A method for three-dimensional structure reconstruction of protein cryo-electron microscopy projection images

Abstract

A method for reconstructing the three-dimensional structure of a protein cryo-electron microscopy projection image comprises the following steps: step 1: initializing an autoencoder neural network model to generate an initial population of N latent vectors; step 2: converting the latent vectors into a three-dimensional model in real space through a decoder; step 3: calculating the projection of the three-dimensional model at the corresponding azimuth angle; step 4: calculating the difference between the calculated projection and the user input projection; step 5: using a genetic algorithm to update the latent vector, and iteratively searching for a three-dimensional model and its azimuth angle with a smaller difference. The present invention provides a method for reconstructing a protein three-dimensional model based on a small number of projections, which requires a small number of projections, has a fast calculation speed, and has a high tolerance to noise. The output result can provide a reference for further understanding the three-dimensional structure of the protein.

Description

Technical Field

[0001] The present invention relates to the field of structural biology, and more specifically, to a method for reconstructing the three-dimensional structure of protein cryo-electron microscopy projection images based on a small number of projections. Background Art

[0002] Cryo-electron microscopy (Cryo-EM) and cryo-electron tomography (Cryo-ET) are two powerful tools that are developing rapidly in structural biology research today. Compared with the traditional research methods of structural biology, X-ray crystallography and nuclear magnetic resonance spectroscopy, Cryo-EM and Cryo-ET have the advantages of not requiring crystallization of biological macromolecules, requiring a small amount of sample, and maintaining the normal physiological state of biological macromolecules. In recent years, with the development of sample preparation technology, imaging devices and reconstruction algorithms, Cryo-EM and Cryo-ET have become two indispensable technologies in structural biology research.

[0003] The basic principle of Cryo-EM and Cryo-ET is to quickly freeze the sample in a glassy ice layer. When the incident electron beam passes through the sample, it interacts with the sample, and the intensity and direction of the incident electron beam may change. It is recorded on the detector, so that the structure of the sample can be analyzed. Since the electron dose that biological samples can withstand is very low, only a small dose can be used for imaging, resulting in images with very low signal-to-noise ratio. For this problem, the current method is to enhance the signal by superposition averaging or averaging after image classification, and reconstruct the model using the expectation maximization algorithm based on Bayesian theory. The low signal of a single image requires a large number of images for signal enhancement, which takes a long time to calculate, and is not conducive to researchers quickly understanding the three-dimensional structure of protein macromolecules.

[0004] In Cryo-EM single particle analysis, reconstruction of a three-dimensional structure often requires tens of thousands to millions of images. On the other hand, the structure of biological macromolecules is flexible, and the same type of molecule is likely to correspond to different states at the time of imaging. Transition states are usually sampled less due to their higher energy, and traditional methods cannot be used to reconstruct them. Moreover, the transition states of biological macromolecules are very important for understanding their working mechanisms. In Cryo-ET, the latest technology re-images frozen cell slices. In this case, the amount of data obtained is smaller than that of traditional Cryo-EM single particle analysis, and new algorithms are also needed to reconstruct the model in three dimensions. Summary of the invention

[0005] In order to overcome the problems of too many projections and long calculation time required by the above methods, the present invention provides a method for reconstructing the three-dimensional structure of a protein based on a small number of projections. The method is based on an autoencoder neural network model and a genetic algorithm. The structure of the individual in the genetic algorithm is determined according to whether the direction of the projection is known. The method requires a small number of projections, has a fast calculation speed, and is highly resistant to noise. The method can help researchers quickly understand the approximate three-dimensional structure of a protein macromolecule based on several projections. At the same time, the output model of the present invention can be used as an initial model for fine reconstruction.

[0006] The present invention provides the following technical solutions:

[0007] A method for reconstructing a three-dimensional structure of a protein cryo-electron microscopy projection image, the method comprising the following steps:

[0008] Step 1: Initialize the autoencoder neural network model to generate N latent vectors of the initial population;

[0009] Step 2: Convert the latent vector into a three-dimensional model in real space through the decoder;

[0010] Step 3: Calculate the projection of the three-dimensional model at the corresponding azimuth angle;

[0011] Step 4: Calculate the difference between the calculated projection and the user input projection;

[0012] Step 5: Use the genetic algorithm to update the latent vector by iteratively searching for the 3D model and its azimuth with the smallest difference.

[0013] Further, the step 1 comprises the following steps:

[0014] 1.1 Randomly generate the initial population for the genetic algorithm according to the distribution law of latent vectors; the distribution law is: the distribution law described by the mean and variance of the parameters of the PISA protein database after the structural information is encoded by the encoder in the 200-dimensional latent space;

[0015] 1.2 If the spatial orientation corresponding to the user input projection is unknown, a corresponding number of quaternions are added to the end of the latent vector according to the number of input projections to describe the spatial orientation.

[0016] In the step 1.1, the distribution law is: the distribution law described by the mean and variance of the parameters of the structural information of the PISA protein database after being encoded by the encoder in the 200-dimensional latent space.

[0017] In step 1.2, if the spatial orientation is known, the individual structure of the population is a 200-dimensional latent space parameter; if the spatial orientation is unknown, a corresponding number of quaternions are added to the end of the 200-dimensional latent vector according to the number of input projections to describe the spatial orientation.

[0018] Furthermore, step 2 includes the following steps:

[0019] 2.1 Input each 200-dimensional latent vector parameter describing the model shape into the decoder and decode it into the corresponding 3D model;

[0020] 2.2 Determine whether the decoded three-dimensional model is a connected structure. If so, the output model remains unchanged; if not, determine its maximum connected structure and replace the original output model.

[0021] The step 3 comprises the following steps:

[0022] 3.1 Get quaternion. When the projection direction is known, it is provided by the user. When the projection direction is unknown, it is obtained by removing the first 200 parameters from the latent vector.

[0023] 3.2 Rotate the three-dimensional model output by the decoder according to each set of quaternions obtained;

[0024] 3.3 Based on the rotated three-dimensional model, the projection is calculated and summed along the predetermined imaging direction.

[0025] In step 3.1, the quaternion is obtained by providing the user when the projection direction is known, and by removing the first 200 parameters from the latent vector when the projection direction is unknown.

[0026] The step 5 comprises the following steps:

[0027] 5.1 Sort the population according to the loss value from small to large, and select the top 20 individuals with the lowest loss value;

[0028] 5.2 Perform selection operation on the remaining population, randomly select individuals with smaller loss values, and update the population;

[0029] 5.3 Perform a crossover operation on the remaining population, taking two adjacent individuals as a group, randomly select points, perform two-point crossover, and update the population;

[0030] 5.4 Perform mutation operation on the remaining updated population. For each individual in the population, set a 5% probability of mutation, re-randomly select values ​​according to the latent vector distribution law, and update the population again;

[0031] 5.5 Merge the population consisting of the first 20 best individuals with the population after the operation to obtain a new population. Decode the population and calculate the projection and compare the difference with the input projection signal. Sort the population from small to large according to the loss value, and determine whether the set maximum number of iterations is reached. If not, go to step 2. If so, select the first 20 best individuals after sorting, calculate the average value and output it.

[0032] The beneficial effects of the present invention are as follows: compared with the existing reconstruction methods, the present invention requires fewer projections, has a fast calculation speed, and has a high tolerance to noise. It can help researchers quickly understand the approximate three-dimensional structure of a protein macromolecule based on several projections of the protein macromolecule. At the same time, the output model of the present invention can be used as an initial model for fine reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is an overall flow chart according to the present invention.

[0034] Figure 2 are three orthogonal direction projections input by the user in the embodiment of the present invention.

[0035] Figure 3 It is the final reconstruction result in the embodiment of the present invention.

[0036] Figure 4 It is a real model in the embodiment of the present invention. DETAILED DESCRIPTION

[0037] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0038] Reference Figure 1 to Figure 4 , a method for reconstructing the three-dimensional structure of a protein cryo-electron microscopy projection image, comprising the following steps:

[0039] Step 1: Initialize the population to 300 200-dimensional latent vectors, and randomly generate the first-generation population for the genetic algorithm according to the latent vector distribution law; the distribution law is: the distribution law described by the mean and variance of the parameters of the structural information of the PISA protein database after being encoded by the encoder in the 200-dimensional latent space; if the spatial orientation is known, the individual structure of the population is the 200-dimensional latent space parameter; if the spatial orientation is unknown, then according to the number of input projections, the corresponding number of quaternions is added to the end of the 200-dimensional latent vector to describe the spatial orientation;

[0040] Step 2: Decode the latent vector into a 3D model, including the following steps:

[0041] 2.1 Input each 200-dimensional latent vector parameter describing the model shape into the decoder and decode it into the corresponding 3D model;

[0042] 2.2 Determine whether the decoded 3D model is a connected structure. If so, the output model remains unchanged. If not, determine its maximum connected structure and replace the original output model.

[0043] Step 3: Calculate the projection of the 3D model, including the following steps:

[0044] 3.1 Get quaternion, the projection direction is known and provided by the user;

[0045] 3.2 Rotate the three-dimensional model output by the decoder according to each set of quaternions obtained;

[0046] 3.3 According to the rotated three-dimensional model, summing and calculating the projection along the predetermined imaging direction;

[0047] In step 3.1, the quaternion is obtained by providing the user when the projection direction is known, and by removing the first 200 parameters from the latent vector when the projection direction is unknown.

[0048] Step 4: Calculate the difference between the calculated projection and the user input projection, recorded as loss;

[0049] Step 5: Use the genetic algorithm to iteratively search for a 3D model with a smaller loss, including the following steps:

[0050] 5.1 Sort the population according to the loss value from small to large, and select the top 20 individuals with the lowest loss value;

[0051] 5.2 Perform selection operation on the remaining population, randomly select individuals with smaller loss values, and update the population;

[0052] 5.3 Perform a crossover operation on the remaining updated population, taking two adjacent individuals as a group, randomly select points, perform two-point crossover, and update the population;

[0053] 5.4 Perform mutation operation on the remaining population. For each individual in the population, set a 5% probability of mutation, re-randomly select values ​​according to the latent vector distribution law, and update the population again;

[0054] 5.5 Merge the population consisting of the first 20 best individuals with the population after the operation to obtain a new population. Decode the population and calculate the projection and compare the difference with the input projection signal. Sort the population from small to large according to the loss value, and determine whether the set maximum number of iterations is reached. If not, go to step 2. If so, select the first 20 best individuals after sorting, calculate the average value and output it.

[0055] In this embodiment, the user provides three mutually orthogonal projections in known directions. Figure 2 It shows three orthogonal direction projections input by the user in an embodiment of the present invention, and the input projection size is 31*31 pixels. Figure 3 The final reconstruction result in the embodiment of the present invention is shown, and the size of the reconstruction result is 31*31*31 voxels. Figure 4 The real model is shown, and it can be seen that the shape structure of the reconstructed model is highly consistent with the real model.

[0056] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for reconstructing the three-dimensional structure of a protein cryo-electron microscopy projection image, characterized in that: The method comprises the following steps: Step 1: Initialize the autoencoder neural network model to generate N latent vectors of the initial population; Step 2: Convert the latent vector into a three-dimensional model in real space through the decoder; Step 3: Calculate the projection of the three-dimensional model at the corresponding azimuth angle; Step 4: Calculate the difference between the calculated projection and the user input projection; Step 5: Use the genetic algorithm to update the latent vector and iteratively search for the 3D model and its azimuth with the smallest difference; The step 1 comprises the following steps: 1.1 According to the distribution law of latent vectors, the initial population used for genetic algorithm is randomly generated; 1.2 If the spatial orientation corresponding to the user input projection is unknown, then the corresponding number of quaternions is added to the end of the latent vector according to the number of input projections to describe the spatial orientation; In the step 1.1, the distribution law is: the distribution law described by the mean and variance of the parameters of the structural information of the PISA protein database after being encoded by the encoder in the 200-dimensional latent space; In step 1.2, if the spatial orientation is known, the individual structure of the population is a 200-dimensional latent space parameter; if the spatial orientation is unknown, a corresponding number of quaternions are added to the 200-dimensional latent space parameter end according to the number of input projections to describe the spatial orientation; The step 5 comprises the following steps: 5.1 Sort the population from small to large according to the projection difference, and select the top n best individuals with the lowest projection difference; 5.2 Perform selection operation on the remaining population, randomly select individuals with small projection difference, and update the population; 5.3 Perform a crossover operation on the remaining updated population, taking adjacent individuals as a group, randomly select points, perform two-point crossover, and update the population again; 5.4 Perform mutation operation on the remaining updated population. For each individual in the population, set a mutation with low probability, re-randomly select values ​​according to the latent vector distribution law, and update the population again. 5.5 Merge the population consisting of n optimal vectors with the operated population to obtain a new population, decode the population and calculate the projection and compare the difference with the input projection signal, sort them from small to large according to the projection difference, and determine whether the set maximum number of iterations is reached. If not, go to step 5.

2. If so, select the first n optimal individuals after sorting, calculate the average value and output it.

2. A method for reconstructing the three-dimensional structure of a protein cryo-electron microscopy projection image according to claim 1, characterized in that: The step 2 comprises the following steps: 2.1 Input each 200-dimensional latent space parameter describing the model shape into the decoder and decode it into the corresponding 3D model; 2.2 Calculate the maximum connected structure domain of the decoded three-dimensional model.

3. A method for reconstructing the three-dimensional structure of a protein cryo-electron microscopy projection image according to claim 1, characterized in that: The step 3 comprises the following steps: 3.1 Get the specific spatial orientation corresponding to the quaternion; 3.2 Rotate the three-dimensional model output by the decoder according to each set of quaternions obtained; 3.3 Calculate the projection based on the rotated three-dimensional model.

4. A method for reconstructing the three-dimensional structure of a protein cryo-electron microscopy projection image according to claim 3, characterized in that: In step 3.1, the quaternion is obtained: when the projection direction is known, it is provided by the user; when the projection direction is unknown, it is obtained by removing the first 200-dimensional latent space parameters from the latent vector.

Images