Forensic physical evidence multi-person identity authentication method

Abstract

The invention relates to a forensic physical evidence multi-person identity authentication method. The method comprises the following steps: S1, optimizing a short tandem repeat map mathematical model; s2, preprocessing the short tandem repeat map mathematical model to obtain sites containing information; s3, using a mean shift algorithm to process peak points in the atlas, converging to a density local maximum point, and outputting the number of modes as the preliminary estimation of the number of contributors of the locus; s4, judging the mixing ratio of each contributor by using a gradient descent algorithm; and S5, generating an initial vector depending on the mixing ratio according to the mixing ratio, performing sliding matching by keeping the contribution amount of the contributor unchanged, finding an allele size value suitable for genome information contributed by each contributor, and generating a genotype by searching an id corresponding to the nearest allele size in an original map. By means of the design, DNA information of multiple persons can be processed, and contributor composition and proportion of all components in mixed data are analyzed.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of DNA identification, and particularly relates to a method for multi-person identity identification of forensic physical evidence. BACKGROUND

[0002] The forensic physical evidence identification of a crime scene is a major issue related to national security, public security, people's life and property safety and social stability. At present, the forensic physical evidence identification relies on the continuous improvement of DNA analysis technology. The core of DNA testing is STR (Short Tandem Repeat) analysis technology. Short Tandem Repeat (STR) is a short DNA sequence composed of 2-6 base pairs, accounting for about 3% of the human genome. The core sequence of STR has individual difference polymorphism, so STR is also called the DNA fingerprint of cells. When evaluated for identification purposes, the number of repeat units varies from person to person, with high discrimination.

[0003] For the DNA evidence in a multi-person criminal case, the analysis of the mixed STR graph is the key to the case investigation. The mixed STR graph contains the DNA information of multiple people, and the contributors and proportions of each component in the mixed data need to be correctly analyzed as one of the case judgment bases. However, due to the mixing of multi-person information, the STR graph analysis of multiple people is difficult. SUMMARY

[0004] In order to solve the problem of existing multi-person STR graph analysis difficulty, the application designs a method for multi-person identity identification of forensic physical evidence, so as to process the STR graph mathematical model by using the mean shift algorithm, and to obtain the matching genotype by using the gradient descent algorithm technology mixed ratio.

[0005] A method for multi-person identity identification of forensic physical evidence, please refer to Figure 1 , comprising the following steps: Step S1, optimizing the short tandem repeat sequence graph mathematical model; Step S2, preprocessing the short tandem repeat sequence graph mathematical model to obtain information-containing sites; Step S3, using the mean shift algorithm to process the peak points in the graph, converging to the local maximum value point of the density, and outputting the mode number as the preliminary estimate of the number of contributors of this locus; Step S4, using the gradient descent algorithm to judge the mixed ratio of each contributor; Step S5, generating an initial vector dependent on the mixed ratio, sliding matching by keeping the contribution amount of the contributor unchanged, finding the appropriate allele size value of the genomic information contributed by each contributor, and then generating the genotype by searching for the id corresponding to the nearest allele size in the original graph.

[0006] Preferably, the step S1 comprises: Step S11, since the peak height and allele size of the alleles approximately satisfy the normal distribution, i.e. ; Wherein, mass(zygote): the contribution of the contributor at this locus; size(zygote): the allele size of the short tandem repeat sequence of the contributor at this locus; std(zygote): the variation degree of the allele size of the short tandem repeat sequence; Step S12, limit the size of std to avoid unrealistic optimization results; Step S13, for a plurality of people maps, mass(zygote, marker) = μ(marker) * mass(zygote); Wherein, μ is a proportional coefficient; The height formed formula is obtained: ; Step S14, use a multi-locus comprehensive analysis method to process the multiple solutions problem of sampling ratio and contributor genotype.

[0007] Preferably, the step S2 comprises: Step S21, set the noise and effective information point demarcation standard; Step S22, sort by height as a library of extracted information points; Step S23, remove the height max point in the library one by one, respectively calculate the Gini index of different slices to obtain the Gini index change sequence, until three points; Step S24, difference the short tandem repeat sequence, find the mutation point with the largest drop, and cut off at this position to divide the information points and noise points.

[0008] Preferably, the step S3 comprises: Step S31, use the mean shift algorithm to process the peak points in the map; Step S32, set a threshold for the Gini index of the locus; Step S33, set the number of people for each locus to be the number of identified people minus one, the number of identified people, and the number of identified people plus one, respectively, and try 10 times of gradient descent optimization in three cases, and select the optimal result as the judgment of the number of people.

[0009] Preferably, in the step S4, the gradient descent algorithm specifically comprises: Step S41, set the number of zygotes required, and evaluate the fitting results using the pair-wise variance: The total pair-wise variance is used as the loss function, and the gradient can be calculated: The gradient in the opposite direction is the maximum descent direction, and the obtained descent vector is divided by the square of its modulus to obtain the expected vector of the pair-wise variance descent 1; Step S42, cluster the ratios of the fitting results to find the correct ratio; set the output content proportion vector, the distance function as the Euclidean distance, and the density as at least the number of markers divided by 3 within the radius, select the radius with the highest cluster number, and perform cluster analysis, and the mean of each cluster is the result.

[0010] Step S43, perform a second gradient descent on the initial vector of dependence for each locus, this time first perform 300 iterations of sliding size and std while keeping mass unchanged, so that the peaks in the map are fully matched, then keep size and std fixed, change height to approach the map, and thus obtain more accurate ratios. Then take the average of the ratios generated at each locus in the original result to obtain the final result.

[0011] Preferably, the specific method of step S5 is: Establish a penalty score, for all loci under a certain ratio, accumulate the absolute value of the difference between the size and the nearest size in the original map to the power of 2: Thus, the ratio and the corresponding genotype can be arranged in order of penalty score from small to large, and the penalty score is the smallest.

[0012] Preferably, it is also necessary to set the noise sites to zero.

[0013] The advantages and effects of the present application are as follows: The method for identifying the identities of multiple persons in forensic physical evidence designed by the application comprises the following steps: S1, optimizing a short tandem repeat sequence map mathematical model; S2, preprocessing the short tandem repeat sequence map mathematical model to obtain information-containing sites; S3, using a mean shift algorithm to process peak points in the map, converging to a local maximum density point, and outputting the mode number as a preliminary estimate of the number of contributors at this locus; S4, using a gradient descent algorithm to judge the mixing ratio of each contributor; S5, generating an initial vector dependent on the mixing ratio, performing sliding matching by keeping the contribution amount of the contributor unchanged, finding the appropriate allele size value of the genomic information contributed by each contributor, and then generating a genotype by searching for the id corresponding to the nearest allele size in the original map. Through the above design, the application can process the DNA information of multiple persons and analyze the composition and proportion of each component in the mixed data.

[0014] The above description is only a summary of the technical scheme of the application. In order to more clearly understand the technical means of the application, the content of the description can be implemented, and in order to make the above and other purposes, features and advantages of the application more obvious and easy to understand, the following will be described in detail with the preferred embodiments of the application and with the help of the accompanying drawings.

[0015] According to the detailed description of the specific embodiments of the application in the following text combined with the accompanying drawings, those skilled in the art will more clearly understand the above and other purposes, advantages and features of the application. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical scheme in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or the prior art description. Obviously, the drawings in the following description are some embodiments of the application, and those skilled in the art can also obtain other drawings according to these drawings without creating any creative labor. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual proportion.

[0017] Figure 1 The flow chart of the method for identifying the identities of multiple persons in forensic physical evidence designed by the application; Figure 2 The Gini index chart for analyzing the uniformity of STR map information points; Figure 3 The Gini index change sequence chart of different slices obtained by each highest point; Figure 4 The effect comparison chart before and after noise reduction; Figure 5 The threshold value change chart for identifying the number of persons with Gini index; Figure 6 Figure of the minimum value of the variance for the fit results for each locus in a run; Figure 7 Figure of the DBSCAN clustering output clusters for a sample as the radius is varied. DETAILED DESCRIPTION

[0018] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of embodiments of the present application, rather than all the embodiments of the present application. In the following description, specific details such as specific configurations and components are provided only to help a comprehensive understanding of the embodiments of the present application. Therefore, it should be apparent to those skilled in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. In addition, in order to be clear and concise, the description of known functions and structures is omitted in the embodiments.

[0019] It should be understood that the "one embodiment" or "the embodiment" mentioned throughout the specification means that the specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present application. Therefore, "one embodiment" or "the embodiment" appearing throughout the specification does not necessarily mean the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner.

[0020] In addition, reference numerals and / or letters can be repeated in different examples in the present application. Such repetition is for the purpose of simplification and clarity, and does not itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0021] The term "and / or" herein is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that A exists alone, B exists alone, and A and B exist together. The term "and" herein is a description of another association relationship of the associated objects, which means that there can be two relationships, for example, A and B can mean that A exists alone and A and B exist together. In addition, the character " / " herein generally means that the associated objects before and after are in an "or" relationship.

[0022] The term "at least one" herein is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, at least one of A and B can mean that A exists alone, A and B exist together, and B exists alone.

[0023] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; for example, as used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It must be noted that as used herein:

[0024] Embodiment 1: This embodiment mainly introduces the establishment of STR profile mathematical model, which specifically includes: I. Single STR gene type on a single locus is processed through amplification, separation, etc. In theory, it should produce a single band, that is, form a single site (size, height) on the picture. However, due to the fact that part of the DNA may be broken and connected during actual amplification, the material is not completely uniform during electrophoresis, and there are errors during reading, etc. The STR profile is "a peak centered on a certain size", which can be approximately considered as a normal distribution of height-size: (The following shows the case of zygote being homozygous, and the case of zygote being heterozygous is completely similar): Note: mass (zygote): the contribution of this contributor at this locus size (zygote): the size of the STR sequence of this contributor at this locus std (zygote): the variation degree of STR sequence size (depends on the operation accuracy of this time sampling) Therefore, for each contributor, we can fit it as a variable about (mass (zygote), size (zygote), std (zygote)) (homozygote) or (mass (zygote), size1 (zygote), std1 (zygote), size2 (zygote), std2 (zygote)) (heterozygote) in this mathematical model to approximate the actual profile.

[0025] It is noted that the size deviation of the STR sequence forms a very narrow peak in practice, so the size of std will be limited in the subsequent optimization part to avoid unrealistic optimization results.

[0026] As for the profile of multiple people, we can assume that mass (zygote, marker) = μ (marker) * mass (zygote) according to hypothesis 4, that is, the ratio of the contributions of different contributors determines the ratio of the peak height in each locus. The proportional coefficient μ depends on the characteristics of different markers and the use of different primers in the amplification process, and the simplified list is as follows: The height formation formula is obtained: II. Data analysis and processing method How to deal with the proportion coefficient μ (marker) which is easily interfered by other factors? Since the absolute contribution amount is generally positively correlated with the sampling amount in an absolute sense, the focus in practice is not the absolute contribution amount of different contributors in the sample, but their contribution rate. Therefore, we can perform relative processing on the sample data to obtain the ratio: From the formula, the ideal processing result is that the contribution rate obtained by analyzing the same contributor at different loci marker is exactly the same. However, in reality, on the one hand, there is interference from experimental errors, and on the other hand, the multiple solution problem must be considered, that is, the same profile of a certain locus (marker) may have multiple contributor combination situations. For example, the two peaks with size-height combination (100, 30) and (120, 40) may be a homozygous contributor with genotype (G100; G100) and another homozygous contributor with genotype (G120; G120) mixed at a sampling ratio of 3:4, or a heterozygous contributor with genotype (G100; G120) and another heterozygous contributor with genotype (G120; G120) mixed at a sampling ratio of 6:1. In order to deal with the multiple solution problem of sampling ratio and contributor genotype, the fitting results of multiple loci need to be analyzed comprehensively.

[0027] The specific method of multi-locus comprehensive analysis is: it is expected that many ratios from different locus fitting results can be found near the correct ratio (correct cluster) (the distance is measured after vectorization). However, since the fitting obtained ratio may occur random aggregation of non-correct ratio, and even the average value of the results near the correct ratio may exist deviation. Therefore, the average value of the high-density candidate ratio screened by clustering is back-substituted, that is, the parameter combination under this ratio is generated for fine optimization, to obtain a more accurate contribution ratio, and then lock this contribution ratio, and perform sliding parameter adjustment, to gradually fine-tune the fragment length and peak width of each zygote, so that the simulated STR peak shape matches the peak position, peak width and height of the experimental profile.

[0028] (If the ratio is correct, the final size of the STR sequence will be very close to the size point in the original profile). Note: STR loci cannot accurately process multiple maps with the same contribution ratio, because the STR identification process requires the original sample to be split to release free fragments, which will disrupt the direct connection between the original DNA fragments, and can only be indirectly calculated by the contribution amount. If the mixing ratio is indeed extremely close, only single-cell sequencing can be attempted to retain the original spatial structure.

[0029] III. Raw image quality evaluation Example 2, based on Example 1, this example mainly introduces preprocessing: obtaining information-containing sites (Gini index), first distinguishing noise from points containing effective information, to facilitate improving the fitting and identification rate and genotype effectiveness and obtaining the variation range of each height.

[0030] I. Set the noise and effective information point demarcation standard 1. Gini index principle Please refer to Figure 2 , the Gini index is a core indicator for measuring the uniformity of data distribution, calculated by the area integral under the Lorenz curve, the formula is Gini index = A / (A+B), the value range is 0~1. Among them: A represents the area between the Lorenz curve and the 45-degree uniform line (non-uniform area); B represents the area under the Lorenz curve.

[0031] When the Lorenz curve coincides with the 45-degree line (A=0), the data distribution is completely uniform (Gini index = 0); if the data has significant differences, the larger the A value, the closer the Gini index to 1.

[0032] Variable meaning: X-axis: sample cumulative proportion sorted by allele peak height (height) from low to high Y-axis: cumulative proportion corresponding to allele peak height (height) Special note: Data containing negative values will cause the Gini index to exceed 1, and the data in this project are non-negative, meeting the calculation requirements.

[0033] Gini index calculation formula (data has been sorted in the formula) 2. Noise characteristics and cause analysis Noise in the experiment is a random and directionless fluctuation, the main causes include: Biological factors: DNA fragmentation and fusion, natural variation of STR length, environmental DNA contamination; Technical factors: During the electrophoresis process, gel non-uniformity, excessive voltage or long electrophoresis time will cause DNA fragment tailing, affecting the clarity of the spectrum.

[0034] 3. Reasonableness of mathematical model selection The reasons for using the Gini coefficient are as follows: Data adaptability: Peak height is a non-negative continuous variable, which meets the calculation premise of the Gini coefficient distribution of 0 to 1; The physical meaning is clear: the difference between noise and effective signal is quantified by distribution uniformity. The more uneven the peak height distribution (the larger the Gini coefficient), the more significant the separation between the effective signal (main peak) and the noise. Threshold interpretability: Taking the maximum value of the decline in the Gini index change series as the threshold for truncation can intuitively distinguish the "effective information-dominated area" from the "noise-dominated area" (according to Hypothesis 6, the distribution means of the two are significantly different).

[0035] 2. Noise reduction problem → Effective information point extraction problem Take each locus as a unit and proceed in sequence. Take the processing of one locus as an example: 1. Sort the copy a by height as a library for extracting information points 2. Remove the height max points in copy a one by one, and calculate the Gini index for different slices to obtain the Gini index change sequence until there are less than 3 points (to prevent unstable jumps of the Gini index caused by too few points).

[0036] 3. Please refer to Figure 3 , perform differential analysis on the sequence, find the mutation point with the largest drop, and truncate at this position to divide the information site from the noise site.

[0037] Principle: The height distribution of the original atlas has "higher sites generated by sampling" and "lower sites generated by noise", so the Gini index is in a median state; when the higher peaks are gradually removed, the total height of the slice is almost entirely provided by a few peaks, and the Gini index gradually increases (if there are some low peaks that are difficult to distinguish from noise, it may decrease slightly); when all peaks are removed, only the height generated by noise remains, so the Gini index becomes a minimum value, which can be used as the basis for division.

[0038] 3. Visualization: Please refer to the comparison between the STR image after noise reduction and before noise reduction. Figure 4 .

[0039] 4. Evaluation of noise reduction effect: From the foregoing noise reduction method, due to the limitation of Gini index itself, at least 3 points are required for each locus to effectively reduce noise. If the sampling result of a certain locus is less than 3 points, we think that these points are all valid sites, which will at most cause 1 noise site not to be reduced, and under the premise of hypothesis 6, it does not interfere with the subsequent processing.

[0040] Example 3, based on examples 1~2, this embodiment mainly introduces how to judge the number of contributors I. Meanshift clustering 1. Introduction of MeanShift clustering: MeanShift (Mean Shift) is a non-parametric clustering and feature space analysis method based on kernel density estimation theory. Its core idea is to calculate the local mean of data points through iteration, so that the data points "drift" along the direction of increasing density, and finally converge to the local maximum value point of density, so as to realize clustering or target tracking and other tasks.

[0041] For data points xi in d-dimensional space, the kernel density estimation function is: (where K(·) is the kernel function, exponential function is used this time, and h is the bandwidth parameter) Given the initial point yj, MeanShift iteratively calculates the weighted mean: Where g(·) is the derivative function of the kernel function K(·). When the contour of the kernel function is convex and monotonically decreasing, the iteration sequence and the density estimation sequence must converge to the local maximum value point, and the density estimation value is monotonically increasing.

[0042] The core advantage of MeanShift is non-parametric density estimation, which can automatically detect "patterns" (i.e. local maximum of density) in feature space. Mixed samples contain multiple contributors, and the same locus may exhibit multiple peaks superimposed. A single individual's STR locus usually corresponds to 1-2 patterns (homozygous / heterozygous). In a mixed sample, if there are k contributors, the same locus may exhibit k or 2k patterns (depending on whether it is heterozygous). MeanShift converges to the center of each pattern through iteration, directly outputting the number of patterns, which can be used as a preliminary estimate of the number of contributors for this locus. Traverse all loci of the sample to estimate the number of people, and select the maximum value as the global estimate of the number of contributors.

[0043] 2. Practical application: In practice, we found that the loci with low Gini index (i.e. the separation between the sampling results and the noise is not high) can easily interfere with the identification of the total number of people, and need to set a threshold to ignore the loci with low Gini index. In addition, due to the diversity of genotypes, the number of people identified often fluctuates slightly around the true number, so we choose to set the number of people for each locus to be the identified number minus one, the identified number, and the identified number plus one respectively, and try 10 times of gradient descent optimization (see the specific method in question two), and select the optimal result as the judgment of the number of people.

[0044] Note: As the number of people increases, the accuracy of the model must increase, so we take the approach of "stopping increasing the number of people once the difference between the fitted results and the actual atlas decreases to the variation of noise (assuming noise satisfies normal distribution, take 0.95 significance upper limit)"; If this condition is not met, determine "whether increasing the number of people by 1 leads to the reduction of the pair correlation to 1 / 4 of the original" (pair correlation is also seen below, i.e. the difference between the fitted and actual atlas is reduced by half), to prevent overfitting.

[0045] II. Visualization of results Please refer to Figure 5 It can be seen that the application of 0.3 as the threshold is the most appropriate.

[0046] Result evaluation: (after further optimization) It can be seen that the identification is more accurate when the number of people is low, but when the number of people is too high, it is easy to be less. In fact, when the number of people is too large, it is difficult to identify, that is, the polymorphism of the STR locus itself is not enough to support the identification of the number of people, at this time, even if the number of people is analyzed manually, it is also impossible to get more accurate results, and other methods need to be used to determine the number of people.

[0047] Example 4, based on example 3, this example mainly introduces the judgment method of the mixing ratio of each contributor I. Gradient descent optimization 1. Introduction to gradient descent optimization Gradient Descent is a commonly used iterative optimization algorithm for minimizing objective functions, widely applied in machine learning, deep learning, and other fields. Its core idea is to calculate the gradient of the objective function at the current parameters, update the parameters in the opposite direction of the gradient (the direction of the fastest decrease in function value), and gradually approach the optimal solution. According to the amount of data used in each update, it is mainly divided into three variants: Batch Gradient Descent (BGD), Stochastic Gradient Descent (SGD), and Mini-Batch Gradient Descent (Mini-Batch GD). This scheme uses an adaptive optimization method based on the Batch Gradient Descent (BGD) framework, which integrates a parameter perturbation strategy to enhance global search capability.

[0048] The working principle of Gradient Descent: Calculate the gradient of the loss function with respect to the model parameters, then update the parameters in the opposite direction of the gradient (i.e., the steepest descent direction). In this way, each iteration will make the loss function value decrease (at least locally), gradually approaching the minimum value of the loss function.

[0049] Set the loss function: a. Initialize parameters: Choose an initial model parameter vector.

[0050] b. Calculate the gradient: Calculate the gradient of the loss function with respect to the parameters, indicating the rate of change of the loss function at the current parameter value.

[0051] c. Update parameters: Update the parameter values in the opposite direction of the gradient with a certain step size, so that the loss function gradually decreases.

[0052] d. Iteration: Repeat steps b and c until the termination condition is met (such as reaching the maximum number of iterations, loss function convergence, etc.).

[0053] 2. Gradient descent method for the model used: In the model used in this case, set the number of zygotes to the number of demand, and use the paired variance to evaluate the fitting result: (Same as only showing the case of homozygotes) Take the total paired variance as the loss function, and calculate the gradient: The opposite direction of the gradient is the maximum descent direction, and the obtained descent vector divided by the modulus square gives the expected vector of paired variance descent 1.

[0054] In actual optimization, we found that since the peak points contribute mainly to the matching variance, the optimization results tend to quickly approach the peak rather than fit the entire spectrum. Therefore, we relativized the loss function, that is, the variance of each site was divided by the square of the height of the site. However, this operation is likely to cause the opposite problem, that is, excessive preference for low-height sites. To this end, we chose to add a penalty height, dividing the variance of non-peak sites by the sum of the square of the height of the site and the square of the noise variation of the locus. The resulting loss function is: (For peak penalty height, it is 0, and for non-peak penalty height, it is the degree of noise variation) The descent vectors are very similar to the original settings and will not be repeated due to space limitations. However, the original pairing variance is still used to evaluate the reliability of the fitting results to measure the difference between the fit and the spectrum.

[0055] Different zygotes in the same fitting result are sorted by mass, and the one with the largest mass is taken as the relative ratio 1 (this setting is for the convenience of subsequent clustering). The relative ratios of other zygotes are calculated and combined to form the ratio vector of the fitting result.

[0056] 3. Specific parameter settings: Initial vector generation: Due to the biological characteristics of STR loci, homozygotes (two identical alleles) are generally less likely in the population than heterozygotes (two different alleles). Assuming an initial probability of homozygosity of 20% and heterozygosity of 80%, we randomly generate initial genotypes at a ratio of 0.8:0.2 (e.g., heterozygotes are labeled (a,b) and homozygotes are labeled (a,a)). Depending on whether the genotype is homozygous or heterozygous, we randomly select the peak parameters (size and height) of one or both alleles as the corresponding values ​​for the zygote. The initial std is set to 1 (because peaks are typically very narrow in practice). Theoretical basis: The heterozygosity of human STR loci is usually between 0.5 and 0.8 (with large differences between different loci), but using a conservative low heterozygosity ratio during initial randomization can reduce interference from extreme values, and subsequent gradient descent optimization is used to approximate the true distribution.

[0057] Please refer to Figure 6, the step size is 0.05 of the current loss function value, the iteration number is 1000 times, if the loss function does not decrease after iteration, the step size is halved, and a small random bias is added to avoid falling into a local optimal solution. Since the number of people involved in this problem is small, the number of attempts for each locus is 2^(global number of people + 2) times to fully explore multiple solutions (if the fitting result collides with the previous fitting result after 2^global number of people times, stop trying). If the number of people is too large, you can use the relative ratio vector of the fitting results of the gene locus to generate an initial vector heuristically after fully trying multiple gene loci, which reduces the computational complexity (the function has been overloaded in the related code to facilitate such expansion if necessary).

[0058] After optimization by this method, most fitting results can easily achieve a pair difference of less than 10^-10: II. DBSCAN clustering analysis: As mentioned above, the ratio of the fitting result needs to be clustered to find the correct ratio. (The fitting results used in the following steps have been filtered out a small number of unreasonable results with a pair difference greater than the square of the noise variation degree multiplied by the number of sites) 1. Introduction to DBSCAN clustering: DBSCAN clustering is a density-based algorithm for discovering clusters in large spatial databases with noise. This clustering method requires less domain knowledge to determine input parameters, and can discover clusters of arbitrary shape with good efficiency on large data. At present, the relationship between environmental variables and the impact of biological molecule detection process is not fully clear. Based on the advantages of DBSCAN, such as few parameters, discovery of arbitrary shape clusters, and high efficiency in processing large-scale data, it can better cope with the complexity of environmental uncertainty in STR profile determination and the randomness brought by gradient descent optimization algorithm. (This project has a large amount of data and various cluster shapes) The core idea of DBSCAN clustering algorithm is that for each point in the cluster, its neighborhood with a given radius must contain at least a minimum number of points, i.e. the density in the neighborhood must exceed a certain threshold. The shape of the neighborhood is determined by the distance function dist(p, q) of two points p and q.

[0059] 2. As described above, the output content ratio vector (parallel to the position of multiple process data), the distance function is set to Euclidean distance, the density is set to at least marker number divided by 3 (rounded down) within the radius, the radius is selected to have the highest cluster number, and the cluster analysis is performed. The mean of each cluster is the result.

[0060] 3. Explanation of parameter selection: the correlation between density and genotype is because the total amount of fitting results is closely related to the total number of markers, but the density cannot be set too high, otherwise it is difficult to find the correct ratio from itself for markers with low variability, such as markers with only 1 or 2 effective sites on the map, resulting in fewer points near the correct ratio, and setting too high a density may result in the correct site being ignored, please refer to Figure 7 In practice, it is found that the number of markers divided by 3 is more appropriate as a setting; selecting as many clusters as possible is to avoid missing the correct ratio and to make the results more accurate, since the maximum distance between vectors is 1*(number of requirements-1)^0.5, and generally when the radius reaches 0.1*(number of requirements-1)^0.5, all data has been clustered into a cluster, therefore, between 0 and 0.1*(number of requirements-1)^0.5, 100 equal divisions are made, and the fitting result with the largest number of clusters is found.

[0061] III. Further data processing: For the multiple results obtained in the previous step, each result is processed: generate an initial vector for each locus (by calculating the total mass of this locus to deduce the mass of each zygote, and select the size of the dependent peak height relationship, that is, give high mass to high zygote) for the second gradient descent, this time first in the state of keeping mass unchanged, 300 iterations of sliding size and std are performed, so that the peaks in the map are fully matched, then keep size and std fixed, change height to approach the map, and thus obtain more accurate ratios. Take the average of the ratios generated at each locus in the original result to obtain the final result.

[0062] But the output result at this time still has multiple possibilities, which need to be judged by the next step whether the genotype part is reasonable, and sorted according to the degree of reasonableness.

[0063] Accuracy evaluation: for most samples, the identification result is relatively accurate, and the MSE is small, but when the number of people is large (up to 5), it is easy to make mistakes, and the second or even third result needs to be considered.

[0064] Example 5, based on example 4, this example mainly introduces the judgment of the genotype of each contributor Genotype reading: a series of relatively accurate contribution ratios have been obtained above, so a series of initial vectors dependent on each ratio can be generated (the method is the same as above), and the appropriate size value of each zygote is found by sliding matching while keeping the mass unchanged, and the genotype is generated by looking up the nearest size corresponding id in the original map.

[0065] Here we can evaluate which ratio is most reasonable with the genotype: we expect the correct ratio to generate a fitting result size that is extremely close to the original map size; in order to highlight the influence of extreme values, we establish a penalty score, which is the sum of the absolute value of the power of 2 of the difference between the size of all loci under a certain ratio and the nearest size of the original map: Thus, the ratio and the corresponding genotype can be arranged in ascending order of the penalty score, and the one with the smallest penalty score is the most reliable result.

[0066] Note: Although multi-locus mutual reference can effectively deal with the multi-solution phenomenon, it still cannot completely rule out multiple possibilities in complex maps, especially the sample STR variability given this time is not sufficient to support a completely determined result, so multiple results ranked at the top in the output should be considered comprehensively. It can be seen that for the preferred (smaller penalty score) results, the genotype recognition accuracy is extremely accurate for high contribution proportion, but is easily disturbed at low contribution proportion, especially the mixing of high proportion contributors.

[0067] Example 6, based on example 5, the main contribution of this embodiment is that when the contribution ratio is greatly different, the genotype of the individual with low contribution ratio is easily disturbed and difficult to identify if the actual application needs to focus on the individual with low contribution rate. Further analysis is needed to compare each locus; in addition, for the case of a single individual contributing to the main map, the program will tend to divide into two or more individuals with close ratios, and in actual application, rough estimation can be made through other information to obtain more accurate results.

[0068] The above only describes the preferred embodiments of the present application and does not limit the protection scope of the present application. For those skilled in the art, the present application can have various changes and variations. Any changes, modifications, replacements, integrations and parameter changes to these embodiments within the spirit and principles of the present application, which can realize the same function without departing from the principles and spirit of the present application, fall within the protection scope of the present application.

Claims

1. A method for multiple-person identification using forensic evidence, characterized in that: The following steps are involved: Step S1, optimizing the mathematical model of short tandem repeat sequence patterns; Step S2, preprocessing the mathematical model of the short tandem repeat sequence map to obtain sites containing information; Step S3: Use the mean shift algorithm to process the peak points in the spectrum, converge to the local maximum point of density, and output the number of patterns as a preliminary estimate of the number of contributors to this locus; Step S4: Using the gradient descent algorithm, determine the mixing ratio of each contributor; Step S5: Generate an initial vector based on the mixing ratio, perform sliding matching by keeping the contribution of the contributor unchanged, find the appropriate allele size value of the genomic information contributed by each contributor, and then generate the genotype by searching the ID corresponding to the nearest allele size in the original map.

2. A method for multiple-person identification based on forensic evidence according to claim 1, characterized in that: The step S1 comprises: Step S11: Since the peak height and allele size of the allele approximately satisfy the normal distribution, that is, ； Among them, mass (zygote): the contribution of this contributor at this locus; size (zygote): the allele size of the short tandem repeat sequence of this contributor at this locus; std (zygote): the degree of variability in the allele size of short tandem repeat sequences; Step S12: limit the size of std to avoid unrealistic optimization results; Step S13: For a multi-person map, mass(zygote, marker)=μ(marker)*mass(zygote); Wherein, μ is the proportional coefficient; The height is formed by the formula: ； Step S14: using a multi-locus comprehensive analysis method to handle the multiple solutions of sampling ratio and contributor genotype.

3. The method for multiple-person identification based on forensic evidence according to claim 1, characterized in that: The step S2 comprises: Step S21, setting the boundary standard between noise and valid information points; Step S22, sorting by height as a library for extracting information points; Step S23: remove the height max points in the library one by one, and calculate the Gini index for different slices to obtain a Gini index change sequence until there are less than 3 points; Step S24: Differentiate the short tandem repeat sequence to find the mutation point with the largest drop, and truncate at this position to divide the informative site from the noise site.

4. The method for multiple-person identification based on forensic evidence according to claim 1, wherein: The step S3 comprises: Step S31: using a mean shift algorithm to process peak points in the spectrum; Step S32: setting a threshold for the Gini index of the locus; Step S33: For each locus, the number of people is set to the number of people identified minus one, the number of people identified, and the number of people identified plus one. Ten gradient descent optimizations are performed for the three conditions, and the optimal result is selected as the number of people to be judged for fine-tuning.

5. The method for multiple-person identification based on forensic evidence according to claim 1, characterized in that: In step S4, the gradient descent algorithm specifically includes: Step S41: Set the required number of zygotes and evaluate the fitting results using the pairing variance: Using the total pairwise variance as the loss function, the gradient can be calculated: The opposite direction of the gradient is the direction of maximum descent. Dividing the obtained descent vector by the square of its modulus gives the expected vector whose pairing variance decreases by 1. Step S42: cluster the ratios of the fitting results to find the correct ratios; set the output content ratio vector, the distance function to the Euclidean distance, set the density to at least the number of markers within the radius divided by 3, select the radius with the highest number of clusters, perform cluster analysis, and take the mean of each cluster as the result. Step S43: Perform a second gradient descent on the initial vector of the dependency generated for each locus. This time, perform 300 iterations of sliding size and std while keeping mass unchanged so that the peaks in the spectrum are fully matched. Then, keep size and std fixed and change height to approximate the spectrum, thereby obtaining a more accurate ratio. Then, take the average of the ratios generated at each locus in the original result to obtain the final result.

6. The method for multiple-person identification based on forensic evidence according to claim 1, characterized in that: The specific method of step S5 is: Establish a penalty score by adding the power of 2 of the absolute value of the difference between all sizes of all loci at a certain ratio and the nearest size of the original map: Thus, the ratios and corresponding genotypes can be arranged from small to large according to the penalty scores, and the one with the smallest penalty score is the most reliable result.

7. The method for multiple-person identification based on forensic evidence according to claim 3, characterized in that: It is also necessary to set the noise points to zero.

Images