a first diffusion model, which may also be referred to as a conditional scoring model. The diffusion model 210 may receive x i And output the scoring function 211 (ie, the first scoring function). The diffusion model 220 may be a second diffusion model, which is also called an unconditional scoring model. The diffusion model 220 may receive x i The target scoring function 202 can be obtained based on at least the scoring function 211 and the scoring function 221. Assuming that the first predetermined weight is γ, the target scoring function 202 can be derived from γ×the scoring function 211 and (1-γ)×the scoring function 221.

[0044] The conditional scoring model and the unconditional scoring model may be collectively referred to as a baseline model, which is described in detail below using examples.

[0045] As an example, the structure prediction model 110 may include an unconditional scoring model (i.e., the second diffusion model) and the serial-conditional model The unconditional model is trained on protein structures without any sequence information and effectively captures the conformational distribution of general proteins. On the other hand, the sequence-conditional scoring model has access to the protein sequence information (denoted as seq) and the corresponding structure x at time t. t .

[0046] Any suitable network architecture can be used to parameterize the corresponding scoring function. For example, the unconditional scoring model can use the sinusoidal embedding of the residue index and the diffusion time t as its single ({s i}) and pairs ({z ij}) embedding, and the conditional scoring model (i.e., the first diffusion model) additionally concatenates its single embedding with a pre-computed representation of the amino acid sequence (e.g., from the folding model). Note that the choice of sequence representation for the conditional scoring model is flexible. Using pre-trained representations from the folding model helps the diffusion model generate reasonable protein structures, while the unconditional scoring model can effectively improve sampling diversity. The two scoring models can be trained using the denoised score matching (DSM) loss function: Where λ(t) is a reweighting function that is inversely proportional to the score norm. During the backward sampling process, a hyperparameter γ (i.e., a first predetermined weight) is used to control the strength of the classifier-free guidance from the conditional score model so that the score function It can be estimated by the following formula: where s θ (x t , t|seq) is the target scoring function, the term is the first scoring function, and the term is the second scoring function. To simplify the notation, the sequence conditional term is omitted in the baseline scoring model, i.e., s θ (x t , t) = s θ (x t , t|seq).

[0047] Continue with the given denoising step. In the given denoising step, after determining the target scoring function, the target scoring function is updated based on the predicted force field for the first intermediate structure. A second intermediate structure is then determined as the output structure of the given step based on the updated target scoring function and the first intermediate structure. If the given denoising step is the last denoising step of the diffusion process, the second intermediate structure is the output of the diffusion process and can be determined as the target structure or used to determine the target structure. If the given denoising step is not the last denoising step, the second intermediate structure is the input structure of the next denoising step.

[0048] Return Reference Figure 2 , the diffusion model 230 may receive x i And output the prediction force field 231 for updating the objective function 202. Based on the updated objective scoring function and the first intermediate structure x i To determine the second intermediate structure x i-1 .

[0049] Starting from the predicted force field, the diffusion model can successfully reweight the generated structures to ensure that they better follow the equilibrium distribution. Figure 3A diagram showing an example generation process of a protein structure according to some embodiments of the present disclosure is shown. In this way, as a force-guided model, the structure prediction model 110 of the present disclosure is generated for multiple conformations of a protein. Using a sequence-based conditional scoring network to guide the unconditional scoring model, the structure prediction model 110 achieves reasonable conformational diversity while ensuring sample quality. On the basis of energy guidance, a novel force-guided sampling strategy is used to estimate the intermediate force function, and then the intermediate force function is embedded in the reverse time sampling process. As shown in the figure above, by combining the sequence-conditional and unconditional scoring models, diverse conformations can be sampled with reasonable quality. As shown in the sample below, by introducing force guidance, structures with lower energy can be generated to better conform to the Boltzmann distribution.

[0050] Some example embodiments regarding force field prediction are described below.

[0051] In some embodiments, force field prediction is performed by determining predicted energies of intermediate structures generated during the diffusion process using a diffusion model for energy prediction (also referred to as a diffusion energy model). A predicted force field for the intermediate structure is derived based on the predicted energies and the intermediate structure.

[0052] Despite the ability of diffusion models to generate diverse structures, these conformations are not always plausible in the sense that they may reside in high-energy regions of the potential energy surface. Due to the limited availability of multi-conformational data within existing protein structure databases, the training data distribution does not conform to an equilibrium distribution, but rather contains only a few data points near the potential energy minimum for each protein sequence. This necessitates the development of generative models that are data-driven and guided by physics-based guidance to generate samples according to a Boltzmann distribution. To this end, energy guidance is introduced, in which a predicted energy function imparts a reward to guide the conformation generation process.

[0053] Given an existing (baseline) diffusion model that generates samples x0~q0(x0), the goal is to sample protein conformations from the equilibrium distribution, in is an intractable normalization constant, ε0 is an energy function (e.g., generated by any suitable molecular simulation tool) that evaluates the potential energy of each generated conformation x0. k is the inverse temperature factor. Given any test function F(x0), p0(x0) gives a more accurate estimate through importance sampling and is defined as

[0054] In the forward diffusion process, the original samples x0~q0(x0) are perturbed according to the SDE described in equation (1), and it is asserted that p t (x t |x0):=qt (x t |x0). What really attracts attention is the marginal distribution p t (x t ). The energy function has the following properties.

[0055] In the first proposal, it is assumed that And for t∈(0,1], p t (x t |x0):=q t (x t |x0). Then, the marginal distribution satisfies where q t (x t ) is based on the marginal distribution of the data, and ε t (x t )satisfy: where ε t (x t ) is called the intermediate energy function, which captures the dynamic changes of energy at various stages of the diffusion process.

[0056] Using the exact formula in Equation (7), the neural network f parameterized by φ φ (x t , t) is used to optimize the problem to approximate the intermediate energy function ε t (x t ),

[0057] Given the optimal satisfy Sampling can then be performed according to a reverse time SDE with a scoring function: where η is a hyperparameter that controls the strength of the guidance, and s θ (x t , t) refers to the baseline scoring model. Note that s θ (x t , t) and intermediate energy network f φ (x t , t) are trained independently.

[0058] If the scoring model only samples protein backbone structures, then all-atom structures can be generated for energy evaluation using suitable molecular simulation tools in any suitable manner. However, there are inherent challenges in normalizing differences in energies calculated by molecular simulation tools. Within limited batch sizes, significant numerical fluctuations have been demonstrated in the estimates. In contrast, the numerical values ​​of atomic forces will be more stable. Therefore, the design of force-guidance strategies tailored to protein conformation generation has proven to be very important. Compared to energy-guidance, force-guidance does not suffer from the energy fluctuation problem and can be directly used as a guide in the reverse sampling process. The following describes how to implement intermediate force-guidance.

[0059] To this end, in some embodiments, force field prediction can be guided by using a diffusion model for force prediction (also referred to as a diffusion force model). In other words, the predicted force field for the intermediate structure generated during the diffusion process can be generated by the diffusion force model.

[0060] In the context of protein conformational modeling, the physics-based energy function ε0(x0) (i.e., the interatomic potential energy) and its gradient have been obtained (i.e., the force on each atom). Compared to the unnormalized potential energy function, the atomic forces are more localized and exhibit better numerical stability, which is also more consistent with the score matching goal. Following the marginal distribution of energy in equation (7), the following properties exist for the force function.

[0061] In the second proposal, suppose p t (x t |x0):=q t (x t |x0) holds true, where 0<t≤1, and the energy function of the intermediate state follows equation (7). The corresponding intermediate force has the following formula: in Represents an implicit distribution The scoring function follows: where η is a hyperparameter that controls the strength of the force guidance.

[0062] The second proposal reveals the exact force at time t, thus advancing the understanding of the intermediate energy function. The intermediate force formula consists of the true potential energy, the diffusion transition kernel, and the marginal scoring function (estimated by the scoring model). When t is close to 0, q t (x0|x t ) converges to the increment function, and the intermediate force Converges to the true value. On the other hand, when t is close to 1, q t (x t |x0) is simplified to qt (x t ), and force disappear.

[0063] In some embodiments, as described above, the objective scoring function can include a translation component and a rotation component. Further, in some embodiments, the translation component can be updated based on the predicted force field and the second predetermined weight, while the rotation component is not updated. In this way, structural stability can be maintained.

[0064] Continue with the given denoising step. In some embodiments, to determine the second intermediate structure as the output structure of the given step, a transition to the first intermediate structure can be determined based on the updated target scoring function and the first intermediate structure. Thus, the second intermediate structure can be derived by applying the transition to the first intermediate structure.

[0065] From SE(3) N Starting with the prior distribution in , an example complete inference process is shown in Algorithm 1 (as shown in Table 1). Scoring Network (i.e., the target scoring function) is a baseline model that mixes sequence conditional scoring and unconditional scoring. During the backward sampling process, the force-guided term h ψ (x t , t) is integrated into the scoring function term at each time step, and its guidance strength is controlled by the hyperparameter η. In order to maintain structural stability, In

[15] , force guidance is applied only to the translational component (i.e., α-carbons). Table 1 Example reasoning process

[0066] As shown in Table 1, in row 1, the initial structure x is obtained. N . Lines 2 to 9 correspond to the diffusion process starting from the initial structure. Specifically, in line 4, based on the conditional scoring function Unconditional scoring function s θ (x i , i, seq) and the first predetermined weight γ, obtain the target score. In the 5th line, obtain the prediction force field H. In the 6th and 7th lines, obtain the transformation and Note that the target scoring function is updated using the predicted force field H In this case, the predicted force field is used to update the transformation component of the target scoring function As mentioned above, this can maintain structural stability. In line 8, the second intermediate structure x is obtained i-1 After multiple denoising steps, the target structure x0 is determined.

[0067] In the above embodiments, the diffusion process is performed by using at least one diffusion model for score estimation (e.g., a conditional score model and an unconditional score model) and another diffusion model for energy prediction or force prediction (e.g., a force prediction model). In some embodiments, one or more score models can be trained separately from the force or energy prediction model.

[0068] In some embodiments, a force or energy prediction model can be trained after one or more scoring models. For example, during the training process of the force or energy prediction model, a scoring function can be generated based on the intermediate structure of the sample protein by using the trained one or more scoring models. A predicted force field can be generated based on the intermediate structure by using the force or energy prediction model. An estimated force field can be derived based on the scoring function, the true structure, and the true energy of the sample protein. The parameters of the force or energy prediction model can then be updated based on the difference between the predicted force field and the estimated force field. For example, the force or energy prediction model can be trained by minimizing a loss function based on the difference between the predicted force field and the estimated force field.

[0069] We now describe an example training procedure. A baseline scoring model is used to generate protein conformations from q0(x0), which are then used to train a separate intermediate force network. That is, in x t Department The tangent space of . Let K be a positive integer representing the size of the training batch. To ensure that the data in the batch follow the same Boltzmann distribution, we first randomly select a protein sequence and then extract K samples from the sequence. The perturbation data x0 is obtained at time t by adding noise to the sampled data after the SDE in equation (1). Make Intermediate force loss function is defined as: in The latter component of equation (11b) represents the exact value of the intermediate force at time t, where is the s in the baseline θ (x t , t) estimates an intractable scoring function, and q t (x t |x0) is a Gaussian distribution that can be handled. It is not calculated over the entire generated dataset. Instead, we use K samples in a batch to estimate the expectation. The force-guided training process is summarized in Algorithm 2 (as shown in Table 2). Table 2 Example training process

[0070] As shown in Table 2, in the third row, the intermediate structure x for time step t is obtained t Then, based on the intermediate structure x t Generate scoring function s θ (x t , t). In lines 5 to 7, the diffusion model for force prediction is used to determine the predicted force field h at time t. ψ (x t , t). Based on the scoring function s θ (x t , t), the true value structure x0 and the true value energy ε0(x0) to obtain the estimated force field In line 8, you can base your To update the parameters of the other diffusion model, the force prediction model is trained in this way.

[0071] Considering that the exact intermediate force formula in Eq. (10) is equal to the MD force field at t = 0 And when t=1, it becomes 0, the network h ψ (x t , t) is constructed as the following interpolation form: where g ψ (x t , t) is a neural network that estimates the intermediate terms within the interpolation conformation. This ensures boundary conditions at t∈{0,1} and has been empirically found to reduce the variance during training.

[0072] In some embodiments, the diffusion process described above can be one of multiple diffusion processes used for structure prediction on a target protein. These diffusion processes can have different initial structures at the beginning of the denoising step. Based on the respective target structures generated by the multiple diffusion processes, a distribution of protein energies related to the protein energy of the target protein can be constructed. In this way, a protein conformational landscape can be constructed and one or more stable protein structures can be determined.

[0073] Finally, embodiments of this disclosure introduce a diffusion model for protein structure prediction. By blending a sequence-conditional scoring network with an unconditional model, a balance between sampling quality and diversity is achieved through classifier-free guidance. Under this baseline model, novel, theoretically guaranteed, physics-based energy and force guidance strategies are proposed that effectively guide the diffusion sampler to generate low-energy conformations that better conform to the underlying Boltzmann distribution. Example Procedure and Equipment

[0074] Figure 44 shows a flow chart of a method 400 for protein structure prediction according to some example embodiments of the present disclosure. The method 400 may be used in Figure 1 The electronic device 120 shown in FIG. At block 410 , an initial structure of the target protein is obtained based on composition information related to the amino acids of the target protein for the diffusion process. At block 420 , a diffusion process is performed on the initial structure based on the composition information and a force field prediction for the target protein. At block 430 , a target structure of the target protein is determined based on the results of the diffusion process.

[0075] In some embodiments, the diffusion process includes multiple denoising steps, and performing a given denoising step among the multiple denoising steps includes: determining a target scoring function based on a first intermediate structure and composition information, wherein the first intermediate structure includes an initial structure or an output structure from a denoising step before the given denoising step; updating the target scoring function based on a predicted force field for the first intermediate structure; and determining a second intermediate structure as the output structure of the given step based on the updated target scoring function and the first intermediate structure.

[0076] In some embodiments, the composition information includes the amino acid sequence of the target protein, and based on the first intermediate structure and the composition information, determining the target scoring function includes: generating a first scoring function based on the first intermediate structure using a first diffusion model conditioned on the amino acid sequence; generating a second scoring function based on the first intermediate structure using a second diffusion model not conditioned on the amino acid sequence; and deriving the target scoring function based on the first scoring function, the second scoring function and a first predetermined weight.

[0077] In some embodiments, the target scoring function includes a translation component and a rotation component, and updating the target scoring function includes updating the translation component based on the predicted force field and the second predetermined weight.

[0078] In some embodiments, determining a second intermediate structure as an output structure for a given step based on the updated target scoring function and the first intermediate structure includes: determining a transition to the first intermediate structure based on the updated target scoring function and the first intermediate structure; and deriving the second intermediate structure by applying the transition to the first intermediate structure.

[0079] In some embodiments, force field prediction is performed by: determining predicted energies for intermediate structures generated during the diffusion process using a diffusion model for energy prediction; and deriving predicted force fields for the intermediate structures based on the predicted energies and the intermediate structures.

[0080] In some embodiments, force field prediction is performed by determining a predicted force field for intermediate structures generated during the diffusion process using a diffusion model for force prediction.

[0081] In some embodiments, the diffusion process is one of multiple diffusion processes used to predict the structure of the target protein, and the method further includes: constructing a distribution of protein energy with respect to the protein structure of the target protein based on the corresponding target structures generated by the multiple diffusion processes.

[0082] In some embodiments, the diffusion process is performed using at least one diffusion model for score estimation and another diffusion model for energy prediction or force prediction, and the another diffusion model is trained in the following manner: using the trained at least one diffusion model, generating a scoring function based on the intermediate structure of the sample protein; using the other diffusion model, determining a predicted force field based on the intermediate structure; deriving an estimated force field based on the scoring function, the true structure and the true energy of the sample protein; and updating parameters of the another diffusion model based on the difference between the predicted force field and the estimated force field.

[0083] In some embodiments of the present disclosure, a non-transient computer program product is provided. The non-transient computer program product includes a computer-readable storage medium. The computer-readable storage medium has program instructions implemented therein. The program instructions can be executed by an electronic device to cause the electronic device to perform a method for protein structure prediction. The method includes: obtaining an initial structure of a target protein based on composition information related to the amino acids of the target protein for a diffusion process; performing a diffusion process on the initial structure based on the composition information and a force field prediction for the target protein; and determining a target structure of the target protein based on the results of the diffusion process. In some embodiments of the present disclosure, the method also includes other steps described in the present disclosure.

[0084] Figure 5 FIG. 5 is a block diagram of an electronic device 500 capable of implementing various embodiments of the present disclosure. It should be understood that Figure 5 The electronic device 500 shown is for illustration purposes only and does not imply any limitation on the functionality and scope of the present disclosure in any way. The electronic device 500 can be used to implement the above method 500. Figure 5 As shown, electronic device 500 may be a general electronic device and may include at least one or more processors or processing units 510 , a memory 520 , a storage unit 530 , one or more communication units 540 , one or more input devices 550 , and one or more output devices 560 .

[0085] Processing unit 510 may be a physical or virtual processor and may implement various processes based on program 525 stored in memory 520. In a multi-processor system, multiple processing units execute computer-executable instructions in parallel to increase the parallel processing capabilities of electronic device 500. Processing unit 510 may also be referred to as a central processing unit (CPU), a microprocessor, a controller, or a microcontroller.

[0086] The electronic device 500 typically includes various computer storage media. Such media can be any media accessible to the electronic device 500, including but not limited to volatile and non-volatile media, or removable and non-removable media. The memory 520 can be a volatile memory (e.g., a register, a cache, a random access memory (RAM)), a non-volatile memory (such as a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM) or flash memory) or any combination thereof. The storage unit 530 can be any removable or non-removable medium and can include machine-readable media such as memory, a flash drive, a disk or other media, which can be used to store information and / or data and can be accessed in the electronic device 500.

[0087] The electronic device 500 may further include additional removable / non-removable, volatile / non-volatile memory media. Figure 5 Although not shown in the figure, a magnetic disk drive for reading from and / or writing to a removable, non-volatile magnetic disk and an optical disk drive for reading from and / or writing to a removable, non-volatile optical disk may be provided. In this case, each drive may be connected to the bus (not shown) via one or more data medium interfaces.

[0088] The communication unit 540 communicates with another electronic device via a communication medium. In addition, the functions of the components in the electronic device 500 can be implemented in a single computing cluster or multiple computing machines that communicate via a communication connection. Thus, the electronic device 500 can operate in a networked environment using logical connections to one or more other servers, networked personal computers (PCs), or another general network node.

[0089] Input device 550 may be one or more of various input devices, such as a mouse, keyboard, trackball, voice input device, etc. Output device 560 may be one or more of various output devices, such as a display, speaker, printer, etc. With the aid of communication unit 540, electronic device 500 may also communicate with one or more external devices (not shown), such as storage devices and display devices, one or more of which enable a user to interact with electronic device 500, or any device (such as a network card, modem, etc.) that enables electronic device 500 to communicate with one or more other electronic devices, if desired. Such communication may be performed via an input / output (I / O) interface (not shown).

[0090] In some embodiments, instead of being integrated into a single device, some or all components of the electronic device 500 may also be arranged in a cloud computing architecture. In a cloud computing architecture, components may be provided remotely and work together to implement the functionality described in this disclosure. In some embodiments, cloud computing provides computing, software, data access, and storage services, which does not require the end user to know the physical location or configuration of the systems or hardware providing these services. In various embodiments, cloud computing uses appropriate protocols to provide services via a wide area network (such as the Internet). For example, a cloud computing provider provides an application via a wide area network that can be accessed through a web browser or any other computing component. The software or components of the cloud computing architecture and the corresponding data may be stored in a remote location. The computing resources in a cloud computing environment may be consolidated or distributed across locations in a remote data center. Cloud computing infrastructure can provide services through shared data centers, although they act as a single access point for users. Therefore, cloud computing architecture can be used to provide the components and functionality described herein from a service provider in a remote location. Alternatively, they can be provided from a traditional server or installed directly or otherwise on a client device.

[0091] The functions described herein may be performed, at least in part, by one or more hardware logic components. For example, but not limited to, illustrative types of hardware logic components that may be used include field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0092] The program code for executing the methods of the subject matter described herein can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0093] In the context of the present disclosure, a machine-readable medium may be any tangible medium that can contain or store a program used by or in conjunction with an instruction execution system, device, or apparatus. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of machine-readable storage media would include an electrical connection with one or more wires, a portable computer floppy disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0094] In addition, although operations are shown in a particular order, this should not be understood as requiring such operations to be performed in the particular order shown or in a continuous order, or to perform all operations shown, to achieve desired results. In some cases, multitasking and parallel processing may be advantageous. Equally, although some specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of the subject matter described herein, but should be interpreted as describing features that can be specific to a particular embodiment. Some features described in the context of a separate embodiment also can be combined in a single implementation. On the contrary, the various features described in a single implementation also can be realized individually or with any suitable subcombination in a plurality of embodiments.

[0095] Although the subject matter has been described in language specific to structural features and / or methodological acts, it is to be understood that the subject matter specified in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing the claims.

[0096] From the foregoing, it should be understood that specific embodiments of the presently disclosed technology have been described herein for illustrative purposes, but various modifications may be made without departing from the scope of the present disclosure. Accordingly, the technology of the present disclosure is not to be limited, except as in the appended claims.

[0097] The embodiments of the subject matter and functional operations described in this disclosure can be implemented in various systems, digital electronic circuits or computer software, firmware or hardware (including the structures disclosed in this specification and their structural equivalents) or a combination of one or more of them. The embodiments of the subject matter described in this specification can be implemented as one or more computer program products, that is, one or more modules of computer program instructions encoded on a tangible and non-transitory computer-readable medium for execution by a data processing device or for controlling the operation of the data processing device. The computer-readable medium can be a machine-readable storage device, a machine-readable storage substrate, a memory device, a material combination that affects a machine-readable propagation signal, or a combination of one or more of them. The term "data processing unit" or "data processing device" covers all devices, equipment and machines for processing data, including, for example, a programmable processor, a computer or multiple processors or computers. In addition to hardware, the device can also include code that creates an execution environment for the computer program in question, for example, code that constitutes processor firmware, a protocol stack, a database management system, an operating system or a combination of one or more of them.

[0098] A computer program (also referred to as a program, software, software application, script, or code) can be written in any form of programming language, including compiled or interpreted languages, and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for a computing environment. A computer program does not necessarily correspond to a file in a file system. A program may be stored as part of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document), in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, subroutines, or code portions). A computer program may be deployed to execute on one computer or multiple computers located at one site or distributed across multiple sites and interconnected by a communications network.

[0099] For example, processors suitable for executing computer programs include general-purpose and special-purpose microprocessors, and any one or more processors of any type of digital computer. Typically, the processor will receive instructions and data from a read-only memory or a random access memory or both. The basic elements of a computer are a processor for executing instructions and one or more memory devices for storing instructions and data. Typically, a computer will also include or be operably coupled to receive data from or transmit data to or to one or more large-capacity storage devices (e.g., magnetic disks, magneto-optical disks, or optical disks) for storing data. However, a computer need not have such devices. Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media, and memory devices, including, for example, semiconductor memory devices, such as EPROM, EEPROM, and flash memory devices. The processor and memory can be supplemented by or incorporated into a dedicated logic circuit system.

[0100] The specification, together with the drawings, are intended to be considered as exemplary only, wherein exemplary means example. As used herein, the use of "or" is intended to include "and / or" unless the context clearly dictates otherwise.

[0101] Although this disclosure contains many details, these details should not be interpreted as limitations on the scope of any disclosure or what may be claimed, but rather as descriptions of features that may be specific to a particular embodiment of a particular disclosure. Certain features described in this disclosure in the context of separate embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented in multiple embodiments individually or in any suitable subcombination. Furthermore, although features may be described above as functioning in certain combinations and even initially claimed as such, in some cases, one or more features from a claimed combination may be removed from the combination, and a claimed combination may involve a subcombination or a variation of a subcombination.

[0102] Similarly, although operations are shown in a particular order in the accompanying drawings, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, in order to achieve the desired results. Furthermore, the separation of various system components in the embodiments described in this disclosure should not be understood as requiring such separation in all embodiments. Only a few embodiments and examples are described, and other embodiments, enhancements, and variations can be made based on what is described and illustrated in this disclosure.

Claims

1. A method for protein structure prediction, comprising: For the diffusion process, based on composition information related to amino acids of the target protein, an initial structure of the target protein is obtained; performing the diffusion process on the initial structure based on the composition information and the force field prediction for the target protein; as well as Based on the results of the diffusion process, the target structure of the target protein is determined.

2. The method of claim 1 , wherein the diffusion process comprises a plurality of denoising steps, and performing a given denoising step of the plurality of denoising steps comprises: determining a target scoring function based on a first intermediate structure and the composition information, wherein the first intermediate structure comprises the initial structure or an output structure from a denoising step prior to the given denoising step; updating the objective scoring function based on the predicted force field for the first intermediate structure; as well as Based on the updated target scoring function and the first intermediate structure, a second intermediate structure is determined as an output structure of the given step.

3. The method of claim 2, wherein the composition information comprises an amino acid sequence of the target protein, and based on the first intermediate structure and the composition information, determining a target scoring function comprises: generating a first scoring function based on the first intermediate structure using a first diffusion model conditioned on the amino acid sequence; generating a second scoring function based on the first intermediate structure using a second diffusion model that is not conditioned on the amino acid sequence; as well as The target scoring function is derived based on the first scoring function, the second scoring function and a first predetermined weight.

4. The method of claim 2, wherein the target scoring function includes a translation component and a rotation component, and updating the target scoring function includes: Based on the predicted force field and a second predetermined weight, the translation component is updated.

5. The method according to claim 2, wherein determining the second intermediate structure as the output structure of the given step based on the updated target scoring function and the first intermediate structure comprises: determining a transition to the first intermediate structure based on the updated target scoring function and the first intermediate structure; as well as By applying the transformation to the first intermediate structure, the second intermediate structure is obtained.

6. The method according to claim 1, wherein the force field prediction is performed by: Determining predicted energies for intermediate structures generated during the diffusion process using a diffusion model for energy prediction; and Based on the predicted energy and the intermediate structure, a predicted force field for the intermediate structure is derived.

7. The method according to claim 1, wherein the force field prediction is performed by: Using a diffusion model for force prediction, a predicted force field is determined for the intermediate structures generated during the diffusion process.

8. The method according to claim 1, wherein the diffusion process is one of a plurality of diffusion processes used for structural prediction of the target protein, and the method further comprises: Based on the corresponding target structures generated by the plurality of diffusion processes, a distribution of protein energy with respect to the protein structure of the target protein is constructed.

9. The method of claim 1 , wherein the diffusion process is performed using at least one diffusion model for score estimation and another diffusion model for energy prediction or force prediction, and the another diffusion model is trained by: generating a scoring function based on the intermediate structure of the sample protein using the at least one trained diffusion model; determining a prediction force field based on the intermediate structure using the other diffusion model; deriving an estimated force field based on the scoring function, the true structure and the true energy of the sample protein; as well as Parameters of the other diffusion model are updated based on the difference between the predicted force field and the estimated force field.

10. An electronic device comprising a computer processor coupled to a computer readable memory unit, the memory unit comprising instructions that, when executed by the computer processor, implement the actions of protein structure prediction, the actions comprising: For the diffusion process, based on composition information related to amino acids of the target protein, an initial structure of the target protein is obtained; performing the diffusion process on the initial structure based on the composition information and the force field prediction for the target protein; as well as Based on the results of the diffusion process, the target structure of the target protein is determined.

11. The apparatus of claim 10, wherein the diffusion process comprises a plurality of denoising steps, and performing a given denoising step of the plurality of denoising steps comprises: determining a target scoring function based on a first intermediate structure and the composition information, wherein the first intermediate structure comprises the initial structure or an output structure from a denoising step prior to the given denoising step; updating the objective scoring function based on the predicted force field for the first intermediate structure; as well as Based on the updated target scoring function and the first intermediate structure, a second intermediate structure is determined as an output structure of the given step.

12. The apparatus of claim 11, wherein the composition information comprises an amino acid sequence of the target protein, and based on the first intermediate structure and the composition information, determining a target scoring function comprises: generating a first scoring function based on the first intermediate structure using a first diffusion model conditioned on the amino acid sequence; generating a second scoring function based on the first intermediate structure using a second diffusion model that is not conditioned on the amino acid sequence; as well as The target scoring function is derived based on the first scoring function, the second scoring function and a first predetermined weight.

13. The apparatus of claim 11, wherein the target scoring function comprises a translation component and a rotation component, and updating the target scoring function comprises: Based on the predicted force field and a second predetermined weight, the translation component is updated.

14. The apparatus of claim 11, wherein determining the second intermediate structure as the output structure of the given step based on the updated target scoring function and the first intermediate structure comprises: determining a transition to the first intermediate structure based on the updated target scoring function and the first intermediate structure; as well as By applying the transformation to the first intermediate structure, the second intermediate structure is obtained.

15. The apparatus of claim 10, wherein the force field prediction is performed by: Determining predicted energies for intermediate structures generated during the diffusion process using a diffusion model for energy prediction; and Based on the predicted energy and the intermediate structure, a predicted force field for the intermediate structure is derived.

16. The apparatus of claim 10, wherein the force field prediction is performed by: Using a diffusion model for force prediction, a predicted force field is determined for the intermediate structures generated during the diffusion process.

17. The apparatus according to claim 10, wherein the diffusion process is one of a plurality of diffusion processes for performing structure prediction on the target protein, and the action further comprises: Based on the corresponding target structures generated by the plurality of diffusion processes, a distribution of protein energy with respect to the protein structure of the target protein is constructed.

18. The apparatus of claim 10, wherein the diffusion process is performed using at least one diffusion model for score estimation and another diffusion model for energy prediction or force prediction, and the another diffusion model is trained by: generating a scoring function based on the intermediate structure of the sample protein using the at least one trained diffusion model; determining a prediction force field based on the intermediate structure using the other diffusion model; deriving an estimated force field based on the scoring function, the true structure and the true energy of the sample protein; as well as Parameters of the other diffusion model are updated based on the difference between the predicted force field and the estimated force field.

19. A computer program product, comprising a non-transitory computer-readable storage medium having program instructions implemented therein, the program instructions being executable by an electronic device to cause the electronic device to perform actions for protein structure prediction, the actions comprising: For the diffusion process, based on composition information related to amino acids of the target protein, an initial structure of the target protein is obtained; performing the diffusion process on the initial structure based on the composition information and the force field prediction for the target protein; as well as Based on the results of the diffusion process, the target structure of the target protein is determined.

20. The computer program product of claim 19, wherein the diffusion process comprises a plurality of denoising steps, and performing a given denoising step of the plurality of denoising steps comprises: determining a target scoring function based on a first intermediate structure and the composition information, wherein the first intermediate structure comprises the initial structure or an output structure from a denoising step prior to the given denoising step; updating the objective scoring function based on the predicted force field for the first intermediate structure; as well as Based on the updated target scoring function and the first intermediate structure, a second intermediate structure is determined as an output structure of the given step.

Images