Protein property prediction method based on fusion of deep learning and intelligent optimization

Abstract

This invention provides a protein property prediction method based on the fusion of deep learning and intelligent optimization, belonging to the field of protein property prediction technology. The method first acquires protein sequence data and converts it into a digital representation, then uses a convolutional neural network to extract protein sequence features. Next, the protein sequence is encoded into protein music based on amino acid classification, isoelectric point, and hydrophilicity index, generating corresponding Mel spectrograms, which are input into a VGG16 network to extract frequency features. The sequence and frequency features of the protein are combined to achieve protein property prediction. Simultaneously, a genetic algorithm with an elite retention strategy is introduced to automatically search for parameters of the protein property prediction model. L1 and L2 regularization coefficients are selected as search parameters, and their ranges are set. The population is randomly initialized, and then the fitness value of the population is calculated sequentially, followed by selection, crossover, and mutation operations, iterating until the termination condition is met. Finally, the model is tested.

Description

Technical Field

[0001] This invention relates to the field of protein property prediction technology, and in particular to a protein property prediction method that converts proteins into music and integrates deep learning and intelligent optimization methods. Background Technology

[0002] Proteins play an extremely important role in living systems. They are essential substances present in all organisms and play a central role in the structure and function of the body. For example, proteins are one of the most important organic molecules in living organisms, constituting the main structural and functional building blocks of cells. Proteins are the executors of biological functions, performing a variety of biological functions in the body, including transporting substances, signal transduction, cell adhesion, immune response, and apoptosis. Most biological reactions depend on enzymes, which are a special type of protein that can accelerate chemical reactions, making metabolic processes in the body more efficient and controllable. Proteins encode genetic information and, through the protein synthesis process (translation), convert genetic information into functional products.

[0003] Understanding the relationship between protein sequences (such as their domains) and their structure or function is a long-standing challenge with profound scientific significance. Mastering protein properties has significant scientific and applied implications in biological science and medical research. For example, research on protein properties has promoted the development of bioinformatics and computational biology, including protein structure prediction, functional annotation, and interaction network analysis; comparing protein properties can help study evolutionary relationships and adaptive changes in biological species, thereby deepening our understanding of biodiversity; drug development also involves the interaction between proteins and drugs, and by studying protein properties in depth, scientists can design and develop more effective drugs to treat various diseases, including cancer, infectious diseases, and autoimmune diseases.

[0004] Protein property prediction based on deep learning is an important bioinformatics task, used to predict information such as protein structure, function, and interactions. Traditional biochemical experimental methods for determining protein properties are typically costly, time-consuming, and low-throughput; therefore, developing efficient and accurate computational methods for protein function prediction is crucial. With the continuous development of artificial intelligence and advancements in biological sequencing technologies, the accumulation of vast amounts of protein data has enabled deep learning methods to achieve significant progress in this field. By applying these technologies, researchers can explore protein characteristics more deeply, contributing to research in biology, medicine, and bioinformatics, and providing new methods and perspectives for solving protein-related problems.

[0005] Early protein characterization learning methods defined different rules to extract physicochemical or statistical features from protein sequences. In 1986, Klein et al. used the physicochemical properties of amino acids to characterize proteins, classifying them into 26 functional categories. The range of properties ranged from simple compositional features, such as average hydrophobicity and net charge number, to amphiphilicity and the tendency of various residues in certain preferred conformations. In 2000, Feng et al. proposed a new algorithm for predicting membrane protein types, which considered information in the amino acid sequence in addition to the protein's amino acid composition, using an autocorrelation function based on a 20-amino acid hydrophobicity index. In 2015, Vijayakumar et al. developed a novel feature descriptor based on amino acid composition—Dipeptide Deviation from Expected Mean (DDE)—to effectively distinguish linear B-cell epitopes from non-epitopes.

[0006] Advances in deep learning and natural language processing have driven the development of protein sequence models to leverage large-scale protein sequence corpora. Due to the similarity between sentences in natural language and protein sequences, many methods in the field of Natural Language Processing (NLP) are currently used to obtain protein representations and predict protein properties, including Convolutional Neural Networks (CNNs), Long Short-Term Memory Networks (LSTMs), word2vec, doc2vec, and Transformers. NLP techniques can be used to obtain efficient protein representations, using text-like encoding methods to improve the understanding and representation of protein sequences, thereby enabling better analysis of protein structure, function, and interactions.

[0007] In 2016, Wan et al. proposed a scheme combining feature embedding techniques with deep learning to predict protein-compound interactions. This scheme used the word2vec method to learn low-dimensional implicit representations of protein features from large amounts of unlabeled protein sequence data, achieving good results. In 2018, Xu et al. proposed Phosconcontext2Vec, a distributed representation of residue-level sequence context for potential phosphorylation sites, and demonstrated its application in predicting general and kinase-specific phosphorylation sites. In 2021, Ahmed et al. trained two autoregressive models (Transformer-XL, XLNet) and four autoencoder models (BERT, Albert, Electra, T5) based on the UniRef and BFD databases, which contain up to 393 billion amino acids. Their results showed that the original protein language model embeddings from unlabeled data captured some biophysical features of protein sequences and achieved good results in protein secondary structure prediction and protein functional property prediction tasks.

[0008] In summary, most current protein characterization learning methods are based on the amino acid sequence of proteins, but they almost entirely ignore the position and arrangement of amino acids within the protein sequence, i.e., the frequency information of amino acids in the protein. Furthermore, current protein sequence-based characterization learning methods often use one-hot encoding, lacking an understanding of the physicochemical properties of the amino acids that make up the protein sequence. These limitations lead to an insufficient understanding of the protein sequence and the relationship between structure and function in the models. Therefore, how to incorporate the physicochemical properties of amino acids into protein property prediction models and how to incorporate the frequency features of the protein sequence as a supplement are issues worthy of consideration. Summary of the Invention

[0009] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing a protein property prediction method based on the fusion of deep learning and intelligent optimization. This solution proposes a novel protein property prediction model that, through thorough learning of protein sequence information, supplements it with frequency information encoded as music by proteins, and further utilizes a genetic algorithm with an elitist preservation strategy to search for model parameters, thereby improving the model's accuracy in protein property prediction tasks.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by this invention is: a protein property prediction method based on the fusion of deep learning and intelligent optimization, comprising the following steps:

[0011] Step 1: Obtain protein sequence data, divide the protein sequence data into training set, validation set and test set, and convert the representation of the protein amino acid sequence into numerical form;

[0012] Step 1.1: Segment the protein sequence into a list of amino acid symbols;

[0013] Step 1.2: Convert the amino acids to their corresponding numbers according to the correspondence between amino acid symbols and numbers;

[0014] Step 1.3: Convert the obtained numerical representation of amino acids into the corresponding one-hot encoded representation;

[0015] Each amino acid in the protein sequence is represented as a 21-dimensional one-hot vector, where the number of bits corresponding to the number is 1 and the remaining bits are 0.

[0016] Step 2: Encode the protein sequence into protein music; obtain the sequence information of the protein, map each amino acid of the protein to a note with different beats and volumes, and encode the protein sequence into protein music.

[0017] Step 2.1: Classify amino acids according to the properties of their R groups and establish a correspondence between amino acids and musical notes;

[0018] Based on the properties of the R group of amino acids, the 20 common amino acids are divided into 5 categories: nonpolar aliphatic R group amino acids, aromatic R group amino acids, polar uncharged R group amino acids, positively charged R group amino acids, and negatively charged R group amino acids.

[0019] Glycine is defined as 60, representing the middle note C; methionine and proline correspond to notes D4 and E4, respectively; alanine, valine, leucine, and isoleucine are ordered according to their hydrophilicity index, corresponding to notes F4, A4, G4, and B4, respectively; aromatic R-group amino acids, including phenylalanine, tyrosine, and tryptophan, correspond to notes B3, G3, and A3, respectively; ordered according to their isoelectric point, aspartic acid and glutamic acid, among the negatively charged R-group amino acids, correspond to notes C5 and D5, respectively; serine, threonine, cysteine, asparagine, and glutamine, among the polar uncharged R-group amino acids, correspond to notes A5, B5, E5, F5, and G5, respectively; and lysine, arginine, and histidine, among the positively charged R-group amino acids, correspond to notes D6, E6, and C6, respectively.

[0020] Step 2.2: Obtain the hydrophilicity index (hydrophilicity_value) of the amino acid. Map the obtained hydrophilicity index (hydrophilicity_value) of the amino acid to the amplitude of the note corresponding to the current amino acid, where 0 represents mute and 127 represents maximum volume. Limit the amplitude mapping result of the amino acid to the note within [amplitude]. min ,amplitude max Between, among which, amplitude min ,amplitude max These are the minimum and maximum values ​​of amplitude, respectively.

[0021] Step 2.3: Obtain the isoelectric point pI_value of the amino acid and map the obtained isoelectric point pI_value of the amino acid to the duration time of the current amino acid note;

[0022] Step 2.4: Set the timbre and tempo of the protein music track;

[0023] Step 2.5: Generate protein music according to the parameters in steps 2.2-2.4 above;

[0024] Step 3: Generate the mel spectrogram of the protein music and extract the frequency features of the protein sequence;

[0025] Step 3.1: Calculate the corresponding spectrogram based on the protein music generated in Step 2;

[0026] Step 3.1.1: Given the following calculation parameters: sampling rate sr, highest frequency fmax, fast Fourier transform window size n_fft, sampling interval between consecutive frames hop_length, number of generated Mel bands n_mels, and spectrogram exponent power;

[0027] Step 3.1.2: Pad the ends of the input protein music, with a padding length of half the window size n_fft / 2, and the padding value is a constant 0;

[0028] Step 3.1.3: Frame the completed protein music to obtain short time frames;

[0029] Step 3.1.4: After slicing the completed protein music into frames, apply a Hann window function to each frame;

[0030] Step 3.1.5: Perform Discrete Fourier Transform on each frame after frame division and window calculation, and concatenate the vectors obtained from the Discrete Fourier Transform of each frame to obtain the result matrix.

[0031] Step 3.1.6: Take the absolute value of the result matrix to obtain its amplitude, and then perform corresponding calculations based on the power of the given spectrogram to obtain the spectrogram of the protein music.

[0032] Step 3.2: Construct a Mel filter bank. The filter bank uses a triangular filter to extract the frequency band of the spectrogram at the Mel scale, thereby obtaining the Mel spectrogram.

[0033] Step 3.3: Convert the obtained Mel spectrogram to the dB scale to obtain the Mel spectrogram of protein music at the dB scale;

[0034] Step 3.4: Save the Mel spectrogram of the protein music at the dB scale, and adjust the Mel spectrogram at the dB scale into a matrix of size [n_channel,width,height], where n_channel,width,height are the number of channels, width, and height of the Mel spectrogram at the dB scale, respectively.

[0035] Step 4: Set the parameters to be searched in the protein property prediction model, and the range of the parameters to be searched; initialize a population of N individuals according to the encoding rules;

[0036] Step 4.1: The model parameters to be searched are the L1 regularization coefficient λ1 and the L2 regularization coefficient λ2;

[0037] Step 4.2: Set the range of parameters λ1 and λ2 to [0,1];

[0038] Step 4.3: Initialize a population of N individuals according to the encoding rules;

[0039] Step 5: Construct a protein property prediction model;

[0040] Step 5.1: Construct a protein sequence feature extraction module to extract the sequence features of proteins; the protein sequence feature extraction module is constructed by combining a one-dimensional convolutional neural network and a linear layer network, and contains three layers, namely two convolutional neural network layers and one linear layer neural network layer.

[0041] Step 5.2: Construct a protein frequency feature extraction module to extract the frequency features of proteins; the protein frequency feature extraction module is constructed by using the feature extraction part of the VGG16 model and then concatenating it with a linear layer network.

[0042] Step 5.3: Construct a protein property prediction module. The protein property prediction module receives the protein sequence features and protein frequency features extracted by the protein sequence feature extraction module and the protein frequency feature extraction module, concatenates the two, and inputs them into a linear layer network for mapping to obtain the protein property prediction results.

[0043] Step 5.4: Define the loss function for the protein property prediction model and select the model optimizer;

[0044] Step 5.4.1: Select the Adam optimizer as the optimizer for the protein property prediction model and set the learning rate parameter;

[0045] Step 5.4.2: To prevent overfitting, L1 regularization and L2 regularization terms are introduced into the loss function of the protein property prediction model;

[0046] The loss function of the protein property prediction model is selected according to the task. For classification tasks, cross-entropy loss is selected as the loss function, and for regression tasks, mean squared error is selected as the loss function.

[0047] Step 6: Forward calculation of the protein property prediction model to extract the protein's sequence features and frequency features to predict the protein's properties;

[0048] Step 6.1: Extract protein sequence features using the protein sequence feature extraction module. The input to this module is the one-hot representation of the protein sequence; the protein sequence features are obtained after calculation through two convolutional layers and one linear layer.

[0049] Step 6.1.1: Pass the one-hot representation of the input protein sequentially through the first layer of a one-dimensional convolutional neural network, the ReLU activation function, and the second layer of a one-dimensional convolutional neural network, the ReLU activation function.

[0050] Step 6.1.2: The protein sequence one-hot representation after passing through the convolutional neural network has a dimension of [batch_size, seq_len, o_channel], where batch_size is the batch size, seq_len is the length of the protein sequence, and o_channel is the number of output channels. That is, the feature dimension of the protein sequence features after passing through the convolutional neural network. Then, the MaxReadout method is used to take the maximum value in the protein sequence length dimension, and the result is calculated as [batch_size, o_channel].

[0051] Step 6.1.3: Finally, the calculation results of step 6.1.2 are mapped to [batch_size, hidden_dim1] through the constructed linear layer network, where hidden_dim1 is the dimension of the protein sequence features;

[0052] Step 6.2: Extract protein frequency features using the protein frequency feature extraction module. The input to this module is the Mel spectrogram of the protein music converted to the dB scale.

[0053] Step 6.2.1: The input to the protein frequency feature extraction module is the Mel spectrogram of the protein music at the dB scale. The input dimension is [batch_size, n_channel, width, height], where n_channel is the number of channels in the Mel spectrogram of the protein music at the dB scale, and width and height are the width and height of the input Mel spectrogram of the protein music at the dB scale, respectively. Then, the features are extracted by the feature extraction part of VGG16.

[0054] Step 6.2.2: After feature extraction via the VGG16 network, the resulting matrix is ​​flattened to [batch_size, hidden_dim2], where hidden_dim2 is the feature dimension after passing through the VGG16 network. Then, the protein frequency features are calculated sequentially through dropout layers and linear layers to obtain the protein frequency features, where the frequency features have the dimension [batch_size, hidden_dim3], where hidden_dim3 is the protein frequency feature dimension.

[0055] Step 6.3: Protein property prediction. By receiving the extracted protein sequence features and protein frequency features, the properties of the protein are predicted.

[0056] Step 6.3.1: By concatenating the received protein sequence features and protein frequency features according to the feature dimension, the protein representation is obtained. The protein representation dimension is [batch_size, hidden_dim1+hid_dim3];

[0057] Step 6.3.2: Input the protein representation into the linear layer to predict protein properties;

[0058] Step 7: Train and validate the protein property prediction model, and obtain fitness values ​​for different individuals;

[0059] Step 7.1: Calculate the model loss according to the loss function defined in the task, backpropagate, and calculate the corresponding gradient;

[0060] Step 7.2: Use the Adam optimizer to update the parameters of the protein property prediction model based on the gradient;

[0061] Step 7.3: Repeat steps 6-7.2 above until the loss of the protein property prediction model on the validation set no longer decreases;

[0062] Step 7.4: The minimum loss of the obtained protein property prediction model on the validation set is used as the fitness value of the current individual;

[0063] Step 8: Determine if the termination condition is met. If the termination condition is met, proceed to step 10; otherwise, continue to step 9.

[0064] Step 9: Population selection, crossover, and mutation operations;

[0065] Step 9.1: Independently select N-1 individuals from the current population;

[0066] Step 9.2: Perform crossover operations independently on these N-1 individuals;

[0067] Step 9.3: Independently mutate these N-1 crossover individuals;

[0068] Step 9.4: Calculate the best individual in the current population and insert it into the first position of the N-1 individuals after crossover mutation to obtain the new generation population;

[0069] Step 9.5: Return to step 5 to construct the protein property prediction model and calculate the fitness value of the new population;

[0070] Step 10: Protein property prediction model testing; Train the model using the optimal combination of model parameters obtained from the search and calculate the model loss and corresponding evaluation metrics on the test set.

[0071] The beneficial effects of adopting the above technical solution are as follows: The protein property prediction method based on the fusion of deep learning and intelligent optimization provided by this invention encodes the protein sequence as music and extracts its Mel spectrogram as a frequency representation. Then, it uses a combination of VGG16 model and MLP model to extract the frequency information as a global information representation of the protein sequence, which makes up for the shortcomings of simply extracting information from the protein sequence. At the same time, it uses a genetic algorithm with an elite preservation strategy to search for the coefficients of L1 and L2 penalty terms in the model training, thereby improving the model prediction performance. Attached Figure Description

[0072] Figure 1 A flowchart illustrating a protein property prediction method based on the fusion of deep learning and intelligent optimization provided in an embodiment of the present invention;

[0073] Figure 2Waveform and Mel spectrogram of protein music provided in embodiments of the present invention;

[0074] Figure 3 A model diagram of the VGG16 network provided in an embodiment of the present invention. Detailed Implementation

[0075] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0076] Considering that current protein feature extraction methods are mainly based on protein sequence information and the protein representation vectors learned through local context information, they rarely consider the frequency distribution information of protein sequences and the physicochemical properties of amino acids in proteins. Therefore, this invention proposes a protein music generation and property prediction model based on the fusion of deep learning and intelligent optimization. Based on the physicochemical properties of amino acids, it realizes the conversion of protein sequences into music, and then extracts the frequency features for feature enhancement. At the same time, in order to find suitable model parameters, a genetic algorithm with an elite preservation strategy is introduced to automatically search for model parameters. The model used in this invention consists of two parts: one part is the feature extracted based on protein sequence information, and the other part is to encode the protein into music using the hydrophilicity index and isoelectric point of amino acids, and then extract the converted spectrogram as frequency features to supplement the protein characterization extracted based on sequence information. Finally, the two are combined to predict protein properties.

[0077] In this embodiment, a protein property prediction method based on the fusion of deep learning and intelligent optimization is used, such as... Figure 1 As shown, it includes the following steps:

[0078] Step 1: Obtain protein sequence data, divide the protein sequence data into training set, validation set and test set, and convert the representation of the amino acid sequence of the protein into digital form for computer processing;

[0079] Step 1.1: Segment the protein sequence into a list of amino acid symbols;

[0080] A protein sequence is composed of a series of amino acid symbols, such as 'GSTTIEEAQNKKYQAEPRSWTKAGRTIGGKNWETEVNRAEASI'. Segmenting a protein sequence divides it into a list of amino acid symbols, such as ['G','S','T','T','I','E','E','A','Q','N','K','K','Y','Q','A','E','P','R','S','W','T','K','A','G','R','T','I','G','G','K','N','W','E','T','E','V','N','R','A','E','A','S','I'].

[0081] Step 1.2: Convert the amino acids to their corresponding numbers according to the correspondence between amino acid symbols and numbers;

[0082] In this embodiment, according to the correspondence between amino acid symbols and numbers in Table 1, amino acids are converted into their corresponding numbers. For example, amino acids ['G','S','T','T','I','E','E','A','Q','N','K','K','Y','Q','A','E','P','R','S','W','T','K','A','G','R','T','I','G','G','K',' The corresponding numbers for N','W','E','T','E','V','N','R','A','E','A','S','I'] are [8,16,17,17,10,7,7,1,6,3,12,12,19,6,1,7,15,2,16,18,17,12,1,8,2,17,10,8,8,12,3,18,7,17,7,20,3,2,1,7,1,16,10];

[0083] Table 1. Amino acid symbols, corresponding numbers, and unique heat codes.

[0084]

[0085]

[0086] Step 1.3: Convert the obtained numerical representation of amino acids into the corresponding one-hot encoded representation;

[0087] In this embodiment, the symbols of 20 amino acids and 1 [unclear] are considered. <unk>symbols, <unk>Symbols are used to represent unknown symbols in the protein sequence. Therefore, the number corresponding to each amino acid in the protein sequence is represented as a 21-dimensional one-hot vector. The number of bits in the vector is 1, and the value of the remaining bits is 0, as shown in Table 1.

[0088] Step 2: Encode the protein sequence into protein music; obtain the sequence information of the protein, map each amino acid of the protein to a note with different beats and volumes, and encode the protein sequence into protein music. Figure 2 The image shown in this embodiment is a waveform and Mel spectrogram of a protein sequence converted into protein music.

[0089] Step 2.1: Classify amino acids according to the properties of their R groups and establish a correspondence between amino acids and musical notes;

[0090] Based on the properties of the R group of amino acids, the 20 common amino acids are divided into 5 categories: nonpolar aliphatic R group amino acids, aromatic R group amino acids, polar uncharged R group amino acids, positively charged (basic) R group amino acids, and negatively charged (acidic) R group amino acids.

[0091] Because glycine is the only amino acid without a chiral carbon atom, it is optically inactive and has the simplest structure among amino acids, falling between polar and nonpolar. Therefore, glycine is defined as 60, representing the middle note C (C4). Methionine and proline, also nonpolar aliphatic R-group amino acids, correspond to notes D4 and E4, respectively. Alanine, valine, leucine, and isoleucine are classified according to their hydrophilicity index. The amino acids are sorted by index and correspond to the notes F4, A4, G4, and B4 respectively. Aromatic R-amino acids, including phenylalanine, tyrosine, and tryptophan, are assigned to the notes B3, G3, and A3 respectively. Then, they are assigned to higher frequency notes in the order of negatively charged (acidic) R-amino acids, polar uncharged R-amino acids, and positively charged (basic) R-amino acids. They are sorted by isoelectric point. Among the negatively charged (acidic) R-amino acids, aspartic acid and glutamic acid are assigned to the notes C5 and D5 respectively. Among the polar uncharged R-amino acids, serine, threonine, cysteine, asparagine, and glutamine are assigned to the notes A5, B5, E5, F5, and G5 respectively. Among the positively charged (basic) R-amino acids, lysine, arginine, and histidine are assigned to the notes D6, E6, and C6 respectively.

[0092] Step 2.2: Obtain the hydrophilicity index (hydrophilicity_value) of the amino acid. Map the obtained hydrophilicity index (hydrophilicity_value) of the amino acid to the amplitude of the corresponding note for the current amino acid, where 0 represents silence and 127 represents maximum volume. We do not want the volume of the note to be too low after the amino acid conversion, so we limit the amplitude mapping result of the amino acid to [amplitude]. min ,amplitude max The calculation formula is as follows:

[0093] amplitude=(hydrophilicity_value-hydrophilicity min ) / (hydrophilicity max -hydrophilicity min )

[0094] *(amplitude max -amplitude min )+amplitude min

[0095] Among them, amplitude min ,amplitude max These are the minimum and maximum values ​​of amplitude, and hydrophilicity, respectively. min and hydrophilicity max These are the minimum and maximum values ​​of hydrophilicity_value, respectively.

[0096] Step 2.3: Obtain the isoelectric point pI_value of the amino acid, and map the obtained isoelectric point pI_value of the amino acid to the duration (beats) time of the current amino acid note. The calculation formula is as follows:

[0097] time = (pI_value - pI_value) min ) / (pI_value max -pI_value min )*(time max -time min )+time min

[0098] Among them, time min time max These are the minimum and maximum values ​​of time, respectively, and pI_value min pI_value max These are the minimum and maximum values ​​of pI_value, respectively.

[0099] Step 2.4: Set the timbre and tempo (i.e., beat) of the protein music track;

[0100] In this embodiment, the timbre of the audio track is selected as Piano 1; each beat is set to 0.5 seconds, that is, 120 beats per minute;

[0101] Step 2.5: Generate protein music according to the parameters in steps 2.2-2.4 above;

[0102] In this embodiment, in order to incorporate the physicochemical properties of amino acids, the isoelectric point and hydrophilicity index of amino acids are considered to represent the number of beats and volume of the generated music during the music generation process; the properties of amino acids and the correspondence between amino acids and musical notes are shown in Table 2.

[0103] Table 2. Isoelectric point, hydrophilicity index, and corresponding musical notes of amino acids.

[0104]

[0105]

[0106] Step 3: Generate the Mel spectrogram of the protein music and extract the frequency features of the protein sequence; after converting the protein into music, the protein music can be feature extracted in the same way as sound feature extraction. The extraction method used in this invention is to calculate the Mel spectrogram of the protein music.

[0107] Step 3.1: Calculate the corresponding spectrogram based on the protein music generated in Step 2;

[0108] Step 3.1.1: Given the following calculation parameters: sampling rate sr, highest frequency fmax, Fast Fourier Transform (FFT) window size n_fft, sampling interval between consecutive frames hop_length, number of generated Mel bands n_mels, and spectrogram exponent power;

[0109] In this embodiment, the sampling rate sr is 22050, the highest frequency fmax is 1024, the fast Fourier transform window size n_fft is 8192, the sampling interval hop_length between consecutive frames is 2048, the number of generated Mel frequency bands n_mels is 96, and the exponent power of the Mel spectrogram is 2.

[0110] Step 3.1.2: Pad the input protein music at both ends. The padding length is half the window size n_fft / 2, and the padding value is a constant 0.

[0111] Step 3.1.3: Frame the completed protein music to obtain short-timeframes;

[0112] The frequencies in a signal change over time, so in most cases, performing a Fourier transform on the entire signal would lose the frequency profile of the signal over time, making it meaningless. To avoid this, we assume that the frequencies in the signal are stationary over a very short time frame. Therefore, by performing a Fourier transform on this short time frame, a good approximation of the signal's frequency profile can be obtained by concatenating adjacent frames.

[0113] Step 3.1.4: After slicing the completed protein music into frames, apply a Hann window function to each frame;

[0114] The main purpose of adding a window function is to better satisfy the periodicity requirements of the time-domain signal in Fast Fourier Transform (FFT) processing and reduce leakage. The formula for calculating the Hann window is shown below:

[0115]

[0116] Where α = 0.5, and M is the window size;

[0117] Step 3.1.5: Perform Discrete Fourier Transform on each frame after frame division and window calculation, and concatenate the vectors obtained from the Discrete Fourier Transform of each frame to obtain the result matrix.

[0118] For an N-point sequence {x[n]} 0≤n≤N Its Discrete Fourier Transform (DFT) is

[0119]

[0120] Where e is the base of the natural logarithm, and i is the imaginary unit;

[0121] When calculating the DFT of pure real number inputs, the output is Hermitian symmetric, meaning that the negative frequency terms are simply the complex conjugates of the corresponding positive frequency terms, thus the negative frequency terms are redundant. Therefore, this embodiment does not calculate the negative frequency terms, and the length of the output transform axis is N / / 2+1;

[0122] Step 3.1.6: The matrix obtained after the discrete Fourier transform is a complex-valued matrix. Since we take the absolute value of the result matrix to obtain its amplitude, we then perform corresponding calculations based on the power exponent of the given spectrogram to obtain the spectrogram of the protein music. In this embodiment, the value of power is 2, so we perform a square calculation.

[0123] Step 3.2: Construct a Mel filter bank. The filter bank uses a triangular filter to extract the frequency band of the spectrogram at the Mel scale, thereby obtaining the Mel spectrogram.

[0124] The Mel scale is designed to mimic the non-linear perception of sound by the human ear, exhibiting stronger discrimination at lower frequencies and weaker discrimination at higher frequencies. We can convert between frequency (f) and Mel using the following formula:

[0125] m = 2595log 10 (1+f / 700)

[0126] f = 700(10) m / 2595 -1)

[0127] Step 3.3: Convert the obtained Mel spectrogram to the dB scale to obtain the Mel spectrogram of protein music at the dB scale. The calculation formula is as follows:

[0128] S_dB = 10 * log 10 (S / ref)

[0129] Where ref is the maximum value in the Mel spectrogram;

[0130] Step 3.4: Save the obtained Mel spectrogram of the protein music at the dB scale, and adjust the Mel spectrogram at the dB scale into a matrix of size [n_channel, width, height], where n_channel, width, and height are the number of channels, width, and height of the Mel spectrogram at the dB scale, respectively; in this embodiment, n_channel = 3, width = height = 216.

[0131] Step 4: Set the parameters to be searched in the protein property prediction model, and the range of the parameters to be searched; initialize a population of N individuals according to the encoding rules; in this embodiment, N = 50;

[0132] Step 4.1: The model parameters to be searched are the L1 regularization coefficient λ1 and the L2 regularization coefficient λ2;

[0133] This invention utilizes the VGG16 model to extract frequency information from protein music converted to Mel spectrograms. Due to the large number of layers and parameters in the model, L1 and L2 regularization are introduced to penalize the VGG16 model parameters and prevent overfitting. The loss function after adding L1 and L2 regularization is as follows: Where L(θ) is the original loss function of the model, θ i These are the model parameters in VGG16. Adding L2 regularization reduces the impact of features on the overall model by decaying the weights, thus preventing overfitting and improving the model's generalization performance on new data. L1 regularization, on the other hand, penalizes the weights, making the parameters more sparse and reducing model complexity. In this embodiment, choosing appropriate penalty coefficients λ1 and λ2 is crucial, as λ1 and λ2 are set as the parameters to be searched.

[0134] Step 4.2: Set the range of parameters λ1 and λ2 to [0,1];

[0135] Step 4.3: Initialize a population of N individuals according to the encoding rules; in this embodiment, the initial population is randomly sampled from the search space;

[0136] Step 5: Construct a protein property prediction model;

[0137] Step 5.1: Construct a protein sequence feature extraction module to extract the sequence features of proteins; the protein sequence feature extraction module is constructed by combining a one-dimensional convolutional neural network and a linear layer network, and contains three layers, namely two convolutional neural network layers and one linear layer neural network layer.

[0138] In this embodiment, the first convolutional neural network has 21 input channels, 1024 output channels, a kernel size of 5, a stride of 1, and padding of 2; the second convolutional neural network has 1024 input channels, 1024 output channels, a kernel size of 5, a stride of 1, and padding of 2; the linear layer has 1024 input features and 512 output features.

[0139] Step 5.2: Construct a protein frequency feature extraction module to extract the frequency features of proteins; the protein frequency feature extraction module is constructed by using the feature extraction part of the VGG16 model and then concatenating it with a linear layer network.

[0140] In this embodiment, the VGG16 model is as follows: Figure 3 As shown, the network layers are as follows:

[0141] A two-dimensional convolutional neural network with 3 input channels, 64 output channels, a kernel size of 3, a stride of 1, and padding of 1.

[0142] ReLU activation function;

[0143] A two-dimensional convolutional neural network with 64 input channels, 64 output channels, a kernel size of 3, a stride of 1, and padding of 1.

[0144] ReLU activation function;

[0145] A two-dimensional max pooling layer with a kernel size of 2, a stride of 2, and a panning of 0;

[0146] A binary convolutional neural network with 64 input channels, 128 output channels, a kernel size of 3, a stride of 1, and padding of 1.

[0147] ReLU activation function;

[0148] A two-dimensional convolutional neural network with 128 input channels, 128 output channels, a kernel size of 3, a stride of 1, and padding of 1.

[0149] ReLU activation function;

[0150] A two-dimensional max pooling layer with a kernel size of 2, a stride of 2, and a panning of 0;

[0151] A two-dimensional convolutional neural network with 128 input channels, 256 output channels, a kernel size of 3, a stride of 1, and padding of 1.

[0152] ReLU activation function;

[0153] A two-dimensional convolutional neural network with 256 input channels, 256 output channels, a kernel size of 3, a stride of 1, and padding of 1.

[0154] ReLU activation function;

[0155] A two-dimensional convolutional neural network with 256 input channels, 256 output channels, a kernel size of 3, a stride of 1, and padding of 1.

[0156] ReLU activation function;

[0157] A two-dimensional max pooling layer with a kernel size of 2, a stride of 2, and a panning of 0;

[0158] A two-dimensional convolutional neural network with 256 input channels, 512 output channels, a kernel size of 3, a stride of 1, and padding of 1.

[0159] ReLU activation function;

[0160] A two-dimensional convolutional neural network with 512 input channels, 512 output channels, a kernel size of 3, a stride of 1, and padding of 1.

[0161] ReLU activation function;

[0162] A two-dimensional convolutional neural network with 512 input channels, 512 output channels, a kernel size of 3, a stride of 1, and padding of 1.

[0163] ReLU activation function;

[0164] A two-dimensional max pooling layer with a kernel size of 2, a stride of 2, and padding of 0;

[0165] A two-dimensional convolutional neural network with 512 input channels, 512 output channels, a kernel size of 3, a stride of 1, and padding of 1.

[0166] ReLU activation function;

[0167] A two-dimensional convolutional neural network with 512 input channels, 512 output channels, a kernel size of 3, a stride of 1, and padding of 1.

[0168] ReLU activation function;

[0169] A two-dimensional convolutional neural network with 512 input channels, 512 output channels, a kernel size of 3, a stride of 1, and padding of 1.

[0170] ReLU activation function;

[0171] A two-dimensional max pooling layer with a kernel size of 2, a stride of 2, and padding of 0;

[0172] Step 5.3: Construct a protein property prediction module. The protein property prediction module receives the protein sequence features and protein frequency features extracted by the protein sequence feature extraction module and the protein frequency feature extraction module, concatenates the two, and inputs them into a linear layer network for mapping to obtain the protein property prediction results.

[0173] Step 5.4: Define the loss function for the protein property prediction model and select the model optimizer;

[0174] Step 5.4.1: Select the Adam optimizer as the optimizer for the protein property prediction model and set the learning rate parameter. In this embodiment, the learning rate is 0.001.

[0175] Step 5.4.2: To prevent overfitting, L1 and L2 regularization terms are introduced into the loss function of the protein property prediction model. Therefore, the loss function can be defined as:

[0176]

[0177] Where L'(θ) is the loss function with L1 regularization and L2 regularization terms introduced, and L(θ) is the model loss function. The choice of L(θ) varies depending on the task. For classification tasks, cross-entropy loss is chosen as the loss function, and for regression tasks, mean square error (MSE) is chosen as the loss function.

[0178] Step 6: Forward calculation of the protein property prediction model to extract the protein's sequence features and frequency features to predict the protein's properties;

[0179] Step 6.1: Extract protein sequence features using the protein sequence feature extraction module. The input to this module is the one-hot representation of the protein sequence; the protein sequence features are obtained after calculation through two convolutional layers and one linear layer.

[0180] The calculation formula for a one-dimensional convolutional neural network is shown below:

[0181]

[0182] The dimension of the input data is (N, C). in The output dimension is (N, C). out ,L out N is the number of input samples, C represents the number of channels, and L is the length of the input sample sequence.

[0183] The transformation of the linear layer is:

[0184] y = xA T +b

[0185] Step 6.1.1: Pass the one-hot representation of the input protein sequentially through the first layer of a one-dimensional convolutional neural network, the ReLU activation function, and the second layer of a one-dimensional convolutional neural network, the ReLU activation function.

[0186] Step 6.1.2: The protein sequence one-hot representation after passing through the convolutional neural network has a dimension of [batch_size, seq_len, o_channel], where batch_size is the batch size, seq_len is the length of the protein sequence, and o_channel is the number of output channels, i.e., the feature dimension of the protein sequence features after passing through the convolutional neural network. Then, the MaxReadout method is used to take the maximum value in the protein sequence length dimension, and the result is calculated as [batch_size, o_channel]. In this embodiment, batch_size is 128 and o_channel is 1024.

[0187] Step 6.1.3: Finally, the calculation results of step 6.1.2 are mapped to [batch_size, hidden_dim1] through the constructed linear layer network, where hidden_dim1 is the dimension of the protein sequence features. In this embodiment, hidden_dim1 = 512.

[0188] Step 6.2: Extract protein frequency features using the protein frequency feature extraction model block. The input to this module is the Mel spectrogram of the protein music converted to the dB scale.

[0189] Step 6.2.1: The input to the protein frequency feature extraction module is the Mel spectrogram of the protein music at the dB scale. The input dimension is [batch_size, n_channel, width, height], where n_channel is the number of channels in the protein music Mel spectrogram at the dB scale, and width and height are the width and height of the input protein music Mel spectrogram at the dB scale. In this embodiment, n_channel = 3, width = height = 216. Then, the feature extraction part of VGG16 extracts the Mel spectrogram features.

[0190] Step 6.2.2: After feature extraction via the VGG16 network, the resulting matrix is ​​flattened to [batch_size, hidden_dim2], where hidden_dim2 is the feature dimension after passing through the VGG16 network. Then, the protein frequency features are calculated sequentially through dropout layers and linear layers to obtain the protein frequency features. The dimension of the frequency features is [batch_size, hidden_dim3], where hidden_dim3 is the dimension of the protein frequency features. In this embodiment, hidden_dim3 = 512.

[0191] Step 6.3: Protein property prediction. By receiving the extracted protein sequence features and protein frequency features, the properties of the protein are predicted.

[0192] Step 6.3.1: By concatenating the received protein sequence features and protein frequency features according to the feature dimension, the protein representation is obtained. The protein representation dimension is [batch_size, hidden_dim1+hid_dim3];

[0193] Step 6.3.2: Input the protein representation into the linear layer to predict protein properties;

[0194] Step 7: Train and validate the protein property prediction model, and obtain fitness values ​​for different individuals;

[0195] Step 7.1: Calculate the model loss according to the loss function defined in the task, backpropagate, and calculate the corresponding gradient;

[0196] Step 7.2: Use the Adam optimizer to update the parameters of the protein property prediction model based on the gradient;

[0197] Step 7.3: Repeat steps 6-7.2 above until the loss of the protein property prediction model on the validation set no longer decreases;

[0198] Step 7.4: The minimum loss of the obtained protein property prediction model on the validation set is used as the fitness value of the current individual;

[0199] Step 8: Determine if the termination condition is met. If the termination condition is met, proceed to step 10; otherwise, continue to step 9. In this embodiment, the termination condition is set to the number of evolutions of the population, MAXGEN = 10.

[0200] Step 9: Population selection, crossover, and mutation operations;

[0201] Step 9.1: Independently select N-1 individuals from the current population;

[0202] Step 9.2: Perform crossover operations independently on these N-1 individuals;

[0203] In this embodiment, a single-point crossover algorithm is used to perform the crossover operation. Single-point crossover means that only one crossover point is randomly set in the individual encoding string, and then parts of the chromosomes of the two ligand individuals are exchanged at that point.

[0204] Step 9.3: Independently mutate these N-1 crossover individuals;

[0205] In this embodiment, the mutation operation selected is the mutation operator in the Breeder genetic algorithm, and each parameter to be searched is expressed as p. m The mutation operation is performed with a probability of 1 / n, where n is the number of parameters to be searched. At least one variable in each individual undergoes mutation. The calculation of variable mutation is: z i =x i ±range i ·δ, where range i To define the range of variation, it is generally set to 0.1 of the corresponding search variable range. α i ∈0,1;

[0206] Step 9.4: Calculate the best individual in the current population and insert it into the first position of the N-1 individuals after crossover mutation to obtain the new generation population;

[0207] Step 9.5: Return to step 5 to construct the protein property prediction model and calculate the fitness value of the new population;

[0208] Step 10: Protein property prediction model testing; train the model using the optimal combination of model parameters obtained from the search and calculate the model loss and corresponding evaluation metrics on the test set (the loss function for regression tasks is MSE, and the evaluation metric is Spearman's Rho; for classification tasks, the loss function is cross-entropy loss, and the evaluation metric is classification accuracy).

[0209] This embodiment conducted experiments on the Beta-lactamase Activity Prediction task and the BinaryLocalization Prediction task to verify the method of the present invention.

[0210] The Beta-lactamase Activity Prediction task aims to predict the activity of primary mutants of the TEM-1 β-lactamase protein. The prediction target is the experimentally tested fitness score (a real number), which records the mutational effect of each mutant. This task is a regression task, with training, validation, and test sets of sizes 4158, 520, and 520, respectively. The evaluation metric for the experimental results is Spearman's Rho (the larger the better). As shown in Table 3, the table displays the results of different models on this task. The model "Our model (base)" in the table represents the result without L1 and L2 regularization, but with the addition of protein music encoding frequency features to the original CNN model. The Spearman's Rho metric improved from 0.781 to 0.823, an improvement of 0.042, achieving a good result. In this embodiment, LSTM, ResNet, Transformer, and ProBERT were selected for comparison. Compared with these models, the protein property prediction model of this invention achieved the best result on the Beta-lactamase Activity Prediction task. In order to further improve the accuracy of the prediction results, after adding a genetic algorithm with an elite preservation strategy to search for L1 and L2 regularization coefficients, it can be seen that the accuracy of the model prediction results was further improved to 0.838.

[0211] Table 3. Experimental results of the Beta-lactamase Activity Prediction task.

[0212] Model Test Spearman's Rho #Params LSTM 0.139 27,080,328 ResNet 0.152 11,300,354 Transformer 0.261 21,545,985 CNN 0.781 6,403,073 ProtBert 0.731 420,981,761 Our model (base) 0.823 30,030,657 Our model 0.838 30,030,657

[0213] The Binary Localization Prediction task predicts whether a protein is "membrane-bound" or "soluble." This is a binary classification task, and the evaluation metric for the experimental results is accuracy, as shown in Table 4. The training sample size was 5161, the validation sample size was 1727, and the test sample size was 1746. On this task, the protein property prediction model of this invention achieved a good result. Although the result is not as good as LSTM and ProtBERT, it is worse than ResNet and Transformer. Because the basic model of this invention is CNN, the model enhanced with the frequency features of protein music improved the accuracy by 1.09% compared to the original CNN model. Further, after using a genetic algorithm to search for the model's parameters, the prediction accuracy was further improved to 85.13%, demonstrating the effectiveness of the method of this invention.

[0214] Table 4. Experimental Results of the Binary Localization Prediction Task

[0215] Model Test Acc #Params LSTM 88.11 27,080,969 ResNet 78.99 11,300,867 Transformer 75.74 21,546,498 CNN 82.67 6,404,098 ProtBert 91.32 420,982,786 Our model (base) 83.76 30,031,682 Our model 85.13 30,031,682

[0216] The above results demonstrate that the model obtained using the method of this invention outperforms the original model that simply extracts features from sequence information, achieving better experimental results. In the Beta-lactamase Activity Prediction task, compared to the protein property prediction model based on convolutional neural networks, the Spearman's Rho metric improved from 0.781 to 0.838, an improvement of 0.057. Furthermore, the model of this invention achieved the best results compared to LSTM, ResNet, Transformer, and ProBERT. In the Binary Localization Prediction task, compared to the protein property prediction model based on convolutional neural networks, the accuracy improved from 82.67% to 85.13%, an improvement of 2.46%.

[0217] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.< / unk> < / unk>

Claims

1. A protein property prediction method based on the fusion of deep learning and intelligent optimization, characterized in that: Includes the following steps: Step 1: Obtain protein sequence data, divide the protein sequence data into training set, validation set and test set, and convert the representation of the protein amino acid sequence into numerical form; Step 2: Encode the protein sequence into protein music; obtain the sequence information of the protein, map each amino acid of the protein to a note with different beats and volumes, and encode the protein sequence into protein music. Step 3: Generate the mel spectrogram of the protein music and extract the frequency features of the protein sequence; Step 4: Set the parameters to be searched in the protein property prediction model, and the range of the parameters to be searched; initialize a population of N individuals according to the encoding rules; Step 5: Construct a protein property prediction model; Step 5.1: Construct a protein sequence feature extraction module to extract the sequence features of proteins; The protein sequence feature extraction module is constructed using a combination of one-dimensional convolutional neural networks and linear layer networks, consisting of three layers: two convolutional neural networks and one linear layer neural network. Step 5.2: Construct a protein frequency feature extraction module to extract the frequency features of proteins; the protein frequency feature extraction module is constructed by using the feature extraction part of the VGG16 model and then concatenating it with a linear layer network. Step 5.3: Construct a protein property prediction module. The protein property prediction module receives the protein sequence features and protein frequency features extracted by the protein sequence feature extraction module and the protein frequency feature extraction module, concatenates the two, and inputs them into a linear layer network for mapping to obtain the protein property prediction results. Step 5.4: Define the loss function for the protein property prediction model and select the model optimizer; Step 5.4.1: Select the Adam optimizer as the optimizer for the protein property prediction model and set the learning rate parameter; Step 5.4.2: To prevent overfitting, L1 regularization and L2 regularization terms are introduced into the loss function of the protein property prediction model; Step 6: Forward calculation of the protein property prediction model to extract the protein's sequence features and frequency features to predict the protein's properties; Step 7: Train and validate the protein property prediction model, and obtain fitness values ​​for different individuals; Step 8: Determine if the termination condition is met. If the termination condition is met, proceed to step 10; otherwise, continue to step 9. Step 9: Population selection, crossover, and mutation operations; Step 10: Protein property prediction model testing; Train the model using the optimal combination of model parameters obtained from the search and calculate the model loss and corresponding evaluation metrics on the test set.

2. The protein property prediction method based on the fusion of deep learning and intelligent optimization according to claim 1, characterized in that: The specific method for step 1 is as follows: Step 1.1: Segment the protein sequence into a list of amino acid symbols; Step 1.2: Convert the amino acids to their corresponding numbers according to the correspondence between amino acid symbols and numbers; Step 1.3: Convert the obtained numerical representation of amino acids into the corresponding one-hot encoded representation; Each amino acid in the protein sequence is represented as a 21-dimensional one-hot vector, with the number of bits corresponding to the number in the vector being 1 and the remaining bits being 0.

3. The protein property prediction method based on the fusion of deep learning and intelligent optimization according to claim 2, characterized in that: The specific method for step 2 is as follows: Step 2.1: Classify amino acids according to the properties of their R groups and establish a correspondence between amino acids and musical notes; Based on the properties of the R group of amino acids, the 20 common amino acids are divided into 5 categories: nonpolar aliphatic R group amino acids, aromatic R group amino acids, polar uncharged R group amino acids, positively charged R group amino acids, and negatively charged R group amino acids. Glycine is defined as 60, representing the middle note C; methionine and proline correspond to notes D4 and E4, respectively; alanine, valine, leucine, and isoleucine are ordered according to their hydrophilicity index, corresponding to notes F4, A4, G4, and B4, respectively; aromatic R-group amino acids, including phenylalanine, tyrosine, and tryptophan, correspond to notes B3, G3, and A3, respectively; ordered according to their isoelectric point, aspartic acid and glutamic acid, among the negatively charged R-group amino acids, correspond to notes C5 and D5, respectively; serine, threonine, cysteine, asparagine, and glutamine, among the polar uncharged R-group amino acids, correspond to notes A5, B5, E5, F5, and G5, respectively; and lysine, arginine, and histidine, among the positively charged R-group amino acids, correspond to notes D6, E6, and C6, respectively. Step 2.2: Obtain the hydrophilicity index (hydrophilicity_value) of the amino acid, and map the obtained hydrophilicity index (hydrophilicity_value) to the amplitude of the note corresponding to the current amino acid, where 0 represents mute and 127 represents maximum volume. Limit the amplitude mapping result of the note corresponding to the amino acid to [...]. , Between, among which, , These are the minimum and maximum values ​​of amplitude, respectively. Step 2.3: Obtain the isoelectric point pI_value of the amino acid and map the obtained isoelectric point pI_value of the amino acid to the duration time of the current amino acid note; Step 2.4: Set the timbre and tempo of the protein music track; Step 2.5: Generate protein music according to the parameters in steps 2.2-2.4 above.

4. The protein property prediction method based on the fusion of deep learning and intelligent optimization according to claim 3, characterized in that: The specific method for step 3 is as follows: Step 3.1: Calculate the corresponding spectrogram based on the protein music generated in Step 2; Step 3.1.1: Given the following calculation parameters: sampling rate sr, highest frequency fmax, fast Fourier transform window size n_fft, sampling interval between consecutive frames hop_length, number of generated Mel bands n_mels, and spectrogram exponent power; Step 3.1.2: Pad the ends of the input protein music, with a padding length of half the window size n_fft / 2, and the padding value is a constant 0; Step 3.1.3: Frame the completed protein music to obtain short time frames; Step 3.1.4: After slicing the completed protein music into frames, apply a Hann window function to each frame; Step 3.1.5: Perform Discrete Fourier Transform on each frame after frame division and window calculation, and concatenate the vectors obtained from the Discrete Fourier Transform of each frame to obtain the result matrix. Step 3.1.6: Take the absolute value of the result matrix to obtain its amplitude, and then perform corresponding calculations based on the power of the given spectrogram to obtain the spectrogram of the protein music. Step 3.2: Construct a Mel filter bank. The filter bank uses a triangular filter to extract the frequency band of the spectrogram at the Mel scale, thereby obtaining the Mel spectrogram. Step 3.3: Convert the obtained Mel spectrogram to the dB scale to obtain the Mel spectrogram of protein music at the dB scale; Step 3.4: Save the Mel spectrogram of the protein music at the dB scale, and adjust the Mel spectrogram at the dB scale into a matrix of size [n_channel, width, height], where n_channel, width, and height are the number of channels, width, and height of the Mel spectrogram at the dB scale, respectively.

5. The protein property prediction method based on the fusion of deep learning and intelligent optimization according to claim 4, characterized in that: The specific method for step 4 is as follows: Step 4.1: The model parameters to be searched are the L1 regularization coefficient λ1 and the L2 regularization coefficient λ2; Step 4.2: Set the range of parameters λ1 and λ2 to [0, 1]; Step 4.3: Initialize a population of N individuals according to the encoding rules.

6. The protein property prediction method based on the fusion of deep learning and intelligent optimization according to claim 5, characterized in that: The loss function of the protein property prediction model is selected according to the task. For classification tasks, cross-entropy loss is selected as the loss function, and for regression tasks, mean squared error is selected as the loss function.

7. The protein property prediction method based on the fusion of deep learning and intelligent optimization according to claim 6, characterized in that: The specific method for step 6 is as follows: Step 6.1: Extract protein sequence features using the protein sequence feature extraction module. The input to this module is the one-hot representation of the protein sequence; the protein sequence features are obtained after calculation through two convolutional layers and one linear layer. Step 6.1.1: Pass the one-hot representation of the input protein sequentially through the first layer of a one-dimensional convolutional neural network, the ReLU activation function, and the second layer of a one-dimensional convolutional neural network, the ReLU activation function. Step 6.1.2: The protein sequence one-hot representation after passing through the convolutional neural network has a dimension of [batch_size, seq_len, o_channel], where batch_size is the batch size, seq_len is the length of the protein sequence, and o_channel is the number of output channels. That is, the feature dimension of the protein sequence features after passing through the convolutional neural network. Then, the MaxReadout method is used to take the maximum value in the protein sequence length dimension, and the result is calculated as [batch_size, o_channel]. Step 6.1.3: Finally, the calculation results of step 6.1.2 are mapped to [batch_size, hid_dim1] through the constructed linear layer network, where hid_dim1 is the dimension of the protein sequence features; Step 6.2: Extract protein frequency features using the protein frequency feature extraction module. The input to this module is the Mel spectrogram of the protein music converted to the dB scale. Step 6.2.1: The input to the protein frequency feature extraction module is the Mel spectrogram of the protein music at the dB scale. The input dimension is [batch_size, n_channel, width, height], where n_channel is the number of channels in the protein music Mel spectrogram at the dB scale, and width and height are the width and height of the input protein music Mel spectrogram at the dB scale, respectively. Then, the feature extraction part of VGG16 extracts the features. Step 6.2.2: After feature extraction via the VGG16 network, the resulting matrix is ​​flattened to [batch_size, hidden_dim2], where hidden_dim2 is the feature dimension after passing through the VGG16 network. Then, the protein frequency features are calculated sequentially through dropout layers and linear layers to obtain the protein frequency features, where the frequency features have the dimension [batch_size, hidden_dim3], where hidden_dim3 is the protein frequency feature dimension. Step 6.3: Protein property prediction. By receiving the extracted protein sequence features and protein frequency features, the properties of the protein are predicted. Step 6.3.1: By concatenating the received protein sequence features and protein frequency features according to the feature dimension, the protein representation is obtained. The protein representation dimension is [batch_size, hid_dim1 + hid_dim3]; Step 6.3.2: Input the protein representation into the linear layer to predict protein properties.

8. The protein property prediction method based on the fusion of deep learning and intelligent optimization according to claim 7, characterized in that: The specific method for step 7 is as follows: Step 7.1: Calculate the model loss according to the loss function defined in the task, backpropagate, and calculate the corresponding gradient; Step 7.2: Use the Adam optimizer to update the parameters of the protein property prediction model based on the gradient; Step 7.3: Repeat steps 6-7.2 above until the loss of the protein property prediction model on the validation set no longer decreases; Step 7.4: The minimum loss of the obtained protein property prediction model on the validation set is used as the fitness value of the current individual.

9. The protein property prediction method based on the fusion of deep learning and intelligent optimization according to claim 8, characterized in that: The specific method for step 9 is as follows: Step 9.1: Independently select N-1 individuals from the current population; Step 9.2: Perform crossover operations independently on these N-1 individuals; Step 9.3: Independently mutate these N-1 crossover individuals; Step 9.4: Calculate the best individual in the current population and insert it into the first position of the N-1 individuals after crossover mutation to obtain the new generation population; Step 9.5: Return to step 5 to construct the protein property prediction model and calculate the fitness value of the new population.

Images