Biological plaque in vitro time prediction method

Abstract

The application discloses a method for predicting the ex vivo time of biological traces. The method is based on deep learning and machine learning to infer the ex vivo time of biological traces. The method mainly screens gene features or microbial OTUs feature variables effective for ex vivo time prediction from original data of biological samples with different ex vivo times through a recurrent neural network and a PCA module. A high-precision lightGBM model training module is used to train a prediction model based on sample data with effective gene features or microbial OTUs feature variables. The ex vivo time of biological traces is predicted through the model. Experiments prove that the method established by the application can predict the ex vivo time of blood samples or saliva samples. Within the applicable time range of the model (0-168 days for blood samples and 0-150 days for saliva samples), the accuracy, precision, recall rate and F1 of the model can reach more than 85% or 90%, the prediction performance of the model is good, and the model meets the actual needs of event processing.

Description

Technical Field

[0001] This invention relates to the field of biotechnology, specifically to a method for predicting the time of biological stains after removal from the body. Background Technology

[0002] With the continuous advancement of informatization and intelligentization in the field of public security, criminal science and technology are facing higher development requirements. Driven by the joint development of life sciences and criminal technology, the increasing efficiency of biomarker discovery and extraction and the sensitivity of DNA testing technology have led to greater emphasis in the field of forensic science on the role of biological trace examination in court proceedings and the improvement of evidence chains. At the same time, this has also placed new and higher demands on the examination, identification, and information analysis of biological traces: providing reliable verification for individual identification and paternity testing, while also providing direction for clarifying the relationship between obtained DNA and organisms, the specificity of biological information at the nucleic acid level and the characterization of personal features, and the inference of spatiotemporal clues from biological traces.

[0003] The inferred time of biomarker retention, like the inferred time of death of an organism, can provide more comprehensive clues and evidence from a temporal perspective. In real-world events, especially those occurring in public places such as shopping malls, hotels, and internet cafes with complex personnel, where surveillance videos have blind spots or ineffective coverage, where there is high population mobility and detection methods cannot effectively identify individuals, or where large-scale investigations are too costly, accurate inference of biomarker retention time not only provides valuable time information but also plays a crucial role in clarifying the correlation between organism information and event occurrence, inferring the event process, and reconstructing the scene. This saves time, reduces investigation costs, enhances the value of evidence, and improves the timeliness of event handling and resolution. Therefore, studying the evolutionary patterns of biomarkers and establishing methods for inferring biomarker retention time can be an effective and direct approach.

[0004] Saliva (stains) and blood (stains) are the most common biological traces, and DNA STR typing methods are mature and have high detection rates. Saliva (stains) can be obtained from samples such as cigarette butts, mineral water bottles, tableware, water cups, and oral swabs. Furthermore, saliva contains a rich resource of human oral microorganisms, and research on the structure of the oral microbiome is well-developed, making it an ideal sample for inferring the time of biological traces. Blood (stains) are often closely related to the course of events. Blood (stains) are the most extensively and deeply studied biological traces in forensic science, serving as crucial evidence for reconstructing events, characterizing the timing of events, and identifying individuals involved. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to predict the time of biological stains outside the body.

[0006] To address the aforementioned technical problems, this invention provides a method for predicting the time of ex vivo of biological stains. The method may include the following steps: sequencing biological stain samples with known ex vivo times at different ex vivo times to obtain gene sequence data and / or microbial sequence data of the samples; performing data analysis on the gene sequence data and / or the microbial sequence data to obtain data analysis results for the samples; using a recurrent neural network to extract the known ex vivo times and data analysis results to obtain sample gene feature variables and / or microbial OTUs feature variables that are effective in predicting the ex vivo time; and using PCA to screen from the gene feature variables and / or the OTUs feature variables to identify important gene feature variables and / or microbial OTUs feature variables that are more effective in predicting the ex vivo time. Or important OTU feature variables, based on the important gene feature variables and / or important OTU feature variables, obtain important genes and / or important OTUs that are effective for predicting ex vivo time; extract the sequence data of the important genes and / or the abundance data of the important OTUs, based on the sequence data of the important genes and / or the abundance data of the important OTUs, use the lightGBM model to train and obtain a model for predicting the ex vivo time of biomarkers, and use the model to predict the ex vivo time of the biomarker sample to be predicted.

[0007] In the method described above, the important genes and / or the important OTUs may be the target genes and / or target OTUs for constructing the predictive biological stain time-out model.

[0008] In the method described above, the biological stain sample can be a saliva sample or a blood sample. The time outside the body for the blood sample can be 0–168 days. The time outside the body for the saliva sample can be 0–150 days.

[0009] In the method described above, the saliva sample can be an oral pharyngeal swab sample from an organism. The microbial sequence data can be the microbial sequence data of the saliva sample. The blood sample can be a peripheral blood sample from an organism. The gene sequence data can be the RNA sequence data of the blood sample.

[0010] In the method described above, the sequencing can be amplicon sequencing or transcriptome sequencing of the 16sDNA V4 region. The data analysis results of the sample can be the relative abundance information of OTUs of the microorganisms in the sample or the gene sequence information and gene expression FPKM information of the sample.

[0011] In the method described above, the gene sequence information may be gene expression sequence information. The gene sequence data and / or the microbial sequence data may be valid data after quality control. The data analysis may be OTU analysis and / or species annotation analysis. The data analysis may also be gene expression level analysis.

[0012] In the method described above, the recurrent neural network can use LSTM as the basic unit. Given an input sequence x, the LSTM passes through an input gate i. t Forgotten Gate t Output gate o t The control model output is specifically described as follows:

[0013] i t =σ(W i ×[h t-1 x t ]+b i )

[0014] f t =σ(W f ×[h t-1 x t ]+b f )

[0015] o t =σ(W o ×[h t-1 x t ]+b o )

[0016] Where σ is the activation function constructed by the Sigmoid function, h t-1 Let C be the current hidden layer, where W and b are the corresponding weight matrix and bias, respectively. The updated memory unit C... t for:

[0017] q t =tanh(W q ×[h t-1 x t ]+b q )

[0018] c = f t c t-1 +i t q t

[0019] Where q t As an alternative memory unit, c t-1为 Existing memory gates, LSTM updated hidden layer h t for:

[0020] h t =o t tanh(c t )

[0021] LSTM uses input x t Update hidden layer h tThe final hidden layer obtained by inputting the entire sequence is the implicit feature of the input.

[0022] In the method described above, the PCA screening process can be as follows:

[0023] 1) Arrange the original data into an n x m matrix X;

[0024] 2) Zero-mean value for each row of X, i.e., subtract the mean of that row;

[0025] 3) Calculate the covariance matrix;

[0026] 4) Find the eigenvalues ​​and corresponding eigenvectors of the covariance matrix;

[0027] 5) Arrange the eigenvectors into a matrix from top to bottom according to the magnitude of their corresponding eigenvalues, and take the first k rows to form matrix P;

[0028] 6) This refers to the data after dimensionality reduction to k dimensions.

[0029] In the model construction method described above, the model acquisition module described in A3 may include the following modules:

[0030] A3-1) Model training module.

[0031] A3-2) Model Validation Module.

[0032] A3-1) The model training module can be established by a method including the following steps: using LightGBM to train the model based on the important genes and their sequence data. During model training, for each node of xgboost, the gains of all the important features at that node during splitting are calculated:

[0033]

[0034] Among them G L G R H represents the first derivative of the results of the left and right subtrees after feature splitting with respect to the loss function (here, softmax). R H L represents the second derivative of the left and right subtrees with respect to the loss function during splitting; the other symbols are parameters.

[0035] In the method described above, the model for predicting the time of biological stain removal can be obtained based on the deep learning computing platform PaddlePaddle 1.8.3 and the machine learning computing tool sklearn 0.24.2.

[0036] To address the aforementioned technical problems, the present invention also provides an apparatus for predicting the time of biological stain removal from the body.

[0037] The device may include the following modules:

[0038] A1. Sequence Data Acquisition Module: Used to obtain gene sequence data and / or microbial sequence data of biological stain samples from samples with known in vitro time and different in vitro time.

[0039] A2. Sequence Data Analysis Module: Used to perform data analysis on the gene sequence data and / or the microbial sequence data to obtain the data analysis results of the sample.

[0040] A3. Recurrent Neural Network Module: Used to extract the known in vitro time and the data analysis results using a recurrent neural network to obtain the genetic feature variables and / or the microbial OTUs feature variables of the sample that are effective for predicting in vitro time.

[0041] A4. PCA Module: Used to use PCA to screen and obtain important gene feature variables and / or important OTU feature variables from the gene feature variables and / or the OTU feature variables, and to obtain important genes or important OTUs based on the important gene feature variables and / or important OTU feature variables.

[0042] A5. LightGBM Model Training Module: Used to extract the sequence data of the important genes and / or the abundance data of the important OTUs, and to use LightGBM to construct a model to predict the time of biological stains to detachment based on the sequence data of the important genes and / or the abundance data of the important OTUs.

[0043] A6. Biomarker Ex-Visit Time Prediction Module: Used to predict the ex-visit time of the biomarker sample to be predicted using the model.

[0044] In the device described above, the biological stain sample may be a saliva sample or a blood sample.

[0045] In the device described above, the saliva sample can be an oral pharyngeal swab sample from an organism. The microbial sequence data can be the microbial sequence data of the saliva sample. The blood sample can be a peripheral blood sample from an organism. The gene sequence data can be the RNA sequence data of the blood sample.

[0046] In the device described above, the gene sequence data may be transcriptome sequencing data. The microbial sequence data may be amplicon sequencing data of the 16sDNA V4 region. The data analysis results may be the relative abundance information of OTUs of the microorganisms in the sample or the gene sequence information and gene expression FPKM information of the sample.

[0047] In the device described above, the gene sequence information may be gene expression sequence information. The gene sequence data and / or the microbial sequence data may be valid data after quality control. The data analysis may be OTU analysis and / or species annotation analysis. The data analysis may also be gene expression level analysis.

[0048] To address the aforementioned technical problems, the present invention also provides a computer-readable storage medium storing a computer program. The computer program enables a computer to perform the steps described above or enables a computer to perform modules of the apparatus described above.

[0049] To address the aforementioned technical problems, the present invention also provides a computer-readable storage medium storing a computer program. The computer program enables a computer to perform the steps of the methods described above, or the computer program enables a computer to perform the steps of a module of the apparatus described above.

[0050] This invention utilizes next-generation sequencing technology to deeply analyze the microbial community structure of saliva (stains) and the transcriptomic characteristics of blood (stains). It can obtain rich biological characteristic data with small samples and single experiments, providing a massive data source for machine learning-based in vitro time-varying analysis of saliva / blood (stains). At the same time, by applying machine learning technology, it establishes a data model with diverse variables and a high-order training set. Using deep learning and convolutional neural networks, it allows the machine to learn and train itself under unsupervised conditions, further exploring the time-varying patterns of in vitro biological stains, screening time-varying marker genes and microbial populations, and providing a method for in vitro time-varying characterization of saliva (stains) and blood (stains). Experiments demonstrate that the biological stain ex vivo time estimation method established in this invention achieves higher accuracy for predicting blood samples with fewer group numbers, ranging from a minimum of 73% to a maximum of 89%. Furthermore, even with a relatively small number of group numbers (e.g., blood stains divided into 10 groups), the sample grouping time periods can meet the needs of most event scene analysis. In this case, the model's accuracy, precision, recall, and F1 score are all above 85%, indicating good model prediction performance and meeting practical requirements. Similarly, the biological stain ex vivo time estimation method established in this invention achieves a minimum accuracy of 80% and a maximum of 92% for predicting saliva samples. Likewise, even with a relatively small number of group numbers (e.g., saliva stains divided into 7 groups), the sample grouping time periods can meet the needs of most event scene analysis. In this case, the model's accuracy, precision, recall, and F1 score are all above 90%, indicating good model prediction performance and meeting practical event handling requirements. Attached Figure Description

[0051] Figure 1 This is a logical architecture diagram of a method for predicting the time of biological traces after they leave the body.

[0052] Figure 2 This is the dilution curve, which represents the species diversity curve for group analysis. The horizontal axis represents the number of randomly selected sequencing reads from the sample, and the vertical axis represents the number of OTUs that can be constructed based on that number of sequencing reads, reflecting the sequencing depth. Different sample groups are represented by curves of different shapes.

[0053] Figure 3 This is a graph showing the distribution of gene expression levels. The horizontal axis represents log10(FPKM+1), and the vertical axis represents gene density.

[0054] Figure 4 This is a violin plot of gene expression levels. The horizontal axis represents the sample (group) name, and the vertical axis represents log10(FPKM+1). Each region of the violin plot corresponds to five statistics (from top to bottom: maximum, upper quartile, median, lower quartile, and minimum). The width of each violin plot represents the number of genes at that expression level.

[0055] Figure 5 Top 50 key characteristic variables for blood sample time-out analysis.

[0056] Figure 6 Top 50 key characteristic variables for saliva stain (oral swab) ex vivo time analysis. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to specific embodiments. The given embodiments are merely illustrative of the invention and not intended to limit its scope. The embodiments provided below can serve as a guide for further improvements by those skilled in the art and do not constitute a limitation on the invention in any way.

[0058] Unless otherwise specified, the experimental methods used in the following examples are conventional methods, performed according to the techniques or conditions described in the literature in this field or according to the product instructions. Unless otherwise specified, the materials and reagents used in the following examples are commercially available.

[0059] Example 1: Establishment of an algorithm for estimating the time of biological traces after removal from the body

[0060] 1. Obtaining sequencing data

[0061] 1.1 Sample preparation:

[0062] Nineteen unrelated Han Chinese healthy volunteers (9 males and 10 females) were selected. Oral swabs were collected to obtain saliva samples, which were then air-dried and stored at room temperature (temperature: 19–28℃, relative humidity: 20–63%). DNA samples were extracted from the samples at 0d (4h after collection), 7d, 14d, 21d, 28d, 60d, 90d, 120d, and 150d for testing. Samples collected at 0d (4h after collection), 7d, 14d, 21d, 28d, 60d, 90d, 120d, and 150d were named D0, D7, D14, D21, D28, D60, D90, D120, and D150, respectively.

[0063] Twenty unrelated healthy Han Chinese volunteers (10 males and 10 females) were selected. Peripheral blood samples were collected, anticoagulated with EDTA, and stored at 4℃. Total RNA was extracted from the samples at 0d (4h after collection), 0.5d (12h after collection), 1d, 2d, 4d, 7d, 14d, 21d, 28d, 56d, 84d, 112d, 140d, and 168d for detection. Samples collected at these times were named D0, D0.5, D1, D2, and D3, respectively. D4, D7, D14, D21, D28, D56, D84, D112, D140 and D168.

[0064] 1.2 Nucleic acid extraction:

[0065] 1.2.1 Extraction of microbial DNA from saliva samples using the CTAB method

[0066] ① Cut two oral swabs and place them in a 2.0 mL EP tube. Add 1000 μL of CTAB lysis buffer and 200 μL of lysozyme. Incubate at 65 °C for 60 min, inverting and mixing several times during the incubation period to ensure complete lysis of the sample.

[0067] ② Centrifuge at 12000 rpm for 3 min, collect 950 uL of the supernatant, add an equal volume of phenol (pH 8.0): chloroform: isoamyl alcohol (25:24:1), mix by inversion, and centrifuge at 12000 rpm for 10 min;

[0068] ③ Take the supernatant, add an equal volume of chloroform:isoamyl alcohol (24:1), mix by inversion, and centrifuge at 12000 rpm for 10 min;

[0069] ④ Transfer the supernatant to a 1.5 mL centrifuge tube, add 3 / 4 volume of isopropanol, shake up and down, and precipitate at -20℃;

[0070] ⑤ Centrifuge at 12000 rpm for 10 min, pour out the liquid (be careful not to pour out the precipitate), and wash twice with 1 mL of 75% ethanol;

[0071] ⑥ Dry in a clean bench or air dry at room temperature;

[0072] ⑦ Add ddH2O and incubate at 55-60℃ for 10 min to dissolve the DNA;

[0073] ⑧ Add RNase A 100L to digest RNA, incubate at 37°C for 15 min to obtain gDNA sample, and store at -20°C for later use.

[0074] DNA extraction results showed that DNA was successfully extracted from a total of 171 samples, all of which met the requirements for library construction, resulting in 171 qualified DNA extraction samples.

[0075] 1.2.2 Trizol method for RNA extraction from blood samples

[0076] ① Add 3 times the volume of Trizol to the blood and let it stand at room temperature for 5 minutes to allow it to fully lyse;

[0077] ② Add chloroform at a ratio of 200 μL chloroform / 1 mL Trizol, shake to mix, and place on ice for 15 min;

[0078] ③ Centrifuge at 16,000g for 15 minutes at 4℃;

[0079] ④ Draw the upper aqueous phase into another centrifuge tube;

[0080] ⑤ Add isopropanol at a ratio of 0.5 mL isopropanol / 1 mL Trizol, mix well, and place on ice for 5–10 minutes;

[0081] ⑥ Centrifuge at 12,000g for 10 min at 4℃, discard the supernatant, and let the RNA settle at the bottom of the tube;

[0082] ⑦ Add 1 mL of 75% ethanol to 1 mL of Trizol, gently shake the centrifuge tube, and suspend the precipitate.

[0083] ⑧ Centrifuge at 75000g for 5 minutes at 4℃, discarding as much of the supernatant as possible; repeat the washing process once.

[0084] ⑨ Obtain total RNA samples by air drying at room temperature or vacuum drying for 5-10 minutes, and store at -80℃ for later use.

[0085] RNA extraction results showed that total RNA was successfully extracted from 280 samples, and the total RNA from 278 samples met the requirements for library construction.

[0086] 1.3 Library Construction and Sequencing:

[0087] 1.3.1 16S amplicon sequencing of saliva DNA samples

[0088] 1.3.1.1 PCR Amplification

[0089] ① Primer sequence:

[0090] Table 1 Primer sequences used

[0091]

[0092] ②PCR reaction system (50μL):

[0093]

[0094] ③PCR amplification procedure:

[0095]

[0096] 1.3.1.2 Mixing and purification of PCR products

[0097] The PCR product was mixed with an equal volume of 1× loading buffer and subjected to 2% agarose gel electrophoresis. The target band of the PCR product (400–450 bp) was selected for subsequent experiments. Equal volumes of PCR product were mixed thoroughly and then detected by 2% agarose gel electrophoresis. The target band was recovered using the QIAEX II Gel Extraction Kit (Qiagen). The recovery procedure for the target band is as follows:

[0098] ① After electrophoresis, cut no more than 250 mg of the target band, remove excess gel, and place it in a 1.5 mL centrifuge tube;

[0099] ② Add 300 μL of Buffer QX1 to every 100 mg of gel and vortex for 30 seconds;

[0100] ③ Add 30 μL of QIAEX II and incubate at 50 °C for 10 min, vortexing once every 2 min during the incubation period;

[0101] ④ Centrifuge for 30 seconds, discard the supernatant, and wash with 500 μL Buffer QX1;

[0102] ⑤ Centrifuge for 30 seconds, discard the supernatant, and wash twice with 500 μL Buffer PE;

[0103] ⑥ Centrifuge for 30 seconds, discard the supernatant, and dry in air for 10-15 minutes until the silica beads precipitate turns white;

[0104] ⑦ Add 20 μL of H2O to resuspend the silica beads and incubate at 50 °C for 10 min;

[0105] ⑧ Centrifuge for 30 seconds, collect the supernatant to obtain the purified DNA fragment, and store at -20℃ for later use.

[0106] 1.3.1.3 Library Construction and Sequencing

[0107] use The purified DNA fragments were used to construct libraries using the DNA PCR-Free Sample Preparation Kit (Illumina, USA). The constructed libraries were then quantified using Qubit 2.0 (ThermoScientific, USA) and Agilent Bioanalyzer 2100 (Agilent, USA). After the libraries were deemed acceptable, the 16S V4 region was sequenced using a HiSeq 2500 PE250 to obtain raw sequencing data. The average sequencing read length was 0.09M and the average data size was 0.05G per sample. Sequencing quality data are shown in Table 2.

[0108] Table 2. Statistics of 16S amplicon sequencing data from saliva DNA samples

[0109]

[0110] 1.3.2 Transcriptome sequencing of blood stain RNA samples

[0111] The starting RNA for library construction was total RNA (the total RNA obtained in step 1.2.2), with a total amount ≥100 ng. The library construction kit used was from Illumina. Ultra TM RNA Library Prep Kit. Ribo-Zero was used. TM The rRNA Removal Kit removes ribosomal RNA to obtain mRNA, which is then used... In the Magnesium RNA Fragmentation Module, the obtained mRNA is randomly fragmented to obtain fragmented mRNA. Using the fragmented mRNA as a template and random oligonucleotides as primers, the first strand of cDNA is synthesized in the M-MuLV reverse transcriptase system. Subsequently, the RNA strand is degraded with RNase H, and the second strand of cDNA is synthesized using dNTPs in the DNA polymerase I system. The purified double-stranded cDNA undergoes end repair, followed by A-tailing and ligation of sequencing adapters. CDNA samples of approximately 200 bp are screened using AMPure XP beads, amplified by PCR, and the PCR products are purified again using AMPure XP beads to finally obtain the transcriptome sequencing library.

[0112] After the transcriptome sequencing library was constructed, preliminary quantification was performed using a Qubit 2.0 Fluorometer, diluting the library to 1 ng / uL. Then, the insert size of the library was checked using an Agilent 2100 Bioanalyzer. Once the insert size met expectations, qRT-PCR was used to accurately quantify the effective library concentration (effective concentration higher than 2 nM) to ensure library quality. After passing the library test, different libraries were pooled according to the effective concentration and target data volume requirements, and then 150 bp paired-end sequencing was performed on an Illumina Novaseq 6000 sequencing platform to obtain raw image data files with a sequencing depth of 6 G raw data. The raw image data files were converted into raw sequencing reads (Raw Data) using CASAVA base calling analysis.

[0113] 2. Sequencing data analysis

[0114] 2.1 Analysis of 16S amplicon sequencing data from saliva samples

[0115] The raw data obtained from sequencing contains a certain proportion of dirty data. To make the information analysis results more accurate and reliable, the raw data is spliced ​​and filtered to obtain effective data. OTU (Operational Taxonomic Units) clustering and species classification analysis are then performed with 97% identity.

[0116] 2.1.1 Sequencing Data Processing

[0117] ①Tags truncation: Truncate the Raw Tags (Raw Data, raw sequencing sequence) from the first low-quality base site where the number of consecutive low-quality bases (default quality threshold is <= 19) reaches the set length (default length is 3);

[0118] ②Tags length filtering: After the tags are truncated, the resulting tags dataset is further filtered to remove tags whose consecutive high-quality base length is less than 75% of the tag length.

[0119] ③ Removal of chimeric sequences: The tag sequences are compared with the database (Gold database, http: / / drive5.com / uchime / uchime_algo.html) using the UCHIME Algorithm (http: / / www.drive5.com / usearch / manual / uchime_algo.html) to detect chimeric sequences, and finally the chimeric sequences are removed to obtain the final effective tags.

[0120] 2.1.2 OTU Analysis and Species Annotation

[0121] Uparse software (Uparse v7.0.1001, http: / / drive5.com / uparse / ) was used to cluster all effective tags of all samples. By default, sequences were clustered into OTUs (Operational Taxonomic Units) with a 97% identity. Representative sequences of OTUs were selected based on the algorithm principles, choosing the sequence with the highest frequency among the OTUs as the representative sequences. Species annotation was performed on the representative sequences of OTUs using the Mothur method and the SSUrRNA database of SILVA (http: / / www.arb-silva.de / ) (with a threshold of 0.8–1). Taxonomic information was obtained, and the community composition of each sample was statistically analyzed at each taxonomic level: kingdom, phylum, class, order, family, genus, and species.

[0122] The number of species represented by the OTUs was counted (i.e., the number of OTUs), and a dilution curve was constructed. The results showed that the dilution curve tended to flatten. Figure 2The sequencing data volume was reasonable, indicating that the sequencing depth could comprehensively reflect the diversity of the sample's microbial community. Based on the OTU annotation results, a summary table of the relative abundance of OTUs at different taxonomic levels of the sample was obtained.

[0123] 2.2 Analysis of Transcriptome Sequencing Data from Blood Samples

[0124] 2.2.1 Sequencing quality assessment

[0125] The sequencing error rate for each base in 278 samples was statistically analyzed. The average error rate for Q20 was 98.21%, and the average error rate for Q30 was 94.83%. The lowest error rate was 0.025%, the highest was 0.03%, and the average error rate was 0.0216%. The error rate for 179 samples was also statistically analyzed, with an average GC content of 56.21%. Therefore, the sequencing data from the 278 samples are of acceptable quality and can be used for subsequent data analysis.

[0126] 2.2.2 Data Quality Control

[0127] The raw sequencing data contains low-quality reads with adapters. To ensure the quality of information analysis, the raw reads must be quality controlled and filtered to obtain clean reads. Subsequent analyses are based on clean reads. The data quality control process is as follows: (1) Remove reads with adapters; (2) Remove reads containing N (N indicates that the base information cannot be determined); (3) Remove low-quality reads (reads with Qphred <= 20 accounting for more than 50% of the total read length).

[0128] 2.2.3 Reference sequence alignment analysis

[0129] Sequencing sequence localization algorithm: Hisat2 v2.0 (http: / / daehwankimlab.github.io / hisat2 / download / ) software was selected to perform genome localization analysis on the quality-controlled clean reads. HISAT can effectively align to split reads in RNA Seq sequencing data and is currently the alignment software with the highest alignment rate and the most accurate alignment. During the analysis, the default parameters of the software were used. The Hisat algorithm is mainly divided into three parts: (1) aligning the entire sequencing sequence to a single exon of the genome; (2) aligning the sequencing sequence into segments to two exons of the genome; (3) aligning the sequencing sequence into segments to three or more exons of the genome.

[0130] The alignment rate of 278 samples was statistically analyzed, with an average of 56,644,226.98 total reads per sample. The lowest percentage of clean reads that could be located on the genome was 95.52%, indicating that the reference genome was appropriately selected. Sequencing data are shown in Tables 3 and 4.

[0131] Table 3. Blood sample transcriptome sequencing data (1)

[0132]

[0133]

[0134] Table 4. Blood sample transcriptome sequencing data (2)

[0135]

[0136] 2.2.4 Gene Expression Level Analysis

[0137] Gene expression levels in each sample were analyzed using HTSeq v0.9.1 (https: / / htseq.readthedocs.io / en / master / index.html) software, with the union model employed. The resulting files statistically analyzed the number of genes at different expression levels and the expression level of individual genes, forming a gene expression FPKM summary table. Finally, gene expression FPKM information for each sample was obtained. An FPKM value > 0.1 was selected as the threshold for determining gene expression. The gene expression FPKM distribution is shown below. Figure 3 and Figure 4 The gene expression density distribution differs among different sample groups.

[0138] 3. Construction of a method for predicting the time to ex vivo of biological stains

[0139] 3.1 Algorithm Principle

[0140] Inferences about the time of biological traces, such as saliva and blood samples, are made based on deep learning and machine learning. Figure 1 As shown, the prediction method mainly consists of three modules: recurrent neural network, PCA, and LightGBM model training.

[0141] The recurrent neural network (RNN) and PCA modules are used for data cleaning, specifically to screen gene features effective for predicting ex vivo time from raw biological samples collected at different ex vivo times. RNNs are used to extract gene features (gene sequence information and gene expression FPKM information) from blood samples collected at different ex vivo times, or microbial OTU features (OTU features obtained from microbial species relative abundance data) from saliva samples collected at different ex vivo times. Then, PCA is used to screen important gene features and / or important OTU features from the obtained gene feature variables and / or microbial OTU feature variables, thereby identifying more effective features for inferring ex vivo time. Based on these important gene feature variables and / or important OTU feature variables, key genes and / or important OTUs for model construction are obtained.

[0142] The high-precision lightGBM model training module is used to train a biological stain time-out prediction model. The LightGBM model training process involves extracting features of key genes and / or key OTUs selected based on PCA from the original sample data, and then using these features to train the lightGBM model. This model can construct multiple decision trees, and the boosting training method can greatly increase the model accuracy. Furthermore, lightGBM can handle outlier samples, enables rapid parallel training, facilitates deployment, and achieves ideal inference and prediction results.

[0143] 3.2 Model Construction and Solution Methods

[0144] The construction and solution were implemented using the deep learning computing platform PaddlePaddle 1.8.3 and the machine learning computing tool sklearn 0.24.2.

[0145] 3.3 Algorithm Calculation Process

[0146] 3.3.1 Recurrent Neural Network Module Operation and Processing

[0147] For the saliva sample group, a recurrent neural network was used to extract characteristic variables of OTUs for screening based on the known in vitro time of the samples and the data analysis results (relative abundance information of OTUs); for the blood sample group, a recurrent neural network was used to extract characteristic variables of genes for screening based on the known in vitro time of the samples and the data analysis results (FPKM information of gene expression).

[0148] When processing sequence input, recurrent neural networks (RNNs) record historical information of the sequence through a hidden layer and recursively extract sequence features, enabling them to extract features according to sequence order. The RNN is composed of an autoencoder consisting of two LSTM layers. The input to the first LSTM layer is the time-varying sequence of genes from ex vivo blood or microbial OTUs from saliva stains, and the output is the feature corresponding to that gene or OTU. The input to the second LSTM layer is the extracted gene or OTU feature, and the output is the input to the first LSTM layer. This method can accurately and effectively extract gene or OTU features. The autoencoder using LSTMs can encode the time series of the input genes or OTUs, obtaining the temporal variation information of each gene or OTU.

[0149] The recurrent neural network module uses LSTM as the basic unit. Given an input sequence x, the LSTM passes it through the input gate i. t Forgotten Gate t Output gate o t The control model output is specifically described as follows:

[0150] i t =σ(W i ×[h t-1 x t ]+b i )

[0151] f t =σ(W f ×[h t-1 x t ]+b f )

[0152] o t =σ(W o ×[h t-1 x t ]+b o )

[0153] Where σ is the activation function constructed by the Sigmoid function, h t-1 Let C be the current hidden layer, where W and b are the corresponding weight matrix and bias, respectively. The updated memory unit C... t for:

[0154] q t =tanh(W q ×[h t-1 x t ]+b q )

[0155] c = f t c t-1 +i t qt

[0156] Where q t As an alternative memory unit, c t-1 For the existing memory gates, the LSTM updates the hidden layer h. t for:

[0157] h t =o t tanh(c t )

[0158] LSTM updates the hidden layer h by input x2. t The final hidden layer obtained by inputting the entire sequence is the implicit feature of the input.

[0159] 3.3.2 PCA Module Calculation and Processing

[0160] PCA is a simple and fast method. It processes the (sequence) features of different genes or OTUs extracted from recurrent neural networks, generating weights for each gene or OTU. These weights characterize the contribution of the gene or OTU to in vitro time estimation; a larger weight indicates a more significant role in in vitro time estimation. PCA can be used to screen for features of important genes or OTUs, thereby identifying more effective feature variables for in vitro sample time estimation and obtaining key genes or OTUs for model construction.

[0161] PCA screening can both speed up the model by reducing the number of features and, to some extent, reduce prediction noise, thereby increasing model accuracy. The PCA process is as follows:

[0162] 1) Arrange the original data into an n x m matrix X;

[0163] 2) Zero-mean value for each row of X, i.e., subtract the mean of that row;

[0164] 3) Calculate the covariance matrix;

[0165] 4) Find the eigenvalues ​​and corresponding eigenvectors of the covariance matrix;

[0166] 5) Arrange the eigenvectors into a matrix from top to bottom according to the magnitude of their corresponding eigenvalues, and take the first k rows to form matrix P;

[0167] 6) This refers to the data after dimensionality reduction to k dimensions.

[0168] 3.3.2 LightGBM Model Training Module

[0169] This module is based on the important genes or OTUs screened by the PCA module. It extracts the corresponding important gene features or important OTU features (i.e., gene sequence information and gene expression FPKM information of important genes, or sequence information or relative abundance information of important OTUs) from the original sample data for model training.

[0170] During LightGBM training, for each node of xgboost, the gains of all features when splitting at that node are calculated:

[0171]

[0172] Among them G L G R H represents the first derivative of the results of the left and right subtrees after feature splitting with respect to the loss function (here, softmax). R H L represents the second derivative of the left and right subtrees with respect to the loss function during splitting; the other symbols are parameters.

[0173] This formula can determine which feature of each node is better for splitting. For each node, a threshold can be obtained. If the feature value is greater than the threshold, it enters the right subtree; if it is less than the threshold, it enters the left subtree. A preliminary result can be obtained from the child node.

[0174] 3.4 Training Results of the Method for Predicting the Time Out of Organisms

[0175] Twenty percent of the experimental samples from both the saliva and blood sample groups were randomly selected as training models. In the saliva sample group, based on the sample data (known in vitro time and relative abundance of microbial OTUs), a method for predicting the in vitro time of biomarkers was applied. Recurrent neural networks and PCA were used to screen important feature variables, and a lightGBM model was used to train the model to predict the in vitro time of biomarkers. In the blood sample group, based on the sample data (known in vitro time, gene sequence information, and gene expression FPKM information), a method for predicting the in vitro time of biomarkers was applied. Recurrent neural networks and PCA were used to screen important feature variables, and a lightGBM model was used to train the model to predict the in vitro time of biomarkers.

[0176] The top 50 rankings of key features obtained from saliva and blood samples are as follows: Figure 5 As shown in Figure 6.

[0177] 3.5 Test Results of the Method for Estimating the Ex-Vivo Time of Biological Marks

[0178] The remaining 80% of samples from both the saliva and blood sample groups were used as test samples. In vitro time was estimated for each group of blood and saliva samples under different analytical modes. A total of 135 saliva test samples were collected: 15 samples each from groups D0, D7, D14, D21, D28, D60, D90, D120, and D150. A total of 224 blood test samples were collected: 16 samples each from groups D0, D0.5, D1, D2, D4, D7, D14, D21, D28, D56, D84, D112, D140, and D168.

[0179] Specifically, the important feature variables obtained from step 3.4 and the established method for predicting the time of biological stains outside the body were used to infer the time of biological stains outside the body under eight grouping analysis modes (Table 6): “blood stains-10 group”, “blood stains-11 group”, “blood stains-12 group”, “blood stains-13 group”, “blood stains-14 group”, “saliva stains-7 group”, “saliva stains-8 group”, and “saliva stains-9 group”. The predicted time of biological stains outside the body for each sample in the eight grouping modes was obtained.

[0180] Calculate accuracy, precision, recall, and F1 using the following formulas (Table 5).

[0181] Accuracy=(TP+TN) / (TP+FP+FN+TN)

[0182] Precision = TP / (TP + FP)

[0183] Recall = TP / (TP + FN)

[0184] F1 = Precision * Recall * 2 / (Precision + Recall)

[0185] Where positive samples are those belonging to a certain class, and negative samples are those not belonging to a certain class; Accuracy is the percentage of positive samples predicted as positive; Precision is the percentage of negative samples predicted as positive; Recall is the percentage of positive samples predicted as negative. TP: the number of positive samples predicted as positive; FP: the number of negative samples predicted as positive; TN: the number of negative samples predicted as negative; FN: the number of positive samples predicted as negative.

[0186] Table 5. Model accuracy under different grouping methods

[0187]

[0188] Table 6. List of different grouping methods

[0189]

[0190]

[0191] The results in Table 5 show that the biological stain ex vivo time estimation method established in this invention achieves higher accuracy for blood samples with fewer group numbers, ranging from a minimum of 73% to a maximum of 89%. It is also noteworthy that even with a relatively small number of group numbers (e.g., blood stains divided into 10 groups), the sample grouping time periods can meet the needs of most event on-site analysis. In this case, the model's accuracy, precision, recall, and F1 score are all above 85%, indicating good predictive performance and meeting the actual needs of event handling. Similarly, the biological stain ex vivo time estimation method established in this invention achieves a minimum accuracy of 80% and a maximum of 92% for saliva samples. Again, it is noteworthy that even with a relatively small number of group numbers (e.g., saliva stains divided into 7 groups), the sample grouping time periods can meet the needs of most event on-site analysis. In this case, the model's accuracy, precision, recall, and F1 score are all above 90%, indicating good predictive performance and meeting the actual needs of event handling.

[0192] It should be noted that the selection of the number of model groups should consider two factors: first, the model accuracy—the higher the accuracy, the more reliable the sample prediction results; and second, the number of model groups—the more groups, the more accurate the time span of the sample prediction results. In this study, the prediction model achieved higher accuracy with fewer groups; that is, when the time span of the sample prediction results was relatively accurate, the model accuracy was relatively lower (the model achieved 80% accuracy for saliva stains (group 9) and 73% accuracy for blood stains (group 10). Therefore, in the application of the model, it is necessary to fully consider the required time span of the sample prediction results and the reliability (i.e., accuracy) of the prediction results, and select different grouping modes accordingly.

[0193] Example 2: Validation of the method for estimating the time of biological stains outside the body

[0194] 1. Estimation of time elapsed since saliva stain was removed from the body

[0195] 1.1 Saliva samples were obtained from four oral swabs collected at 7, 14, 21, 60, and 120 days after ex vivo collection, as well as four fresh (0 days) swabs. Microbial DNA was extracted from the samples for testing. The samples collected at 0, 7, 14, 21, 60, and 120 days were named D0, D7, D14, D21, D60, and D120, respectively. Wherein, D0 includes four samples: D0.1, D0.6, D0.10, and D0.15; D7 includes four samples: D7.4, D7.8, D7.13, and D7.17; D14 includes four samples: D14.4, D14.8, D14.13, and D14.17; D21 includes four samples: D21.2, D21.6, D21.11, and D21.13; D60 includes four samples: D60.4, D60.10, D60.14, and D60.18; and D120 includes four samples: D120.4, D120.7, D120.11, and D120.14.

[0196] 1.216sDNA amplification and sequencing

[0197] Using the microbial DNA of each sample in each of the above groups as templates, the primers shown in Table 1 were used for amplification to obtain the amplicon of the 16SDNA V4 region of the sample DNA.

[0198] The above 16SDNA V4 region amplicon was used The DNA PCR-Free Sample Preparation Kit (Illumina, USA) was used to construct the library. The constructed library was then quantified using Qubit 2.0 (ThermoScientific, USA) and Agilent Bioanalyzer 2100 (Agilent, USA). After the library was deemed acceptable, the 16S V4 region was sequenced using a HiSeq 2500 PE250.

[0199] 1.3 OTU Analysis and Species Annotation

[0200] The Uparse software (Uparse v7.0.1001, http: / / drive5.com / uparse / ) was used to cluster all effective tags of all samples. By default, the sequences were clustered into OTUs (Operational Taxonomic Units) with a consistency of 97%. At the same time, representative sequences of OTUs were selected. According to its algorithm principle, the sequence with the highest frequency in OTUs was selected as the representative sequence of OTUs.

[0201] Species annotation was performed on the representative sequences of OTUs. Species annotation analysis was conducted using the Mothur method and the SSUrRNA database of SILVA (http: / / www.arb-silva.de / ) (with thresholds set at 0.8–1) to obtain taxonomic information. Community composition of each sample was statistically analyzed at each taxonomic level: kingdom, phylum, class, order, family, genus, and species. Rapid multiple sequence alignment was performed using MUSCLE (Version 3.8.31, http: / / www.drive5.com / muscle / ) software to obtain the phylogenetic relationships of all representative sequences of OTUs for each sample.

[0202] 1.4 Estimation of time outside the body

[0203] The in vitro time estimation method for biological stains constructed in step 3.4 of Implementation 1 was applied. Following the important characteristic variables in step 3.5 of Example 1, under the "saliva stain-9 groups" model (i.e., divided into 9 groups) (resulting in more accurate prediction time spans and a prediction accuracy of over 80%), the in vitro time estimation was performed using the data analysis results (OTU relative abundance information) of saliva samples. The estimation results are shown in Table 7, and the in vitro time estimation results for each sample were correct.

[0204] Table 7 Results of inference of time out of body for saliva samples

[0205]

[0206] 2. Estimation of time since blood stains left the body

[0207] 2.1 Four blood samples were collected from each of the four days prior to the events described in Example 1, at times of 1 day, 4 days, 7 days, 28 days, 84 days, 112 days, and 168 days. Total RNA was extracted from these samples for testing. The samples collected at 1 day, 4 days, 7 days, 28 days, 84 days, 112 days, and 168 days were named D1, D4, D7, D28, D84, and D112, respectively. Among them, D1 includes four samples: D1B1, D1B2, D1B3, and D1B6; D4 includes four samples: D4B1, D4B2, D4B3, and D4B4; D7 includes four samples: D7B2, D7B3, D7B4, and D7B5; D28 includes four samples: D28B1, D28B2, D28B3, and D28B5; D84 includes four samples: D84B1, D84B2, D84B3, and D84B5; D112 includes four samples: D112B1, D112B3, D112B4, and D112B5; and D168 includes four samples: D168B1, D16883, D168B4, and D168B5.

[0208] 2.2 Transcriptome Sequencing

[0209] The starting RNA for library construction was total RNA, with a total volume ≥100 ng. The library construction kit used was from Illumina. Ultra TM RNA Library Prep Kit. Ribo-Zero was used. TM The rRNA Removal Kit removes ribosomal RNA, followed by application The mRNA obtained in the Magnesium RNA Fragmentation Module is randomly fragmented. After the library passes the library test, different libraries are pooled according to the effective concentration and target data volume requirements, and then 150bp paired-end sequencing is performed on the Illumina Novaseq 6000 sequencing platform, with a sequencing depth of 6G raw data.

[0210] 2.3 Gene Expression Level Analysis

[0211] Gene expression levels in each sample were analyzed using HTSeq v0.9.1 software, with the union model employed. The results files statistically analyzed the number of genes at different expression levels and the expression level of individual genes, forming a comprehensive gene expression FPKM table.

[0212] 2.4 Estimation of time outside the body

[0213] The in vitro time estimation method for biological stains established in step 3.4 of Implementation 1 was applied. Following the important characteristic variables in step 3.5 of Example 1, the in vitro time estimation was performed using the data analysis results (gene sequence information and gene expression FPKM information) of blood samples in the "blood stain-14 group" (i.e., divided into 14 groups) mode. The inference results for the in vitro time of each sample are shown in Table 8, and all inference results are correct.

[0214] Table 8 Results of Blood Sample Ex-Vivo Time Estimation

[0215]

[0216]

[0217] The above results demonstrate that the method for predicting the time of biological stain removal established in this invention, which utilizes a biological stain removal time analysis model constructed based on recurrent neural networks, PCA, and LightGBM (Light Gradient Boosting Machine), can accurately predict the time of removal from the body for saliva and blood samples. Within the experimental sample range, all prediction results are correct. The experimental method is mature, and the prediction results are accurate, providing a good method and reference for criminal technology incident handling practice and related research in the field of forensic science.

[0218] Example 3: Validation Comparative Example of the Method for Estimating the Ex-Vivo Time of Biological Marks

[0219] 1. Estimation of time elapsed since saliva stain was removed from the body

[0220] 1.1 Saliva samples were collected from three healthy Han Chinese volunteers unrelated to the healthy volunteers in Example 1 at in vitro collection times of 7, 14, 21, 28, 60, 90, 120, and 150 days, as well as from three fresh (0-day) cigarette butt samples. Microbial DNA was extracted from these samples for testing. Samples collected at 0, 7, 14, 21, 28, 60, 90, 120, and 150 days were named TD0, TD7, TD14, TD21, TD28, TD60, TD90, TD120, and TD150, respectively. Among them, TD0 includes three samples: TD0.1, TD0.2, and TD0.3; TD7 includes three samples: TD7.1, TD7.2, and TD7.3; TD14 includes three samples: TD14.1, TD14.2, and TD14.3; TD21 includes three samples: TD21.1, TD21.2, and TD21.3; TD28 includes three samples: TD28.1, TD28.2, and TD28.3; TD60 includes three samples: TD60.1, TD60.2, and TD60.3; TD90 includes three samples: TD90.1, TD90.2, and TD90.3; TD120 includes three samples: TD120.1, TD120.2, and TD120.3; and TD150 includes three samples: TD150.1, TD150.2, and TD150.3.

[0221] 1.2 Amplification and Sequencing of 16SDNA

[0222] Using the microbial DNA of each sample in each of the above groups as templates, the primers shown in Table 1 were used for amplification to obtain the amplicon of the 16SDNA V4 region of the sample DNA.

[0223] The above 16SDNA V4 region amplicon was used The DNA PCR-Free Sample Preparation Kit (Illumina, USA) was used to construct the library. The constructed library was then quantified using Qubit 2.0 (ThermoScientific, USA) and Agilent Bioanalyzer 2100 (Agilent, USA). After the library was deemed acceptable, the 16S V4 region was sequenced using a HiSeq 2500 PE250.

[0224] 1.3 OTU Analysis and Species Annotation

[0225] The Uparse software (Uparse v7.0.1001, http: / / drive5.com / uparse / ) was used to cluster all effective tags of all samples. By default, the sequences were clustered into OTUs (Operational Taxonomic Units) with a consistency of 97%. At the same time, representative sequences of OTUs were selected. According to its algorithm principle, the sequence with the highest frequency in OTUs was selected as the representative sequence of OTUs.

[0226] Species annotation was performed on the representative sequences of OTUs. Species annotation analysis was conducted using the Mothur method and the SSUrRNA database of SILVA (http: / / www.arb-silva.de / ) (with thresholds set at 0.8–1) to obtain taxonomic information. Community composition of each sample was statistically analyzed at each taxonomic level: kingdom, phylum, class, order, family, genus, and species. Rapid multiple sequence alignment was performed using MUSCLE (Version 3.8.31, http: / / www.drive5.com / muscle / ) software to obtain the phylogenetic relationships of all representative sequences of OTUs for each sample.

[0227] 1.4 Estimation of time outside the body

[0228] The method for estimating the time out of saliva stains constructed in step 3.4 of Implementation 1, and the key characteristic variables in step 3.5 of Example 1, were used to estimate the time out of saliva stains using the data analysis results (OTU relative abundance information) of saliva samples in the "saliva stains - divided into 9 groups" mode. The estimation results are shown in Table 9. Only 9 samples in groups TD7 to TD21 (accounting for 33.33% of the total test samples) had a prediction error of less than or equal to ±7 days. The prediction results indicate that the method for estimating the time out of saliva stains based on the relative abundance of microbial species in oral swabs is not ideal in cigarette butt samples. This may be related to the differences in microbial communities between oral swabs and cigarette butts. This suggests that in estimating the time out of saliva stains, it is necessary to establish basic data prediction models based on the relative abundance of microbial species in different types of saliva stains (such as oral swabs, cigarette butts, wiping materials from catering utensils, etc.) to achieve more accurate predictions of the time out of saliva stains.

[0229] Table 9. Results of inference of time spent in vitro on saliva samples.

[0230]

[0231] 2. Estimation of time since blood stains left the body

[0232] 2.1 Three blood samples were collected from three Han Chinese healthy volunteers unrelated to the healthy volunteers in Example 1 at in vitro time points of 0 days (4 hours after collection), 0.5 days, 1 day, 2 days, 4 days, 7 days, 14 days, 21 days, 28 days, and 56 days. Total RNA was extracted from the samples for detection. The samples collected at 0 days (4 hours after collection), 0.5 days, 1 day, 2 days, 4 days, 7 days, 14 days, 21 days, and 56 days were named TD0, TD0.5, TD1, TD2, TD4, TD7, TD14, TD21, TD28, and TD56, respectively. Among them, TD0 includes three samples: TD0B7, TD0B8, and TD0B9; TD0.5 includes three samples: TD0.5B7, TD0.5B8, and TD0.5B9; TD1 includes three samples: TD1B7, TD1B8, and TD1B9; TD2 includes three samples: TD2B7, TD2B8, and TD2B9; TD4 includes three samples: TD4B7, TD4B8, and TD4B9; TD7 includes three samples: TD7B7, TD7B8, and TD7B9; TD14 includes three samples: TD14B7, TD14B8, and TD14B9; TD21 includes three samples: TD21B7, TD21B8, and TD21B9; TD28 includes three samples: TD28B7, TD28B8, and TD28B9; and TD56 includes three samples: TD56B7, TD56B8, and TD56B9.

[0233] 2.2 Transcriptome Sequencing

[0234] The starting RNA for library construction was total RNA, with a total volume ≥100 ng. The library construction kit used was from Illumina. Ultra TM RNA Library Prep Kit. Ribo-Zero was used. TM The rRNA Removal Kit removes ribosomal RNA, followed by application The mRNA obtained in the Magnesium RNA Fragmentation Module is randomly fragmented. After the library passes the library test, different libraries are pooled according to the effective concentration and target data volume requirements, and then 150bp paired-end sequencing is performed on the Illumina Novaseq 6000 sequencing platform, with a sequencing depth of 6G raw data.

[0235] 2.3 Gene Expression Level Analysis

[0236] Gene expression levels in each sample were analyzed using HTSeq v0.9.1 software, with the union model employed. The results files statistically analyzed the number of genes at different expression levels and the expression level of individual genes, forming a comprehensive gene expression FPKM table.

[0237] 2.4 Estimation of time outside the body

[0238] The biological stain ex vivo time estimation method established in step 3.4 of Implementation 1 was applied, and the ex vivo time estimation was performed using the data analysis results (gene expression FPKM information) of blood samples in the "blood stain - divided into 14 groups" mode according to the important characteristic variables in step 3.5 of Implementation 1. The ex vivo time estimation results for each sample are shown in Table 10. Among them: (1) the prediction error of 17 samples is less than or equal to ±7 days (one week); (2) the prediction error of 7 samples is 0; (3) the prediction error of 14 to 28 days (TD14 to TD28) is relatively small, with 77.78% of the samples having a prediction error of less than or equal to ±7 days. The results show that, under the objective premise that there are individual and environmental differences between independent samples and model samples, relatively ideal prediction results can still be obtained for 14 to 28 days of samples. This model can provide important information for the judgment and identification of some cases.

[0239] Table 10 Results of Blood Sample Exit Time Estimation

[0240]

[0241] The present invention has been described in detail above. For those skilled in the art, the invention can be practiced in a wide range of ways with equivalent parameters, concentrations, and conditions without departing from its spirit and scope, and without requiring unnecessary experiments. Although specific embodiments have been given, it should be understood that further modifications can be made to the invention. In summary, according to the principles of the invention, this application is intended to include any changes, uses, or improvements to the invention, including changes made using conventional techniques known in the art that depart from the scope disclosed herein. Some of the essential features can be applied within the scope of the following appended claims.

Claims

1. A method for predicting the time of biological stain removal from the body, characterized in that: The method includes the following steps: performing transcriptome sequencing on blood samples from organisms at different known times of in vitro exposure to obtain gene sequence data of the blood samples; performing amplicon sequencing of the 16sDNA V4 region on saliva samples from organisms at different known times of in vitro exposure to obtain microbial sequence data of the saliva samples; and performing data analysis on the gene sequence data and the microbial sequence data to obtain data analysis results of the samples. A recurrent neural network is used to extract the known ex vivo time and data analysis results to obtain sample gene characteristic variables and microbial OTUs characteristic variables that are effective in predicting the ex vivo time. PCA is used to screen out important gene characteristic variables and important OTUs characteristic variables that are more effective in predicting the ex vivo time from the gene characteristic variables and the OTUs characteristic variables. Based on the important gene characteristic variables and important OTUs characteristic variables, important genes and important OTUs that are effective in predicting the ex vivo time are obtained. The sequence data of the important genes and the abundance data of the important OTUs are extracted. Based on the sequence data of the important genes and the abundance data of the important OTUs, the LightGBM model is used to train a model to predict the time of biological stain removal. The model is then used to predict the time of biological stain removal of the sample to be predicted. The data analysis results of the samples include the relative abundance information of OTUs of microorganisms in the saliva sample and the gene sequence information and gene expression FPKM information of the blood sample.

2. The method according to claim 1, characterized in that: The model for predicting the time of biological stain removal from the body was obtained based on the deep learning computing platform PaddlePaddle 1.8.3 and the machine learning computing tool sklearn 0.24.

2.

3. A device for predicting the time of biological stain removal from the body, characterized in that: The device includes the following modules: A1, Sequence data acquisition module: used to acquire gene sequence data of blood samples with known in vitro time and microbial sequence data of saliva samples with known in vitro time and different in vitro time; A2. Sequence Data Analysis Module: Used to perform data analysis on the gene sequence data and the microbial sequence data to obtain the data analysis results of the sample; A3. Recurrent Neural Network Module: Used to extract the known in vitro time and the data analysis results using a recurrent neural network to obtain the gene feature variables and microbial OTUs feature variables of the sample that are effective for predicting in vitro time; A4. PCA Module: Used to use PCA to screen and obtain important gene feature variables and important OTU feature variables from the gene feature variables and the OTU feature variables, and to obtain important genes or important OTUs based on the important gene feature variables and important OTU feature variables; A5. LightGBM Model Training Module: Used to extract the sequence data of the important genes and the abundance data of the important OTUs, and to use LightGBM to construct a model to predict the time to in vitro biomarker based on the sequence data of the important genes and the abundance data of the important OTUs; A6. Biomarker Time to In Vitro Prediction Module: Used to use the model to predict the time to in vitro biomarker sample to be predicted; The gene sequence data is transcriptome sequencing data; The microbial sequence data is amplicon sequencing data of the 16SDNA V4 region; The data analysis results include the relative abundance information of microorganisms OTUs in the saliva sample and the gene sequence information and gene expression FPKM information of the blood sample.

4. A computer-readable storage medium storing a computer program that causes a computer to establish the steps of the method as claimed in claim 1 or 2, or the computer program causes a computer to establish a module of the apparatus as claimed in claim 3.

5. A computer-readable storage medium storing a computer program that causes a computer to perform the steps of the method as claimed in claim 1 or 2, or the computer program causes a computer to perform the steps of a module of the apparatus as claimed in claim 3.

Images