Protein anomaly detection method based on ultra-large-scale evolutionary algorithm and hardware acceleration

Abstract

The present invention discloses a protein anomaly detection method based on a super-large-scale evolutionary algorithm and hardware acceleration, comprising: 1. collecting mass spectrometry feature data; 2. generating a mass spectrometry feature selection scheme population and external archive, and setting parameters; 3. updating the external archive and rapidly clustering and grouping the mass spectrometry features; 4. after performing mating pool selection, simultaneously generating offspring mass spectrometry feature selection schemes in the original space and the reduced space after grouping; 5. adaptively adjusting algorithm parameters using the offspring feature selection scheme, then merging the offspring and parent mass spectrometry feature selection scheme populations for environmental selection, iterating the above process to select high-quality feature selection schemes, and ultimately obtaining the optimal protein anomaly detection feature selection scheme. The present invention can efficiently handle the problem of super-large-scale protein anomaly detection, quickly obtain high-quality feature selection schemes in a super-large-scale mass spectrometry feature set, and perform protein anomaly detection to improve detection accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of feature selection for ultra-large-scale multi-objective problems, and in particular to a protein anomaly detection method based on an ultra-large-scale evolutionary algorithm and hardware acceleration. Background Art

[0002] In recent years, the field of protein abnormality detection has attracted increasing attention. By detecting and classifying the abundance of a large number of proteins in human serum at given mass values, it can be used in applications such as assisting abnormality judgment and drug treatment.

[0003] Detecting protein anomalies can help us understand human health. With the continuous advancement of biotechnology, the collection and analysis of protein features is increasing, allowing for the extraction of a large number of features. However, as the scale of data increases, the difficulty of selecting feature selection schemes for protein anomaly detection also increases. In the problem of protein anomaly detection, a large amount of feature data, on the one hand, allows us to obtain more information and more accurately identify whether a protein is abnormal. On the other hand, due to the mutual influence of features and the increase of redundant features, it may also lead to increased computational effort and decreased accuracy. Therefore, how to remove redundant features and select high-quality mass spectrometry feature selection schemes has become the focus of feature selection technology for protein anomaly detection.

[0004] However, most current protein anomaly detection methods often require excessive computation, consume a long time, and occupy a large amount of memory on large-scale mass spectrometry feature sets due to the high number of dimensional features, making it impossible to efficiently obtain high-quality protein anomaly detection feature selection schemes. Although existing methods of grouping and dimensionality reduction of features can effectively reduce the amount of computation and time consumption, they do not take into account the interactions between features, and it is easy to lose high-quality protein mass spectrometry features, thereby generating inaccurate mass spectrometry feature selection schemes and leading to incorrect detection of whether proteins are abnormal. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, the present invention proposes a mass spectrometry data classification feature selection method based on a very large-scale evolutionary algorithm and hardware acceleration, in order to efficiently handle the problem of very large-scale protein anomaly detection, quickly obtain high-quality feature selection schemes in very large-scale mass spectrometry data feature sets and perform protein anomaly detection, thereby improving the accuracy of detection and classification.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

[0007] The protein anomaly detection method based on a super-large-scale evolutionary algorithm and hardware acceleration of the present invention is characterized by the following steps:

[0008] Step 1: Use the SELDI method to obtain an X×D-dimensional mass spectral feature set F of serum, which is used to characterize the abundance of proteins with given mass values in serum, where X represents the number of samples and D represents the number of mass spectral features corresponding to each sample;

[0009] Step 2: Set the population size to N, and define the population set as P = {p1, p2, ···, p i ,···,p N}, p i represents the i-th individual in the population, and each individual represents a mass spectral feature selection scheme consisting of selecting different mass spectral features from D mass spectral features, and p i ={p i1 ,p i2 ,···,p ij ,···,p iD}, p ij represents the i-th individual p i Whether the jth mass spectrum feature is selected into the mass spectrum feature selection scheme, if p ij =1, it means selected, if p ij =0, it means not selected; j∈[1,D]; the search space where the individuals in the population P are located is defined as the original space, and the dimension of the original space is D;

[0010] Step 2.1, calculate the score of each mass spectrometric feature as the importance of protein abnormality detection;

[0011] Step 2.1.1. Generate a set of D mass spectrum feature selection schemes Q = {q1,q2,···,q k ,···,q D},q k represents the kth individual in the set Q, and q k ={q k1 ,q k2 ,···,q kj ,···,q kD},q kj represents the kth individual q k Whether the jth mass spectrum feature of is selected into the mass spectrum feature selection scheme, and q k Only the kth mass spectrum feature is selected, that is, q kk =1, the rest q kj =0, k≠j;

[0012] Step 2.1.2: Calculate the kth individual q using formula (1) k The classification error rate indicator error(q k ):

[0013] error(qk )=EN(q k ) / X (1)

[0014] In formula (1), EN(q k ) means using the kth individual q k The number of samples misclassified when the mass spectral feature selection scheme represented by k The number of mass spectral features selected in the mass spectral feature selection scheme represented by k );

[0015] Step 2.1.3, use the error rate of each individual in the set Q and the number of selected mass spectrum features to perform non-dominated sorting on all individuals, and take the frontier number where the individual is located as the score of the individual, recorded as Score = {score1, score2, ···, score k ,···,score D}, where score k represents the kth individual q k The score of , i.e., the importance of the k-th mass spectrum feature;

[0016] Step 2.2, initialization;

[0017] Step 2.2.1. Define and initialize the number of iterations t = 1, and the probability of the tth generation generating offspring in the original space is ρ t , the number of groups in the tth generation is K t , the maximum number of iterations is T; initialize the external archive A of the mass spectrometry feature selection scheme t Empty set; external archive A t The default value is N A ;

[0018] Step 2.2.2, initialize the t-th generation population in, represents the t-th generation population P t The i-th individual of , that is, the i-th mass spectrum feature selection scheme, and initialize each individual with a full 0 vector of length D;

[0019] Step 2.2.3, initialize i=1;

[0020] Step 2.2.4, generate the i-th random number rand i ;

[0021] Step 2.2.5, from the i-th individual Randomly select two mass spectrum features from the equation, and record the indexes of the two selected mass spectrum features as m and n respectively. represents the i-th individual Whether the mth mass spectrum feature is selected, represents the i-th individual Whether the nth mass spectrum feature is selected; if score m <score n , then let Indicates that the mass spectrometry feature selection scheme individual The mth mass spectrum feature is selected, otherwise, let Indicates that the feature selection scheme individual The nth mass spectrum feature is selected;

[0022] Step 2.2.6, repeat rand i ×D times step 2.2.5, thus obtaining the updated i-th individual and assigning it to

[0023] Step 2.2.6, after assigning i+1 to i, determine whether i>N is true. If so, it means that the updated t-th generation population is obtained and assigned to P t Then use it as the initial population; otherwise, return to step 2.2.4;

[0024] Step 3: Update external archive A t and quickly cluster and group mass spectral features;

[0025] Step 3.1, the tth generation population P t and external archive A t Merge and perform non-dominated sorting on the merged t-th generation set, assigning the first sorted subset to the external archive A t ;

[0026] If the external archive A after assignment t The number of mass spectral feature selection schemes in exceeds N A , then calculate A t Crowding distance for mass spectral feature selection, removing A t The mass spectrum feature scheme with the smallest crowding distance until A t The number of mass spectrometry feature schemes does not exceed N A So far, the updated external archive is obtained and assigned to A t ;

[0027] Step 3.2: Based on external archive A t and the number of groups K in the tth generation t Clustering the mass spectrum features in the original space;

[0028] Step 3.2.1, from A t All mass spectrometry feature selection schemes are obtained as a set and Indicates A t The situation vector of the jth mass spectrum feature being selected in different individuals, where |A t | indicates A t The number of mass spectral feature selection schemes, Indicates that the jth mass spectrum feature is in A t Is the ath individual in the selected? If It means not selected; a∈[1,|A t ];

[0029] Using formula (2), we can obtain the situation vector of the jth mass spectrum feature being selected in different individuals: Sparsity Thus we get the sparsity of the tth generation t ;

[0030]

[0031] In formula (2), Representation vector The number of elements that are 1;

[0032] Step 3.2.2: The mass spectrum feature with the sparsity closest to 0.5 among all the mass spectrum features selected is used as the reference mass spectrum feature. The reference mass spectrum feature index is denoted as c. The similarity between the bth mass spectrum feature and the reference mass spectrum feature is calculated using formula (3). Thus, all mass spectrometry features are sorted according to similarity to obtain the similarity ranking of the tth generation t ;

[0033]

[0034] In formula (3), Indicates A t The number of feature selection schemes in which the b-th mass spectral feature is not selected and the reference mass spectral feature is selected; Indicates A t The number of feature selection schemes in which the b-th mass spectral feature is selected and the reference mass spectral feature is not selected; Indicates A t The number of feature selection schemes in which the b-th mass spectrum feature is selected and the reference mass spectrum feature is selected;

[0035] Step 3.2.4, Sparsity t The mass spectrum feature selection vectors with a median value of 1 are placed in the same group The mass spectrum feature selection vector with a value of 0 is placed in From rank tRemove and After the corresponding mass spectrum features are obtained, a new similarity ranking is obtained and assigned to rank t ;

[0036] rank t The mass spectral features in K t -2 groups, thus obtaining the index set of grouped mass spectral features, recorded as in, Represents the index set of mass spectrum features of group g, g∈[3,K t ],in Respectively The indices of the first and last mass spectral features in , Indicates the size of the group; |rank t | indicates similarity ranking t length;

[0037] Step 4: After performing the mating pool selection, the progeny mass spectrum feature selection scheme is generated in both the grouped reduced space and the original space;

[0038] Step 4.1: For the t-th generation population P t Perform non-dominated sorting and crowding distance calculation, and then use the binary tournament method to select the population P according to the ranking and crowding distance. t Select 2N mass spectrum feature selection schemes and use them as the t-th generation parent set P′ t ,and in, is the t-th generation parent set P′ t The sth mass spectrum feature selection scheme in ; s∈[1,2N];

[0039] Step 4.2, generate a random vector rand={r1,r2,···,r s ,···,r 2N}, where r s Represents the sth random number;

[0040] Define and initialize the partition vector index t is a vector of all zeros of length 2N; and in, Represents the t-th generation parent set P′ t The sth mass spectrometry feature selection scheme It is the parent individual that generates offspring in the original space or the parent individual that generates offspring in the reduced space;

[0041] If r s <ρ t, then let Otherwise, let Thus, the updated partition vector is obtained and assigned to index t ,in, Represents the t-th generation parent set P′ t The sth mass spectrometry feature selection scheme is the parent individual that generates offspring in the original space, Represents a mass spectrometry feature selection scheme The parent individuals that generate offspring in the reduced space;

[0042] Step 4.3, according to the partition vector index t The t-th generation parent set P′ t The subset of parent individuals that generate offspring in the original space is recorded as PB t ,and in, Represents the parent individual subset PB that generates offspring in the original space t The e-th mass spectrum feature selection scheme in , and Represents the e-th mass spectrum feature selection scheme Whether the jth mass spectrum feature in |PB is selected; t |The number of parents that generate children for the original space, e∈[1,|PB t ];

[0043] According to the partition vector index t The parent set P′ t The parent individual subset that generates offspring in the reduced space is recorded as PB′ t ,and in, Represents the parent individual subset PB′ that generates offspring in the reduced space t The fth mass spectrum feature selection scheme in , and Indicates the fth mass spectrum feature selection scheme Whether the jth mass spectrum feature is selected; |PB′ t |To reduce the space to generate the number of parents of children, f∈[1,|PB′ t ];

[0044] Step 4.4: Group the index set based on the grouped mass spectrum features t PB′ t Individuals are mapped to K t dimensional reduced space and group them into t The mass spectrum characteristics corresponding to each group in the is regarded as a dimension, thereby mapping the D-dimensional individuals in the original space to the reduced space into K t dimensional individuals;

[0045] Use formula (4) to calculate the fth individual in the original space The g-th group The probability that the g-th position in the reduced space is 1 for the corresponding mass spectrum feature selection

[0046]

[0047] In formula (4), Indicates the g-th group of the t-th generation Number of indexes in the

[0048] The generated dimension is |PB′ t |×K t The random matrix rand′, if Then let the parent set PB in the reduction space *t The fth individual The g-th bit of is set to 1, otherwise, the g-th bit is set to 0, thus obtaining the parent set PB in which the original space is mapped to the reduced space. *t ; where rand′ fg represents the gth random number of the fth random vector in rand′;

[0049] Step 4.5: Based on the parent set PB in the original space t Generate the offspring set OB in the original space t ;

[0050] Step 4.5.1. Initialize the offspring set OB t is an empty set;

[0051] Step 4.5.2: From the parent individual subset PB of the original space t Randomly select two individuals as parents u t and parent v t ,and Indicates whether the jth mass spectrum feature is selected, Indicates opt-in, Indicates not to be selected; Indicates whether the jth mass spectrum feature is selected;

[0052] Initialize the child individual o to the parent u;

[0053] Step 4.5.3. Generate a random number rand″1. If rand1″<0.5, randomly select two mass spectral features from the mass spectral features selected in parent u and not selected in parent v, compare the scores of the two mass spectral features, and set the mass spectral feature with the larger score corresponding to the mass spectral feature in offspring individual o as not selected. Otherwise, randomly select two mass spectral features from the mass spectral features selected in parent u and selected in parent v, compare the scores of the two mass spectral features, and set the mass spectral feature with the smaller score corresponding to the mass spectral feature in offspring individual o as selected.

[0054] Step 4.5.4, generate a random number rand″2. If rand″2 is less than 0.5, randomly select two mass spectral features from the mass spectral features selected by the offspring o, and compare the scores of the two mass spectral features. The mass spectral feature with the larger score in the offspring individual o is set as not selected; otherwise, randomly select two mass spectral features from the mass spectral features not selected by the offspring individual o, compare the scores of the two mass spectral features, and set the mass spectral feature with the smaller score in the offspring individual o as selected; thereby adding the offspring individual o to the offspring set OB t middle;

[0055] Step 4.5.5: Set PB from the parent t After deleting the parent u and parent v, if the parent set PB t If it is not empty, return to step 4.5.2 to execute, otherwise, it means that the t-th generation offspring set OB in the original space is obtained t ;

[0056] Step 4.6: For the parent set PB in the reduced space *t Use single-point crossover and bitwise mutation to obtain the offspring set in the reduced space in, Represents the offspring set OB *t The hth individual in Represents the t-th generation offspring set OB *t The hth individual In the gth position of the reduced space; |OB *t | represents the offspring set OB *t the number of individuals in

[0057] Step 4.7, the offspring individuals The offspring mapped into the original space is denoted by Represents a feature selection scheme Whether the jth mass spectrum feature in the original space is selected; if the jth mass spectrum feature in the original space Belongs to the gth group Then the offspring individuals in the space will be reduced The gth position Assign to Thus, the set of descendants of the reduced space mapped to the original space is obtained Denotes the offspring set OB′ t The hth individual in |OB′ t | represents the offspring set OB′ t The number of individuals in OB *t |=|OB′ t |;

[0058] Step 4.8, the original space offspring set OB t and the sub-set OB′ of the reduced space mapped to the original space t The union of the t-th generation offspring population O t , let the t-th generation offspring population O t The scale is N;

[0059] Step 5: Use the t-th generation offspring population O t Adaptive adjustment of the number of groups K t and probability ρ t , the t-th generation offspring population O t and the tth generation population P t Environmental selection is performed after the merger;

[0060] Step 5.1: Calculate the number of mass spectrum feature groups K of the t+1 generation using equations (5) and (6) respectively. t+1 and the probability ρ of the t+1th generation t+1 :

[0061]

[0062]

[0063] In formula (5), s 1,t represents the t-th generation population P t The number of children generated in the reduced space, ns 1,t represents the t-th generation population P t The number of non-dominated solutions in the offspring generated in the reduced space, s 1,t-1 represents the t-1 generation population P t-1 The number of children generated in the reduced space, ns 1,t-1 represents the t-1 generation population P t-1 The number of non-dominated solutions in the offspring generated in the reduced space; when t = 1, the initial population is assigned to the t-1 generation population P t-1 ;

[0064] In formula (6), s 2,t represents the t-th generation population P tThe number of children generated in the original space, ns 2,t represents the t-th generation population P t The number of non-dominated solutions in the offspring generated in the original space;

[0065] Step 5.2: Merge the t-th generation population P t and the offspring population O t , and perform non-dominated sorting and crowding distance calculation on the merged 2N individuals after deduplication, so as to select the first N individuals after sorting as the t+1 generation population and assign them to P t ;

[0066] Step 5.3: After assigning t+1 to t, determine whether t>T is true. If so, output the optimal feature selection scheme population P. t , and the optimal feature selection scheme population P t The mass spectrometry features selected in the step are used as the basis for detecting whether the protein in the serum is abnormal; otherwise, the process returns to step three and executes the sequence.

[0067] Compared with the existing technology, the beneficial effects of the present invention are embodied in:

[0068] 1. This paper proposes an efficient initialization strategy and offspring generation method, which takes into account the interaction between mass spectrometry features and their influence on anomaly detection. On this basis, a parameter adaptation strategy is designed to generate high-quality mass spectrometry feature selection schemes, thereby improving the accuracy of protein anomaly detection.

[0069] 2. The present invention designs a fast and effective feature grouping strategy, which divides the original large-scale mass spectrometry features into several groups and maps them into a reduced space with lower dimensionality. While consuming less computing resources, it efficiently adjusts and optimizes the mass spectrometry feature subsets to quickly obtain a mass spectrometry feature selection scheme for protein anomaly detection.

[0070] 3. All calculations of the evolutionary method in the present invention are based on matrices. Even when the number of mass spectrometry features is large, high-quality mass spectrometry feature selection schemes for protein anomaly detection can still be quickly obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 is a flow chart of the method of the present invention;

[0072] Figure 2 This is a schematic diagram of the rapid grouping method according to the present invention. DETAILED DESCRIPTION

[0073] In this embodiment, a protein anomaly detection method based on a super-large-scale evolutionary algorithm and hardware acceleration is provided. Figure 1As shown in the figure, a high-quality mass spectrometry feature selection scheme for protein anomaly detection is obtained through the designed initialization strategy, offspring generation strategy, and parameter adaptation strategy. The mass spectrometry features are grouped and reduced in dimension, which greatly reduces the time and space consumed in the process of generating the mass spectrometry feature selection scheme. Specifically, the steps are as follows:

[0074] Step 1: Use the SELDI method to obtain an X×D-dimensional mass spectral feature set F of serum, which is used to characterize the abundance of proteins with given mass values in serum, where X represents the number of samples and D represents the number of mass spectral features corresponding to each sample. Table 1 shows a mass spectral feature set containing 10 mass spectral features and 3 samples. The first 10 columns represent the abundance of different proteins with given mass values, and the 11th column represents the protein abnormality detection category corresponding to the sample, where 1 represents abnormality and 0 represents normality.

[0075] Table 1 Mass spectrometry feature set

[0076] Feature Index 1 2 3 4 5 6 7 8 9 10 11 Sample 1 123 171 424 257 164 16 39 278 40 233 1 Sample 2 126 131 356 153 201 9 40 261 39 241 0 Sample 3 91 162 441 109 100 3 41 267 37 230 1

[0077] Step 2: Set the population size to N, and define the population set as P = {p1, p2, ···, p i ,···,p N}, p i represents the i-th individual in the population, and each individual represents a mass spectral feature selection scheme consisting of selecting different mass spectral features from D mass spectral features, and p i ={p i1 ,p i2 ,···,p ij ,···,p iD}, p ij represents the i-th individual p i Whether the jth mass spectrum feature is selected into the mass spectrum feature selection scheme, if p ij =1, it means selected, if p ij =0, it means not selected; j∈[1,D]; the search space where the individual in the population P is located is defined as the original space, and the dimension of the original space is D;

[0078] Step 2.1, calculate the score of each mass spectrometric feature as the importance of protein abnormality detection;

[0079] Step 2.1.1. Generate a set of D mass spectrum feature selection schemes Q = {q1,q2,···,q k ,···,q D},q k represents the kth individual in the set Q, and q k ={q k1 ,q k2 ,···,qkj ,···,q kD},q kj represents the kth individual q k Whether the jth mass spectrum feature of is selected into the mass spectrum feature selection scheme, and q k Only the kth mass spectrum feature is selected, that is, q kk =1, the rest q kj =0, k≠j;

[0080] Step 2.1.2: Calculate the kth individual q using formula (1) k The classification error rate indicator error(q k ):

[0081] error(q k )=EN(q k ) / X (1)

[0082] In formula (1), EN(q k ) means using the kth individual q k The number of samples misclassified when the mass spectral feature selection scheme represented by k The number of mass spectral features selected in the mass spectral feature selection scheme represented by k );

[0083] Step 2.1.3, use the error rate of each individual in the set Q and the number of selected mass spectrum features to perform non-dominated sorting on all individuals, and take the frontier number where the individual is located as the score of the individual, recorded as Score = {score1, score2, ···, score k ,···,score D}, where score k represents the kth individual q k The score of , i.e., the importance of the k-th mass spectrum feature;

[0084] Step 2.2, initialization;

[0085] Step 2.2.1. Define and initialize the number of iterations t = 1, and the probability of the tth generation generating offspring in the original space is ρ t , the number of groups in the tth generation is K t , the maximum number of iterations is T; initialize the external archive A of the mass spectrometry feature selection scheme t Empty set; external archive A t The default value is N A ;

[0086] Step 2.2.2, initialize the t-th generation population in, represents the t-th generation population Pt The i-th individual of , that is, the i-th mass spectrum feature selection scheme, and initialize each individual with a full 0 vector of length D;

[0087] Step 2.2.3, initialize i=1;

[0088] Step 2.2.4, generate the i-th random number rand i ;

[0089] Step 2.2.5, from the i-th individual Randomly select two mass spectrum features from the equation, and record the indexes of the two selected mass spectrum features as m and n respectively. represents the i-th individual Whether the mth mass spectrum feature is selected, represents the i-th individual Whether the nth mass spectrum feature is selected; if score m <score n , then let Indicates that the mass spectrometry feature selection scheme individual The mth mass spectrum feature is selected, otherwise, let Indicates that the feature selection scheme individual The nth mass spectrum feature is selected;

[0090] Step 2.2.6, repeat rand i ×D times step 2.2.5, thus obtaining the updated i-th individual and assigning it to Step 2.2.6, after assigning i+1 to i, determine whether i>N is true. If so, it means that the updated t-th generation population is obtained and assigned to P t As the initial population; otherwise, return to step 2.2.4; As shown in Table 2, in this embodiment, the population size is set to 5, and the value of each dimension in the individual indicates whether the mass spectrum feature is selected into the mass spectrum feature selection scheme. For example, if individual 1 is {1, 0, 1, 1, 0, 1, 0, 0, 1, 0}, the mass spectrum feature selection scheme represented is {1, 3, 4, 6, 9}, that is, the 1st, 3rd, 4th, 6th, and 9th mass spectrum features are selected;

[0091] Table 2 Mass spectrometry feature selection scheme population

[0092] Feature Index 1 2 3 4 5 6 7 8 9 10 Individual 1 1 0 1 1 0 1 0 0 1 0 Individual 2 1 1 0 1 0 1 0 0 0 0 Individual 3 0 1 1 1 0 0 0 1 0 1 Individual 4 0 0 1 0 0 1 0 0 0 0 Individual 5 1 0 1 0 0 1 0 0 0 1

[0093] Step 3: Update external archive A t and quickly cluster and group mass spectral features;

[0094] Step 3.1, the tth generation population P t and external archive At Merge and perform non-dominated sorting on the merged t-th generation set, assigning the first sorted subset to the external archive A t ;

[0095] If the external archive A after assignment t The number of mass spectral feature selection schemes in exceeds N A , then calculate A t Crowding distance for mass spectral feature selection, removing A t The mass spectrum feature scheme with the smallest crowding distance until A t The number of mass spectrometry feature schemes does not exceed N A So far, the updated external archive is obtained and assigned to A t ;

[0096] Step 3.2: Based on external archive A t and the number of groups K in the tth generation t Clustering the mass spectrum features in the original space;

[0097] Step 3.2.1, from A t All mass spectrometry feature selection schemes are obtained as a set and Indicates A t The situation vector of the jth mass spectrum feature being selected in different individuals, where |A t | indicates A t The number of mass spectral feature selection schemes, Indicates that the jth mass spectrum feature is in A t Is the ath individual in the selected? If It means not selected; a∈[1,|A t ];

[0098] Using formula (2), we can obtain the situation vector of the jth mass spectrum feature being selected in different individuals: Sparsity Thus we get the sparsity of the tth generation t ;

[0099]

[0100] In formula (2), Representation vector The number of elements in is 1; Figure 2 The example shown is a cluster grouping of 10 individuals. The value of each dimension of the individual indicates whether the mass spectrum feature is selected into the mass spectrum feature selection scheme. The sparsity calculation of the fourth mass spectrum feature selection vector is:

[0101] Step 3.2.2: The mass spectrum feature with the sparsity closest to 0.5 among all the mass spectrum features selected is used as the reference mass spectrum feature, and the reference mass spectrum feature index is denoted as c, as follows: Figure 2 As shown in the figure, the sixth mass spectrum feature is selected as the reference mass spectrum feature, and the similarity between the bth mass spectrum feature and the reference mass spectrum feature is calculated using formula (3): Thus, all mass spectrometry features are sorted according to similarity to obtain the similarity ranking of the tth generation t ;

[0102]

[0103] In formula (3), Indicates A t The number of feature selection schemes in which the bth mass spectral feature is not selected and the reference mass spectral feature is selected; Indicates A t The number of feature selection schemes in which the bth mass spectral feature is selected and the reference mass spectral feature is not selected; Indicates A t The number of feature selection schemes in which the bth mass spectrum feature is selected and the reference mass spectrum feature is selected;

[0104] Step 3.2.4, Sparsity t The mass spectrum feature selection vectors with a median value of 1 are placed in the same group The mass spectrum feature selection vector with a value of 0 is placed in From rank t Remove and After the corresponding mass spectrum features are obtained, a new similarity ranking is obtained and assigned to rank t ;

[0105] rank t The mass spectral features in K t -2 groups, thus obtaining the index set of grouped mass spectral features, recorded as in, Represents the index set of mass spectrum features of group g, g∈[3,K t ],in Respectively The indices of the first and last mass spectral features in , Indicates the size of the group; |rank t | indicates similarity ranking t Length; such as Figure 2As shown, the first and second mass spectrum features are grouped as one, the ninth and tenth mass spectrum features are grouped as one, and the remaining third to eighth mass spectrum features are evenly divided into K t -2 groups;

[0106] Step 4: After performing the mating pool selection, the progeny mass spectrum feature selection scheme is generated in both the grouped reduced space and the original space;

[0107] Step 4.1: For the t-th generation population P t Perform non-dominated sorting and crowding distance calculation, and then use the binary tournament method to select the population P according to the ranking and crowding distance. t Select 2N mass spectrum feature selection schemes and use them as the t-th generation parent set P′ t ,and in, is the t-th generation parent set P′ t The sth mass spectrum feature selection scheme in ; s∈[1,2N];

[0108] Step 4.2, generate a random vector rand={r1,r2,···,r s ,···,r 2N}, where r s Represents the sth random number;

[0109] Define and initialize the partition vector index t is a vector of all zeros of length 2N; and in, Represents the t-th generation parent set P′ t The sth mass spectrometry feature selection scheme It is the parent individual that generates offspring in the original space or the parent individual that generates offspring in the reduced space;

[0110] If r s <ρ t , then let Otherwise, let Thus, the updated partition vector is obtained and assigned to index t ,in, Represents the t-th generation parent set P′ t The sth mass spectrometry feature selection scheme is the parent individual that generates offspring in the original space, Represents a mass spectrometry feature selection scheme is the parent individual that generates offspring in the reduced space; as shown in Table 3, the index with a length of 10 tA vector, where the value of each dimension indicates whether the individual is selected as the parent individual for generating offspring in the original space. A partition vector of {1,0,0,1,0,0,1,1,1,0} indicates that the 1st, 4th, 7th, 8th, and 9th individuals are selected as the parent individuals for generating offspring in the original space, and the remaining individuals are selected as the parent individuals for generating offspring in the reduced space.

[0111] Table 3 Partition vector

[0112] Individual Index 1 2 3 4 5 6 7 8 9 10 Divide vector 1 0 0 1 0 0 1 1 1 0

[0113] Step 4.3, according to the partition vector index t The t-th generation parent set P′ t The subset of parent individuals that generate offspring in the original space is recorded as PB t ,and in, Represents the parent individual subset PB that generates offspring in the original space t The e-th mass spectrum feature selection scheme in , and Represents the e-th mass spectrum feature selection scheme Whether the jth mass spectrum feature in |PB is selected; t |The number of parents that generate children for the original space, e∈[1,|PB t ];

[0114] According to the partition vector index t The parent set P′ t The parent individual subset that generates offspring in the reduced space is recorded as PB′ t ,and in, Represents the parent individual subset PB′ that generates offspring in the reduced space t The fth mass spectrum feature selection scheme in , and Indicates the fth mass spectrum feature selection scheme Whether the jth mass spectrum feature is selected; |PB′ t |To reduce the space to generate the number of parents of children, f∈[1,|PB′ t ];

[0115] Step 4.4: Group the index set based on the grouped mass spectrum features t PB′ t Individuals are mapped to K t dimensional reduced space and group them into t The mass spectrum characteristics corresponding to each group in the is regarded as a dimension, thereby mapping the D-dimensional individuals in the original space to the reduced space into K t dimensional individuals;

[0116] Use formula (4) to calculate the fth individual in the original space The g-th group The probability that the g-th position in the reduced space is 1 for the corresponding mass spectrum feature selection

[0117]

[0118] In formula (4), Indicates the g-th group of the t-th generation Number of indexes in the

[0119] The generated dimension is |PB′ t |×K t The random matrix rand′, if Then let the parent set PB in the reduction space *t The fth individual The g-th bit of is set to 1, otherwise, the g-th bit is set to 0, thus obtaining the parent set PB in which the original space is mapped to the reduced space. *t ; where rand′ fg represents the gth random number of the fth random vector in rand′;

[0120] Step 4.5: Based on the parent set PB in the original space t Generate the offspring set OB in the original space t ;

[0121] Step 4.5.1. Initialize the offspring set OB t is an empty set;

[0122] Step 4.5.2: From the parent individual subset PB of the original space t Randomly select two individuals as parents u t and parent v t ,and Indicates whether the jth mass spectrum feature is selected, Indicates opt-in, Indicates not to be selected; Indicates whether the jth mass spectrum feature is selected;

[0123] Initialize the child individual o to the parent u;

[0124] Step 4.5.3. Generate a random number rand″1. If rand1″<0.5, randomly select two mass spectral features from the mass spectral features selected in parent u and not selected in parent v, compare the scores of the two mass spectral features, and set the mass spectral feature with the larger score corresponding to the mass spectral feature in offspring individual o as not selected. Otherwise, randomly select two mass spectral features from the mass spectral features selected in parent u and selected in parent v, compare the scores of the two mass spectral features, and set the mass spectral feature with the smaller score corresponding to the mass spectral feature in offspring individual o as selected.

[0125] Step 4.5.4, generate a random number rand″2. If rand″2 is less than 0.5, randomly select two mass spectral features from the mass spectral features selected by the offspring o, and compare the scores of the two mass spectral features. The mass spectral feature with the larger score in the offspring individual o is set as not selected; otherwise, randomly select two mass spectral features from the mass spectral features not selected by the offspring individual o, compare the scores of the two mass spectral features, and set the mass spectral feature with the smaller score in the offspring individual o as selected; thereby adding the offspring individual o to the offspring set OB t middle;

[0126] Step 4.5.5: Set PB from the parent t After deleting the parent u and parent v, if the parent set PB t If it is not empty, return to step 4.5.2 to execute, otherwise, it means that the t-th generation offspring set OB in the original space is obtained t ;

[0127] Step 4.6: For the parent set PB in the reduced space *t Use single-point crossover and bitwise mutation to obtain the offspring set in the reduced space in, Represents the offspring set OB *t The hth individual in Represents the t-th generation offspring set OB *t The hth individual In the gth position of the reduced space; |OB *t | represents the offspring set OB *t the number of individuals in

[0128] Step 4.7, the offspring individuals The offspring mapped into the original space is denoted by Represents a feature selection scheme Whether the jth mass spectrum feature in the original space is selected; if the jth mass spectrum feature in the original space Belongs to the gth group Then the offspring individuals in the space will be reduced The gth position Assign to Thus, the set of descendants of the reduced space mapped to the original space is obtained Denotes the offspring set OB′ t The hth individual in |OB′ t | represents the offspring set OB′ t The number of individuals in OB *t |=|OB′ t |;

[0129] Step 4.8, the original space offspring set OB t and the sub-set OB′ of the reduced space mapped to the original space t The union of the t-th generation offspring population O t , let the t-th generation offspring population O t The scale is N;

[0130] Step 5: Use the t-th generation offspring population O t Adaptive adjustment of the number of groups K t and probability ρ t , the t-th generation offspring population O t and the tth generation population P t Environmental selection is performed after the merger;

[0131] Step 5.1: Calculate the number of mass spectrum feature groups K of the t+1 generation using equations (5) and (6) respectively. t+1 and the probability ρ of the t+1th generation t+1 :

[0132]

[0133]

[0134] In formula (5), s 1,t represents the t-th generation population P t The number of children generated in the reduced space, ns 1,t represents the t-th generation population P t The number of non-dominated solutions in the offspring generated in the reduced space, s 1,t-1 represents the t-1 generation population P t-1 The number of children generated in the reduced space, ns 1,t-1 represents the t-1 generation population P t-1 The number of non-dominated solutions in the offspring generated in the reduced space; when t = 1, the initial population is assigned to the t-1 generation population P t-1 ;

[0135] In formula (6), s 2,t represents the t-th generation population P tThe number of children generated in the original space, ns 2,t represents the t-th generation population P t The number of non-dominated solutions in the offspring generated in the original space;

[0136] Step 5.2: Merge the t-th generation population P t and the offspring population O t , and perform non-dominated sorting and crowding distance calculation on the merged 2N individuals after deduplication, so as to select the first N individuals after sorting as the t+1 generation population and assign them to P t ;

[0137] Step 5.3: After assigning t+1 to t, determine whether t>T is true. If so, output the optimal feature selection scheme population P. t , and the optimal feature selection scheme population P t The mass spectrometry features selected in the step are used as the basis for detecting whether the protein in the serum is abnormal; otherwise, the process returns to step three and executes the sequence.

Claims

1. A protein anomaly detection method based on ultra-large-scale evolutionary algorithm and hardware acceleration, characterized by Proceed as follows: Step 1: Use the SELDI method to obtain an X×D-dimensional mass spectral feature set F of serum, which is used to characterize the abundance of proteins with given mass values in serum, where X represents the number of samples and D represents the number of mass spectral features corresponding to each sample; Step 2: Set the population size to N and define the population set as P = {p1, p2, ..., p i ,…,p N }, p i represents the i-th individual in the population, and each individual represents a mass spectral feature selection scheme consisting of selecting different mass spectral features from D mass spectral features, and p i ={p i1 ,p i2 ,…,p ij ,…,p iD }, p ij represents the i-th individual p i Whether the jth mass spectrum feature is selected into the mass spectrum feature selection scheme, if p ij =1, it means selected, if p ij =0, it means not selected; j∈[1,D]; the search space where the individuals in the population P are located is defined as the original space, and the dimension of the original space is D; Step 2.1, calculate the score of each mass spectrometric feature as the importance of protein abnormality detection; Step 2.1.

1. Generate a set of D mass spectrum feature selection schemes Q = {q1,q2,…,q k ,…,q D },q k represents the kth individual in the set Q, and q k ={q k1 ,q k2 ,…,q kj ,…,q kD },q kj represents the kth individual q k Whether the jth mass spectrum feature of is selected into the mass spectrum feature selection scheme, and q k Only the kth mass spectrum feature is selected, that is, q kk =1, the rest q kj =0, k≠j; Step 2.1.2: Calculate the kth individual q using formula (1) k The classification error rate indicator error(q k ): error(q k )=EN(q k ) / X (1) In formula (1), EN(q k ) means using the kth individual q k The number of samples misclassified when the mass spectral feature selection scheme represented by k The number of mass spectral features selected in the mass spectral feature selection scheme represented by k ); Step 2.1.3, use the error rate of each individual in the set Q and the number of selected mass spectrum features to perform non-dominated sorting on all individuals, and take the frontier number where the individual is located as the score of the individual, recorded as Score = {score1, score2, ..., score k ,…,score D }, where score k represents the kth individual q k The score of , i.e., the importance of the k-th mass spectrum feature; Step 2.2, initialization; Step 2.2.

1. Define and initialize the number of iterations t = 1, and the probability of the tth generation generating offspring in the original space is ρ t , the number of groups in the tth generation is K t , the maximum number of iterations is T; initialize the external archive A of the mass spectrometry feature selection scheme t Empty set; external archive A t The default value is N A ; Step 2.2.2, initialize the t-th generation population in, represents the t-th generation population P t The i-th individual of , that is, the i-th mass spectrum feature selection scheme, and initialize each individual with a full 0 vector of length D; Step 2.2.3, initialize i=1; Step 2.2.4, generate the i-th random number rand i ; Step 2.2.5, from the i-th individual Randomly select two mass spectrum features from the equation, and record the indexes of the two selected mass spectrum features as m and n respectively. represents the i-th individual Whether the mth mass spectrum feature is selected, represents the i-th individual Whether the nth mass spectrum feature is selected; if score m <score n , then let Indicates that the mass spectrometry feature selection scheme individual The mth mass spectrum feature is selected, otherwise, let Indicates that the feature selection scheme individual The nth mass spectrum feature is selected; Step 2.2.6, repeat rand i ×D times step 2.2.5, thus obtaining the updated i-th individual and assigning it to Step 2.2.6, after assigning i+1 to i, determine whether i>N is true. If so, it means that the updated t-th generation population is obtained and assigned to P t Then use it as the initial population; otherwise, return to step 2.2.4; Step 3: Update external archive A t and quickly cluster and group mass spectral features; Step 3.1, the tth generation population P t and external archive A t Merge and perform non-dominated sorting on the merged t-th generation set, assigning the first sorted subset to the external archive A t ; If the external archive A after assignment t The number of mass spectral feature selection schemes in exceeds N A , then calculate A t Crowding distance for mass spectral feature selection, removing A t The mass spectrum feature scheme with the smallest crowding distance until A t The number of mass spectrometry feature schemes does not exceed N A So far, the updated external archive is obtained and assigned to A t ; Step 3.2: Based on external archive A t and the number of groups K in the tth generation t Clustering the mass spectrum features in the original space; Step 3.2.1, from A t All mass spectrometry feature selection schemes are obtained as a set and Indicates A t The situation vector of the jth mass spectrum feature being selected in different individuals, where |A t | indicates A t The number of mass spectral feature selection schemes, Indicates that the jth mass spectrum feature is in A t Is the ath individual in the selected? If It means not selected; a∈[1,|A t |]; Using formula (2), we can obtain the situation vector of the jth mass spectrum feature being selected in different individuals: Sparsity Thus we get the sparsity of the tth generation t ; In formula (2), Representation vector The number of elements that are 1; Step 3.2.2: The mass spectrum feature with the sparsity closest to 0.5 among all the mass spectrum features selected is used as the reference mass spectrum feature. The reference mass spectrum feature index is denoted as c. The similarity between the bth mass spectrum feature and the reference mass spectrum feature is calculated using formula (3). Thus, all mass spectrometry features are sorted according to similarity to obtain the similarity ranking of the tth generation t ; In formula (3), Indicates A t The number of feature selection schemes in which the b-th mass spectral feature is not selected and the reference mass spectral feature is selected; Indicates A t The number of feature selection schemes in which the b-th mass spectral feature is selected and the reference mass spectral feature is not selected; Indicates A t The number of feature selection schemes in which the b-th mass spectrum feature is selected and the reference mass spectrum feature is selected; Step 3.2.4, Sparsity t The mass spectrum feature selection vectors with a median value of 1 are placed in the same group The mass spectrum feature selection vector with a value of 0 is placed in From rank t Remove and After the corresponding mass spectrum features are obtained, a new similarity ranking is obtained and assigned to rank t ; rank t The mass spectral features in K t -2 groups, thus obtaining the index set of grouped mass spectral features, recorded as in, Represents the index set of mass spectrum features of group g, g∈[3,K t ],in Respectively The index of the first and last mass spectral features in t =[|rank t | / (K t -2)] indicates the size of the group; |rank t | indicates similarity ranking t length; Step 4: After performing the mating pool selection, the progeny mass spectrum feature selection scheme is generated in both the grouped reduced space and the original space; Step 4.1: For the t-th generation population P t Perform non-dominated sorting and crowding distance calculation, and then use the binary tournament method to select the population P according to the ranking and crowding distance. t Select 2N mass spectrum feature selection schemes and use them as the t-th generation parent set P′ t ,and in, is the t-th generation parent set P′ t The sth mass spectrum feature selection scheme in ; s∈[1,2N]; Step 4.2, generate a random vector rand = {r1, r2, ..., r s ,…,r 2N }, where r s Represents the sth random number; Define and initialize the partition vector index t is a vector of all zeros of length 2N; and in, Represents the t-th generation parent set P′ t The sth mass spectrometry feature selection scheme It is the parent individual that generates offspring in the original space or the parent individual that generates offspring in the reduced space; If r s <ρ t , then let Otherwise, let Thus, the updated partition vector is obtained and assigned to index t ,in, Represents the t-th generation parent set P′ t The sth mass spectrometry feature selection scheme is the parent individual that generates offspring in the original space, Represents a mass spectrometry feature selection scheme The parent individuals that generate offspring in the reduced space; Step 4.3, according to the partition vector index t The t-th generation parent set P′ t The subset of parent individuals that generate offspring in the original space is recorded as PB t ,and in, Represents the parent individual subset PB that generates offspring in the original space t The e-th mass spectrum feature selection scheme in , and Represents the e-th mass spectrum feature selection scheme Whether the jth mass spectrum feature in |PB is selected; t |The number of parents that generate children for the original space, e∈[1,|PB t |]; According to the partition vector index t The parent set P′ t The parent individual subset that generates offspring in the reduced space is recorded as PB′ t ,and in, Represents the parent individual subset PB′ that generates offspring in the reduced space t The fth mass spectrum feature selection scheme in , and Indicates the fth mass spectrum feature selection scheme Whether the jth mass spectrum feature is selected; |PB′ t |To reduce the space to generate the number of parents of children, f∈[1,|PB′ t |]; Step 4.4: Group the index set based on the grouped mass spectrum features t PB′ t Individuals are mapped to K t dimensional reduced space and group them into t The mass spectrum characteristics corresponding to each group in the is regarded as a dimension, thereby mapping the D-dimensional individuals in the original space to the reduced space into K t dimensional individuals; Use formula (4) to calculate the fth individual in the original space The g-th group The probability that the g-th position in the reduced space is 1 for the corresponding mass spectrum feature selection In formula (4), Indicates the g-th group of the t-th generation Number of indexes in the The generated dimension is |PB′ t |×K t The random matrix rand′, if Then let the parent set PB in the reduction space *t The fth individual The g-th bit of is set to 1, otherwise, the g-th bit is set to 0, thus obtaining the parent set PB in which the original space is mapped to the reduced space. *t ; where rand′ fg represents the gth random number of the fth random vector in rand′; Step 4.5: Based on the parent set PB in the original space t Generate the offspring set OB in the original space t ; Step 4.5.

1. Initialize the offspring set OB t is an empty set; Step 4.5.2: From the parent individual subset PB of the original space t Randomly select two individuals as parents u t and parent v t ,and Indicates whether the jth mass spectrum feature is selected, Indicates opt-in, Indicates not to be selected; Indicates whether the jth mass spectrum feature is selected; Initialize the child individual o to the parent u; Step 4.5.

3. Generate a random number rand″1. If rand″1 is less than 0.5, randomly select two mass spectral features from the mass spectral features selected in parent u and not selected in parent v, compare the scores of the two mass spectral features, and set the mass spectral feature with the larger score corresponding to the mass spectral feature in offspring individual o as not selected. Otherwise, randomly select two mass spectral features from the mass spectral features selected in parent u and selected in parent v, compare the scores of the two mass spectral features, and set the mass spectral feature with the smaller score corresponding to the mass spectral feature in offspring individual o as selected. Step 4.5.4, generate a random number rand″2. If rand″2 is less than 0.5, randomly select two mass spectral features from the mass spectral features selected by the offspring o, and compare the scores of the two mass spectral features. The mass spectral feature with the larger score in the offspring individual o is set as not selected; otherwise, randomly select two mass spectral features from the mass spectral features not selected by the offspring individual o, compare the scores of the two mass spectral features, and set the mass spectral feature with the smaller score in the offspring individual o as selected; thereby adding the offspring individual o to the offspring set OB t middle; Step 4.5.5: Set PB from the parent t After deleting the parent u and parent v, if the parent set PB t If it is not empty, return to step 4.5.2 to execute, otherwise, it means that the t-th generation offspring set OB in the original space is obtained t ; Step 4.6: For the parent set PB in the reduced space *t Use single-point crossover and bitwise mutation to obtain the offspring set in the reduced space in, Represents the offspring set OB *t The hth individual in Represents the t-th generation offspring set OB *t The hth individual In the gth position of the reduced space; |OB *t | represents the offspring set OB *t the number of individuals in Step 4.7, the offspring individuals The offspring mapped into the original space is denoted by Represents a feature selection scheme Whether the jth mass spectrum feature in the original space is selected; if the jth mass spectrum feature in the original space Belongs to the gth group Then the offspring individuals in the space will be reduced The gth position Assign to Thus, the set of descendants of the reduced space mapped to the original space is obtained Denotes the offspring set OB′ t The hth individual in |OB′ t | represents the offspring set OB′ t The number of individuals in OB *t |=|OB′ t |; Step 4.8, the original space offspring set OB t and the sub-set OB′ of the reduced space mapped to the original space t The union of the t-th generation offspring population O t , let the t-th generation offspring population O t The scale is N; Step 5: Use the t-th generation offspring population O t Adaptive adjustment of the number of groups K t and probability ρ t , the t-th generation offspring population O t and the tth generation population P t Environmental selection is performed after the merger; Step 5.1: Calculate the number of mass spectrum feature groups K of the t+1 generation using equations (5) and (6) respectively. t+1 and the probability ρ of the t+1th generation t+1 : In formula (5), s 1,t represents the t-th generation population P t The number of children generated in the reduced space, ns 1,t represents the t-th generation population P t The number of non-dominated solutions in the offspring generated in the reduced space, s 1,t-1 represents the t-1 generation population P t-1 The number of children generated in the reduced space, ns 1,t-1 represents the t-1 generation population P t-1 The number of non-dominated solutions in the offspring generated in the reduced space; when t = 1, the initial population is assigned to the t-1 generation population P t-1 ; In formula (6), s 2,t represents the t-th generation population P t The number of children generated in the original space, ns 2,t represents the t-th generation population P t The number of non-dominated solutions in the offspring generated in the original space; Step 5.2: Merge the t-th generation population P t and the offspring population O t , and perform non-dominated sorting and crowding distance calculation on the merged 2N individuals after deduplication, so as to select the first N individuals after sorting as the t+1 generation population and assign them to P t ; Step 5.3: After assigning t+1 to t, determine whether t>T is true. If so, output the optimal feature selection scheme population P. t , and the optimal feature selection scheme population P t The mass spectrometry features selected in the step are used as the basis for detecting whether the protein in the serum is abnormal; otherwise, the process returns to step three and executes the sequence.

Images