A tumor screening method and system based on centromere region features

Abstract

The application discloses a tumor screening method and system based on centromere region characteristics, and belongs to the technical field of artificial intelligence tumor screening. Sequencing read data of peripheral blood free DNA of a subject is obtained, a centromere region alignment result is extracted through genome alignment, three types of sequencing representation indexes, i.e., coverage depth, single nucleotide polymorphism count and soft splicing count, are counted, valid sites are screened through two-dimensional paired statistical tests of original values and coverage depth normalized proportion values, and sequencing depth interference sites are eliminated. A multi-channel feature matrix is constructed by integrating original indexes and multiple derived features, the feature matrix is input into a residual convolutional neural network classification model integrated by multiple sub-models, and a tumor risk probability value of the subject is output. The application fully mines sequencing features of the centromere region, combines a multi-channel feature matrix and an integrated learning model, and improves prediction accuracy and robustness.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence tumor screening technology, and more specifically to a tumor screening method and system based on centromere region characteristics. Background Technology

[0002] Currently, non-invasive tumor screening technology based on peripheral blood cfDNA high-throughput sequencing has been widely developed. Existing technologies mainly rely on two pathways for tumor screening: targeted gene panel sequencing and whole-genome routine fragment feature analysis. By extracting sequencing information from routine regions of the genome and combining it with machine learning models to achieve sample differentiation screening, it has become an important research direction for early non-invasive tumor detection.

[0003] However, existing mainstream cancer screening programs still have objective technical shortcomings: targeted sequencing relies on fixed known gene loci, resulting in a narrow scope of genomic information utilization; conventional whole-genome analysis only utilizes information from ordinary regional fragments, resulting in low overall information utilization; at the same time, traditional screening programs have high sequencing costs, making it difficult to adapt to large-scale population screening, and there is still room for improvement in overall screening performance.

[0004] Therefore, how to propose a tumor screening method and system based on centromere region characteristics, while making full use of effective genomic information, to reduce screening costs and improve overall screening performance, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a tumor screening method and system based on centromere region features. By relying on centromere region-specific sequencing features combined with multi-dimensional feature fusion and integrated modeling, the accuracy and robustness of tumor screening are improved, and the screening cost is reduced.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: On one hand, this invention discloses a tumor screening method based on centromere region features, comprising the following steps: S1. Obtain sequencing reads from peripheral blood cell-free DNA samples of the subjects; S2. Align the sequencing read data to the reference genome and extract the alignment results of the centromere region in the reference genome; S3. Based on the alignment results, statistical analysis of sequencing characterization indicators for each point in the centromere region; S4. Screening of each point in the centromere region according to the sequencing characterization indicators; S5. Construct a multi-channel feature matrix from the sequencing characterization indicators corresponding to the retained sites after screening; S6. Input the multi-channel feature matrix into the pre-trained classification model to obtain and output the tumor risk probability value of the subject.

[0007] Preferably, S2 includes: The sequencing read data is aligned to a reference genome to pinpoint the physical matching location of each sequencing read on the genome. Based on the preset centromere region coordinates, sequencing reads whose matching positions fall within the centromere region are selected as alignment results.

[0008] Preferably, the sequencing characterization metrics include coverage depth, single nucleotide polymorphism count, and soft splicing count.

[0009] Preferably, S4 includes: For each site, paired statistical tests are performed based on the original value of the soft splice count and the soft splice ratio value normalized by the coverage depth to obtain the p-value of the original value test and the p-value of the normalized ratio test. Sites whose original value test p-value is less than a preset threshold and whose normalized ratio test p-value is greater than or equal to the preset threshold are removed, while sites whose normalized ratio test p-value is less than the preset threshold are retained.

[0010] Preferably, when constructing the multi-channel feature matrix in S5, the sequencing characterization index and the derived features of the sequencing characterization index are used together as multiple channels of the multi-channel feature matrix. The derived features include the ratio of single nucleotide polymorphism count to coverage depth, the ratio of soft splice count to coverage depth, the logarithmic transformation value of coverage depth, the logarithmic transformation value of single nucleotide polymorphism count, and the logarithmic transformation value of soft splice count.

[0011] Preferably, the classification model adopts an ensemble learning strategy, consisting of multiple sub-models trained based on different initialization parameters, and the final output tumor risk probability value is the average of the output probabilities of each sub-model.

[0012] Preferably, the sub-model is a binary classification model based on a residual convolutional neural network.

[0013] On the other hand, the present invention also discloses a tumor screening system based on centromere region features, comprising: The data preprocessing module is used to acquire sequencing reads from peripheral blood cell-free DNA samples of the subjects; The centromere alignment module is used to align the sequencing read data to a reference genome and extract the alignment results of the centromere regions in the reference genome. The feature extraction module is used to statistically analyze the sequencing characterization indicators of each point in the centromere region based on the alignment results. The site selection module is used to select each site in the centromere region according to the sequencing characterization indicators; A multi-channel matrix construction module is used to construct a multi-channel feature matrix from the sequencing characterization indicators corresponding to the retained sites after screening. The risk assessment output module is used to input the multi-channel feature matrix into a pre-trained classification model to obtain and output the tumor risk probability value of the subject.

[0014] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a tumor screening method and system based on centromere region characteristics, which has the following beneficial effects: 1. This is the first time that sequence features of the centromere, a highly complex region of the genome, have been applied to tumor screening, broadening the sources of genomic features available for liquid biopsy; 2. Paired statistical tests effectively distinguish between sequencing depth fluctuations and actual changes in soft splicing ratios, reducing false positive signals and improving the reliability of feature screening; 3. A multi-channel feature matrix containing original and derived features was constructed. Combined with a residual convolutional neural network ensemble model, experimental results show that the accuracy on the training set reached 91.94% and the accuracy on the independent test set reached 85.71%, demonstrating good generalization ability and clinical application potential. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 A flowchart of the method provided by the present invention; Figure 2 This is a flowchart of the Wilcoxon screening process of the present invention; Figure 3 Flowchart for constructing a multi-channel feature matrix; Figure 4 This is a diagram of the sub-model architecture. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The tumor risk probability values ​​output by this invention are intended to provide intermediate reference information for tumor screening and should not be used as independent clinical diagnostic criteria. The final diagnostic conclusion should be made by a physician based on the subject's complete clinical information.

[0019] The samples included in this example are from the Chinese Cancer Prostate Genome and Epigenome Atlas (CPGEA) project. This cohort comprises 208 patients with primary prostate cancer who underwent radical prostatectomy at Changhai Hospital, Second Military Medical University. Freshly frozen tumor tissue and paired histologically normal prostate tissue were collected from each patient. All 208 tumor-normal paired samples underwent whole-genome sequencing (WGS), with some samples further subjected to whole-genome bisulfite sequencing (187 pairs), RNA sequencing (134 tumors), and microRNA sequencing (105 tumors), generating a total of 1,268 sequencing datasets. This example uses a portion of the WGS data.

[0020] First, 200M reads were sequentially extracted from each paired-end FASTQ file pair of each sample as subsequent sequencing data to standardize the total sequencing volume for each sample and simulate a tumor screening scenario. Since the sequencing process is randomized, this operation does not cause read distribution shifts across the genome. 180 pairs were divided into a training dataset, and 28 pairs were divided into a test dataset.

[0021] On one hand, embodiments of the present invention disclose a tumor screening method based on centromere region features, comprising the following steps: S1. Obtain sequencing reads of peripheral blood cell-free DNA samples from the subjects.

[0022] The peripheral blood cell-free DNA sample was obtained as cell-free DNA from ex vivo plasma. Plasma was prepared using a double centrifugation method, followed by separation and purification via silica gel membrane centrifugation column method or magnetic bead method. The specific procedures are as follows: Plasma preparation: Collected whole blood is centrifuged at 1,600-2,000×g for 10 minutes at 4°C within 4-8 hours after collection using EDTA or Streck BCT anticoagulation blood collection tubes. After transferring the supernatant, it is centrifuged at 16,000-20,000×g for 10 minutes to remove residual blood cells and cell debris.

[0023] cfDNA extraction: Take 1-2 mL of plasma processed by two-step centrifugation and extract cfDNA using the Qiagen QIAamp Circulating Nucleic Acid Kit based on the principle of silica membrane adsorption.

[0024] During the procedure, proteinase K and lysis buffer containing vector RNA provided in the kit were first added to the plasma. After mixing, the mixture was incubated at 56°C to degrade proteins and release nucleic acids. Subsequently, binding buffer was added to adjust the solution environment, and all liquid was transferred to a centrifuge column. cfDNA selectively adsorbed onto the inner membrane of the column under high-salt conditions. After two centrifugations with washing buffer to remove impurities, the purified cfDNA was recovered with 30–50 μL of elution buffer. After extraction, the cfDNA concentration was determined using the Qubit fluorescent dye method, which is specifically designed for high-sensitivity quantification of double-stranded DNA. The fragment length distribution was assessed using an Agilent 2100 bioanalyzer based on microfluidic chip electrophoresis. The main peak of healthy donor cfDNA was typically located at 166–170 bp.

[0025] The purified cfDNA should be aliquoted and stored at -20°C or -80°C, avoiding repeated freeze-thaw cycles to prevent degradation.

[0026] Purified cfDNA was used to prepare sequencing libraries using commercially available cfDNA-specific library construction kits (such as the NEXTFLEX CellFree DNA-Seq Kit 2.0 or the Twist cfDNA Library Preparation Kit). These kits typically include core steps such as end repair and A addition, adapter ligation, and high-fidelity PCR amplification. The starting DNA input can be as low as 5–10 ng, and they are specifically optimized for highly fragmented nucleic acids such as cfDNA, enabling efficient capture of short nucleic acid fragments of approximately 166–170 bp in length.

[0027] After library construction, the library concentration was determined using a fluorescent dye-based nucleic acid quantification instrument, and the fragment size distribution was evaluated using a microfluidic chip electrophoresis system to ensure the library quality met the requirements for sequencing. After passing quality control, the library was placed on a next-generation sequencing platform (such as the Illumina NovaSeq 6000 series) for low-depth whole-genome sequencing in paired-end sequencing mode. The optimal sequencing depth was approximately 10× (i.e., each genomic locus was read approximately 10 times on average) to obtain full-genome reads at a relatively low sequencing cost. Raw FASTQ format reads were obtained after sequencing.

[0028] S2. Align the sequencing reads to the reference genome and extract the alignment results from the centromere regions of the reference genome, including: Sequencing read data is aligned to a reference genome to pinpoint the physical matching location of each sequencing read on the genome.

[0029] Specifically, the obtained sequencing read data were aligned to the reference genome using Nvidia Parabricks fq2bam alignment software. The container version was NVIDIA_parabricks_4.6.0.sif, and the reference genome was T2T-YAO1.1, resulting in a BAM file of the whole genome alignment results.

[0030] Based on the preset centromere region coordinates, sequencing reads whose matching positions fall within the centromere region are selected as alignment results.

[0031] The BAM files were indexed and constructed using the samtools software, and centromere regions were extracted. Precise centromere region localization was based on the results of the Cenmap software package on the T2T-YAO1.1 reference genome. The centromere region localization BED files cover 23 chromosomal regions (chr1–chr22 and chrX), with a total length of 160,616,524 bp. For alignment reads crossing boundaries, a preservation strategy was adopted, ultimately yielding BAM alignment files containing only the centromere regions of the 23 chromosomes.

[0032] S3. Based on the alignment results, the sequencing characterization indicators of each point in the centromere region are statistically analyzed.

[0033] From the obtained centromere region alignment results BAM file, a custom Python script was used to extract preset features. The data filtering rules were as follows: unmapped, duplicate, qcfail, and secondary reads were removed; no additional alignment quality (MAPQ) or base quality (BaseQ) thresholds were applied. The following three sequencing characterization metrics were statistically analyzed: Coverage depth: The number of reads covering this site.

[0034] Single nucleotide polymorphism count (SNP_Count): The number of reads at this site where the bases do not match the reference bases and neither is N.

[0035] Softclip_Count: Only counts the reference coordinates corresponding to the CIGAR opcode 4 (soft clipping) at the beginning or end of the read segment.

[0036] Ultimately, 23 Parquet files were obtained, each corresponding to a centromere region of one of the 23 chromosomes. The statistical results included the following fields: chromosome name (Chrom), reference genomic location (Pos, 1-based), reference base (Ref), coverage depth (Depth), single nucleotide polymorphism count (SNP_Count), and softclip count (Softclip_Count).

[0037] S4. Screening of each site in the centromere region based on sequencing characterization indicators.

[0038] This embodiment uses a nonparametric Wilcoxon signed-rank test under a paired design to screen centromere sites with tumor-normal differentiation capabilities, such as... Figure 2 As shown, the specific operation is as follows: Using the statistical results of a sufficient number of tumor-normal paired samples extracted in step S3 (208 pairs in this embodiment), the Depth and Softclip_Count for each locus were extracted. First, loci with a Depth greater than 2 for both normal and tumor samples were selected in 90% (187 pairs in this embodiment) or higher of the sample pairs. Next, a paired Wilcoxon test was performed on the raw count (Softclip_Count) and its depth-normalized proportion (Softclip_Ratio = Softclip_Count / Depth for each locus), yielding two p-values ​​for each locus. Using p < 0.05 as the significance criterion, the sources of locus differences were categorized into four types: those with significant raw counts but insignificant normalization were attributed to depth interference, denoted as Depth; those with significant normalized proportions but insignificant raw counts were attributed to changes in the true soft splicing proportion, denoted as Clip; those with both significant were denoted as Both; and those with neither significant were denoted as None. This classification strategy aims to eliminate false positive signals introduced by depth fluctuations. Ultimately, 95,132 significant sites attributable to changes in the soft splicing ratio (i.e., Clip and Both classes) were identified from over 160 million original sites. To further reduce dimensionality, the top 3,000 sites with the smallest p-value of the Wilcoxon test for Softclip_Ratio were selected for subsequent training.

[0039] S5. Construct a multi-channel feature matrix from the sequencing characterization indicators corresponding to the retained sites after screening.

[0040] In the multi-channel feature matrix, the original three basic features, Depth, SNP_Count, and Softclip_Count, are expanded into an eight-channel representation. Two additional ratio features (specifically SNP_Count / Depth and Softclip_Count / Depth) and log1p logarithmic transformations of each original feature (specifically ln(1+Depth), ln(1+SNP_Count), and ln(1+Softclip_Count)) are constructed, resulting in a total of 3+2+3=8 channels for enhanced representation, thereby strengthening the ability to represent nonlinear patterns.

[0041] The dataset was divided into training and validation subsets in an 8:2 ratio, maintaining a balanced proportion of positive and negative samples during the partitioning process. Then, Z-score standardization was applied to each of the 8 channels based on the training subset statistics (specifically, the mean and standard deviation). (Refer to...) Figure 3 .

[0042] S6. Input the multi-channel feature matrix into the pre-trained classification model to obtain and output the tumor risk probability value of the subject.

[0043] The classification model employs an ensemble learning strategy, consisting of multiple sub-models trained with different initialization parameters. The final output tumor risk probability is the average of the probabilities output by each sub-model.

[0044] like Figure 4 As shown, the sub-model employs a one-dimensional convolutional structure containing two levels of residual blocks. Each residual block consists of a double convolutional layer, batch normalization, and a GELU activation function. Shortening branches align the dimensions via identity mapping or 1×1 convolution. Each residual block is followed by max pooling and Dropout regularization with a probability of 0.15. After adaptive average pooling compression, the binary classification logit value is output by a fully connected layer. Training uses binary cross-entropy loss and the AdamW optimizer, with an initial learning rate of 1×10⁻⁶. -3 Weight decay 1×10 -5 The batch size is 32. If the validation set AUC does not improve for 10 consecutive rounds, an early stop is triggered and the optimal parameters are rolled back. During the prediction phase, the validation set probabilities are iterated within the range of 0.01 to 0.99, and an adaptive threshold that maximizes accuracy is selected.

[0045] To improve the robustness of the model, an ensemble strategy is adopted: five models are trained in parallel using five different random number seeds (43, 123, 425, 777, 2026), and the outputs of these five models are averaged to obtain the final output. For the test sample, its centromere region features are processed sequentially through steps S1-S5 to obtain a multi-channel feature matrix, which is then input into the ensemble model to obtain and output the average tumor risk probability value. When the output probability value is greater than or equal to the adaptive threshold (0.4900 in this embodiment), the subject is determined to be at high risk of tumor; when the output probability value is less than the preset threshold, the subject is determined to be normal. It should be noted that this determination result is only intermediate reference information and should not be used as an independent basis for clinical diagnosis.

[0046] Experimental results show that the accuracy on the training set reached 91.94%, and the accuracy on the independent test set reached 85.71%, with only one false negative reported according to the confusion matrix. The adaptive threshold of 0.4900 on the validation set further balanced sensitivity and specificity, indicating that this embodiment possesses good generalization ability and clinical application potential in tumor detection tasks.

[0047] On the other hand, the present invention also discloses a tumor screening system based on centromere region features, comprising: The data preprocessing module is used to acquire sequencing reads from peripheral blood cell-free DNA samples of the subjects; The centromere alignment module is used to align sequencing read data to the reference genome and extract the alignment results of centromere regions in the reference genome; The feature extraction module is used to statistically analyze the sequencing characterization indicators of each point in the centromere region based on the alignment results. The site selection module is used to select sites in the centromere region based on sequencing characterization indicators; The multi-channel matrix construction module is used to construct a multi-channel feature matrix from the sequencing characterization indicators corresponding to the selected and retained sites. The risk assessment output module is used to input the multi-channel feature matrix into a pre-trained classification model to obtain and output the tumor risk probability value of the subject.

[0048] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0049] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A tumor screening method based on centromere region features, characterized in that, Includes the following steps: S1. Obtain sequencing reads from peripheral blood cell-free DNA samples of the subjects; S2. Align the sequencing read data to the reference genome and extract the alignment results of the centromere region in the reference genome; S3. Based on the alignment results, statistical analysis of sequencing characterization indicators for each point in the centromere region; S4. Screening of each point in the centromere region according to the sequencing characterization indicators; S5. Construct a multi-channel feature matrix from the sequencing characterization indicators corresponding to the retained sites after screening; S6. Input the multi-channel feature matrix into the pre-trained classification model to obtain and output the tumor risk probability value of the subject.

2. The tumor screening method based on centromere region characteristics according to claim 1, characterized in that, S2 include: The sequencing read data is aligned to a reference genome to determine the physical matching position of each sequencing read on the genome. Based on the preset centromere region coordinates, sequencing reads whose matching positions fall within the centromere region are selected as alignment results.

3. The tumor screening method based on centromere region characteristics according to claim 1, characterized in that, The sequencing characterization metrics include coverage depth, single nucleotide polymorphism count, and soft splicing count.

4. The tumor screening method based on centromere region characteristics according to claim 3, characterized in that, S4 include: For each site, paired statistical tests are performed based on the original value of the soft splice count and the soft splice ratio value normalized by the coverage depth to obtain the p-value of the original value test and the p-value of the normalized ratio test. Sites whose original value test p-value is less than a preset threshold and whose normalized ratio test p-value is greater than or equal to the preset threshold are removed, while sites whose normalized ratio test p-value is less than the preset threshold are retained.

5. A tumor screening method based on centromere region features according to claim 3, characterized in that, When constructing the multi-channel feature matrix in S5, the sequencing characterization index and the derived features of the sequencing characterization index are used together as multiple channels of the multi-channel feature matrix. The derived features include the ratio of single nucleotide polymorphism count to coverage depth, the ratio of soft splice count to coverage depth, the logarithmic transformation value of coverage depth, the logarithmic transformation value of single nucleotide polymorphism count, and the logarithmic transformation value of soft splice count.

6. The tumor screening method based on centromere region features according to claim 1, characterized in that, The classification model employs an ensemble learning strategy, consisting of multiple sub-models trained based on different initialization parameters. The final output tumor risk probability is the average of the output probabilities of each sub-model.

7. A tumor screening method based on centromere region features according to claim 6, characterized in that, The sub-model is a binary classification model based on residual convolutional neural networks.

8. A tumor screening system based on centromere region features, characterized in that, include: The data preprocessing module is used to acquire sequencing reads from peripheral blood cell-free DNA samples of the subjects; The centromere alignment module is used to align the sequencing read data to a reference genome and extract the alignment results of the centromere regions in the reference genome. The feature extraction module is used to statistically analyze the sequencing characterization indicators of each point in the centromere region based on the alignment results. The site selection module is used to select each site in the centromere region according to the sequencing characterization indicators; A multi-channel matrix construction module is used to construct a multi-channel feature matrix from the sequencing characterization indicators corresponding to the retained sites after screening. The risk assessment output module is used to input the multi-channel feature matrix into a pre-trained classification model to obtain and output the tumor risk probability value of the subject.

Images