Neural network-based cryo-em protein model building method and storage medium

Abstract

The application provides a cryo-EM protein model building method based on a neural network and a storage medium, the storage medium stores a deep neural network, the deep neural network comprises a cryo-conversion module stack, the cryo-conversion module stack comprises a plurality of cryo-conversion modules Cryoformer; the Cryoformer comprises an encoder and a decoder; the decoder is used for learning the matching of sequence-related representation and three-dimensional space information of a cryo-EM density map, and cross-fusion is performed on the sequence-related representation and the three-dimensional space information output by the encoder. The deep neural network can be used for processing the cryo-EM density map, obtaining the atomic model of the corresponding protein complex structure, being efficient and accurate, and being capable of processing low-resolution density maps, so that the application range of automatic cryo-EM model building is greatly expanded.

Description

Technical Field

[0001] This invention belongs to the field of structural biology image processing technology, and specifically relates to a computer storage medium, a computer system, a cryo-electron microscopy density map processing method, and a method for building cryo-electron microscopy protein models based on neural networks. Background Technology

[0002] With the breakthrough development of single-particle cryo-electron microscopy technology, especially the continuous innovation in hardware and software, cryo-electron microscopy (cryo-EM) has become a key method for resolving the structures of macromolecules and cellular machines with important biological significance, especially for protein complex structures.

[0003] While advanced machine learning-based structure prediction algorithms, such as AlphaFold and RoseTTAFold, are changing the way we analyze the 3D structure of individual proteins, automatically constructing 3D atomic structure models from electron cloud density maps generated by cryo-electron microscopy is a highly challenging task, especially for protein complexes composed of multiple proteins. This requires the model builder to have a high level of understanding of the protein's structural features and side chain conformations, particularly in regions with significant density differences. This necessitates considerable time for refinement and may even require additional experiments to obtain higher-resolution density maps. However, obtaining high-resolution density maps requires good homogeneity of the resolved complex and highly advanced cryo-electron microscopy equipment, making it a significant challenge, if not impossible.

[0004] In summary, existing methods for building atomic models suffer from problems such as high requirements for image quality, high technical requirements for implementers, and poor accuracy of results.

[0005] Therefore, there is an urgent need to develop a solution that provides high accuracy, full automation, and supports the construction of protein complex structural models for cryo-electron microscopy at medium and low resolutions. Summary of the Invention

[0006] To address the above problems, the present invention provides a computer storage medium storing a deep neural network.

[0007] The deep neural network includes a cryoformer module stack, which includes multiple cryoformer modules.

[0008] Cryoformer includes encoders and decoders;

[0009] The decoder is used to learn the matching of three-dimensional spatial information of sequence correlation representation and cryo-electron microscopy density map, and cross-fusion the three-dimensional spatial information output by the sequence correlation representation and the encoder.

[0010] Furthermore, the decoder takes the output of the sequence branch of the deep neural network, the output of the encoder, and the three-dimensional position encoding of the cryo-electron microscopy density map as input, and generates cross-sequence representations through self-attention modules and cross-attention modules.

[0011] Furthermore, each Cryoformer includes N enc Encoders and N dec One decoder;

[0012] Sequence-related characterization includes multiple sequence characterization and amino acid pairing characterization;

[0013] Each decoder adds the multiple sequence representation and the amino acid pairing representation to the crossed single sequence representation after passing through a linear layer, and then adds them to each other after passing through a LayerNorm layer to form a new single sequence representation;

[0014] The new single-sequence representation is input into the self-attention module;

[0015] The output of the self-attention module, the amino acid embedding characterization, the encoder output, and the 3D position encoding of the density map are all input into the cross-attention module to match cryo-electron microscopy density map features and sequence features.

[0016] Furthermore, the cross-attention module uses Q... c K c and V c Three variables are used as inputs, where Q c K is the result of adding the output of the self-attention module to the amino acid embedding characterization. c V is the result of adding the density map representation output by the encoder to the 3D position encoding of the density map. c This represents the density map output of the encoder.

[0017] Furthermore, the output of the cross-attention module is added to the output of the self-attention module and then input into the third LayerNorm layer. The output of the third LayerNorm layer is processed by a linear layer and then superimposed with the output of the third LayerNorm layer. This superposition is then input into the fourth LayerNorm layer for further processing, and a new, separate sequence representation of the cross is output.

[0018] Furthermore, the deep neural network employs the CryoFold model, which incorporates Cryoformer.

[0019] CryoFold includes the sequence branching described above for learning sequence-related characterizations of protein evolution from protein sequences, including multiple sequence characterization and amino acid pairing characterization.

[0020] Furthermore, the sequence branching includes an encoding module and an embedding representation learning module.

[0021] The encoding module is used to encode amino acid sequences, multiple sequence alignment (MSA) data, and structural templates.

[0022] The embedding characterization learning module is used to learn the embedding of encoded amino acid sequences, MSAs, and structural templates to generate multiple sequence characters and amino acid pairing characters.

[0023] The sequence branch also includes the Evoformer stack, which is used to learn multiple sequence representations and amino acid pairing representations, and output new multiple sequence representations and amino acid pairing representations.

[0024] Furthermore, the deep neural network employs the CryoFold model, which incorporates Cryoformer.

[0025] CryoFold includes a cryo-electron microscopy density map branch, which comprises a three-dimensional residual neural network used to map high-dimensional features into low-dimensional density map representations.

[0026] Furthermore, the cryo-electron microscopy density map branch takes the cryo-electron microscopy density map as input, passes through a 3D convolutional neural network layer, a batch normalization layer, a modified linear unit (ReLU), and a max pooling layer, and then sequentially inputs it into four 3D residual convolutional modules. Finally, it is processed by a 3D convolutional neural network layer before being output.

[0027] The present invention also provides a computer system, comprising:

[0028] One or more processors and one or more non-transitory computer-readable media, the storage of which is configured to process the aforementioned deep neural network of cryo-electron microscopy density maps.

[0029] The present invention also provides a method for processing cryo-electron microscopy density maps, which uses the aforementioned deep neural network to process the cryo-electron microscopy density maps.

[0030] The present invention also provides a method for constructing a cryo-electron microscopy protein model based on a neural network, comprising: processing the cryo-electron microscopy density map using the aforementioned deep neural network to obtain an atomic model of the corresponding protein complex structure.

[0031] This invention proposes a method for building cryo-electron microscopy protein models based on neural networks, as well as storage media. It employs an end-to-end deep learning network model (referred to as CryoFold in this invention, a method or model for determining protein structure from cryo-electron microscopy density maps). By combining advanced methods based on protein sequence structure prediction, it parses atomic models of protein complex structures from low-resolution cryo-electron microscopy density maps. This method is highly accurate, easy to use, and has low requirements for image resolution, thus expanding its applicability.

[0032] On a benchmark dataset of 317 protein complexes, CryoFold... On the low-resolution density map, the TM-score reaches 0.91. On high-resolution density maps, the TM-score reached 0.95. Comparison with AlphaFold-Multimer, a sequence-based protein complex prediction method, revealed that CryoFold achieved a 25% improvement over AlphaFold-Multimer with the aid of cryo-electron microscopy density maps. Furthermore, CryoFold also demonstrated significant advantages compared to other similar methods. CryoFold will greatly accelerate the process of protein complex structure analysis, especially for uncaptured heteroconformal states and low-resolution densities in PDBs (Protein Databases), including in-situ structures.

[0033] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 A schematic diagram of the CryoFold network framework according to an embodiment of the present invention is shown;

[0036] Figure 2 A schematic diagram of the branch structure of the cryo-electron microscopy density map of the CryoFold network according to an embodiment of the present invention is shown.

[0037] Figure 3A schematic diagram of the Cryoformer encoder structure according to an embodiment of the present invention is shown;

[0038] Figure 4 A schematic diagram of the Cryoformer decoder structure according to an embodiment of the present invention is shown;

[0039] Figure 5(a) shows a schematic diagram of the cryo-electron microscopy density map processing procedure according to an embodiment of the present invention;

[0040] Figure 5(b) illustrates a schematic diagram of the EMPIAR downsampling dataset construction process according to an embodiment of the present invention;

[0041] Figure 5(c) shows a schematic diagram of the low-pass filtering data set process according to an embodiment of the present invention;

[0042] Figure 5(d) illustrates a schematic diagram of a simulated dataset process according to an embodiment of the present invention;

[0043] Figure 5(e) shows the data distribution of each dataset in each resolution range according to an embodiment of the present invention;

[0044] Figure 6 The data distribution diagrams for various resolution ranges according to embodiments of the present invention are shown;

[0045] Figure 7 A schematic diagram showing a comparison between the results of CryoFold predictions and the published structures in the database, according to an embodiment of the present invention, is provided.

[0046] Figure 8 A schematic diagram illustrating the performance metrics of CryoFold's prediction results on data at various resolution ranges according to an embodiment of the present invention is shown.

[0047] Figure 9 A comparative schematic diagram of CryoFold with other related methods according to embodiments of the present invention is shown;

[0048] Figure 10 A comparative schematic diagram of CryoFold and AlphaFold-Multimer according to an embodiment of the present invention is shown;

[0049] Figure 11 The distribution of CryoFold and AlphaFold-Multimer results on Chain-match is shown in the figure according to an embodiment of the present invention;

[0050] Figure 12 A schematic diagram illustrating the effects of CryoFold and AlphaFold-Multimer on a protein complex structure (PDB ID: 6q0t) according to an embodiment of the present invention is shown. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] This invention combines advanced 3D image recognition and protein structure prediction technologies to provide an end-to-end deep neural network, CryoFold (a cryo-folding model). CryoFold predicts protein complex structures by combining cryo-electron microscopy density maps, amino acid sequences, multiple sequence alignment (MSA), and structural templates. CryoFold includes multiple core neural network modules, referred to in this invention as Cryoformer modules, forming a Cryoformer stack (cryo-folding module stack). Utilizing CryoFold containing cryo-folding modules, and combining 3D information from the density map, evolutionary information from the MSA, and homology information from the structural template, the main chain and side chain representations of the protein complex's structural model can be effectively learned, such as... Figure 1 As shown. Furthermore, to ensure the geometric constraints between bond lengths and bond angles in the protein structure, a final structural model with three-dimensional coordinates is generated using structural modules in CryoFold.

[0053] Without loss of generality, the deep neural network of this invention can be stored in a computer storage medium, such as RAM, ROM, EEPROM, EPROM, flash memory, disk, and combinations thereof. When CryoFold is invoked, it can execute a cryo-electron microscopy density map processing method.

[0054] This invention also provides a computer system, including one or more processors and one or more non-transitory computer-readable media storing a deep neural network model, CryoFold, configured to process cryo-electron microscopy density maps. Further, the cryo-electron microscopy images are processed to obtain atomic models of the corresponding protein complex structures. The CryoFold network model of this invention can not only build protein complex models from high-resolution cryo-electron microscopy density maps, but also automatically build protein complex models from low-resolution cryo-electron microscopy density maps, expanding the application scope of automatically building models using neural networks.

[0055] The structure of the cryofold model CryoFold, an embodiment of the present invention, is described below by way of example.

[0056] like Figure 1 As shown, the CryoFold network framework includes two input branches, a cryo-electron microscopy (cryo-EM) conversion module stack, a structure module, and multiple output modules. The two input branches are the cryo-EM density map branch and the sequence branch. The cryo-EM density map branch includes a three-dimensional residual neural network used to learn amino acid information, secondary structure, and protein backbone information from the cryo-EM density map.

[0057] Sequence branching is used to learn sequence-related representations (including MSA representations and amino acid pairing representations) related to protein evolution from protein sequences. Sequence branching includes a coding module and an embedding representation learning module. The coding module encodes input information such as amino acid sequences, MSAs, and templates. The embedding representation learning module performs embedding learning on the encoded amino acid sequences, MSAs, and templates to generate MSA representations (multiple sequence representations) and amino acid pairing representations (referred to as pairing representations). Sequence branching also includes the Evoformer stack (see...). Figure 4 The Evoformer stack is used to learn MSA (Massively Interactions) and inter-amino acid (IIA) pairing characteristics. After learning, it outputs new MSA and IIA pairing characteristics. The amino acid sequence, MSA, and structural templates can be generated based on the input protein complex sequence. Preferably, in this embodiment of the invention, not only are multiple iterations performed during the generation of the protein atomic structure, but also a cryo-electron microscopy (cryo-EM) three-dimensional density map is simulated for the generated structural model. The simulated map is then added as input to the cryo-EM density map branch for iterative optimization.

[0058] The cryo-electron microscopy density map branch takes a cryo-electron microscopy density map of shape W×H×L as input. This input passes through a 3D convolutional neural network layer, a batch normalization layer, a ReLU (Revised Linear Unit) layer, and a max-pooling layer. Then, it is sequentially fed into four 3D residual convolutional modules (ResBlock). Finally, a 3D convolutional neural network layer with a kernel size of 1 maps the high-dimensional features to a low-dimensional density map representation, which serves as the first density map representation. The MaxPooling layer reduces the length, width, and height dimensions of the density map to half their original size, resulting in a shape of W / 2×H / 2×L / 2. The four residual network modules have the same architecture; the second residual network module has a stride of 2, while the other three 3D residual convolutional networks have a stride of 1. Therefore, after the second residual network module, the feature map shape becomes W / 4×H / 4×L / 4. Finally, the density map feature map is mapped to a 384-dimensional feature map by a 3D convolutional neural network layer with a kernel size of 1.

[0059] The cryoformer module stack consists of multiple (e.g., 8) Cryoformers. Cryoformer is the key module of CryoFold, and each Cryoformer includes N... enc Encoders and N dec The encoder is used to learn the global three-dimensional spatial information of amino acids from the cryo-electron microscopy density map, and the decoder is used to learn the matching between the multiple sequence characterization and the three-dimensional spatial information of the density map, that is, to cross-fuse the multiple sequence characterization output by the sequence branch and the three-dimensional spatial information output by the encoder.

[0060] The Cryoformer encoder (or simply encoder) takes the output of the cryo-electron microscopy density map processed by a 3D residual neural network, i.e., the first density map representation, as input. After flattening, it is added to the 3D positional encoding of the density map, and then input into a multi-head self-attention module. It then sequentially passes through a first LayerNorm layer, a linear layer, and a second LayerNorm layer to generate a new density map representation, i.e., the second density map representation, as shown below. Figure 3 As shown. The self-attention module enables direct information interaction between voxel points throughout the density map, thereby obtaining a global representation based on the semantic features and 3D positional information of the entire density map. This further improves the recognition rate of information such as amino acid types, secondary structures, protein backbones, protein topology, domain interactions, and the orientation of the entire protein complex from the density map. The second density map representation serves as the input to the next encoder. N enc The encoders have the same structure, but their parameters are not shared.

[0061] The Cryoformer decoder (or simply decoder) takes the output of the sequence branch, the output of the encoder, and the 3D positional encoding of the density map as input. It generates crossed single representations through self-attention modules, cross-attention modules, LayerNorm layers, and linear layers, respectively. Figure 4 As shown. To preserve the effective evolutionary information learned from the multiple sequences, each decoder layer adds the multiple sequence representation and pairing representation output from the Evoformer stack to the crossed single sequence representation after passing through a linear layer, and then adds them together after passing through a LayerNorm layer to form a new single sequence representation. That is, the new single sequence representation integrates the amino acid embedding representation, the multiple sequence representation, the pairing representation, and the crossed single sequence representation output from the previous decoder. Next, the new single sequence representation will be Q... s K s and V sThe inputs are in the form of three variables, Q, and are fed into the self-attention module. s and K s Both are the result of adding novel single-sequence characterization and amino acid embedding characterization, V s This is a novel single-sequence characterization method. To enhance the strength of amino acid type characterization, amino acid embedding characterization is also incorporated into Q. s and K s In the new single-sequence representation. Specifically, the output of the Evoformer stack is added to the amino acid embedding representation after passing through the LayerNorm layer, and the result of the addition is compared with Q. s and K s The outputs of the self-attention module, amino acid embedding characterization, encoder output, and 3D position encoding of the density map are all input into the cross-attention module for matching cryo-electron microscopy density map features and sequence features.

[0062] The cross-attention module in the Cryoformer decoder is crucial for matching the three-dimensional spatial information of sequence correlation representations (including multiple sequence representations and matching representations) and density maps in the neural network space. The cross-attention module uses Q... c K c and V c Three variables are used as inputs, where Q c K is the result of adding the output of the self-attention module to the amino acid embedding characterization. c V is the result of adding the density map representation output by the encoder to the 3D position encoding of the density map. c The density map representation is output by the encoder. Through a cross-attention module, sequence-related representations are fused with 3D spatial information from the 3D cryo-electron microscopy density map, providing the source of 3D coordinate position information for each amino acid in the sequence. This ensures that the final generated protein's all-atom coordinates are based on an atomic model derived from the cryo-electron microscopy density map. The output of the cross-attention module is added to the output of the self-attention module and then input into a LayerNorm layer (the third LayerNorm layer). The output of the third LayerNorm layer is processed by a linear layer and then superimposed with the output of the third LayerNorm layer before being input into a fourth LayerNorm layer for further processing, resulting in a new, separate cross-representation of the sequence.

[0063] The CryoFold network model employs multiple loss functions for end-to-end training. These loss functions are associated with tasks including density map-based amino acid type recognition, density map-based secondary structure type recognition, density map-based amino acid semantic segmentation, mask recognition for multiple sequence alignment, residue distance prediction, regression of all-atom coordinates, prediction of side chain torsion angles, and prediction of inter-atom collisions, among others. Correspondingly, the density map-based amino acid type recognition uses the cross-entropy loss Lc for amino acid type recognition. CLS The density map-based secondary structure type identification uses a loss function called cross-entropy loss L for secondary structure type identification. SS The loss function used in the density map-based amino acid semantic segmentation is the cross-entropy loss L of amino acid semantic segmentation. seg The loss function used for mask recognition in multi-sequence alignment is the cross-entropy loss L of mask recognition. MSA The loss function used for residue distance prediction is the cross-entropy loss L of residue distance prediction. dist The loss function used for full atomic coordinate regression is the Frame Aligned Point Error (FAPE) loss associated with full atomic coordinate regression. FAPE and root mean square error loss L RMSD Protein main chain frame prediction uses FAPE loss L FAPE-BF and root mean square error loss L RMSD-BF The loss function used for side chain torsion angle prediction is the loss L for side chain torsion angle prediction. angle The loss function used for inter-atomic collision prediction is the inter-atomic collision prediction L. clash In addition, a correlation loss Li based on the predicted structure simulation density map and the input density map was also adopted. density .

[0064] The expression for the loss function above can be obtained from existing technology and will not be elaborated further.

[0065] The CryoFold training process consists of three stages. The first stage aims to learn the mask features of multiple sequence alignments. Therefore, the cross-entropy loss for mask recognition in multiple sequence alignments has a larger weight, while the two root mean square error losses have smaller weights. Empirically, the weight of the cross-entropy loss for mask recognition in multiple sequence alignments is 160, while the root mean square of the protein main chain frame prediction and the root mean square of the regression of all-atom coordinates are 0.1. Furthermore, to ensure training stability, the weights of the side chain torsion angle prediction loss, the inter-atom collision loss, and the density map correlation loss are all 0. The training process ends when the cross-entropy loss of the mask features of multiple sequence alignments reaches a stable state. The total loss function for the first stage is as follows:

[0066] L = LCLS +L SS +L dist +L seg +160L MSA +0.1L RMSD +L FAPE +0.1L RMSD-BF +L FAPE-BF

[0067] The goal of the second stage is to train the coordinates of the main chain atoms of the protein structure; therefore, the two root mean square error losses are set to 1.0. The total loss function for the second stage is as follows:

[0068] L = L CLS +L SS +L dist +L seg +160L MSA +L RMSD +L FAPE +L RMSD-BF +L FAPE-BF

[0069] The third stage aims to train for accurate prediction of the entire atomic structure. Therefore, three loss functions are added: side chain torsion angle prediction, inter-atom collisions, and density map correlation. Based on experience, the weights for these three are 1.0, 0.1, and 1.0, respectively. The total loss function for the third stage is as follows:

[0070] L = L CLS +L SS +L dist +L seg +160L MSA +L RMSD +L FAPE +L RMSD-BF +L FAPE-BF +L angle +0.1L clash +L density

[0071] For example, training was performed using 24 NVIDIA 40G A100 GPUs, with the three phases taking 3 days, 7 days, and 30 days, respectively. Adam was used as the optimizer with an initial learning rate of 0.001, and the learning rate was decayed in a stepwise manner, decreasing by an order of magnitude every 10,000 steps.

[0072] The cryo-electron microscopy density map processing method of this invention uses the aforementioned CryoFold model to process the cryo-electron microscopy density map to obtain intermediate or final output products of a deep neural network. Without loss of generality, this invention also provides a method for building a cryo-electron microscopy protein model based on a neural network, which uses the aforementioned deep neural network to process the cryo-electron microscopy density map to obtain an atomic model of the corresponding protein complex structure. Both the cryo-electron microscopy density map processing method and the neural network-based cryo-electron microscopy protein model building method further include data processing and model training before using CryoFold.

[0073] The data processing and model training process is illustrated below as an example.

[0074] Cryo-electron microscopy 3D density maps and corresponding published atomic models were collected from the EMDB and PDB databases, respectively. These samples can be filtered out in the following cases:

[0075] The release date is after the specified date;

[0076] The resolution of the PDB structure is greater than

[0077] The reconstruction method is not based on single-particle cryo-electron microscopy (SPA) analysis.

[0078] No valid protein sequence was found. A valid protein sequence is defined as one that is at least 25 amino acids long and contains fewer than 30% unknown residues.

[0079] The correlation coefficient between the density maps and atomic models was less than 0.5. The experimental analysis dataset consisted of 9150 cryo-electron microscopy 3D density maps. In this example, 20 density maps without protein molecules, 30 density maps with multiple related atomic structures, and 123 density maps that showed poor structural consistency with the atomic structures during manual inspection were removed. After this process, 8977 density maps were retained. For some symmetric proteins, the .cif file from the PDB contains only the atomic coordinates of one asymmetric unit. Therefore, ChimeraX was used to apply a symmetry operation (_pdbx_struct_oper_list) based on a symmetric unit to obtain a .pdb format file containing all atomic coordinates.

[0080] Since the original cryo-electron microscopy density maps are typically much larger than the bounding box of the structural model, for density maps outside the structural model, the structural model (with asymmetry applied) is first clipped using Phenix.map_box to reduce the size of the density map. The cryo-electron microscopy density map is then reshaped to a specific voxel size of 0.6667 using spline interpolation. The density values ​​are then normalized to [0,2] based on interval partitioning. All density map samples are saved as .mrc files. The processing procedure for cryo-electron microscopy density maps is shown in Figure 5(a).

[0081] The sequence data processing procedure is as follows: The first step in running CryoFold is to process the input sequence. This process includes taking one or more sequences as input and generating input features. In this embodiment, the AlphaFold2 data flow is used to generate features for each chain's sequence, and all sequences from 8977 structural models are processed according to the following description. The specific data processing flow can be described as follows:

[0082] Multiple sequence alignments (MSA) are searched from sequence databases. HHblits is used to search the BFD and UniRef30 (2020_02 version) databases. JackHMMER is used to search the UniRef90, MGnify, Metaeuk, and MGY databases. For homologous sequences from different sources, they are sorted according to their similarity to the query sequence, and duplicate sequences are removed from the MSA.

[0083] Homologous templates were searched from PDB70. Using hhsearch with UniRef90's MSA profile as input, the PDB70 database was searched to obtain PDBIDs and chain IDs. The corresponding mmCIF files were then retrieved from a pre-prepared local PDB database. The chain sequences were aligned with the mmCIF residues according to the canonical sequence to resolve atomic 3D coordination (dimensions [chain length, 37, 3]) and masks (dimensions [chain length, 37]). Template residue types, atomic positions, and atomic masks were extracted for use as template features in the following analyses. A maximum of 20 templates were retained for subsequent analyses.

[0084] Combining Multiple Chains: If multiple chains exist in the sample, features from each chain are combined. Features with chain length as the first dimension are directly concatenated, including aatype, residual_index, between_segment_residues, seq_length, sequence, and num_alignments. For features with ordinal number as the first dimension, including msa and deletion_matrix_int, the first dimension is first filled with the number of maximum values ​​with zero ordinal numbers. Then, these two features from different chains are concatenated along the second (chain length) dimension. Template features are processed similarly to MSA features. The first dimension (≤20), the number of templates, is first filled with the number of maximum template numbers across all chains. After concatenation along the chain length dimension, all features from the 8977 samples are saved as a compressed pickle file.

[0085] Splitting the Training and Validation Sets: The purpose of splitting the training and validation sets is to divide the 8977 density maps into two groups: a training dataset and a test dataset with low homology. First, 40% sequence identity clustering files are downloaded from the RCSB PDB database. These clustering files define many chain clusters with sequence identity greater than 40%. To construct the test set, one PDB model (potentially containing multiple chains) is randomly selected from all samples each time, and any other PDB model with chains having a sequence identity greater than 40% with the sampled PDB model is also added to the test set. This process is repeated until the test data size reaches 317. Finally, the training set consists of 8660 density maps and PDB pairs, used to train the CryoFold model. The test set consists of 317 density maps and PDB pairs, used to evaluate the CryoFold model. Figure 6 As shown.

[0086] Data Augmentation: To improve the performance of the CryoFold model, this embodiment of the invention performs three augmentation methods on the training set. The first is the EMPIAR downsampling dataset, which downsamples the two-dimensional particle images of each EMDB density map to reconstruct multiple density maps at a lower resolution. The second is the low-pass filtered dataset, which uses high-resolution (higher than) data from the EMDB (Electron Microscopy Database) to... The density maps are low-pass filtered to achieve multiple resolution levels. The third is a simulated dataset, which simulates cryo-electron microscopy density maps of protein complexes that do not have density maps in the PDB dataset.

[0087] The steps for constructing the EMPIAR (Electron Microscopy Public Image Archive) downsampling dataset are shown in Figure 5(b). Eighty-eight image datasets were extracted from EMPIAR, and the corresponding data processing procedures were reproduced. These procedures were obtained from relevant papers in the archive. In this embodiment, a total of 112 density maps were reconstructed, with the number of grain images ranging from 14,262 to 730,118.

[0088] The particle images were resampled multiple times, and a new density map was reconstructed using each subset. These density maps all have the same atomic structure model as the original density map.

[0089] In this embodiment of the invention, images with resolutions lower than [specified resolution] are discarded. The density maps were generated, and a dataset consisting of 19,887 density maps and 112 atomic structures was compiled.

[0090] Low-pass filtered dataset: High resolution (higher than) in EMDB The density map is subjected to a low-pass filter. This embodiment of the invention employs the low-pass filtering method in RELION (a software for cryo-electron microscopy 3D reconstruction) and user-defined parameters to perform low-pass filtering on the high-resolution data, including using different thresholds. The data is processed to generate a large amount of low-resolution data, and then cropped and resized to reshape the voxel size, as shown in Figure 5(c).

[0091] Simulated Dataset: In the PDB, approximately 90% of the protein complex structures were obtained via X-ray methods, most of which lack cryo-electron microscopy density maps. To train on this large dataset, over 100,000 cryo-electron microscopy density maps were simulated on these PDBs. Multiple simulations were performed within the resolution range to obtain a large amount of simulated density map data, as shown in Figure 5(d).

[0092] Ultimately, the four datasets, totaling more than 500,000 samples, formed a complex dataset spanning multiple resolutions, as shown in Figure 5(e).

[0093] The trained CryoFold model can directly perform whole-atom model inference by inputting cryo-electron microscopy density maps and protein complex sequences. For example... Figure 7As shown, the protein complex structure is a model constructed by CryoFold based on the cryo-electron microscopy density map (EMD-7770) generated in the experiment. It can be seen from the figure that CryoFold's results match very well with the structure published in the PDB data (PDB: 6cvm), and the side chains also match the density map very well.

[0094] like Figure 8 As shown, on the benchmark dataset of 317 protein complexes, the data points to the left of the dashed line are high-resolution. The sample size is 138, representing the high-resolution dataset. To the right of the gray dashed line are the data points representing lower resolution. The dataset contains 179 samples, representing a low-resolution dataset. It can be seen that CryoFold achieves high performance on both the template-alignment-required metric TM-score and the template-match-free metric Chain-match. Specifically, it performs well on the low-resolution dataset. On the cryo-electron microscopy density map, the average TM-score of the CryoFold prediction results was 0.91. At high resolution... The figure shows that CryoFold's average TM-score for predictions is 0.95. Furthermore, on the chain-match metric, the average chain-match scores for the high-resolution and low-resolution datasets are 0.92 and 0.87, respectively. It can also be seen that CryoFold performs well at high resolutions. The protein backbone built on the dataset C α The root mean square error is In low resolution Dataset on These results demonstrate that CryoFold can accurately construct atomic models of protein complexes from low-resolution cryo-electron microscopy density maps.

[0095] On the aforementioned low-resolution cryo-electron microscopy density map dataset, comparisons were made with commonly used methods such as Phenix, DeepTracer, and ModelAngelo. Figure 9 The results show that CryoFold outperforms other methods. On the chain-match metric, CryoFold's average score is 0.87, while Phenix's is 0.03 and ModelAngelo's is 0.41. For Seq-match (sequence matching score), which assesses the accuracy of amino acid type matching, the target type of amino acid is found within a radius of... In terms of the seq-match (within the ball) metric, CryoFold also outperforms these methods, with an average seq-match of 0.94, far exceeding Phenix's 0.05, ModelAngelo's 0.43, and DeepTracer's 0.40.

[0096] To compare with AlphaFold-Multimer, embodiments of this invention tested protein complexes with 174 residues less than 2500. Figure 10 As shown, the results indicate that CryoFold outperforms AlphaFold-Multimer across all metrics, including Chain-match, TM-score, and GDT-TS. CryoFold's average Chain-match is 0.85, TM-score is 0.87, and GDT-TS is 0.73, while AlphaFold-Multimer's average Chain-match is 0.36, TM-score is 0.57, and GDT-TS is 0.31. Figure 11 As can be seen from the scatter plot, cryo-electron microscopy density mapping significantly improves the accuracy of CryoFold in constructing protein complex structures. Furthermore, CryoFold's accuracy varies with the resolution of the cryo-electron microscopy density map; higher resolution results in higher accuracy. While AlphaFold2 can accurately predict the structures of most single-chain proteins, its performance in predicting protein complexes still has considerable room for improvement. CryoFold, by simultaneously inputting the cryo-electron microscopy density map and the sequence into a neural network, can accurately construct atomic models of protein complexes.

[0097] Figure 12 The example demonstrates CryoFold and AlphaFold-Multimer in one instance (EMD: 20552). The effect on the protein complex structure (PDBID: 6q0t) is shown. It consists of 5 protein chains and 1322 modeling residues. The resolution of the cryo-electron microscopy density map is [resolution missing]. The prediction results of AlphaFold-Multimer and CryoFold are 0.104 and 0.791 on the TM-score, and 0.400 and 0.783 on the Chain-match. It can be seen that CryoFold shows a significant advantage over AlphaFold-Multimer.

[0098] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A computer storage medium, characterized in that, It stores deep neural networks. The deep neural network includes a cryoformer module stack, which includes multiple cryoformer modules. Cryoformer includes encoders and decoders; The decoder is used to learn the matching of three-dimensional spatial information of sequence correlation representation and cryo-electron microscopy density map, and cross-fuses the three-dimensional spatial information output by the sequence correlation representation and the encoder. The decoder takes the output of the sequence branch of the deep neural network, the output of the encoder, and the three-dimensional position encoding of the cryo-electron microscopy density map as input, and generates a cross-sequence representation through a self-attention module and a cross-attention module. Each Cryoformer includes N enc Encoders and N dec One decoder; Sequence-related characterization includes multiple sequence characterization and amino acid pairing characterization; Each decoder adds the multiple sequence representation and the amino acid pairing representation to the crossed single sequence representation after passing through a linear layer, and then adds them to each other after passing through a LayerNorm layer to form a new single sequence representation; The new single-sequence representation is input into the self-attention module; The output of the self-attention module, the amino acid embedding characterization, the encoder output, and the 3D position encoding of the density map are all input into the cross-attention module to match cryo-electron microscopy density map features and sequence features. Cross-attention module with Q c K c and V c Three variables are used as inputs, where Q c K is the result of adding the output of the self-attention module to the amino acid embedding characterization. c V is the result of adding the density map representation output by the encoder to the 3D position encoding of the density map. c This represents the density map output of the encoder.

2. The computer storage medium according to claim 1, characterized in that, The output of the cross-attention module is added to the output of the self-attention module and then input into the third LayerNorm layer. The output of the third LayerNorm layer is processed by a linear layer and then superimposed with the output of the third LayerNorm layer. This superposition is then input into the fourth LayerNorm layer for further processing, and a new, separate sequence representation of the cross is output.

3. The computer storage medium according to claim 1 or 2, characterized in that, The deep neural network employs the CryoFold model, which incorporates Cryoformer. CryoFold includes the sequence branching described above for learning sequence-related characterizations of protein evolution from protein sequences, including multiple sequence characterization and amino acid pairing characterization.

4. The computer storage medium according to claim 3, characterized in that, The sequence branch includes an encoding module and an embedding representation learning module; The encoding module is used to encode amino acid sequences, multiple sequence alignment (MSA) data, and structural templates. The embedding characterization learning module is used to learn the embedding of encoded amino acid sequences, MSAs, and structural templates to generate multiple sequence characters and amino acid pairing characters. The sequence branch also includes the Evoformer stack, which is used to learn multiple sequence representations and amino acid pairing representations, and output new multiple sequence representations and amino acid pairing representations.

5. The computer storage medium according to claim 1 or 2, characterized in that, The deep neural network employs the cryo-folding model CryoFold, which includes Cryoformer. CryoFold includes a cryo-electron microscopy density map branch, which includes a three-dimensional residual neural network used to map high-dimensional features into low-dimensional density map representations.

6. The computer storage medium according to claim 5, characterized in that, The cryo-electron microscopy density map branch takes the cryo-electron microscopy density map as input, passes through a 3D convolutional neural network layer, a batch normalization layer, a modified linear unit (ReLU), and a max pooling layer, and then sequentially inputs it into four 3D residual convolutional modules. Finally, it is processed by a 3D convolutional neural network layer before being output.

7. A computer system, characterized in that, include: One or more processors and one or more non-transitory computer-readable media, the storage of which is configured to process cryo-electron microscopy density maps of a deep neural network as described in any one of claims 1-6.

8. A method for processing cryo-electron microscopy density maps, characterized in that, The cryo-electron microscopy density map is processed using a deep neural network as described in any one of claims 1-6.

9. A method for constructing a cryo-electron microscopy protein model based on a neural network, characterized in that, include: By processing the cryo-electron microscopy density map using a deep neural network as described in any one of claims 1-6, an atomic model of the corresponding protein complex structure is obtained.

Images