Protein multi-conformation model training method and protein three-dimensional structure prediction method

Abstract

The invention provides a training method of a protein multi-conformation model and a prediction method of a protein three-dimensional structure. The training method comprises the following steps: acquiring a plurality of initial protein structure prediction models and a plurality of single conformation samples; training each initial protein structure prediction model based on the plurality of single conformation samples to obtain a plurality of teacher protein structure prediction models corresponding to the plurality of initial protein structure prediction models; obtaining a to-be-treated amino acid sequence; respectively predicting the three-dimensional structure of the amino acid sequence to be processed based on the plurality of teacher protein structure prediction models to obtain a plurality of protein three-dimensional structures respectively corresponding to the plurality of teacher protein structure prediction models; forming a multi-conformation sample by the amino acid sequence to be treated and each protein three-dimensional structure in the plurality of protein three-dimensional structures; and training a to-be-trained protein multi-conformation model based on the multi-conformation sample. Therefore, a multi-conformation sample can be generated, and the multi-conformation prediction capability of the model is improved.

Description

Technical Field

[0001] The present disclosure relates to the technical field of protein prediction, and in particular to a method for training a protein multi-conformation model, a method for predicting a protein three-dimensional structure, and a method for determining a multi-conformation sample. Background Art

[0002] Antibodies play a vital role in defending against pathogens by recognizing and binding to specific foreign substances. Due to the flexibility of the complementarity determining region (CDR) loops, these antibodies often have multiple conformations, which means that there are multiple three-dimensional structures for a given amino acid sequence. Although regression-based methods like AlphaFold2 have made significant progress in protein structure prediction, they are limited to predicting a single three-dimensional structure for a given amino acid sequence, and cannot predict multiple conformations. Another problem that needs to be faced when training multi-conformational structure predictions that can predict multiple conformations is that there is less multi-conformational data in existing common protein databases or antibody databases, which is not enough to train a protein multi-conformational prediction model with sufficient accuracy. Summary of the invention

[0003] The purpose of the embodiments of the present disclosure is to provide a method and device for training and predicting a protein multi-conformation model, so as to alleviate the problem that there is insufficient multi-conformation data to train a protein multi-conformation prediction model with sufficient accuracy.

[0004] In a first aspect, an embodiment of the present disclosure provides a method for training a protein multi-conformation model, wherein the protein multi-conformation model is used to predict multiple structural conformations of a given amino acid sequence; the method comprises: obtaining multiple initial protein structure prediction models and multiple single conformation samples, each initial protein structure prediction model corresponding to different initial parameters, and each single conformation sample comprising at least one amino acid sequence and a corresponding three-dimensional protein structure; training each initial protein structure prediction model based on multiple single conformation samples to obtain multiple teacher protein structure prediction models corresponding to the multiple initial protein structure prediction models; obtaining an amino acid sequence to be processed; predicting the three-dimensional structure of the amino acid sequence to be processed based on multiple teacher protein structure prediction models to obtain multiple three-dimensional protein structures corresponding to the multiple teacher protein structure prediction models; combining the amino acid sequence to be processed and each of the multiple three-dimensional protein structures into a multi-conformation sample, so that for one multi-conformation sample, there is another sample having the same amino acid sequence as the multi-conformation sample and a different protein three-dimensional structure; training the protein multi-conformation model to be trained based on the multi-conformation samples.

[0005] In a second aspect, an embodiment of the present disclosure provides a method for predicting a three-dimensional structure of a protein, using the protein multi-conformation prediction model obtained by training in the first aspect to predict multiple structural conformations of the protein; the method comprises: obtaining an amino acid sequence of the protein to be predicted; randomly generating multiple different noise backbones; for each of the noise backbones, in at least one diffusion cycle, using the protein multi-conformation model obtained by training, using the embedding representation corresponding to the current diffusion cycle to update the backbone initial value corresponding to the current diffusion cycle, and obtaining a predicted clean backbone; determining the predicted three-dimensional structure corresponding to the noise backbone according to the predicted clean backbone obtained in the last diffusion cycle of at least one diffusion cycle; wherein, in the first diffusion cycle of at least one diffusion cycle, the noise backbone is used as the backbone initial value corresponding to the current diffusion cycle; in the subsequent diffusion cycles of at least one diffusion cycle, the backbone initial value corresponding to the current diffusion cycle is determined by a diffusion module according to the backbone initial value corresponding to the previous diffusion cycle and the predicted clean backbone obtained in the previous diffusion cycle; the embedding representation corresponding to the current diffusion cycle is determined according to the amino acid sequence of the protein to be predicted; and the predicted three-dimensional structure corresponding to each noise backbone is used as the multiple structural conformations of the protein to be predicted.

[0006] In a third aspect, the present disclosure provides a method for determining a multi-conformation sample, wherein the multi-conformation sample is used to train a protein multi-conformation prediction model, and the protein multi-conformation model is used to predict multiple structural conformations of a given amino acid sequence. The method comprises: obtaining multiple initial protein structure prediction models and multiple single conformation samples, each initial protein structure prediction model corresponding to different initial parameters, and each single conformation sample comprising at least one amino acid sequence and a corresponding three-dimensional protein structure; training each initial protein structure prediction model based on the multiple single conformation samples to obtain multiple teacher protein structure prediction models corresponding to the multiple initial protein structure prediction models; obtaining an amino acid sequence to be processed; predicting the three-dimensional structure of the amino acid sequence to be processed based on the multiple teacher protein structure prediction models to obtain multiple three-dimensional protein structures corresponding to the multiple teacher protein structure prediction models; and combining the amino acid sequence to be processed and each of the multiple three-dimensional protein structures into a multi-conformation sample, so that for one multi-conformation sample, there is another sample having the same amino acid sequence as the multi-conformation sample and a different three-dimensional protein structure.

[0007] In a fourth aspect, the present disclosure provides an electronic device, comprising: a processor and a storage device; the storage device stores a computer program, and when the computer program is executed by the processor, the method as described in any of the above aspects is executed.

[0008] In a fifth aspect, the present disclosure provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the method of any of the aforementioned aspects.

[0009] In a sixth aspect, the present disclosure provides a computer program product, comprising computer instructions, wherein the computer instructions, when executed by a processor, implement the steps of the method of any of the aforementioned aspects.

[0010] The embodiments of the present disclosure provide a method for training a protein multi-conformation model, a method for predicting a protein three-dimensional structure, and a method for determining a multi-conformation sample. The method comprises the following steps: obtaining a plurality of initial protein structure prediction models and a plurality of single conformation samples, training each initial protein structure prediction model based on the plurality of single conformation samples, and obtaining a plurality of teacher protein structure prediction models corresponding one-to-one to the plurality of initial protein structure prediction models; then, predicting the three-dimensional structure of an amino acid sequence to be processed based on the plurality of teacher protein structure prediction models, and obtaining a plurality of protein three-dimensional structures corresponding one-to-one to the plurality of teacher protein structure prediction models; combining the amino acid sequence to be processed and each of the plurality of protein three-dimensional structures to form a multi-conformation sample; and using richer and more comprehensive multi-conformation samples to train the protein multi-conformation model to be trained can enhance the prediction capability of the protein multi-conformation model, so that it can predict multiple structural conformations of a given amino acid sequence. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings required for use in the embodiments of the present disclosure will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present disclosure and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0012] Figure 1 A flowchart of a method for training a protein multi-conformation model provided in an embodiment of the present disclosure;

[0013] Figure 2A An example of a method for training a protein multi-conformation model provided in an embodiment of the present disclosure;

[0014] Figure 2B A schematic diagram of a process for updating the backbone initial value corresponding to the current diffusion cycle using a structure module and superimposed features and iterative paired features provided in an embodiment of the present disclosure.

[0015] Figure 3 A flowchart of a method for predicting a protein multi-conformation model provided in an embodiment of the present disclosure;

[0016] Figure 4 A flow chart of a method for determining a protein conformation sample provided by an embodiment of the present disclosure;

[0017] Figure 5 A schematic diagram of the structure of an electronic device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present disclosure will be described below in conjunction with the accompanying drawings in the embodiments of the present disclosure.

[0019] Figure 1 A flowchart of a method 100 for training a protein multi-conformation model provided in an embodiment of the present disclosure. The protein multi-conformation model is used to predict multiple structural conformations of a given amino acid sequence; Figure 1 As shown, the method includes:

[0020] S110, obtaining a plurality of initial protein structure prediction models and a plurality of single conformation samples, each initial protein structure prediction model corresponding to a different initial parameter, each single conformation sample including at least one amino acid sequence and a corresponding protein three-dimensional structure.

[0021] The architecture of the initial protein structure prediction model here can be an existing general protein structure prediction model or antibody structure prediction model (when the amino acid sequence is the amino acid sequence of an antibody) similar to AlphaFold2 or xTrimoABfold. Exemplarily, the architecture of each initial protein structure prediction model can be the same, the architecture of the initial protein structure prediction model and the architecture of the protein multi-conformation can be the same, and the initial parameters corresponding to each initial protein structure prediction model are different. It is understandable that the model parameters of the teacher protein structure prediction model obtained by training different initial protein structure prediction models are different, and the three-dimensional protein structures predicted by different teacher protein structure prediction models based on a given amino acid sequence are also different. In this way, a plurality of different three-dimensional protein structures corresponding to a given amino acid sequence can be obtained.

[0022] The single conformation sample here can be data from PDB (Protein Data Bank) and / or OAS (Observed Antibody Space database). The three-dimensional structure of the protein here can include the coordinates (and rotation angles) of the backbone and side chain atoms. Exemplarily, a single conformation sample can include an amino acid sequence and a three-dimensional protein structure corresponding to the amino acid sequence. Most of the existing protein structure databases are single conformation samples, and there are not enough multi-conformation samples that can be used to train protein multi-conformation models, and multi-conformation samples of antibodies are even rarer. In the case of insufficient samples, it is difficult to train a protein multi-conformation model that can accurately predict multiple structural conformations of the same amino acid sequence.

[0023] S120, training each initial protein structure prediction model based on multiple single conformation samples to obtain multiple teacher protein structure prediction models corresponding to the multiple initial protein structure prediction models.

[0024] Among them, the amino acid sequence included in the single conformation sample is used as the input data for training, and the protein three-dimensional structure included in the single conformation sample can be used as a label. Based on the difference between the output data of the initial protein structure prediction model (predicted protein three-dimensional structure) and the label, the parameters of the initial protein structure prediction model are optimized until the training termination condition is met (the training termination condition here can be that the loss reaches the expected number, or the number of training times reaches the preset number), and the teacher protein structure prediction model is obtained.

[0025] It is understandable that each initial protein structure prediction model corresponds to a trained teacher protein structure prediction model. Even if the model architectures of the initial protein structure prediction models are the same, due to their different initial parameters, the parameters of the trained protein structure prediction models corresponding to the initial protein structure prediction models are different.

[0026] S130, obtaining an amino acid sequence to be processed;

[0027] It is understandable that if the protein multi-conformation model is expected to predict multiple structural conformations of an antibody, the amino acid sequence to be processed is the amino acid sequence of the antibody; if the protein multi-conformation model is expected to predict multiple structural conformations of a general protein, the amino acid sequence to be processed is the amino acid sequence of a general protein. Accordingly, the initial protein structure prediction model in step 110 is an antibody structure prediction model or a general protein structure prediction model.

[0028] The amino acid sequence to be processed here can be an amino acid sequence in a single conformation sample. The protein three-dimensional structure corresponding to the amino acid sequence to be processed here can be unknown or known. The known protein three-dimensional structure corresponding to the amino acid sequence to be processed can be one or more, but the number of known protein three-dimensional structures corresponding to the amino acid sequence to be processed should be less than the number of required protein three-dimensional structures. For example, there are 2 known protein three-dimensional structures corresponding to the amino acid sequence to be processed, but in the training data used to train the protein multi-conformation model, one amino acid sequence is required to correspond to 10 different protein three-dimensional structures, so it is necessary to continue to generate 8 protein three-dimensional structures that are different from the known 2 protein three-dimensional structures through subsequent steps.

[0029] S140, predicting the three-dimensional structure of the amino acid sequence to be processed based on multiple teacher protein structure prediction models, respectively, to obtain multiple protein three-dimensional structures corresponding to the multiple teacher protein structure prediction models;

[0030] Through this step, multiple protein three-dimensional structures can be predicted for the same amino acid sequence to be processed, thereby providing enough multi-conformation samples as training data for the protein multi-conformation model.

[0031] For example, the three-dimensional structure of the amino acid sequence to be processed is unknown, and 10 teacher protein structure prediction models are used to predict the three-dimensional structure of the amino acid sequence to be processed, and 10 different protein three-dimensional structures are obtained. For another example, there are two known three-dimensional structures of the amino acid sequence to be processed, and 10 teacher protein structure prediction models are used to predict the three-dimensional structure of the amino acid sequence to be processed, and 10 different protein three-dimensional structures are obtained, and the two known three-dimensional structures constitute 12 different protein three-dimensional structures.

[0032] S150, respectively composing a multi-conformation sample with the amino acid sequence to be processed and each of the multiple protein three-dimensional structures, so that for one multi-conformation sample, there exists another sample having the same amino acid sequence as the multi-conformation sample and a different protein three-dimensional structure;

[0033] Continuing with the previous example, the amino acid sequence to be processed and each of the 10 protein 3D structures are respectively combined into a multi-conformation sample, so that 10 multi-conformation samples can be obtained. A multi-conformation sample includes an amino acid sequence and a protein 3D structure, and multiple multi-conformation samples with the same amino acid sequence can be regarded as a group of multi-conformation samples.

[0034] The difference between multi-conformation samples and single-conformation samples is that for a multi-conformation sample A, another multi-conformation sample B can be found, and multi-conformation sample A and multi-conformation sample B include the same amino acid sequence and different protein three-dimensional structures. For a single-conformation sample C, another single-conformation sample D cannot be found, so that single-conformation sample C and single-conformation sample D include the same amino acid sequence and different protein three-dimensional structures.

[0035] S160, training the multi-conformation model of the protein to be trained based on the multi-conformation samples.

[0036] Exemplarily, the architecture of the protein multi-conformation model can be the same as the architecture of the initial protein structure prediction model.

[0037] After determining the multi-conformation samples, the multi-conformation samples can be used to train the multi-conformation model of the protein to be trained, so that the multi-conformation model of the protein can learn that one amino acid sequence can correspond to multiple different protein three-dimensional structures, so as to enhance the ability of the multi-conformation model of the protein to predict multiple structural conformations of a given amino acid sequence.

[0038] It is understandable that the multi-conformation samples used for training the protein multi-conformation model include not only the multi-conformation samples generated by step S150, but also the multi-conformation samples measured by experimental means.

[0039] The disclosed embodiments provide a method for training a protein multi-conformation model, which uses a single conformation sample to train multiple different teacher protein structure prediction models, and uses multiple teacher protein structure prediction models to predict the three-dimensional structure of a to-be-processed amino acid sequence, so as to obtain multiple three-dimensional structures of the to-be-processed amino acid sequence; the to-be-processed amino acid sequence and each of the multiple protein three-dimensional structures form a multi-conformation sample; thus, the training data used to train the protein multi-conformation model is greatly enriched. Using richer and more comprehensive multi-conformation samples to train the protein multi-conformation model to be trained can enhance the prediction ability of the protein multi-conformation model, so that it can predict multiple structural conformations of a given amino acid sequence.

[0040] In some embodiments, the above step S160 includes the following steps, as shown below: Figure 2A Provide explanation.

[0041] Step 1), select a multi-conformation sample as the current sample;

[0042] It is understandable that existing protein structure databases such as PDB also contain a small number of multi-conformation samples. The data for training the protein multi-conformation model can come from the multi-conformation samples obtained in step 150, and on this basis can also include multi-conformation samples from existing protein structure databases.

[0043] Step 2), in each diffusion cycle of at least one diffusion cycle, using the protein multi-conformation model to be trained, using the embedding representation corresponding to the current diffusion cycle to update the initial value of the backbone corresponding to the current diffusion cycle, and obtain a predicted clean backbone;

[0044] Exemplarily, the architecture of the protein multi-conformation model can be shaped like the architecture of Alphafold. In Alphafold, multiple cycles are performed through a recycling mechanism. In each cycle, the initial value of the backbone is updated with the embedded representation corresponding to the current cycle to obtain the predicted backbone. Among them, the initial value of the backbone in each cycle is the origin, that is, the atomic coordinates are 0, and the rotation angle is the unit matrix. Different from this, the initial value of the backbone used in each diffusion cycle in the embodiment of the present disclosure is different, and the initial value of the backbone corresponding to the current diffusion cycle is generated by the diffusion module.

[0045] Exemplarily, in the first diffusion cycle of at least one diffusion cycle, the backbone initial value corresponding to the current diffusion cycle is generated by adding noise to the three-dimensional structure of the protein in the current sample through the diffusion module; Exemplarily, in the subsequent diffusion cycles other than the first diffusion cycle in at least one diffusion cycle, the backbone initial value corresponding to the current diffusion cycle is obtained by removing at least part of the noise from the backbone initial value corresponding to the previous diffusion cycle through the diffusion module. That is, in the first diffusion cycle, noise is gradually added to the three-dimensional structure of the protein in the current sample through the diffusion module, and the backbone with the most noise added is used as the backbone initial value of the first diffusion cycle; in the subsequent diffusion cycles, noise is gradually removed from the backbone initial value of the first diffusion cycle through the diffusion module, and the backbone after each noise removal is used as the backbone initial value corresponding to the subsequent diffusion cycle. It should be noted that the number of diffusion cycles is at least one. When the number of diffusion cycles is only one, there is no subsequent diffusion cycle, and the entire diffusion cycle only includes the first diffusion cycle.

[0046] Exemplarily, the diffusion module does not include learnable parameters, and the results of stepwise noise addition and stepwise denoising are obtained by solving a stochastic differential equation (SDE) or an equivalent ordinary differential equation (ODE, for denoising, ODE can be solved).

[0047] Exemplarily, the embedded representation corresponding to the current diffusion cycle is determined based on the amino acid sequence in the current sample. For the first diffusion cycle, the embedded representation corresponding to the current diffusion cycle is determined only based on the amino acid sequence. Exemplarily, the embedded representation includes sequence features (such as MSA representation or single representation) and pair features (pair representation) that characterize the relationship between amino acid residues. There are many methods for extracting embedded representations. For example, the embedded representation can be obtained based on multiple sequence alignment of the amino acid sequence in the current sample, or the embedded representation is obtained by multiple sequence alignment and template search of the amino acid sequence in the current sample, or the embedded representation is obtained by inputting the amino acid sequence in the current sample into a protein language model (refer to the method in xTrimoABfold). For subsequent diffusion cycles, the embedded representation corresponding to the current diffusion cycle can be determined based on the amino acid sequence in the current sample and the results obtained in the previous diffusion cycle, so that in each subsequent diffusion cycle, the embedded representation corresponding to the current diffusion cycle can be updated using the results obtained in the previous diffusion cycle.

[0048] It is understandable that the initial value of the backbone and the predicted value of the backbone may both include the backbone atomic coordinates X and the rotation relationship R between atoms.

[0049] Step 3), determining the predicted three-dimensional structure of the protein based on the predicted clean backbone obtained from the last diffusion cycle in at least one diffusion cycle.

[0050] The specific process of determining the three-dimensional structure of a protein from its backbone can be referred to in the prior art and is not limited here.

[0051] Exemplarily, in step 2), not only a clean backbone but also a clean side chain is predicted, and determining the predicted three-dimensional structure of the protein based on the predicted clean backbone of the last subsequent cycle may include determining the predicted three-dimensional structure of the protein based on the predicted clean backbone and the predicted clean side chains of the last subsequent cycle.

[0052] Step 4), calculating the loss value according to the protein three-dimensional structure in the current sample and the predicted three-dimensional structure;

[0053] According to the three-dimensional structure in the current sample (i.e. Figure 2A Label in the text) and predict the three-dimensional structure (i.e. Figure 2A Prediction) calculates the loss value;

[0054] Step 5), updating the protein multi-conformation prediction model to be trained according to the loss value.

[0055] It is understandable that updating the protein multi-conformation prediction model can be updating the learnable parameters in the protein multi-conformation prediction model by gradient descent method.

[0056] In one embodiment, the number of at least one diffusion cycle is a target cycle number, and the diffusion module includes a forward diffusion module and a reverse diffusion module; wherein the forward diffusion module is used to add noise, and the reverse diffusion module is used to remove noise.

[0057] Exemplarily, in the first diffusion cycle, the backbone initial value corresponding to the current diffusion cycle is generated by adding noise to the protein three-dimensional structure in the current sample through the forward diffusion module via a target number of diffusion steps;

[0058] Exemplarily, in the subsequent diffusion cycle, the backbone initial value corresponding to the Nth diffusion cycle is obtained from the backbone initial value corresponding to the first diffusion cycle via N-1 diffusion steps, and each diffusion step gradually removes noise. For example, the backbone initial value corresponding to the second diffusion cycle is obtained from the backbone initial value corresponding to the first diffusion cycle via one diffusion step, the backbone initial value corresponding to the third diffusion cycle is obtained from the backbone initial value corresponding to the first diffusion cycle via two diffusion steps, and the backbone initial value corresponding to the last diffusion step is obtained from the backbone initial value corresponding to the first diffusion cycle via the target number of cycles-1 diffusion steps. It is understandable that when determining the backbone initial value corresponding to the subsequent diffusion cycle, it is not necessary to first solve the noise amount of the noise removed in the diffusion step corresponding to the current diffusion cycle, and then subtract this noise amount from the backbone initial value corresponding to the previous diffusion cycle, but it can be directly determined by the reverse diffusion module according to the backbone initial value corresponding to the previous diffusion cycle and the predicted clean backbone obtained by the previous diffusion cycle.

[0059] In one embodiment, in step 2), using the protein multi-conformation model to be trained, and using the embedded representation corresponding to the current diffusion cycle to update the backbone initial value corresponding to the current diffusion cycle, comprises:

[0060] Using the protein multi-conformation model to be trained, the backbone initial value corresponding to the current diffusion cycle is updated using the embedded representation corresponding to the current diffusion cycle and the cycle feature representing the current cycle number; wherein the cycle feature is learnable.

[0061] Exemplarily, an initial cycle feature can be determined according to the current cycle number (for example, by looking up a mapping table of the current cycle number-initial cycle feature), and the initial cycle feature is processed by a linear layer containing learnable parameters to obtain a cycle feature characterizing the current cycle number. In step 5), while updating the protein multi-conformation prediction model to be trained according to the loss value, the learnable parameters containing the linear layer are also updated.

[0062] In this way, adding a cycle feature that can indicate the current cycle number to the input of the protein multi-conformation model explicitly explains the information of the current cycle number, which can improve the accuracy of the protein multi-conformation model in predicting the clean backbone.

[0063] In one embodiment, the protein multi-conformation model includes an embedding iteration module and a structure module;

[0064] Using the protein multi-conformation model to be trained, the backbone initial value corresponding to the current diffusion cycle is updated using the embedded representation corresponding to the current diffusion cycle and the cycle feature representing the current cycle number, including:

[0065] Iterate the embedding representation corresponding to the current diffusion cycle using the embedding iteration module to generate an iterative representation corresponding to the current diffusion cycle;

[0066] By utilizing the structural module, the backbone initial value corresponding to the current diffusion cycle is updated using the iteration representation corresponding to the current diffusion cycle and the cycle feature representing the current cycle number.

[0067] Exemplarily, the protein multi-conformation model adopts an architecture similar to Alphafold, including an embedding iteration module and a structure module. Exemplarily, the embedding iteration module may include multiple iteration units, each iteration unit may share a weight; the structure module may include multiple structure units, each structure unit may share a weight. The embedding representation corresponding to the current diffusion cycle and the backbone initial value corresponding to the current diffusion cycle are both input into the first unit among the multiple units. Specifically, the embedding representation corresponding to the current diffusion cycle is input into the first iteration unit among the multiple iteration units, the output of the first iteration unit is input into the next iteration unit, and the output of the last iteration unit is input into the structure unit; the backbone initial values ​​corresponding to the current diffusion cycle are both input into the first structure unit among the multiple structure units, the output of the first structure unit is input into the next structure unit, and the output of the last structure unit is used as the output of the entire protein multi-conformation model.

[0068] For example, since the process of predicting a clean backbone already includes multiple diffusion cycles, if the number of iterative units in a single diffusion cycle is large, the time required for backbone prediction will be greatly increased. Therefore, the inventors of the present disclosure conducted multiple experiments and determined that the minimum number of iterative units that does not significantly reduce the structure prediction performance is 2.

[0069] The embedded representation corresponding to the current diffusion cycle is iterated using the embedded iteration module to generate an iterated representation corresponding to the current diffusion cycle. Exemplarily, the embedded representation includes a sequence feature and a pairing feature. The embedded iteration module is used to iterate the embedded representation corresponding to the current diffusion cycle to generate an iterative representation, and the iterative representation includes an iterative sequence feature and an iterative pairing feature. Afterwards, the structural module is used to update the backbone initial value corresponding to the current diffusion cycle using the iterative representation corresponding to the current diffusion cycle and the cycle feature representing the current cycle number. Specifically, the cycle feature can be added to the iterative sequence feature (the two are set to be of the same size so that they can be directly added, and the size remains unchanged after addition) to obtain a superposition feature, and the structural module is used to update the backbone initial value corresponding to the current diffusion cycle using the superposition feature and the iterative pairing feature.

[0070] The specific process of using the structure module to update the initial value of the backbone corresponding to the current diffusion cycle using superposition features and iterative pairing features can be found in Figure 2B In the first structural unit of the structural module, in the invariant point attention (IPA) module of the structural module, the superposition features are updated with the backbone initial value corresponding to the current diffusion cycle and the iterative paired features to obtain updated sequence features; the predicted relative translation and rotation are determined by the updated sequence features, and the backbone initial value corresponding to the current diffusion cycle is updated with the relative translation and rotation to obtain an updated backbone; the updated sequence features and the updated backbone are input into the next structural unit.

[0071] As described above, for subsequent diffusion cycles other than the first diffusion cycle in at least one diffusion cycle, the embedding representation corresponding to the current diffusion cycle can be determined based on the amino acid sequence in the current sample and the result obtained in the previous diffusion cycle. For example, in the subsequent diffusion cycle, the embedding representation corresponding to the current diffusion cycle is determined based on the embedding representation corresponding to the first diffusion cycle, the predicted clean backbone obtained in the previous diffusion cycle, and / or the iterative representation generated in the previous diffusion cycle (i.e., the output of the embedding iteration module). The specific determination process can be performed by the recycling embedder. Specifically, the sequence features in the iterative representation of the previous cycle are layer-normalized to obtain the recovered sequence features; the pairing features in the iterative representation of the previous cycle are layer-normalized to obtain the pairing normalized features; the atomic coordinates in the predicted clean backbone obtained in the previous cycle are transformed by relative distance to obtain a probability histogram; the pairing normalized features are added to the probability histogram in a positional manner to obtain the recovered pairing features; the sequence features in the embedded representation corresponding to the first diffusion cycle are added to the recovered sequence features in a positional manner to obtain the sequence features in the embedded representation corresponding to the current diffusion cycle in the subsequent diffusion cycle; the pairing features in the embedded representation corresponding to the first diffusion cycle are added to the recovered pairing features in a positional manner to obtain the pairing features in the embedded representation corresponding to the current diffusion cycle in the subsequent diffusion cycle. Among them, the layer normalization operation contains learnable parameters, which can be updated while updating the parameters of the protein multi-conformation model in step 5).

[0072] In one embodiment, step S160 further includes: randomly determining a target number of cycles from a specified positive integer, where the specified positive integer is a positive integer less than or equal to a quotient of a diffusion distance and a diffusion step length.

[0073] As mentioned above, the number of at least one diffusion cycle is the target cycle number. Before executing step 2), the target cycle number needs to be determined.

[0074] Exemplarily, the diffusion distance is 1. When adding noise, after the entire diffusion distance, the three-dimensional structure of the protein becomes noise; when removing noise, after the entire diffusion distance, the noise becomes a clean backbone. The diffusion step length Δt can be a certain or only a limited number of changes during the entire training process. Δt can be 0.1, 0.01, etc. The diffusion step length can be an input parameter of the forward diffusion module and the reverse diffusion module. If the diffusion distance is 1 and Δt is 0.1, the target number of cycles can be randomly determined from a positive integer between 1 and 10. Figure 2A The case where the target number of cycles is 4 is shown.

[0075] The purpose of randomly determining the target number of cycles is to hope that the protein multi-conformation model can learn the ability to predict a clean backbone from a backbone with different degrees of noise added. When the target number of cycles is 1 / Δt, the protein multi-conformation model can learn the ability to predict a clean backbone from a backbone with the maximum degree of noise added; when the target number of cycles is less than 1 / Δt, the protein multi-conformation model can learn the ability to predict a clean backbone from a backbone with non-maximal noise added. For example, when the target number of cycles is 1, the protein multi-conformation model can learn the ability to predict a clean backbone from a backbone with a small amount of noise added. At the same time, using the same multi-conformation sample and different target number of cycles for training actually plays a role in amplifying the training samples.

[0076] In some embodiments, training a protein multi-conformation model to be trained based on multi-conformation samples includes a first training stage and a second training stage performed successively, the diffusion distances of the first training stage and the second training stage are the same, and the diffusion step length of the first training stage is greater than the diffusion step length of the second training stage.

[0077] It is understandable that the diffusion step size can change during the entire cycle, and the frequency of change of the diffusion step size can be lower than the frequency of change of the target number of cycles. For example, the target number of cycles is re-determined in each training iteration, and the diffusion step size only changes 2-3 times during the entire training process. For example, the entire training process can be divided into two stages, with a diffusion step size Δt=0.1 in the first training stage and a diffusion step size Δt=0.01 in the second training stage. In the first training stage, before each execution of step 2), the target number of cycles is randomly determined from a positive integer of 1-1 / Δt (i.e., randomly selected from a positive integer of 1-10); in the second training stage, before each execution of step 2), the target number of cycles is randomly determined from a positive integer of 1-1 / Δt (e.g., randomly selected from a positive integer of 1 to 100). In this way, rough adjustment can be performed first, and then fine adjustment can be performed to improve the training efficiency of the model.

[0078] A specific embodiment of step S160 is described below.

[0079] Step a, selecting a multi-conformation sample as the current sample;

[0080] Step b, randomly determining a target number of cycles from a specified positive integer, where the specified positive integer is a positive integer less than or equal to the quotient of the diffusion distance and the diffusion step length;

[0081] Step c, determining an embedded representation corresponding to the first diffusion cycle according to the amino acid sequence in the current sample; determining a cycle feature corresponding to the first diffusion cycle;

[0082] Step d, adding noise to the three-dimensional structure of the protein in the current sample through a forward diffusion module via a target number of diffusion steps to generate a backbone initial value corresponding to the first diffusion cycle;

[0083] Step e, in the first diffusion cycle, using the protein multi-conformation model to be trained, using the embedded representation corresponding to the current diffusion cycle and the cycle feature corresponding to the current diffusion cycle to update the initial value of the backbone corresponding to the current diffusion cycle, and obtain the predicted clean backbone;

[0084] Step f, determining the embedding representation corresponding to the current diffusion cycle based on the embedding representation corresponding to the first diffusion cycle, the predicted clean backbone obtained from the previous diffusion cycle, and the intermediate results generated from the previous diffusion cycle (e.g., the output of the embedding iteration module in the protein multi-conformation model); determining the cycle characteristics corresponding to the current diffusion cycle;

[0085] Step g, determining the trunk initial value corresponding to the current diffusion cycle through the reverse diffusion module according to the trunk initial value corresponding to the previous diffusion cycle and the predicted clean trunk obtained from the previous diffusion cycle;

[0086] Step h, in the subsequent diffusion cycle, using the protein multi-conformation model to be trained, using the embedded representation corresponding to the current diffusion cycle and the cycle feature corresponding to the current diffusion cycle to update the initial value of the backbone corresponding to the current diffusion cycle, and obtain the predicted clean backbone;

[0087] Repeat steps f, g, and h until the number of diffusion cycles reaches the target number of cycles.

[0088] Step i, determining a predicted three-dimensional structure of the protein based on a predicted clean backbone obtained from the last diffusion cycle in at least one diffusion cycle;

[0089] Step j, calculating the loss value according to the protein three-dimensional structure in the current sample and the predicted three-dimensional structure;

[0090] Step k, updating the learnable parameters used in the protein multi-conformation prediction model to be trained and other modules according to the loss value (for example, when determining the cycle characteristics in steps c and f, the learnable parameters are used when determining the embedded representation corresponding to the current diffusion cycle in step f).

[0091] Figure 3 1 is a flow chart of a method 300 for predicting a protein three-dimensional structure according to an embodiment of the present disclosure. The protein multi-conformation prediction model obtained by training in the training method 100 is used for prediction; the method 300 comprises:

[0092] S310, obtaining the amino acid sequence of the protein to be predicted;

[0093] The protein three-dimensional structure of the amino acid sequence of the protein to be predicted may be unknown, or the protein three-dimensional structure may be known, but the number of multiple conformations does not meet the requirements. For example, if one protein three-dimensional structure corresponding to a certain amino acid sequence is known, and the other two need to be predicted, the amino acid sequence can be used as the amino acid sequence of the protein to be predicted.

[0094] S320, randomly generate a plurality of different noise trunks.

[0095] The noise stem may be, for example, Gaussian noise.

[0096] It is understandable that as many different structural conformations as desired are generated, as many different noise backbones are used as starting points for structural conformation prediction. Continuing with the previous example, if two other structural conformations need to be predicted, two different noise backbones are randomly generated. In this way, multiple starting points can be generated to generate different three-dimensional conformations from multiple starting points.

[0097] S330, for each noise trunk, execute step S3301 and step S3302:

[0098] S3301, in at least one diffusion cycle, using the trained protein multi-conformation model, using the embedding representation corresponding to the current diffusion cycle to update the initial value of the backbone corresponding to the current diffusion cycle, to obtain a predicted clean backbone;

[0099] S3302, determining a predicted three-dimensional structure corresponding to the noise backbone according to a predicted clean backbone obtained in the last diffusion cycle in at least one diffusion cycle;

[0100] Among them, in the first diffusion cycle of at least one diffusion cycle, the noise backbone is used as the backbone initial value corresponding to the current diffusion cycle; in the subsequent diffusion cycles of at least one diffusion cycle, the backbone initial value corresponding to the current diffusion cycle is determined by the diffusion module according to the backbone initial value corresponding to the previous diffusion cycle and the predicted clean backbone obtained by the previous diffusion cycle; the embedded representation corresponding to the current diffusion cycle is determined according to the amino acid sequence of the protein to be predicted.

[0101] It is understandable that the diffusion module in step S330 is a reverse diffusion module. Different from the training phase of the protein multi-conformation model, the forward diffusion module is no longer used to add noise in the inference phase, but the reverse diffusion module removes noise from the noise trunk.

[0102] As mentioned above, the target number of cycles in the training phase can be randomly determined from a positive integer of [1 1 / Δt]. Unlike the training phase, which hopes that the protein multi-conformation model learns to predict the clean backbone from the backbone with different degrees of noise added, the inference phase always predicts the clean backbone from the noise, so the number of diffusion cycles in the inference phase can be 1 / Δt and is no longer randomly determined. Exemplarily, the diffusion step length Δt of the inference phase can be the same as the diffusion step length of the last training phase.

[0103] The specific process of step S330 can be found in steps a to i, which will not be repeated here.

[0104] S340, taking the predicted three-dimensional structures corresponding to the noise backbones as multiple structural conformations of the protein to be predicted.

[0105] The specific process of step S330 can be found in steps a to i, which will not be repeated here.

[0106] For example, if two different noise trunks are generated in step S320, then two different predicted three-dimensional structures are obtained in step S340, and these two predicted three-dimensional structures constitute two structural conformations of the same given amino acid sequence. Thus, in this embodiment, multiple structural conformations corresponding to the given amino acid sequence are predicted based on the given amino acid sequence.

[0107] Figure 4 4 is a flow chart of a method 400 for determining a multi-conformation sample according to an embodiment of the present disclosure. The multi-conformation sample is used to train a protein multi-conformation prediction model, and the protein multi-conformation model is used to predict multiple structural conformations of a given amino acid sequence. The method 400 includes:

[0108] S410, obtaining a plurality of initial protein structure prediction models and a plurality of single conformation samples, each of the initial protein structure prediction models corresponding to different initial parameters, each single conformation sample including at least one amino acid sequence and a corresponding protein three-dimensional structure;

[0109] S420, training each initial protein structure prediction model based on multiple single conformation samples to obtain multiple teacher protein structure prediction models corresponding to the multiple initial protein structure prediction models;

[0110] S430, obtaining an amino acid sequence to be processed;

[0111] S440, predicting the three-dimensional structure of the amino acid sequence to be processed based on multiple teacher protein structure prediction models, respectively, to obtain multiple protein three-dimensional structures corresponding to the multiple teacher protein structure prediction models;

[0112] S450, the amino acid sequence to be processed and each of the three-dimensional protein structures of the plurality of three-dimensional protein structures are respectively combined into a multi-conformation sample, so that for one multi-conformation sample, there is another sample having the same amino acid sequence as the multi-conformation sample and a different three-dimensional protein structure.

[0113] For the description of steps S410 - S450 , please refer to the description of steps S110 - S150 , which will not be repeated here.

[0114] The multi-conformation sample determination method provided by the embodiment of the present disclosure obtains multiple initial protein structure prediction models and multiple single conformation samples, trains each initial protein structure prediction model based on the multiple single conformation samples, and obtains multiple teacher protein structure prediction models corresponding to the multiple initial protein structure prediction models; then, predicts the three-dimensional structure of the amino acid sequence to be processed based on the multiple teacher protein structure prediction models, and obtains multiple protein three-dimensional structures corresponding to the multiple teacher protein structure prediction models; the amino acid sequence to be processed and each protein three-dimensional structure in the multiple protein three-dimensional structures are combined into a multi-conformation sample; subsequently, richer and more comprehensive multi-conformation samples can be used to train the protein multi-conformation model to be trained, which can improve the prediction ability of the protein multi-conformation model and enable it to predict multiple structural conformations of a given amino acid sequence.

[0115] Figure 5 A structural schematic diagram of an electronic device provided for an embodiment of the present disclosure, the electronic device 80 includes: a processor 81, a memory 82, a bus 83 and a communication interface 84, wherein the processor 81, the communication interface 84 and the memory 82 are connected via the bus 83; the processor 81 is used to execute an executable module stored in the memory 82, such as a computer program.

[0116] The memory 82 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 84 (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used.

[0117] The bus 83 may be an ISA bus, a PCI bus, or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 5 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0118] Among them, the memory 82 is used to store programs, and the processor 81 executes the program after receiving the execution instruction. The method executed by the device for flow process definition disclosed in any of the embodiments of the present disclosure can be applied to the processor 81 or implemented by the processor 81.

[0119] The processor 81 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit or software instructions in the processor 81. The above processor 81 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present disclosure can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in conjunction with the embodiments of the present disclosure can be directly embodied as a hardware decoding processor for execution, or a combination of hardware and software modules in the decoding processor for execution. The software module can be located in a storage medium mature in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 82, and the processor 81 reads the information in the memory 82 and completes the steps of the above method in combination with its hardware.

[0120] The computer program product of the readable storage medium provided in the embodiments of the present disclosure includes a computer-readable storage medium storing program code, and the instructions included in the program code can be used to execute the methods described in the previous method embodiments. The specific implementation can be found in the previous method embodiments, which will not be repeated here.

[0121] In the embodiments provided in the present disclosure, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0122] In addition, the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0123] Furthermore, the functional modules in the various embodiments of the present disclosure may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0124] It should be noted that if the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present disclosure is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the various embodiments of the present disclosure. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM) random access memory (RAM), disk or optical disk and other media that can store program code.

[0125] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0126] The above are only embodiments of the present disclosure and are not intended to limit the scope of protection of the present disclosure. For those skilled in the art, the present disclosure may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the scope of protection of the present disclosure.

Claims

1. A method for training a protein multi-conformation model, characterized in that: The protein multi-conformation model is used to predict multiple structural conformations of a given amino acid sequence; the method comprises: Acquire multiple initial protein structure prediction models and multiple single conformation samples, each of the initial protein structure prediction models has different initial parameters, and each of the single conformation samples includes at least one amino acid sequence and a corresponding protein three-dimensional structure; Training each of the initial protein structure prediction models based on the multiple single conformation samples to obtain multiple teacher protein structure prediction models corresponding to the multiple initial protein structure prediction models; Obtaining an amino acid sequence to be processed; Predicting the three-dimensional structure of the amino acid sequence to be processed based on the multiple teacher protein structure prediction models, respectively, to obtain multiple protein three-dimensional structures corresponding to the multiple teacher protein structure prediction models; The amino acid sequence to be processed and each of the three-dimensional protein structures in the plurality of three-dimensional protein structures are respectively combined into a multi-conformation sample, so that for one multi-conformation sample, there is another sample having the same amino acid sequence as the multi-conformation sample and a different three-dimensional protein structure; The protein multi-conformation model to be trained is trained based on the multi-conformation samples.

2. The method according to claim 1, characterized in that Training the protein multi-conformation model to be trained based on the multi-conformation samples includes: Select a multi-conformation sample as the current sample; In each diffusion cycle of at least one diffusion cycle, using the protein multi-conformation model to be trained, using the embedding representation corresponding to the current diffusion cycle to update the initial value of the backbone corresponding to the current diffusion cycle, to obtain a predicted clean backbone; wherein the initial value of the backbone corresponding to the current diffusion cycle is generated by a diffusion module; the embedding representation corresponding to the current diffusion cycle is determined according to the amino acid sequence in the current sample; in the first diffusion cycle of the at least one diffusion cycle, the initial value of the backbone corresponding to the current diffusion cycle is generated by adding noise to the three-dimensional structure of the protein in the current sample by the diffusion module; determining a predicted three-dimensional structure of the protein based on the predicted clean backbone obtained from the last diffusion cycle of the at least one diffusion cycle; Calculating a loss value according to the three-dimensional structure of the protein in the current sample and the predicted three-dimensional structure; The protein multi-conformation prediction model to be trained is updated according to the loss value.

3. The method according to claim 2, characterized in that In subsequent diffusion cycles other than the first diffusion cycle in the at least one diffusion cycle, the trunk initial value corresponding to the current diffusion cycle is obtained by removing at least part of noise from the trunk initial value corresponding to the previous diffusion cycle by the diffusion module.

4. The method according to claim 2 or 3, characterized in that: The number of the at least one diffusion cycle is a target cycle number, and the diffusion module includes a forward diffusion module and a reverse diffusion module; In the first diffusion cycle, the trunk initial value corresponding to the current diffusion cycle is generated by adding noise to the three-dimensional structure of the protein in the current sample through the forward diffusion module via the target number of diffusion steps; and / or, in the subsequent diffusion cycle, the trunk initial value corresponding to the current diffusion cycle is determined by the reverse diffusion module based on the trunk initial value corresponding to the previous diffusion cycle and the predicted clean trunk obtained in the previous diffusion cycle, and the trunk initial value corresponding to the current diffusion cycle is used to characterize the noise trunk obtained after removing the noise corresponding to one diffusion step from the trunk initial value corresponding to the previous diffusion cycle.

5. The method according to any one of claims 2 to 4, using the protein multi-conformation model to be trained, and using the embedded representation corresponding to the current diffusion cycle to update the backbone initial value corresponding to the current diffusion cycle, comprising: Using the protein multi-conformation model to be trained, updating the backbone initial value corresponding to the current diffusion cycle using the embedded representation corresponding to the current diffusion cycle and the cycle feature representing the current cycle number; Wherein, the cyclic features are learnable.

6. The method according to claim 5, wherein the protein multi-conformation model comprises an embedded iteration module and a structure module; The method of using the protein multi-conformation model to be trained and updating the backbone initial value corresponding to the current diffusion cycle using the embedded representation corresponding to the current diffusion cycle and the cycle feature representing the current cycle number includes: Iterating the embedding representation corresponding to the current diffusion cycle using the embedding iteration module to generate an iterated representation corresponding to the current diffusion cycle; Using the structural module, updating the backbone initial value corresponding to the current diffusion cycle using the iteration characterization corresponding to the current diffusion cycle and the cycle feature characterizing the current cycle number; Among them, in subsequent diffusion cycles other than the first diffusion cycle in the at least one diffusion cycle, the embedded representation corresponding to the current diffusion cycle is determined based on the embedded representation corresponding to the first diffusion cycle, the predicted clean trunk obtained in the previous diffusion cycle and / or the iterative representation generated by the previous diffusion cycle.

7. The method according to claim 4, characterized in that The training of the protein multi-conformation model to be trained based on the multi-conformation samples also includes: randomly determining the target number of cycles from a specified positive integer, where the specified positive integer is a positive integer less than or equal to the quotient of the diffusion distance and the diffusion step length.

8. The method according to claim 7, characterized in that The training of the protein multi-conformation model to be trained based on the multi-conformation samples includes a first training stage and a second training stage performed successively, the diffusion distances of the first training stage and the second training stage are the same, and the diffusion step length of the first training stage is greater than the diffusion step length of the second training stage.

9. A method for predicting a protein three-dimensional structure, characterized in that: The method uses a protein multi-conformation prediction model trained by the training method according to any one of claims 1 to 8 to predict multiple structural conformations of a protein; the method comprises: Obtaining the amino acid sequence of the protein to be predicted; Randomly generate multiple different noise backbones; For each of the noise trunks, In at least one diffusion cycle, using the protein multi-conformation model obtained by the training, using the embedding representation corresponding to the current diffusion cycle to update the initial value of the backbone corresponding to the current diffusion cycle, to obtain a predicted clean backbone; Determine a predicted three-dimensional structure corresponding to the noise trunk according to the predicted clean trunk obtained in the last diffusion cycle of the at least one diffusion cycle; wherein, in the first diffusion cycle of the at least one diffusion cycle, the noise trunk is used as the trunk initial value corresponding to the current diffusion cycle; in the subsequent diffusion cycles of the at least one diffusion cycle, the trunk initial value corresponding to the current diffusion cycle is determined by the diffusion module according to the trunk initial value corresponding to the previous diffusion cycle and the predicted clean trunk obtained in the previous diffusion cycle; and the embedded representation corresponding to the current diffusion cycle is determined according to the amino acid sequence of the protein to be predicted; The predicted three-dimensional structures corresponding to the noise backbones are used as multiple structural conformations of the protein to be predicted.

10. A method for determining a multi-conformation sample, wherein the multi-conformation sample is used to train a protein multi-conformation prediction model, and the protein multi-conformation model is used to predict multiple structural conformations of a given amino acid sequence, characterized in that: The method comprises: Acquire multiple initial protein structure prediction models and multiple single conformation samples, each of the initial protein structure prediction models has different initial parameters, and each of the single conformation samples includes at least one amino acid sequence and a corresponding protein three-dimensional structure; Training each of the initial protein structure prediction models based on the multiple single conformation samples to obtain multiple teacher protein structure prediction models corresponding to the multiple initial protein structure prediction models; Obtaining an amino acid sequence to be processed; Predicting the three-dimensional structure of the amino acid sequence to be processed based on the multiple teacher protein structure prediction models, respectively, to obtain multiple protein three-dimensional structures corresponding to the multiple teacher protein structure prediction models; The amino acid sequence to be processed and each of the multiple protein three-dimensional structures are respectively combined into a multi-conformation sample, so that for one multi-conformation sample, there is another sample with the same amino acid sequence as the multi-conformation sample and a different protein three-dimensional structure.

11. An electronic device, characterized in that: The electronic device comprises: a processor and a storage device; The storage device stores a computer program, and when the computer program is executed by the processor, the method according to any one of claims 1 to 10 is executed.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the method according to any one of claims 1 to 10.

13. A computer program product comprising computer instructions, wherein: When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

Images