Computer program products and applications for data processing devices for ctDNA variant detection

Abstract

This invention discloses a computer program product and its application for data processing devices in the field of bioinformatics for ctDNA variant detection. The technical problem this invention aims to solve is how to detect ctDNA variants in early-stage cancer or postoperative minimal residual disease (MRD) under conditions of no UMI library construction and moderate sequencing depth. This invention constructs a set of supporting sequences for candidate variant sites, generating a sequence feature tensor and a fragment physical feature vector (including normalized fragment length). The former is input into a first neural network branch to extract sequence representation, and the latter into a second neural network branch to extract physical representation. A fusion module combines the sequence representation and physical representation, and a gating unit calculates the gating weight based on the physical representation and dynamically adjusts the contribution of the sequence representation, outputting the probability of the true ctDNA variant. This invention utilizes physical laws to suppress sequencing noise and can be applied to monitor MRD under conditions without molecular barcodes.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to computer program products and applications for data processing devices used for ctDNA variant detection. Background Technology

[0002] With the deepening development of precision medicine, liquid biopsy has become an important tool for early cancer screening, companion diagnostics, and postoperative recurrence monitoring. Circulating tumor DNA (ctDNA), as a characteristic biomarker released from tumor cells into the bloodstream, carries mutational information about the tumor genome. In patients with advanced cancer, ctDNA abundance is high, and conventional next-generation sequencing (NGS) technology can meet the detection requirements. However, in scenarios such as early cancer screening or postoperative minimal residual disease (MRD) monitoring, the content of ctDNA in the plasma cell-free DNA (cfDNA) background is extremely low, and its variant allele frequency (MAF) is often less than 0.1% or even 0.01%.

[0003] Within this critical detection range, traditional bioinformatics analysis methods face severe physical and statistical challenges. Conventional variant detection algorithms (such as Mutect2 or VarScan) are primarily based on Bayesian inference or binomial / Poisson distribution models, relying on the number of support reads and base quality at the variant site. However, NGS sequencing inherently contains systematic noise, including polymerase slip errors introduced by PCR amplification, optical signal crosstalk from the sequencer, and phasing / pre-phasing effects. These systematic errors typically occur between 0.1% and 0.5%, precisely covering the target signal range monitored by MRD. When the frequency of true variants is lower than the background noise level, relying solely on sequence statistical characteristics is insufficient to effectively distinguish true mutations from sequencing artifacts, leading to a sharp decline in detection sensitivity or uncontrollable false positive rates.

[0004] To address this signal-to-noise ratio issue, the industry has developed molecular barcoding (UMI) technology, which uses physical tags to remove duplicates and correct errors in original molecules. While UMI technology can push the detection limit down to 0.01%, its complex library preparation process and extremely high sequencing depth requirements (typically >30,000×) result in high testing costs, making it difficult to widely implement in inclusive clinical surveillance. Therefore, how to overcome the statistical limit of the Poisson distribution through algorithmic innovation under conditions of routine library preparation (without UMI) and moderate sequencing depth (e.g., 5000×) has become an urgent need in the industry.

[0005] In recent years, deep learning technology has been introduced into the field of genomic data analysis. Methods such as Google's DeepVariant transform sequence alignment information into images and use convolutional neural networks (CNNs) to identify error patterns in sequence texture. However, these methods are primarily designed for germline variants and focus mainly on sequence context, neglecting the unique physical and biological characteristics of ctDNA. In fact, ctDNA fragments originate from tumor cell apoptosis and are protected by nucleosome mechanisms; their fragment length (approximately 132-145 bp) is significantly shorter than cfDNA (approximately 167 bp) derived from normal leukocytes, and their end-cutting positions on the genome exhibit specific biases.

[0006] Several explorations have been made in the utilization of fragmentomics features in existing technologies. For example, patent document CN119360984A proposes a deep learning-based method for screening target tumor DNA fragments. This scheme uses model localization analysis to screen tumor DNA fragments, aiming to reduce interference in training samples. However, a deeper analysis reveals that this scheme is essentially a data preprocessing or "cleaning" strategy, focusing on removing interfering data before model training, rather than verifying the authenticity of single base sites during the variant reading stage. While this method improves the training efficiency of subsequent models, it does not construct a real-time fusion mechanism for sequence texture and physical features at the "single read level," and therefore cannot directly solve the micro-error correction problem in variant detection.

[0007] Another representative technology, such as Grail's EP4127231A1, proposes using genomic region modeling for cancer classification. This approach groups fragments into genomic regions and trains regional models to output cancer probability scores. However, the underlying logic of this approach is macro-level statistics at the "sample level" or "region level," aiming to determine whether a subject has cancer (classification), rather than whether a specific sequencing read carries a real mutation (variant calling). Mechanistically, this approach aggregates fragment features into counts or vectors, losing the one-to-one correspondence between "specific mutation sites" and "the physical properties of the specific fragment carrying that mutation." For example, when a low-frequency mutation happens to appear on a 300bp long fragment, based on biological common sense, this is highly likely to be a PCR error (because tumor fragments are usually very short), but the macro-statistical model of EP4127231A1 cannot use this micro-physical contradiction to eliminate this specific false positive.

[0008] In addition, the academic community has also seen the emergence of multimodal ensemble strategies, such as the paper "Deep learning-based multimodal ensemble algorithm...", which attempts to combine mutation density and fragment features. However, such methods usually employ "late fusion" or model ensemble strategies, that is, training two models separately and averaging the results. This loose coupling method has a logical flaw: when the sequence model gives a high-confidence positive prediction due to PCR errors, while the fragment model gives a negative prediction due to physical features, the simple averaging operation will lead to ambiguous results, failing to achieve the "one-vote veto" or "gating inhibition" of sequence errors by physical features.

[0009] In summary, existing technologies are either limited by the high cost of UMI (Unified Microarray Analysis) or only utilize macroscopic fragment statistical features for sample classification, lacking a mutation detection mechanism that can delve into the single-read dimension and deeply and organically integrate microscopic sequence texture with fragment physical properties. Without UMI assistance, how to use deep learning models to capture the subtle biophysical principle that "short fragments carry true mutations, while long fragments carry PCR errors," and transform this into quantifiable error correction capabilities, remains a key unresolved technical bottleneck. Summary of the Invention

[0010] The technical problem to be solved by this invention is how to detect trace amounts of ctDNA variants in early cancer screening or postoperative minimal residual disease (MRD) and / or how to detect ctDNA variants in early cancer screening or postoperative minimal residual disease (MRD) under conventional library preparation (without UMI) and medium sequencing depth (e.g., 5000×) conditions and / or how to construct a deep learning model that can simultaneously utilize sequence context information and fragment omics physical features, and through a deep fusion mechanism of feature layers, achieve precise suppression of sequencing noise using physical laws, thereby significantly improving detection sensitivity and specificity without increasing sequencing costs.

[0011] In view of this, the core technical problem to be solved by the present invention is how to construct a deep learning model that can simultaneously utilize sequence context information and fragment omics physical features to address the problems of low signal-to-noise ratio and high false positive rate in the detection of ultra-low frequency (MAF < 0.1%) variants without UMI sequencing data. Through the deep fusion mechanism of the feature layer, the model can accurately suppress sequencing noise by utilizing physical laws, thereby significantly improving detection sensitivity and specificity without increasing sequencing costs.

[0012] To address the aforementioned technical problems, the present invention provides a data processing apparatus, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor can execute the computer program to perform the following steps: 1) Construct a set of supporting sequences containing supporting sequences, wherein the supporting sequences are derived from sequencing reads containing variant sites in the sequencing data of plasma ctDNA of the test subjects, and the sequencing reads do not contain molecular barcodes; 2) Construct a sequence feature tensor and a fragment physical feature vector for each of the supported sequences based on the set of supported sequences; the sequence feature tensor includes the number of samples (N), the sampling window length (L), and the number of feature channels (C_seq) of the supported sequences; the fragment physical feature vector includes the normalized fragment length, the normalized end distance, and the end motif matching degree; the feature channels include the base category (channel), the base quality (channel), and the chain orientation channel; 3) Input the sequence feature tensor into the first neural network branch to extract the sequence representation vector, and input the fragment physical feature vector into the second neural network branch to extract the physical representation vector; 4) Fuse the sequence representation vector and the physical representation vector to output the authenticity probability of the mutation site.

[0013] The dimension of the sequence feature tensor mentioned above can be N×L×C_seq.

[0014] The mutation sites may be derived from the results of comparing the sequencing data with the reference genome data of the corresponding species of the test subject and detecting the mutations.

[0015] The molecular barcode refers to a short sequence (usually around 10 nt) tag added to each original DNA or cDNA molecule during the NGS library construction process to uniquely identify the molecule.

[0016] The number of sampled sequences N pairs can range from 32 to 256, and in one specific embodiment of the invention, it is 64. The sampling window length L can range from 15 bp to 51 bp, and in one specific embodiment of the invention, it is 33 or 51. The fragment physical feature vector can include at least one of the following features: normalized fragment length, (normalized) end distance, end motif matching degree, and alignment quality. The normalized fragment length can be calculated by dividing the actual fragment length by a preset base, where the preset base is a value selected between 200 bp and 400 bp. The (normalized) end distance can be calculated by calculating the number of nucleotides from the mutation site in the Read to the nearest end of the Read, dividing by the preset base to obtain the (normalized) end distance, which is 150 in one specific embodiment of the invention. The method for calculating the terminal motif matching degree can be as follows: extract the last 4 bp sequence of the sequencing read, encode it as a numerical feature, and then calculate the cosine similarity between the numerical feature and the preset ctDNA dominant terminal motif reference vector. The result is used as the terminal motif matching degree. The method for obtaining the preset ctDNA dominant terminal motif reference vector is as follows: for sequencing reads confirmed to be of ctDNA origin (obtained based on analysis of publicly available cfDNA sequencing data from healthy controls and known cancer patients), the frequency-weighted average of the one-hot encoded vectors of the last 4 bp motif is calculated to form a 16-dimensional reference vector, which is the preset ctDNA dominant terminal motif reference vector.

[0017] The data processing device described above can be used to detect ctDNA variations, and the data processing device may further include the following steps: The authenticity probability is used to determine whether the mutation site is a true ctDNA mutation of the subject being tested.

[0018] In the data processing device described above, the number of sequencing reads supported by the variant site in the sequencing data can be greater than or equal to 2.

[0019] The number of sequencing reads supported is the number of sequencing reads containing the aforementioned variant sites.

[0020] In the data processing device described above, the fusion described in 4) may include a gating unit; the gating unit may be configured to calculate gating weights using an activation function based on the physical representation vector, and dynamically adjust the contribution of the sequence representation vector to the authenticity probability based on the gating weights.

[0021] The fusion described above can be feature fusion, which can be concatenation or summation. In one specific embodiment of the present invention, the fusion can be achieved by using a fully connected layer to combine the weighted sequence representation vector and the physical representation vector. The fusion can also be achieved through other feature fusion methods. The implementation of the fusion described above can be referred to the calculation method of Equation 1 in the embodiments.

[0022] In one specific embodiment of the present invention, the threshold for determining the ctDNA mutation mentioned in 5) is a probability of authenticity greater than or equal to 0.95.

[0023] In the data processing device described above, the first neural network branch in 3) can adopt a convolutional neural network structure; the second neural network branch can adopt a multilayer perceptron structure.

[0024] In one specific embodiment of the present invention, the convolutional neural network structure may be a ResNet-18 network model or / and an Inception network module.

[0025] To address the aforementioned technical problems, the present invention also provides an apparatus for detecting ctDNA variants, the apparatus comprising the following modules: B1) Used to construct a set of supporting sequences containing supporting sequences, wherein the supporting sequences are derived from sequencing reads containing variant sites in the sequencing data of plasma ctDNA of the subject to be tested, and the sequencing reads do not contain molecular barcodes; B2) is used to construct a sequence feature tensor and a fragment physical feature vector for each of the supported sequences based on the set of supported sequences; the sequence feature tensor includes the number of samples, the sampling window length, and the number of feature channels of the supported sequences; the fragment physical feature vector includes the normalized fragment length, the normalized end distance, and the end motif matching degree; the feature channels include base type (channel), base quality (channel), and chain orientation channel; B3) is used to input the sequence feature tensor into the first neural network branch to extract the sequence representation vector, and to input the fragment physical feature vector into the second neural network branch to extract the physical representation vector; B4) is used to fuse the sequence representation vector and the physical representation vector to output the authenticity probability of the mutation site; B5) is used to determine whether the mutation site is a true ctDNA mutation of the subject under test based on the true probability.

[0026] The variant site supports a number of sequencing reads greater than or equal to 2 in the sequencing data.

[0027] The dimension of the sequence feature tensor can be N×L×C_seq.

[0028] The mutation sites may be derived from the results of comparing the sequencing data with the reference genome data of the corresponding species of the test subject and detecting the mutations.

[0029] The molecular barcode refers to a short sequence (usually around 10 nt) tag added to each original DNA or cDNA molecule during the NGS library construction process to uniquely identify the molecule.

[0030] The number of sampled sequences N pairs can range from 32 to 256, and in one specific embodiment of the invention, it is 64. The sampling window length L can range from 15 bp to 51 bp, and in one specific embodiment of the invention, it is 33 or 51. The fragment physical feature vector can include at least one of the following features: normalized fragment length, (normalized) end distance, end motif matching degree, and alignment quality. The normalized fragment length can be calculated by dividing the actual fragment length by a preset base (300 bp in one specific embodiment of the invention), where the preset base is selected from a value between 200 bp and 400 bp. The (normalized) end distance can be calculated by calculating the number of nucleotides from the mutation site in the Read to the nearest end of the Read, dividing it by the preset base to obtain the (normalized) end distance, which is 150 in one specific embodiment of the invention. The end motif matching degree can be calculated by extracting the last 4 bp sequence of the sequencing fragment, encoding it as a numerical feature, and using cosine similarity to calculate the end motif matching degree based on the numerical feature.

[0031] In the above apparatus, the fusion described in B4) may include a gating unit; the gating unit may be configured to calculate gating weights using an activation function based on the physical representation vector, and dynamically adjust the contribution of the sequence representation vector to the authenticity probability based on the gating weights.

[0032] The fusion described above can be feature fusion, which can be concatenation or summation. In a specific embodiment of the present invention, the fusion is achieved by using a fully connected layer to combine the weighted sequence representation vector and the physical representation vector. The fusion can also be achieved through other feature fusion methods. The implementation of the fusion described above can be referred to the calculation method of Equation 1 in the embodiment.

[0033] In one specific embodiment of the present invention, the threshold for determining the ctDNA mutation described in B5) is a probability of authenticity greater than or equal to 0.95.

[0034] In B3), the first neural network branch can adopt a convolutional neural network structure. The second neural network branch can adopt a multilayer perceptron structure.

[0035] In one specific embodiment of the present invention, the convolutional neural network structure is a ResNet-18 network model and / or an Inception network module.

[0036] To address the aforementioned technical problems, the present invention also provides a computer program product, comprising a computer program that, when executed by a processor, implements steps 1)-4) or steps 1)-5) described above.

[0037] To address the aforementioned technical problems, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, can implement steps 1)-4) or any of steps 1)-5) described above.

[0038] To address the aforementioned technical problems, the present invention also provides the application of a substance in the preparation of products for early cancer screening or postoperative minimal residual disease (MRD) screening, wherein the substance may be any of the following: 1) The data processing device described above, 2) The device described above, 3) The computer program products mentioned above, 4) The computer-readable storage medium described above.

[0039] This invention provides a method for detecting ctDNA variants based on multimodal deep learning. The method first constructs a set of supporting sequences for candidate variant sites. These supporting sequences can be derived from sequencing data that does not contain molecular barcodes, or from molecular barcode data analysis results. To ensure detection sensitivity, the screening criterion for candidate variant sites is set to a number of supporting reads greater than or equal to 2.

[0040] Based on the constructed support sequence set, this method further constructs a sequence feature tensor and a fragment physical feature vector. The sequence feature tensor has a dimension of N×L×C_seq, with specific parameters set as follows: the number of support sequence samples N is 32 to 256, the sampling window length L is 15bp to 51bp, and the number of feature channels C_seq includes base sequence information, base quality information, and chain orientation information. The fragment physical feature vector includes at least one feature among normalized fragment length, end-to-end distance, end-to-end motif matching degree, and alignment quality. To accurately reflect the fragment omics characteristics, the normalized fragment length is calculated by dividing the actual fragment length by a preset base, which is selected from a value between 200bp and 400bp.

[0041] In the feature extraction and fusion stage, this method inputs the sequence feature tensor into the first neural network branch to extract sequence representations, and inputs the fragment physical feature vectors into the second neural network branch to extract physical representations. The first neural network branch adopts a convolutional neural network structure, specifically a ResNet-18 network model or an Inception network module; the second neural network branch adopts a multilayer perceptron structure. Subsequently, the fusion module combines the sequence representations and physical representations to output the real probability of candidate mutation sites. This fusion module includes a gating unit configured to calculate gating weights based on the physical representations and dynamically adjust the contribution of the sequence representations to the real probability using these weights. Specifically, the gating weights are generated using the Sigmoid activation function (within the range of 0~1), and the fusion module generates fused features by concatenating or summing the weighted sequence representations with the physical representations. In the model training stage, Focal Loss is used as the loss function to optimize model performance.

[0042] Furthermore, this invention also provides a ctDNA variant detection device based on multimodal deep learning. This device, as the main structure, integrates: a data construction module configured to construct a set of supporting sequences and corresponding sequence feature tensors and fragment physical feature vectors; a feature extraction module configured to input the sequence feature tensors into a first neural network branch to extract sequence representations and input the fragment physical feature vectors into a second neural network branch to extract physical representations; and a fusion detection module configured to combine sequence representations and physical representations to output the authenticity probability of candidate variant sites.

[0043] Finally, the present invention also provides a computer-readable storage medium and an electronic device. The storage medium stores a computer program, which, when executed by a processor, implements the above-described method; the electronic device includes a processor and a memory, which, when executed by the processor, implements the above-described method.

[0044] The computer program products mentioned above can be software products that primarily implement their solutions through computer programs.

[0045] The computer-readable storage medium refers to a carrier for storing data, which may be magnetic tape, disk, floppy disk, optical disk, magneto-optical disk, ROM, PROM, VCD, DVD, hard disk, flash memory, USB flash drive, CF card, SD card, MMC card, SM card, Memory Stick, or xD card, etc.

[0046] The applications or methods described above are not for disease diagnosis. They are not intended to directly obtain disease diagnoses or health status results from living humans or animals.

[0047] The above applications or methods are for non-disease treatment purposes. They are not intended to restore or restore health or reduce suffering in living human or animal bodies.

[0048] The beneficial effects of this invention are: The method for detecting ultra-micro ctDNA variants based on multimodal deep learning and fragment omics feature fusion provided by this invention has the following significant advantages compared to existing technologies: 1) Breaking the low-frequency detection limit without UMI: By introducing fragment omics physical features as "soft UMI," this invention can effectively identify and eliminate PCR amplification errors and sequencing noise on long fragments. Experimental data show that at a conventional sequencing depth of 5000×, the detection sensitivity of this method for ultra-low-frequency variants with 0.05% MAF can reach over 68.5%, far exceeding the traditional Mutect2 method (approximately 18%) and pure sequence deep learning methods, approaching the theoretical detection limit of the Poisson distribution.

[0049] 2) A deep error correction mechanism at the feature level is implemented: Unlike the common model ensemble or late-stage averaging strategies in existing technologies, the gated fusion of this invention occurs during the feature extraction stage. Weights are directly generated from physical features to suppress the activation of sequence features, achieving a "one-vote veto" error correction and avoiding false positives caused by high-confidence sequence errors in later fusion raising the overall average score. This mechanism effectively solves the contradictory situation of "serial sequence variations with questionable physical origins," significantly reducing the false positive rate, achieving a specificity better than 99.99%, and greatly improving the clinical reliability of the results.

[0050] 3) Reduced sequencing costs and lower barriers to clinical application: The algorithm significantly improves the signal-to-noise ratio, enabling detection accuracy at lower sequencing depths (e.g., 3000×-5000×) that previously required ultra-high depths (>30,000×) or UMI library preparation. This directly reduces sequencing costs by approximately 30%-50% and simplifies experimental procedures, providing economically feasible technical support for the widespread adoption of ctDNA liquid biopsy in large-scale clinical scenarios such as postoperative MRD monitoring.

[0051] 4) Enhanced model generalization ability: By incorporating physical features (such as fragment length distribution), the model reduces overfitting to specific sequencer error patterns (such as systematic errors of specific motifs). Because physical features reflect the biological essence (the degradation patterns of tumor DNA) rather than the engineering defects of the sequencer, this method exhibits better robustness and transferability across different sequencing platforms and library preparation kits. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the core logic judgment and processing flow of the mutation detection method described in Embodiment 1 of the present invention.

[0053] Figure 2 This is a graph comparing the detection sensitivity and specificity of Example 1 of the present invention with existing technologies at different mutation frequencies. The horizontal axis represents MAF (logarithmic scale), and the vertical axis represents the detection index. Curve A represents Example 2, curve B represents Comparative Example 1, and curve C represents Comparative Example 2. A shows the trend of the detection sensitivity performance of Example 2 (after adjusting the process parameters of the method in Example 1) compared with Comparative Example 1 (deleted fragment physical characteristics) and Comparative Example 2 (existing statistical method Mutect2) at different mutation allele frequencies (MAF); B shows the trend of the specificity performance of Example 2 (after adjusting the process parameters of the method in Example 1) compared with Comparative Example 1 (deleted fragment physical characteristics) and Comparative Example 2 (existing statistical method Mutect2) at different mutation allele frequencies (MAF). Detailed Implementation

[0054] 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.

[0055] 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.

[0056] To better understand the technical solution of this invention, the key technical parameters involved in the implementation of this invention and their preferred ranges are listed below. These parameter ranges are set based on extensive experimental verification and aim to balance detection sensitivity, specificity, and computational efficiency.

[0057] Example 1. Flowchart of the ctDNA variant detection method based on multimodal deep learning of the present invention.

[0058] This embodiment provides a ctDNA variant detection method based on multimodal deep learning, which can be applied to the MRD monitoring scenario after cancer surgery.

[0059] 1. Data preparation and preprocessing.

[0060] 1.1 Obtain cfDNA data.

[0061] The data used in the model construction of this embodiment comes from plasma cfDNA sequencing projects in publicly available databases. Specifically, it is the dbGaP project phs003170, which contains whole-genome sequencing data of plasma cfDNA from healthy individuals and cancer patients, used for fragment length, fragment end sequences, and related fragment omics characterization analysis.

[0062] After downloading the raw cfDNA sequencing data from the public database, the following steps are performed in sequence: quality control of the raw sequencing data to obtain quality-controlled sequencing data, alignment of the quality-controlled sequencing data to the reference genome, screening of (candidate) variant sites, read extraction, and encoding of fragment physical features and sequence features.

[0063] The validation and test sets consist of plasma cfDNA sequencing data from cancer patients and plasma cfDNA sequencing data from healthy controls in publicly available databases.

[0064] 1.2 Obtaining candidate variant sites.

[0065] The raw sequencing data obtained in step 1.1 was quality-controlled and the Reads were aligned to the human reference genome (hg19) using BWA software to obtain the alignment file.

[0066] Preliminary variant screening: The samtools mpileup software is used to traverse the alignment file (BAM) and screen for variant sites containing non-reference genomic bases. If a variant site has ≥2 supporting reads, it is marked as a (candidate) variant site.

[0067] 2. Feature construction.

[0068] For each candidate variant site, extract all supporting reads covering that site and construct a set of supporting sequences.

[0069] 2.1 Construction of Sequence Feature Tensor

[0070] The sequence feature tensor has a dimension of N×L×C_seq, and the specific parameters are set as follows: the number of samples N supported is 32 to 256, the sampling window length L is 15bp to 51bp, and the feature channels C_seq include (base type, base quality, and chain orientation channels).

[0071] Table 1. Key technical parameters of the process.

[0072]

[0073] The specific parameter settings in this embodiment are as follows: In this embodiment, the number of supported sequence samples N is set to 64. If the actual number of supported reads is less than 64, it is padded with zeros; if it exceeds 64, the first 64 reads are truncated according to the alignment quality (Table 1).

[0074] For each sequence in the support sequence set, with the mutation site as the center point, extend 16bp to the left and right (upstream and downstream of the center point) to extract a sequence window with a total length L=33bp. In this embodiment, a total of 64 sequence windows of 33bp length are obtained from the 64 support sequences.

[0075] Construct a tensor with dimensions of 64×33×6. The six feature channels are as follows: (1) Four base channels: A, T, C, and G (One-hot encoding); (2) Base quality channel (Phred score normalized to 0-1); (3) Chain direction channel (justification chain is 1, antithesis chain is -1).

[0076] 2.2 Construction of fragment physical feature vectors.

[0077] For each Read in the support sequence set, its physical features (including at least one of the following: normalized fragment length, end distance, end motif matching degree, and alignment quality) are extracted and statistical values ​​are calculated to form a physical feature vector. This embodiment selects the following features: (1) Normalized fragment length: Extract the Insert Size of the fragment to which the Read belongs (i.e., the length of the DNA insertion fragment corresponding to a pair of Reads on the reference genome in paired-end sequencing), and divide it by the base number 300bp. For example, if the fragment length is 140bp, the normalized value is 140 / 300≈0.46.

[0078] (2) Normalized end distance: Calculate the number of bases (nucleotides) that the variant site in the Read is closest to the end of the Read, and divide it by 150. For example, if the number of bases that the variant site is closest to the end of the Read is 10 bp, the normalized value is 10 / 150 ≈ 0.067.

[0079] (3) End motif matching degree: Extract the 4bp sequence from the end of the fragment (the end of the fragment closest to the candidate variant site in the DNA insertion fragment to which the Read belongs), and use one-hot encoding as a numerical feature (encode each base in the 4bp sequence at the end as a 4-dimensional vector according to A, T, C, and G respectively, and then represent the 4bp end motif as a 16-dimensional numerical feature vector); calculate the cosine similarity with the preset ctDNA dominant end motif reference vector, and the result is used as the end motif matching degree. The preset ctDNA dominant end motif reference vector is obtained as follows: For the sequencing reads (in the training set) that have been confirmed to be from ctDNA (based on the analysis of publicly available cfDNA sequencing data from healthy controls and known cancer patients), the frequency weighted average of the one-hot encoded vectors of the 4bp motif at the end is calculated to form a 16-dimensional reference vector, which is the preset ctDNA dominant end motif reference vector. The specific steps are: 1) Data source: Use publicly available cfDNA sequencing data from healthy controls and known cancer patients (such as the dbGaP project phs003170) to calculate the frequency distribution of the 4bp motif at the end of tumor-derived cfDNA fragments. 2) Statistical Method: For fragments confirmed to originate from ctDNA (marked by known tumor mutation sites), the frequency distribution of their terminal 4 bp sequences is statistically analyzed. High-frequency terminal motif patterns are selected and encoded using one-hot encoding to form a 16-dimensional reference vector. 3) Specific Values: There are 4^4 = 256 possible combinations of the 4 bp terminal motif. Each combination is encoded into a 16-dimensional vector using one-hot encoding. The preset ctDNA dominant terminal motif reference vector is essentially a weighted frequency vector of these combinations, reflecting the terminal motif distribution characteristics of ctDNA fragments under nucleosome cleavage preference.

[0080] The above features are statistically analyzed (mean and variance) on N (number of support sequence samples) sequences and concatenated into a one-dimensional vector.

[0081] 3. Model architecture and reasoning.

[0082] 3.1 Construct a two-stream neural network model.

[0083] (1) First branch (sequence stream): A high-dimensional sequence representation vector is obtained using a convolutional neural network structure. In this embodiment, a ResNet-18 residual network structure is used. The 64×33×6 sequence feature tensor obtained in step 2.1 is input, and after processing through convolutional layers and residual blocks, a high-dimensional sequence representation vector V_seq is extracted. The ResNet-18 structure specifically includes: an input layer (inputting a 64×33×6 sequence feature tensor), an initial convolutional layer (7×7 convolutional kernel, stride 2), a max pooling layer (maxpool, 3×3 pooling kernel, stride 2), residual block group (conv2–conv5) processing (containing several basic residual blocks, in this embodiment there are 4 basic residual blocks, each basic residual block contains two 3×3 convolutional layers), global average pooling, and a fully connected layer (fc).

[0084] (2) Second branch (physical flow): A multilayer perceptron containing 3 fully connected layers (MLP) is used to extract the physical representation vector V_phy from the fragment physical feature vector obtained in step 2.2.

[0085] 3.2 Gating fusion.

[0086] The fusion module in this step includes a gating unit. This gating unit is configured to calculate gating weights based on the physical representation and dynamically adjust the contribution of the sequence representation vector to the authenticity probability based on these gating weights, specifically: The physical representation vector V_phy obtained from the second branch in step 3.1 is input into the Sigmoid activation function to generate the gating weights. .

[0087] The fusion features are calculated using Equation 1 below. Formula 1, In Equation 1, V_fusion represents the fusion feature. This indicates a concatenation operation, where V_seq represents the high-dimensional sequence representation vector, W_gate represents the gate weights, and V_phy represents the physical representation vector. The concatenation operation involves joining vectors along the feature dimension, specifically concatenating the weighted sequence representation vector and the physical representation vector in the last dimension to form a fused feature vector.

[0088] Finally, the probability P that the candidate site is a true mutation is output by inputting the fused features into the fully connected layer. During the model training phase, Focal Loss is used as the loss function to optimize model performance.

[0089] 4. Result determination.

[0090] If P ≥ 0.95, the site is determined to be a real somatic mutation (ctDNA variation); otherwise, it is determined to be sequencing noise.

[0091] Example 2. The process parameters in Example 1 were adjusted for application in cancer screening.

[0092] This embodiment adjusts some parameters based on the method flow of Embodiment 1 to adapt to the characteristics of different cancer types. The application scenario for this embodiment is early screening for colorectal cancer. The process parameters are adjusted as follows: (1) Sampling window length L: set to 51bp to capture a longer range of sequence context information (sequence feature tensor construction parameters).

[0093] (2) Normalization base: set to 250bp, because the distribution of ctDNA fragments in colorectal cancer is slightly different (fragment physical feature vector acquisition parameters).

[0094] (3) Screening criteria: Supports ≥4 Reads to further improve specificity (supports obtaining parameters from sequence sets).

[0095] (4) Network structure: In the process of obtaining high-dimensional sequence representation vectors using a convolutional neural network structure in the first branch, the Inception module is used to replace the ResNet-18 residual network structure, and local and long-distance sequence features are extracted simultaneously using convolutional kernels of different sizes (model architecture and inference parameters).

[0096] The remaining steps are the same as the ctDNA mutation detection method in Example 1.

[0097] To verify the effectiveness of the method of this invention, this embodiment uses a simulated dataset of known mutations (In-silicoSpike-in) as the test set (NCBI Sequence Read Archive (SRA) database) and a real publicly available sample set as the validation set to perform ctDNA variant detection and analysis. The test set is based on publicly available healthy human cfDNA sequencing data (SRR10296599) from the NCBI SRA database, constructed by simulating the incorporation of known COSMIC somatic mutations at allele frequencies of 0.01% to 1%. The validation set is derived from plasma cfDNA whole-genome sequencing data of cancer patients and healthy controls in the EGA database project EGAS00001006490, where healthy controls are used to assess specificity and establish the background distribution of normal cfDNA fragment omics. The above test set and validation set are different independent datasets from the dbGaP project phs003170 used for model construction in Example 1.

[0098] Test results as follows Figure 2As shown, in the ultra-low frequency range of MAF (Minor Allele Frequency) = 0.05%, the detection sensitivity of ctDNA variants after adjusting the process parameters of the method of the present invention in Example 1 reached 68.5%. Figure 2 Curve A in section A), the specificity remains above 99.99% ( Figure 2 Curve A in section B). Specificity here refers to the variation relative to background noise. Specifically, it is defined as: Equation 2.

[0099] That is, the proportion of candidate sites that are not true mutations (including false mutations caused by PCR amplification errors, sequencer noise, etc.) that the model correctly identifies as not true mutations.

[0100] True negative in Equation 2: Candidate sites generated by background noise, which the model correctly outputs as negative (P < 0.95). False positives in Equation 2: Candidate sites generated by background noise, which the model incorrectly outputs as positive (P≥0.95). Since healthy individuals should not have tumor somatic cell mutations, any "variation" detected in healthy human samples (healthy human plasma cfDNA data) is a false positive.

[0101] In the tests of this embodiment, it was observed that although some sites had high sequence feature scores (appearing to be true mutations), the weight W_gate generated by the gating unit was close to 0 because the reads they belonged to generally had long fragment lengths (>200bp) and uniform end distance distributions (not in line with ctDNA end preferences). This effectively suppressed the activation of sequence branching, resulting in a negative output. This demonstrates the "veto" ability of physical features to eliminate sequence errors.

[0102] Example 3. Comparative Example 1 for verifying the importance of fragment physical characteristics.

[0103] This comparative example aims to verify the importance of the physical characteristics of the fragments.

[0104] The ctDNA mutation detection method in this embodiment is basically the same as that in Embodiment 1, except that: (1) Remove the second neural network branch: In the dual-stream neural network model constructed in step 3.1 of Example 1, only the first neural network branch, namely the sequence feature extraction branch (ResNet-18), is retained.

[0105] (2) Gated fusion: does not include the gated fusion step in step 3.2 of Example 1, and directly inputs the sequence representation into the classifier (fully connected layer).

[0106] The remaining steps are the same as the ctDNA mutation detection method in Example 1.

[0107] That is, this is a deep learning model based purely on sequence textures (a simplified version of DeepVariant).

[0108] Using the same test and validation set data as in Example 2, ctDNA variant detection and analysis were performed following the workflow of this example. The results are as follows: Figure 2 As shown, in the ultra-low frequency range where MAF = 0.05%, the sensitivity of this embodiment (i.e., Comparative Example 1) (without physical features) is 45.2%. Figure 2 (Curve B in A), but at this sensitivity, the false positive rate is high (specificity drops to 99.90%). Figure 2 (Curve B in A). Analysis revealed that a large number of PCR errors on long fragments were misclassified as true variants because their sequence textures were highly similar to those of true variants, making it impossible for the model to distinguish them.

[0109] Example 4. Comparative Example 2 using traditional statistical variation detection methods.

[0110] This comparative study uses traditional statistical methods for detecting variation.

[0111] The same batch of sequencing data was analyzed using the industry-standard Mutect2 software (GATK version 4.1).

[0112] The parameter settings adopt the officially recommended somatic cell mutation detection mode, and the molecular barcode (UMI) related modules are not enabled.

[0113] Using the same test and validation set data as in Example 2, ctDNA variant detection and analysis were performed following the workflow of this example. The results are as follows: Figure 2 As shown, in the ultra-low frequency range where MAF = 0.05%, the sensitivity of this embodiment (i.e., Comparative Example 2, Mutect2) is only 18.3% ( Figure 2 (Curve C in A) Traditional statistical models have difficulty distinguishing between signals and background noise in the low-frequency range and tend to filter out variations with low support.

[0114] The present invention has been described in detail above. Those skilled in the art will recognize that 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. While specific embodiments have been provided, 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.

Claims

1. A data processing apparatus, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to perform the following steps: 1) Construct a set of supporting sequences containing supporting sequences, wherein the supporting sequences are derived from sequencing reads containing variant sites in the sequencing data of plasma ctDNA of the subject to be tested, and the sequencing reads do not contain molecular barcodes; 2) Construct a sequence feature tensor and a fragment physical feature vector for each of the supported sequences based on the set of supported sequences; the sequence feature tensor includes the number of samples, the sampling window length, and the number of feature channels of the supported sequences; the fragment physical feature vector includes the normalized fragment length, the normalized end distance, and the end motif matching degree; the feature channels include base type, base quality, and chain orientation channels; 3) Input the sequence feature tensor into the first neural network branch to extract the sequence representation vector, and input the fragment physical feature vector into the second neural network branch to extract the physical representation vector; 4) Fuse the sequence representation vector and the physical representation vector to output the authenticity probability of the mutation site.

2. The data processing apparatus according to claim 1, characterized in that: The data processing device is used to detect ctDNA mutations, and the data processing device further includes the following steps: 5) Determine whether the mutation site is a true ctDNA mutation of the subject under test based on the true probability.

3. The data processing apparatus according to claim 1 or 2, characterized in that: The mutation site supports a sequencing read count greater than or equal to 2 in the sequencing data.

4. The data processing apparatus according to claim 1 or 2, characterized in that: The fusion described in 4) includes a gating unit; the gating unit is configured to calculate gating weights using an activation function based on the physical representation vector, and to dynamically adjust the contribution of the sequence representation vector to the authenticity probability based on the gating weights.

5. The data processing apparatus according to claim 1 or 2, characterized in that: In step 3), the first neural network branch adopts a convolutional neural network structure; the second neural network branch adopts a multilayer perceptron structure.

6. An apparatus for detecting ctDNA variants, the apparatus comprising the following modules: B1) Used to construct a set of supporting sequences containing supporting sequences, wherein the supporting sequences are derived from sequencing reads containing variant sites in the sequencing data of plasma ctDNA of the subject to be tested, and the sequencing reads do not contain molecular barcodes; B2) is used to construct a sequence feature tensor and a fragment physical feature vector for each of the supported sequences based on the set of supported sequences; the sequence feature tensor includes the number of samples, the sampling window length, and the number of feature channels of the supported sequences; the fragment physical feature vector includes the normalized fragment length, the normalized end distance, and the end motif matching degree; the feature channels include base type, base quality, and chain orientation channels; B3) is used to input the sequence feature tensor into the first neural network branch to extract the sequence representation vector, and to input the fragment physical feature vector into the second neural network branch to extract the physical representation vector; B4) is used to fuse the sequence representation vector and the physical representation vector to output the authenticity probability of the mutation site; B5) Used to determine whether the mutation site is a true ctDNA mutation of the subject under test based on the true probability; The mutation site supports a sequencing read count greater than or equal to 2 in the sequencing data.

7. The apparatus according to claim 6, characterized in that: The fusion described in B4) includes a gating unit; the gating unit is configured to calculate gating weights using an activation function based on the physical representation vector, and to dynamically adjust the contribution of the sequence representation vector to the authenticity probability based on the gating weights.

8. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by the processor, it implements steps 1)-4) of any one of claims 1-5 or steps 1)-5) of any one of claims 1-5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements steps 1)-4) of any one of claims 1-5 or steps 1)-5) of any one of claims 1-5.

10. The application of a substance in the preparation of products for early cancer screening or postoperative minimal residual disease screening, characterized in that: The substance is any one of the following: 1) The data processing apparatus according to any one of claims 1-5, 2) The apparatus of claim 6 or 7, 3) The computer program product of claim 8, 4) The computer-readable storage medium of claim 9.

Images