A method for constructing a protein phase separation characteristic detection model and application

Abstract

The application discloses a kind of protein phase separation characteristic detection model construction method and application, belong to protein phase separation characteristic detection technical field;RNN network is used as main network, and combined with feature local attention mechanism to help model pay more attention to important position, and the features of multiple levels are fused and spliced in the way of residual connection, to fully learn the features of amino acid sequence;On this basis, considering that each protein amino acid sequence does not present equal length distribution, before inputting to model, first fill each protein amino acid sequence in batch to maximum amino acid sequence length, and in order to reduce the influence of padding value on result, before using RNN network and attention mechanism to extract features, first according to sequence length, remove padding by compression, after extracting features, fill back in reverse, to ensure that the dimension of input and output is same, to realize the fusion of feature information in different stages, greatly improve the accuracy of model detection.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of protein phase separation property detection, and more particularly relates to a method for constructing a protein phase separation property detection model and application thereof. BACKGROUND

[0002] Cells, as the basic units of biological structure and function, carry out various tasks to maintain normal operation of life. Inside the cell, various biological macromolecules such as proteins, DNA, RNA, etc. play a crucial role in the complex biochemical reactions and regulatory networks. Various structures exist inside the cell, and the environment is chaotic and disordered. In order to enable the components inside the cell to participate in normal activities at the correct time and space, the cell has evolved a series of organelles, including membrane-enclosed organelles (such as mitochondria, nucleus, lysosomes, etc.) and non-membrane-enclosed organelles (nucleolus, etc.). Unlike the physiological mechanism of biological macromolecules outside the cell, which rely on random collision in liquid to react, the process of cell internal activity depends on the separation of membrane organelles and non-membrane compartments to be completed smoothly. Due to the presence of the membrane on the membrane organelle, the required substances can be naturally wrapped and absorbed inside to perform functions, while many irrelevant substances are isolated outside the membrane, which greatly facilitates the normal operation of internal reaction and regulation mechanism. However, for other reactions inside the cell, due to the lack of compartmentalization of membrane organelles, how biological macromolecules avoid interference from other substances and successfully complete normal reactions in a chaotic environment has become an important research direction in the biological community.

[0003] Currently, some studies have shown that the problem of how to generate compartments in non-membrane regions to facilitate normal cell reactions is based on the liquid-liquid phase separation mechanism to allow biological molecules inside the cell to gather together, thereby bringing a certain order to the internal environment of the cell. In the cell, several weakly interacting molecules such as proteins, DNA, and RNA, etc. gather together through internal forces, and when the concentration of molecules reaches a certain level, phase separation occurs. Similar molecules gather together to speed up the reaction, and irrelevant substances are excluded to reduce interference.

[0004] Due to the existence of the liquid-liquid phase separation mechanism, the substances inside the cell are separated from each other, and various cell activities can be carried out in an orderly manner, which is also helpful for the normal physiological function of the cell. However, the liquid-liquid phase separation process is not long-lasting and stable, and once an abnormality or excessive situation occurs, it is likely to change from the original droplet structure to a more stable and more difficult to reverse gel-like structure, thereby causing disorder of the normal physiological function of the cell, and even inducing serious diseases. In view of the important role of the liquid-liquid phase separation mechanism in the normal physiological function of the cell, it is of great research value to detect the phase separation properties of biological macromolecules, especially proteins, and locate the positions in the amino acid sequence that play a key role.

[0005] The detection of early protein phase separation characteristics is completed by relying on biochemical experiments. Researchers judge the phase separation characteristics of proteins according to the biochemical experimental results of the proteins. This method realizes the detection of protein phase separation characteristics to some extent, provides valuable data resources and solid experimental basis for subsequent research. However, since this method needs to conduct separate experiments on each protein, the experimental process is tedious and requires a large investment of material and financial resources. In addition, the environment in which the protein is placed under each experiment may not be the same as the environment provided in the human body, which will also bring errors to the true experimental results.

[0006] With the continuous progress of the protein phase separation characteristic detection technology by using biochemical experiments, researchers have gradually summarized many biological characteristics of phase separation proteins according to experimental phenomena. Researchers can give a more accurate judgment result of the phase separation characteristics of the target protein according to the biological characteristics of the protein. Compared with the early detection method relying on biological experiments, this method is more convenient, does not need to conduct separate experiments, is more efficient, and has certain improvement in recognition effect. However, this method relies solely on a certain biological characteristic in the recognition process, and each biological characteristic inevitably has certain use limitations, which also leads to the limitation of the use of this method and cannot be successfully applied to any protein.

[0007] Due to the rise of artificial intelligence technology, the introduction of machine learning algorithms into the field of protein phase separation detection has also attracted widespread attention and research from researchers. At present, the protein phase separation recognition method based on machine learning is to combine the existing biological characteristics of proteins and the protein sequence itself, and to train using the existing machine learning algorithm. This method fuses the information of the amino acid sequence, statistical characteristics and various biological characteristics of the protein as the model input, and has a significant improvement in recognition accuracy. However, this method is not an end-to-end model, and still needs to manually extract the sequence information and biological characteristics of the protein before using the model for detection, and the performance is not efficient enough.

[0008] Although these methods have improved the detection ability to some extent, these methods are dependent on prior biological knowledge to some extent. Therefore, it is very necessary to support the experience of experts, and it has a high requirement for professional knowledge reserve. In addition, these methods are not an end-to-end model, and involve multiple processes when used, which is not convenient and the efficiency still needs to be further improved. Since the amino acid sequence of the protein itself contains sufficient information and there are data characteristics of different lengths, the above methods do not fully consider the data characteristics for algorithm design, or directly ignore the information contained in the amino acid sequence of the protein itself, or only convert the amino acid sequence of the protein of different lengths into the same length as the input. This has lost the information represented by the amino acid sequence of the protein itself to some extent, and further limits the improvement of the detection performance of the model, and the detection accuracy is low. SUMMARY

[0009] In view of the above defects or improvement needs of the prior art, the present application provides a method for constructing a protein phase separation characteristic detection model and an application, to solve the technical problem of low accuracy of the prior art protein phase separation characteristic detection.

[0010] In order to achieve the above-mentioned purpose, in a first aspect, the present application provides a method for constructing a protein phase separation characteristic detection model, comprising: inputting a batch set into a protein phase separation characteristic detection model for training;

[0011] Wherein, the batch set includes N protein sample sequences, and is constructed based on the amino acid sequences of the N proteins: for the amino acid sequences of the N proteins, the longest amino acid sequence is taken as a benchmark, and the remaining amino acid sequences are padded at the end to obtain a batch of sample sequences with a length of l max The longest amino acid sequence; each amino acid is assigned a corresponding number, and different amino acids correspond to different numbers; the i-th value in the amino acid sequence is the number corresponding to the i-th amino acid in the protein; i=1,2,…,l max ; N≥2;

[0012] The protein phase separation characteristic detection model comprises:

[0013] An embedding layer for encoding the N sample sequences in the input batch set respectively to obtain N encoding vector sequences;

[0014] The variable-length RNN module is used for overall compression of the N encoded vector sequences, and after removing the encoded vectors at the padding positions, the encoded vectors are input into an RNN network to extract context feature information, so as to obtain a compressed feature sequence; the compressed feature sequence is inversely padded to obtain a first feature sequence of each sample sequence, which has the same size as the corresponding encoded vector sequence; the first feature sequence of each sample sequence is spliced with the encoded vector sequence in the feature dimension to obtain a corresponding intermediate feature sequence;

[0015] The variable-length attention module is used for removing the feature items corresponding to the padding values in each intermediate feature sequence, performing channel-level global pooling operation, inputting the pooling operation result into an attention unit to obtain an influence score of each channel, multiplying each intermediate feature sequence with the corresponding influence score according to the corresponding channel to obtain a corresponding second feature sequence, and superimposing each second feature sequence with the corresponding intermediate feature sequence in a residual manner to obtain a corresponding comprehensive feature sequence.

[0016] The classification module is used for mapping each comprehensive feature sequence to determine whether the corresponding protein has phase separation characteristics.

[0017] Further preferably, the variable-length RNN module is used for splicing the encoded vectors at the same positions in the N encoded vector sequences respectively, removing the encoded vectors at the padding positions, obtaining a plurality of spliced vectors, and inputting the spliced vectors into an RNN network to extract context feature information, so as to obtain a compressed feature sequence; the compressed feature sequence is inversely decomposed and inversely padded to obtain N first feature sequences with the same dimension as the encoded vector sequence.

[0018] Further preferably, the variable-length RNN module is used for splicing the encoded vectors at the same positions in the N encoded vector sequences respectively, removing the encoded vectors at the padding positions, obtaining J spliced vectors which are not empty; the J spliced vectors are respectively output to J RNN units of the RNN network to obtain a compressed feature sequence composed of J compressed features with the same size as the corresponding spliced vectors; the compressed feature sequence is split, the features corresponding to the same encoded sequence in the compressed feature sequence are recombined according to the original positions, and inversely padded to obtain N first feature sequences with the same dimension as the encoded vector sequence; J≥1.

[0019] Further preferably, the RNN network is a BiLSTM network; and the RNN unit is a BiLSTM unit.

[0020] Further preferably, the attention unit is an ECA network.

[0021] In a second aspect, the present application provides a protein phase separation characteristic detection method, comprising:

[0022] inputting a to-be-tested batch set into the protein phase separation characteristic detection model to determine whether each protein in the to-be-tested batch set has the phase separation characteristic;

[0023] The protein phase separation characteristic detection model is constructed by using the method for constructing a protein phase separation characteristic detection model provided in the first aspect of the present application.

[0024] The to-be-tested batch set includes sample sequences of a plurality of proteins, and the method for obtaining the sample sequences includes: padding the end of each amino acid sequence in the batch set based on the longest amino acid sequence to obtain a batch of sample sequences with the same length as the longest amino acid sequence in the to-be-tested batch set; each amino acid is assigned a corresponding number, and different amino acids correspond to different numbers; the i-th value in the amino acid sequence is the number corresponding to the i-th amino acid in the protein; i = 1, 2, …, L; and L is the length of the longest amino acid sequence in the to-be-tested batch set.

[0025] In the second aspect, the present application provides a protein phase separation characteristic detection system, which includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to execute the protein phase separation characteristic detection method provided in the first aspect of the present application.

[0026] In the third aspect, the present application further provides a computer readable storage medium, which includes a stored computer program, wherein when the computer program is run by a processor, the device where the storage medium is located is controlled to execute the method for constructing a protein phase separation characteristic detection model provided in the first aspect of the present application and / or the protein phase separation characteristic detection method provided in the second aspect of the present application.

[0027] Overall, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0028] 1. This invention provides a method for constructing a protein phase separation characteristic detection model. An end-to-end model is constructed, using an RNN network as the backbone to extract contextual information from amino acid sequences. A feature local attention mechanism is combined to help the model focus on important positions. Residual connections are used to fuse and splice features from multiple levels to fully learn the features of the amino acid sequences, resulting in a comprehensive and accurate representation of the features. Furthermore, considering that the amino acid sequences of proteins in a batch are not of equal length, the amino acid sequences in the batch are padded to the maximum amino acid sequence length before being input into the model. To reduce the impact of the padding value on the results and better utilize the features inherent in the protein's amino acid sequence itself, the feature extraction part of the model includes a variable-length RNN module and a variable-length attention module. Before using the RNN network and attention mechanism for feature extraction, the padded sequence is compressed according to its length. After feature extraction, the padded sequence is backfilled to ensure that the input and output dimensions are the same, thus achieving the fusion of feature information at different stages. This invention, by fully utilizing the features inherent in the protein's amino acid sequence itself, achieves a comprehensive and accurate representation of the features, greatly improving the accuracy of protein phase separation characteristic detection.

[0029] 2. Furthermore, the method for constructing the protein phase separation characteristic detection model provided by the present invention involves a variable-length RNN module that, during the overall compression of N coding vector sequences, concatenates the coding vectors at the same positions in the N coding vector sequences and removes the coding vectors at the padding positions to obtain multiple concatenated vectors. These concatenated vectors are then sequentially input into each RNN unit of the RNN network for contextual feature extraction. This method avoids interference from padding values ​​on the phase separation characteristics of the original sequence, ensuring that the input to the model is a sequence of true length. It also guarantees that the sequence can still follow the original arrangement order during compression, facilitating the model to extract feature information at each position of the sequence and sequentially expand the compressed sequence back to the original dimension.

[0030] 3. Furthermore, in the method for constructing the protein phase separation characteristic detection model provided by the present invention, the RNN network is a BiLSTM network. The BiLSTM structure is very good at extracting features from sequences, and its bidirectional structure ensures that the model can cover more contextual information in the process of extracting features from sequences. At the same time, the model does not rely on the support of large amounts of data, making it very suitable for extracting sequence features from small amounts of data.

[0031] 4. Furthermore, in the method for constructing the protein phase separation characteristic detection model provided by the present invention, the attention unit is an ECA network. This network uses a lightweight, non-dimensionality-reduction local cross-channel interaction strategy, which involves only a small number of parameters and can bring about performance improvement. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 Structure diagram of a protein phase separation property detection model provided for embodiment 1 of the present application;

[0033] Figure 2 Encoding flowchart of an embedding layer provided for embodiment 1 of the present application;

[0034] Figure 3 Schematic diagram of overall compression of N encoding vector sequences by a variable-length RNN module provided for embodiment 1 of the present application;

[0035] Figure 4 Structure diagram of an LSTM unit module provided for embodiment 1 of the present application;

[0036] Figure 5 Structure diagram of a BiLSTM network provided for embodiment 1 of the present application;

[0037] Figure 6 Structure diagram of a variable-length attention module provided for embodiment 1 of the present application;

[0038] Figure 7 Schematic diagram of channel-level global pooling operation in a variable-length attention module provided for embodiment 1 of the present application. DETAILED DESCRIPTION

[0039] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0040] Embodiment 1

[0041] A method for constructing a protein phase separation property detection model, comprising: inputting a batch set into the protein phase separation property detection model for batch training;

[0042] Wherein, the batch set includes sample sequences of N proteins, and is constructed based on amino acid sequences of the N proteins: for the amino acid sequences of the N proteins, in order to ensure that the model can be trained in batches, the longest amino acid sequence is taken as a benchmark, and padding is performed at the end of the remaining amino acid sequences to obtain a batch of sample sequences with a length of l max Each amino acid is assigned a corresponding number, and different amino acids correspond to different numbers; the i-th value in the amino acid sequence is the number corresponding to the i-th amino acid in the protein; i = 1, 2, …, l maxN≥2;

[0043] Specifically, the number of amino acid categories in the statistical data set is counted, and an "amino acid letter-number" mapping table representing the letters and corresponding numbers of each category of amino acid is obtained, according to which the protein amino acid letter sequence is converted into a digital sequence to obtain the above-mentioned amino acid sequence. In this embodiment, the proteins in a batch set (one batch) are sorted from long to short according to the length of the amino acid sequence at the input end of the model, and the amino acid sequences of other proteins are padded to the same length by zero padding at the end, and the true length of each protein amino acid sequence is recorded.

[0044] It should be noted that the amino acid sequence of the protein is selected as the input of the model. Since existing protein phase separation characteristic recognition methods more or less involve biological knowledge, the reliability of biological knowledge directly determines the performance of the model, which also leads to the fact that the model is highly dependent on expert experience. In order to minimize the excessive dependence on expert experience, the present application directly extracts the internal characteristic information contained in the amino acid sequence of the protein.

[0045] The present application designs an end-to-end model structure, which is more efficient and has higher accuracy. Specifically, the protein phase separation characteristic detection model constructed in this embodiment is referred to as the RECA model, which specifically includes:

[0046] An embedding layer is used to encode N sample sequences in the input batch set respectively to obtain N encoding vector sequences;

[0047] A variable-length RNN module is used to compress the N encoding vector sequences as a whole, remove the encoding vectors at the padding positions, and input them into the RNN network to extract context feature information to obtain a compressed feature sequence. The compressed feature sequence is inversely padded to obtain a first feature sequence of each sample sequence, which has the same size as the corresponding encoding vector sequence. The first feature sequence of each sample sequence is spliced with its encoding vector sequence in the feature dimension to obtain a corresponding intermediate feature sequence (including three dimensions: batch size, length, and feature dimension). The first feature sequence and the corresponding encoding sequence are spliced in the feature dimension to obtain the intermediate feature sequence, which fuses the feature information at different stages. In an optional implementation, since the conventional pooling function used in the channel-level global pooling operation in the subsequent variable-length attention module can only perform pooling operation on the feature dimension, the length and feature dimension in the intermediate feature sequence are exchanged by transposition operation to use the conventional pooling function to perform pooling operation on the length dimension.

[0048] In an optional embodiment, the variable-length RNN module respectively concatenates the encoding vectors at the same positions in the N encoding vector sequences, removes the encoding vectors at the padding positions, obtains a plurality of concatenated vectors, and inputs the plurality of concatenated vectors into the RNN network to extract context feature information, to obtain a compressed feature sequence; and the compressed feature sequence is inversely disassembled and inversely padded to obtain N first feature sequences with the same dimension as the encoding vector sequence. Specifically, the variable-length RNN module respectively concatenates the encoding vectors at the same positions in the N encoding vector sequences, removes the encoding vectors at the padding positions, and obtains J non-empty concatenated vectors; the J concatenated vectors are respectively output to J RNN units of the RNN network, to obtain a compressed feature sequence composed of J compressed features with the same size as the corresponding concatenated vectors; the compressed feature sequence is disassembled, the features corresponding to the same encoding sequence in the compressed feature sequence are recombined according to the original positions, and inverse padding is performed, to obtain N first feature sequences with the same dimension as the encoding vector sequence; J is greater than or equal to 1.

[0049] The present application removes the influence of the padding value by using compression and decompression: the encoding vector sequence in a batch is removed from the end of the sequence length dimension and compressed into a low-dimensional encoding vector sequence. Specifically, one timestep corresponds to one position of the amino acid, and the encoding vector sequence in a batch follows the original arrangement order and according to the recorded protein amino acid sequence length, the single amino acid at the corresponding position is spliced and filtered one by one to obtain a compressed low-dimensional feature sequence. At the same time, the number of vectors corresponding to the compressed sequence at each timestep of the original sequence is counted. The compressed feature vector sequence is expanded and padded in reverse according to the number of amino acids at the corresponding position to obtain a feature sequence with the same data dimension as the output of the embedding layer. Specifically, according to the number of amino acids at the corresponding position, the compressed feature vector sequence is re-split, reversely spliced in the original order, and padded at the end to generate a vector sequence with the same data dimension as before compression.

[0050] The variable-length attention module is used to remove the feature items corresponding to the padding values in each intermediate feature sequence, perform channel-level global pooling operation, input the pooling operation result into the attention unit to obtain the influence score of each channel, multiply each intermediate feature sequence with the corresponding influence score according to the corresponding channel to obtain the corresponding second feature sequence, and superimpose each second feature sequence with the corresponding intermediate feature sequence in a residual manner to obtain the corresponding comprehensive feature sequence.

[0051] The classification module is used to map each comprehensive feature sequence to determine whether the corresponding protein has phase separation property.

[0052] It should be noted that the RNN network in the above variable-length RNN module can be an LSTM network, a BiLSTM network, a GRU network, a BiGRU network, etc. Preferably, the RNN network is a BiLSTM network, and the bidirectional structure of the BiLSTM network ensures that more context information can be covered in the process of extracting features from the sequence, and at the same time, the model does not rely on large data support, which is very suitable for small data sequence feature extraction. The RNN unit is a BiLSTM unit. In an optional implementation, the variable-length RNN module removes the tail padding value and compresses the encoding vector sequence in a batch into a low-dimensional encoding sequence according to a single timestep, and according to the sequence length, the number of vectors corresponding to the original sequence at each timestep and the compressed sequence is counted; the BiLSTM network extracts context feature information from the positive and negative directions according to the low-dimensional encoding vector sequence and the number of vectors corresponding to the original sequence at each timestep, and outputs a compressed feature sequence of the same data format; the compressed feature vector sequence is reversely filled according to the corresponding position amino acid number information to obtain a feature sequence with the same data dimension as the encoding sequence; the encoding sequence and the feature sequence are spliced in the feature dimension, and the last two dimensions are transposed to obtain an intermediate feature sequence.

[0053] Further, in the process of extracting sequence features, the feature dimension of each position of the sequence will be very high, and directly screening valuable features from it may be too rough. The application further adopts a channel attention mechanism to enable the model to focus on important parts of the sequence, and the fusion operation of the attention can enable the information of multiple channels of the model to interact with each other, helping to improve the feature extraction capability of the model. It should be noted that the attention unit in the above variable-length attention module can be an SE network, an ECA network, a CBAM network, etc. Preferably, in an optional implementation, the attention unit is an ECA network, which uses a lightweight non-reduced local cross-channel interaction strategy, involving only a small number of parameters, to bring performance improvement; specifically, the variable-length attention module compresses the intermediate feature sequence according to the sequence length, removes the padding value at the end, and calculates the global average pooling result of the feature sequence at the channel level, and inputs it into the ECA network; in the ECA network, one-dimensional convolution is used to perform convolution operation on each channel and its surrounding neighbors to capture local cross-channel interaction information, and the convolution operation result is mapped to a value between 0 and 1 through a sigmoid function, the intermediate feature sequence input into this part is multiplied by the influence score according to the corresponding channel to produce a feature sequence fused with local channel interaction information, and the feature sequence fused with local channel interaction information is superimposed with the intermediate feature sequence in a residual connection manner to obtain a comprehensive feature sequence.

[0054] It should be noted that the above classification module can be a multilayer perceptron, a classification network composed of fully connected layers or convolutional layers, etc.

[0055] In order to further illustrate the construction method of the protein phase separation characteristic detection model provided by the present application, a specific embodiment is described in detail below:

[0056] The specific implementation process of the present embodiment is shown in Figure 1

[0057] (1) Data preprocessing

[0058] The number of amino acid categories appearing in the statistical data is counted, and a "amino acid letter-number" mapping table representing the letter of each type of amino acid and its corresponding number is obtained. According to the mapping table, the protein amino acid sequences in the training, validation and test data sets are converted into corresponding digital sequences X i = x1, x2, …, x l , where i represents the sequence number, and l represents the length of the sequence.

[0059] (2) RECA model

[0060] The RECA model mainly consists of four parts, among which the embedding layer (Embedding layer) is responsible for encoding the input amino acid sequences of different lengths; the variable-length RNN module is responsible for extracting the feature information implied in the encoded amino acid sequences; the variable-length attention module is responsible for focusing on the most valuable part from the extracted sequence features and realizing information interaction and fusion between local channels; the classification module is responsible for mapping each comprehensive feature sequence to determine whether the corresponding protein has phase separation characteristics.

[0061] (2.1) Embedding layer

[0062] The amino acid sequence of the protein is preprocessed to obtain the corresponding digital sequence X i = x1, x2, …, x l , where i represents the sequence number, and l represents the length of the sequence. Since the amino acid sequences of proteins have different lengths, in order to ensure that the model can be trained in batches, the present application sorts the proteins in a batch from long to short according to the sequence length at the input end, and fills the amino acid sequences of other proteins to the same length by zero padding at the end based on the longest amino acid sequence, that is, the amino acid sequence length of the proteins in a batch is consistent, and the input sequence of uniform length is obtained b represents the size of batchsize, and l max represents the length of the longest sequence in a batch. At the same time, the true length information L = l 1 ​l 2 ,…,l b .

[0063] As shown in Figure 2 , the input batch amino acid sequence is encoded by the embedding layer to output its corresponding encoded feature vector. The feature vector dimension of the embedding layer is set to 2048, that is, the amino acid at each position in the original protein amino acid sequence is expanded from the original 1D to a feature dimension of 2048. The calculation formula is shown in formula (2-1), and the encoded feature sequence h e represents the feature vector dimension of the embedding layer.

[0064] Y e = Embed (X) (2-1)

[0065] (2.2) Variable-length RNN module

[0066] In the process of padding the original protein amino acid sequence to a uniform length and encoding, since there are zero-padded regions in the sequence, directly inputting the encoded sequence Y e into the RNN network (BiLSTM network in this embodiment) will not only waste computing resources and time, but also affect the accuracy of the encoded sequence in representing the original information, resulting in a decrease in model performance. In order to solve the above problems, the encoded feature sequence needs to be compressed first to reduce the interference caused by the zero-padded region.

[0067] Figure 3 represents the compression process, which compresses the encoded vector sequence Y e in a batch into a low-dimensional encoded vector sequence Y' max in the sequence length l e dimension according to a single timestep to remove the tail padding value and compress it into a low-dimensional encoded vector sequence Y' e . Specifically, a timestep corresponds to an amino acid at a position, and the encoded vector sequence in a batch follows the original arrangement order and according to the recorded protein amino acid sequence length L, the single amino acid at the corresponding position is spliced and filtered for the tail padding value one by one to obtain the compressed low-dimensional feature sequence. At the same time, the number of vectors corresponding to each timestep of the original sequence Y and the compressed sequence Y'

[0068] Y e ',N = Packed (Y e , L) (2-2)

[0069] Then, the compressed low-dimensional vector sequence Y' eThe number of original sequences corresponding to each time step, N, is input into the BiLSTM network to realize feature extraction and information fusion of context sequence relationships.

[0070] Since the BiLSTM network is composed of a plurality of repeated neural unit modules linked together, the application can only consider one LSTM unit module. The LSTM unit module is as shown in Figure 4 The core components are three gate structures and a cell state. The opening and closing states of the three gates determine whether the transmitted information is saved or forgotten. The input of the current LSTM unit module includes the state value c t-1 , the output value h t-1 and the current input value h t of the previous LSTM unit module. The output includes the state value c t and the output value h t of the current unit module. The three gate structures include a forgetting gate f t , an input gate i t and an output gate o t .

[0071] The function used by the three gate states is the Sigmoid function, which is represented by formula (2-3) as follows:

[0072]

[0073] This function limits the output value to be between 0 and 1, which helps the gate state to update or forget the data. Any value multiplied by 0 is equal to 0, i.e. the information is forgotten, and any value multiplied by 1 is equal to 1, i.e. the information is retained. Therefore, the Sigmoid function is used to realize the operation of retaining or forgetting the transmitted information by the gate state of the LSTM unit module.

[0074] The forgetting gate determines how much of the previous time unit state c t-1 needs to be retained to the current time c t , which is represented by formula (2-4) as follows:

[0075] f t =σ(W f [h t-1 ,x t ]+b f ) (2-4)

[0076] In the above formula, W f represents the weight matrix of the forgetting gate obtained by training the model, and b f represents the bias value of the forgetting gate obtained by training the model. [h t-1 ,x tmeans concatenating the two vectors at the matrix level, and σ is a representation of the Sigmoid function. Specifically, the output value h t-1 from the LSTM unit at the previous time and the input x t at the current time are combined and input into the Sigmoid function after linear transformation to obtain the output value f t of the forget gate state, which is multiplied by the unit state c t-1 at the previous time to achieve control over the preservation or forgetting of the unit state value at the previous time.

[0077] The input gate determines how much of the input data x t at the current time needs to be saved to the unit state c t , which is represented by equation (2-5) as follows:

[0078] i t = σ(W i [h t-1 ,x t ]+b i ) (2-5)

[0079] In the above equation, W i represents the weight matrix of the input gate obtained by training the model, and b i represents the bias value of the input gate obtained by training the model. Specifically, the output value h t-1 from the unit at the previous time and the input x t at the current time are combined and input into the Sigmoid function after linear transformation to obtain the output value i t of the input gate state, which is multiplied by the current input unit state c' t to achieve control over the preservation or forgetting of the current input unit state value.

[0080] The current input unit state c' t is represented by equation (2-6) as follows:

[0081] c t '= tanh(W c′ [h t-1 ,x t ]+b c′ ) (2-6)

[0082] In the above equation, W c' represents the weight matrix of the current input unit state obtained by training the model, b c' represents the bias value of the current input unit state obtained by training the model, and the tanh function is represented by equation (2-7) as follows:

[0083]

[0084] Specifically, the cell output value h t-1 from the previous time unit and the current time unit input x t are combined and input into a tanh function after linear transformation to obtain the current input cell state value c' t .

[0085] The forget gate state value f t is multiplied by the cell state value c t-1 from the previous time unit, and the result is added to the product of the input gate state value i t and the current input cell state value c' t , so that the current time unit cell state c t is obtained, which is represented by formula (2-8) as follows:

[0086] c t = f t *c t-1 + i t *c t ' (2-8)

[0087] Specifically, the important information of the previous cell module can be saved by using the control of the forget gate, and irrelevant information of the current input can be avoided by using the control of the input gate.

[0088] The output gate controls how much of the current cell state c t needs to be output to the current output value h t . This is represented by formula (2-9) as follows:

[0089] o t = σ(W o [h t-1 ,x t ]+b o ) (2-9)

[0090] In the above formula, W o represents the weight matrix of the output gate obtained by training the model, and b o represents the bias value of the output gate obtained by training the model. Specifically, the cell output value h t-1 from the previous time unit and the current time unit input x t are combined and input into a Sigmoid function after linear transformation to obtain the output gate state output value o t , which is multiplied by the tanh operation result of the current time unit cell state value c t , so that the current time unit output h t is obtained, which is represented by formula (2-10) as follows:

[0091] h t = o ttanh(c t ) (2-10)

[0092] As Figure 5 shown, the BiLSTM network is based on the LSTM network structure and increases the reverse operation channel. The context information of the sequence can be fully considered by extracting the features of the sequence in the forward and reverse directions. The output result of the network is the stacking of the bidirectional feature sequence. The number of layers of the BiLSTM network layer is set to 1, the number of hidden layer neurons is 128, and the many-to-many output structure is used. The dropout method is introduced, and the value is set to 0.2, which enhances the model generalization performance and robustness. As shown in equation (2-11), the BiLSTM network extracts the context feature information of the sequence according to the input low-dimensional compressed feature vector sequence Y e and the number information N corresponding to the original sequence at each timestep, and outputs the compressed feature sequence of the same data format. At this time, the compressed feature vector sequence is expanded and padded at the end according to the corresponding position amino acid number information N, and the feature sequence with the same data dimension as the output of the embedding layer is obtained h b represents the neuron dimension of the BiLSTM network. Specifically, according to the corresponding position amino acid number information, the compressed feature vector sequence is re-split, reversedly spliced in the original order, and padded at the end to generate a vector sequence with the same data dimension as before compression.

[0093] Y b = Padding(BiLSTM(Y e ',N)) (2-11)

[0094] Since the encoded feature vector sequence and the feature sequence extracted by BiLSTM both contain the feature information of the original protein, the two sequences are spliced and the last two dimensions are transposed to obtain the intermediate feature sequence After that, the model uses residual connection to further fuse the features generated by different channels by inputting this intermediate feature and its input into the subsequent variable-length attention module, so as to obtain the final representation information of the original protein, that is, the feature sequence and the encoded sequence are spliced in the feature dimension layer and input into the variable-length attention module in the form of residual connection to further realize information fusion between different channels.

[0095] (2-3) Variable-length attention module

[0096] The structure of the variable-length attention module is shown in Figure 6 Since the feature sequence extracted by the variable-length BiLSTM network and the encoded sequence output by the encoding layer both involve end padding operations, this part also avoids the interference of the padding area on the original sequence by splicing the feature sequence Ym According to its corresponding sequence length L, the padding value at the end is removed, and the global average pooling result of the feature sequence at the channel level is calculated The calculation formula is shown in formula (2-12), and the calculation process is shown in Figure 7 .

[0097] y m = PackedGAP(Y m , L) (2-12)

[0098] In order to improve the ability of the model to cross-channel interaction with only a small number of parameters, the present application uses one-dimensional convolution with a convolution kernel size k of 3 to perform convolution operation on each channel and its 3 neighbors to capture local cross-channel interaction information, wherein the convolution kernel size reflects the coverage of local cross-channel interaction. As shown in formula (2-13), the result of the convolution operation is mapped to a value ω between 0 and 1 through a sigmoid function, and the size of the value also reflects the influence of the corresponding channel. The variable-length attention module uses one-dimensional convolution operation to perform appropriate cross-channel interaction, which can significantly reduce the complexity of the model while improving the performance of the model. Finally, as shown in formula (2-14), the feature sequence Ym input into the module is multiplied by the influence score mapped to a value between 0 and 1 to generate a feature sequence that integrates local cross-channel interaction information

[0099] w = σ(C1D k (y m )) (2-13)

[0100] Y eca = Y m *w (2-14)

[0101] (2-4) Classification module

[0102] The above feature sequence Y eca and the sequence feature information Y m input into the variable-length attention module are superimposed in a residual connection manner to obtain the total feature information, and the final protein phase separation detection result is given through the classification module-Multi-Layer Perceptron Head (MLPHead) which includes global maximum pooling and linear layer. The calculation formula is shown in formula (2-15).

[0103] Y = MLP(GMP(Y m + Y eca )) (2-15)

[0104] (3) Model training parameters

[0105] In the experiment, due to the small memory occupied by the amino acid sequence of a single protein, the parameter batch size is set to 32, and the training period of the model is set to 80 epochs. The weight decay technique is introduced in the experiment to facilitate the model to easily reach the global optimal value, that is, when the model detects that the accuracy on the validation set no longer improves, the learning rate will be reduced by 10 times in order to expect to achieve further improvement of the accuracy. After completing all the training periods, the data on the test set is selected to test and evaluate the model trained at the end, and the accuracy, sensitivity and specificity of the model on the test set are calculated, and these results are used as the final recognition rate and other performance indicators of the model in this experiment. The cross-entropy loss function between the predicted value and the true value is selected as the loss function of the model training, and the Adam optimizer is selected to update the network parameters, and the initial learning rate is set to 0.0001. The parameters of the Adam optimizer are adjusted, including: β1=0.9, β2=0.999, ε=1e-08.

[0106] Considering that the size of the data set is small, in order to ensure the reliability of the experimental results, the adverse effects brought by random data division are minimized, and ten-fold cross-validation is used for experiment, and the average result is selected as the final experimental result. The evaluation indexes include accuracy, sensitivity and specificity, and the calculation formulas are represented by formulas (3-1, 3-2, 3-3) as follows:

[0107]

[0108]

[0109]

[0110] In the above formula, TP is a true case, TN is a true negative case, FP is a false positive case, and FN is a false negative case.

[0111] The application is a completely data-driven method, which does not require any expert experience or external knowledge support, can effectively adapt to the phase separation property detection of protein amino acid sequences of various lengths, has high classification accuracy, and is an end-to-end structure which can greatly simplify the complexity of the model.

[0112] Embodiment 2

[0113] A protein phase separation property detection method comprises:

[0114] The batch set to be tested is input into the protein phase separation property detection model to determine whether each protein in the batch set to be tested has the phase separation property;

[0115] The protein phase separation property detection model is constructed by using the construction method of the protein phase separation property detection model provided in Embodiment 1.

[0116] The sample sequence set of the to-be-tested batch includes a plurality of proteins, and the obtaining method includes: taking the longest amino acid sequence as a benchmark, padding the end of the remaining amino acid sequences, and obtaining a batch of sample sequences with a length of the longest amino acid sequence in the to-be-tested batch; each amino acid is assigned a corresponding number, and different amino acids correspond to different numbers; the i-th value in the amino acid sequence is the number corresponding to the i-th amino acid in the protein; i = 1, 2, …, L; L is the length of the longest amino acid sequence in the to-be-tested batch.

[0117] It should be noted that in the detection stage, the number of sample sequences of proteins in the to-be-tested batch can be one or more; when the number is one, since padding is not required when inputting, the corresponding length operation is not performed in the protein phase separation property detection model, and only encoding, feature extraction and classification operations are performed.

[0118] The related technical solutions are the same as those in Embodiment 1, and will not be repeated here.

[0119] Embodiment 3

[0120] A protein phase separation property detection system includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to execute the protein phase separation property detection method provided in Embodiment 2 of the present application.

[0121] The related technical solutions are the same as those in Embodiment 2, and will not be repeated here.

[0122] Embodiment 4

[0123] A computer readable storage medium includes a stored computer program, wherein when the computer program is run by a processor, the storage medium controls the device where the storage medium is located to execute the construction method of the protein phase separation property detection model provided in Embodiment 1 of the present application and / or the protein phase separation property detection method provided in Embodiment 2 of the present application.

[0124] The related technical solutions are the same as those in Embodiments 1 and 2, and will not be repeated here.

[0125] Those skilled in the art will readily understand that the above description is only a preferred embodiment of the present application and is not intended to limit the present application, and any modifications, equivalent replacements and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for constructing a protein phase separation characteristic detection model, characterized in that, include: The batch set was input into the protein phase separation characteristic detection model for training; The batch set includes sample sequences of N proteins, which are constructed based on the amino acid sequences of the N proteins: for the amino acid sequences of the N proteins, the longest amino acid sequence is used as the reference, and the remaining amino acid sequences are padded to the end to obtain a batch of sample sequences with the same length as the longest amino acid sequence. ; The protein phase separation characteristic detection model includes: An embedding layer is used to encode the N sample sequences in the batch set respectively to obtain N encoded vector sequences; A variable-length RNN module is used to compress N encoded vector sequences as a whole. After removing the encoded vectors at the padding positions, the compressed feature sequence is input into the RNN network to extract contextual feature information. The compressed feature sequence is then inversely padded to obtain the first feature sequence of each sample sequence, which has the same size as the corresponding encoded vector sequence. The first feature sequence of each sample sequence is then concatenated with its encoded vector sequence along the feature dimension to obtain the corresponding intermediate feature sequence. The variable-length attention module removes the feature terms corresponding to the padding values ​​in each intermediate feature sequence, performs channel-level global pooling, and inputs the pooling result into the attention unit to obtain the influence score of each channel. Each intermediate feature sequence is multiplied by the corresponding influence score according to the corresponding channel to obtain the corresponding second feature sequence. Each second feature sequence is superimposed with the corresponding intermediate feature sequence in the form of residuals to obtain the corresponding comprehensive feature sequence. The classification module is used to map each comprehensive feature sequence to determine whether the corresponding protein has phase separation characteristics.

2. The method for constructing a protein phase separation characteristic detection model according to claim 1, characterized in that, The variable-length RNN module is used to concatenate the encoded vectors at the same position in N encoded vector sequences, remove the encoded vectors at the padding positions, and obtain multiple concatenated vectors. These concatenated vectors are then input into the RNN network to extract contextual feature information, resulting in a compressed feature sequence. The compressed feature sequence is then decomposed and padded in reverse to obtain N first feature sequences with the same dimension as the encoded vector sequences.

3. The method for constructing a protein phase separation characteristic detection model according to claim 1, characterized in that, The variable-length RNN module is used to concatenate the encoding vectors at the same position in N encoding vector sequences, and remove the encoding vectors at the padding positions to obtain... J A non-empty concatenated vector; J Each concatenated vector corresponds to an output fed into the RNN network. J In each RNN unit, we obtain J A compressed feature sequence is formed by compressed features of the same size as the corresponding concatenated vector; the compressed feature sequence is split, and the features corresponding to the same encoding sequence in the compressed feature sequence are merged back according to their original positions and then filled in reverse to obtain N first feature sequences with the same dimension as the encoding vector sequence. .

4. The method for constructing a protein phase separation characteristic detection model according to claim 3, characterized in that, The RNN network is a BiLSTM network; the RNN unit is a BiLSTM unit.

5. The method for constructing a protein phase separation characteristic detection model according to any one of claims 1-3, characterized in that, The attention unit is an ECA network.

6. A method for detecting protein phase separation characteristics, characterized in that, include: The batch set to be tested is input into the protein phase separation characteristic detection model to determine whether each protein in the batch set to be tested has phase separation characteristics. The protein phase separation characteristic detection model is constructed using the method described in any one of claims 1-5. The batch set to be tested includes sample sequences of several proteins. The method for obtaining the sequence includes: for the amino acid sequence of each protein in the batch set, using the longest amino acid sequence as a reference, padding is performed at the end of the remaining amino acid sequences to obtain a batch of sample sequences whose length is the length of the longest amino acid sequence in the batch set to be tested.

7. A protein phase separation characteristic detection system, characterized in that, include: A memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the protein phase separation characteristic detection method of claim 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed by a processor, it controls the device where the storage medium is located to perform the method for constructing a protein phase separation characteristic detection model according to any one of claims 1-5 and / or the protein phase separation characteristic detection method according to claim 6.

Images