Artificial intelligence-based methods for cancer diagnosis and cancer type prediction
By extracting nucleic acids from biological samples to generate reads for comparison and vectorization, and then inputting the data into an artificial intelligence model, the invasiveness and accuracy problems of existing cancer diagnosis are solved, enabling early, highly sensitive, and accurate cancer diagnosis and type prediction.
Patent Information
- Application Number
- CN202180089826.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2020-11-27
- Filing Date
- 2021-11-15
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2041-11-15
AI Technical Summary
Existing cancer diagnostic methods are highly invasive, and the accuracy of screening based on tumor markers is limited. The results of liquid biopsy studies are not definitive enough, making it difficult to diagnose and differentiate cancer types in the early stages.
By extracting nucleic acids from biological samples, generating reads of sequence information, comparing them, generating vectorized data, and inputting them into a trained artificial intelligence model, the output values are analyzed to diagnose cancer and predict its type.
It achieves non-invasive, highly sensitive, and highly accurate cancer diagnosis and type prediction, enabling early detection of cancer and accurate differentiation of cancer types.
Smart Images

Figure CN117083680B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an artificial intelligence-based method for diagnosing cancer and predicting cancer types, and more specifically, to an artificial intelligence-based method for diagnosing cancer and predicting cancer types by extracting nucleic acids from biological samples to obtain sequence information, generating vectorized data based on aligned reads, and analyzing calculated values obtained by inputting the vectorized data into a trained artificial intelligence model. Background Technology
[0002] In clinical practice, cancer diagnosis is typically performed via tissue biopsy following a history check, physical examination, and clinical evaluation. Clinical trial-based cancer diagnosis is only feasible when the number of cancer cells is 1 billion or more and the diameter of the cancer is 1 cm or larger. In this case, the cancer cells already have the potential to metastasize, and at least half of them have already metastasized. Furthermore, tissue biopsy is invasive, adversely causing considerable discomfort to patients and is often not tolerated by cancer treatment. Additionally, tumor markers are used to monitor substances directly or indirectly produced by cancer for cancer screening. However, the accuracy of tumor markers is limited because more than half of tumor marker screening results are normal even in the presence of cancer, and tumor marker screening results are often positive even in the absence of cancer.
[0003] Recently, in response to demands for cancer diagnostic methods that are relatively easy, non-invasive, highly sensitive, and highly specific, liquid biopsies using bodily fluids from patients have been widely used for cancer diagnosis and follow-up. Liquid biopsies, as a non-invasive diagnostic method, have attracted considerable attention as an alternative to traditional invasive diagnostic and examination methods. However, the results of large-scale studies have not confirmed the effectiveness of liquid biopsies as a method for diagnosing cancer, and there are no research results regarding the diagnosis of unclear cancers or the differentiation of unclear cancer types using liquid biopsies.
[0004] Extensive research has been dedicated to cancer diagnosis and treatment to mitigate the health impacts of cancer. In particular, SMCT (Somatic Mutation-Based Cancer Subtyping) is one of the most important research areas. SMCT enables the development of treatment plans based on the cancer type / subtype determined by a patient's somatic gene mutations. Recent reductions in the cost of DNA sequencing have led to a rapid increase in DNA sequencing data, which has greatly facilitated the development of SMCT. Unlike conventional cancer subtyping methods that typically rely on tumor morphology or gene expression levels (i.e., mRNA or protein profiles), SMCT can distinguish tumors with similar histopathological appearances, thus better reflecting the cancer microenvironment and contributing to more accurate cancer classification results (Sun, Y. et al., Sci Rep Vol. 9, 17256, 2019).
[0005] Recently, in addition to SMCT, methods for predicting cancer type using abnormalities in the three-dimensional structure or copy number of chromosomes have been reported (Yuan et al., BMC Genomics, Vol.19(Suppl 6), pp.565, 2018, 10-2019-0036494).
[0006] Meanwhile, research is actively underway to apply effective pattern recognition methods available to humans to real computers as a solution to the problem of classifying input patterns into specific groups, which is frequently encountered in engineering fields.
[0007] In many computer-based studies, there is research on artificial neural networks (ANNs), which are obtained by engineering-modeling the structure of human brain cells that perform effective pattern recognition. To solve the problem of classifying input patterns into specific groups, ANNs use algorithms that simulate human learning abilities. Through these algorithms, ANNs can create a mapping between input and output patterns, which is represented as a learning-capable ANN. Furthermore, ANNs possess generalization capabilities, allowing them to generate relatively correct outputs based on training results for input patterns not used for learning. Due to these two typical properties of learning and generalization, ANNs are being used to solve problems that are difficult to address using traditional sequential programming methods. ANNs have wide applications and are being actively applied in fields such as pattern classification, continuous mapping, nonlinear system recognition, nonlinear control, and robot control.
[0008] Meanwhile, artificial neural networks (ANNs) are computational models implemented in software or hardware that use a large number of artificial neurons connected by interconnects to simulate the computational capabilities of biological systems. ANNs use artificial neurons, which represent the functions of biological neurons in a simplified form. ANNs perform human cognitive or learning processes by interconnecting artificial neurons with interconnections of appropriate connection strengths. The term "connection strength" is interchangeable with "connection weight" and refers to the predetermined value of the connection. ANN learning can be categorized into supervised learning and unsupervised learning. Supervised learning is a method that provides input data and corresponding output data to the neural network and updates the connection strengths of the interconnections, thereby outputting output data corresponding to the input data. Representative learning algorithms include the delta rule and backpropagation learning. Unsupervised learning is a method in which the artificial neural network independently learns connection strengths using only input data without any target values. Unsupervised learning updates connection weights based on the correlation between input patterns.
[0009] Applying large amounts of data to machine learning leads to the so-called "curse of dimensionality," caused by the increase in complexity and dimensionality. In other words, as the dimensionality of the required data approaches infinity, the distance between any two points also approaches infinity, and the data volume (i.e., density) becomes lower in high-dimensional space, failing to accurately reflect the characteristics of the data (Richard Bellman, Dynamic Programming, 2003, Chapter 1). Recent advancements in deep learning, with their hidden layers between the input and output layers, have been reported to significantly improve the performance of classifiers on high-dimensional data such as images, videos, and signals by processing linear combinations of variable values transmitted from the input layer using nonlinear functions (Hinton, Geoffrey, et al., IEEE Signal Processing Magazine Vol. 29.6, pp. 82-97, 2012).
[0010] Various patents (KR 10-2017-0185041, KR 10-2017-0144237 and KR 10-2018-124550) describe the use of artificial neural networks in the biological field, but there is a lack of research on methods for predicting cancer types through artificial neural network analysis based on sequencing information of cell-free DNA (cfDNA).
[0011] Therefore, as a result of extensive and diligent efforts to solve the aforementioned problems and develop artificial intelligence-based methods for diagnosing cancer and predicting cancer types with high sensitivity and accuracy, the inventors discovered that cancer diagnosis and cancer type prediction can be achieved with high sensitivity and accuracy by generating vectorized data based on reads aligned to chromosomal regions and analyzing the data using a trained artificial intelligence model, and based on this discovery, the present invention was completed. Summary of the Invention
[0012] Therefore, one object of the present invention is to provide a method for diagnosing cancer and predicting cancer type based on artificial intelligence.
[0013] Another objective of this invention is to provide a device for diagnosing cancer and predicting cancer types based on artificial intelligence.
[0014] Another object of the present invention is to provide a computer-readable storage medium comprising instructions configured to be executed by a processor for diagnosing cancer and predicting cancer type by the methods described above.
[0015] According to one aspect of the invention, the above and other objectives can be achieved by providing an artificial intelligence-based method for diagnosing cancer and predicting cancer type, the method comprising (a) extracting nucleic acids from a biological sample to obtain sequence information, (b) aligning the sequence information (reads) with a reference genome database, (c) generating vectorized data using nucleic acid fragments based on the aligned sequence information (reads), (d) inputting the generated vectorized data into a trained artificial intelligence model, analyzing the resulting output values, and comparing the resulting output values with cut-off values to determine the presence of cancer, and (e) predicting the cancer type by comparing the output values.
[0016] According to another aspect of the present invention, a method for diagnosing cancer and predicting cancer type based on artificial intelligence is provided, the method comprising (a) extracting nucleic acids from a biological sample to obtain sequence information, (b) aligning the sequence information (reads) with a reference genome database, (c) generating vectorized data using nucleic acid fragments based on the aligned sequence information (reads), (d) inputting the generated vectorized data into a trained artificial intelligence model, analyzing the obtained output value, and comparing the obtained output value with a truncation value to determine whether cancer is present, and (e) predicting the cancer type by comparing the output value.
[0017] According to another aspect of the present invention, an apparatus for diagnosing cancer and predicting cancer type based on artificial intelligence is provided, the apparatus comprising: a decoder configured to extract nucleic acids from a biological sample and decode sequence information; an alignment unit configured to align the decoded sequence with a reference genome database; a data generator configured to generate vectorized data using nucleic acid fragments based on the aligned sequence information (reads); a cancer diagnosis unit configured to input the generated vectorized data into a trained artificial intelligence model, analyze the data, and compare the result value with a truncation value to determine whether cancer is present; and a cancer type predictor for analyzing the output results and thereby predicting the cancer type.
[0018] According to another aspect of the invention, a computer-readable storage medium is provided, comprising instructions configured to be executed by a processor for diagnosing cancer and predicting cancer type by including the following steps: (a) extracting nucleic acids from a biological sample to obtain sequence information; (b) aligning the obtained sequence information (reads) with a reference genome database; (c) generating vectorized data using nucleic acid fragments based on the aligned sequence information (reads); (d) inputting the generated vectorized data into a trained artificial intelligence model, analyzing the resulting output values, and comparing the resulting output values with truncation values to determine the presence of cancer; and (e) predicting the cancer type by comparing the output values. Attached Figure Description
[0019] Figure 1 This is a flowchart illustrating the overall process of determining chromosomal abnormalities based on artificial intelligence according to the present invention.
[0020] Figure 2 This is an example of a GC graph, which is an image obtained by vectorizing NGS data.
[0021] Figure 3 This is a schematic diagram illustrating the configuration of a CNN model according to an implementation method.
[0022] Figure 4 The diagram shows (A) the accuracy of a deep learning model that has learned from the generated GC graph image data in determining the presence of cancer, and (B) the probability distribution for each dataset.
[0023] Figure 5 The results show (A) the accuracy of cancer type predictions by a deep learning model that has learned from the generated GC graph image data, and (B) the probability distribution for each dataset. Detailed Implementation
[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by one of ordinary skill in the art to which this invention pertains. Generally, the terms used herein are well-known in the art and are commonly used.
[0025] This invention discovers that by comparing sequencing data obtained from a sample with a reference genome, generating vectorized data based on the aligned nucleic acid fragments, calculating DPI using a trained artificial intelligence model, and performing analysis, cancer can be diagnosed and cancer type predicted with high sensitivity and accuracy.
[0026] In other words, in one embodiment of the present invention, a method is developed that includes sequencing DNA extracted from blood, aligning the sequencing data with a reference genome, calculating the distance between nucleic acid fragments or the amount of nucleic acid fragments in each predetermined chromosome bin, generating vectorized data, wherein the chromosome bins are located on the X-axis and the distance between nucleic acid fragments or the amount of nucleic acid fragments are located on the Y-axis, training a deep learning model to calculate DPI, comparing the DPI with a cutoff value to determine whether cancer has occurred, and identifying the cancer type showing the highest DPI among the DPIs calculated for various corresponding cancer types as the cancer type of the sample. Figure 1 ).
[0027] On one hand, the present invention relates to a method for providing information based on artificial intelligence to diagnose cancer and predict cancer type, the method comprising:
[0028] (a) Extracting nucleic acids from biological samples to obtain sequence information;
[0029] (b) Align the sequence information (reads) with a reference genome database;
[0030] (c) Vectorized data is generated using nucleic acid fragments based on the aligned sequence information (reads);
[0031] (d) Input the generated vectorized data into a trained artificial intelligence model, analyze the output values, and compare the output values with the cutoff values to determine the presence of cancer; and
[0032] (e) Predict cancer type by comparing output values.
[0033] In this invention, any nucleic acid fragment can be used without restriction, as long as it is a nucleic acid fragment extracted from a biological sample, and the nucleic acid fragment is preferably a cell-free nucleic acid (CVA) or an intracellular nucleic acid fragment, but is not limited thereto.
[0034] In this invention, nucleic acid fragments can be obtained through direct sequencing, next-generation sequencing, or nonspecific whole-genome amplification sequencing.
[0035] In this invention, when using next-generation sequencing, nucleic acid fragments can be read.
[0036] In this invention, the cancer can be a solid tumor or a hematologic malignancy, and is preferably selected from the group consisting of: non-Hodgkin lymphoma, Hodgkin lymphoma, acute myeloid leukemia, acute lymphoblastic leukemia, multiple myeloma, head and neck cancer, lung cancer, glioblastoma, colorectal / rectal cancer, pancreatic cancer, breast cancer, ovarian cancer, melanoma, prostate cancer, thyroid cancer, gastric cancer, gallbladder cancer, bile duct cancer, bladder cancer, small bowel cancer, cervical cancer, unknown primary cancer, kidney cancer, and mesothelioma, but the cancer is not limited thereto.
[0037] In this invention,
[0038] Step (a) includes:
[0039] (ai) Obtaining nucleic acids from blood, semen, vaginal cells, hair, saliva, urine, oral cells, amniotic fluid containing placental cells or fetal cells, tissue cells and mixtures thereof;
[0040] (a-ii) Use salting out, column chromatography, or bead method to remove proteins, fats, and other residues from the collected nucleic acids to obtain purified nucleic acids;
[0041] (a-iii) Prepare single-end or double-end sequencing libraries from purified nucleic acids or nucleic acids randomly fragmented by enzymatic digestion, pulverization or hydroshear methods;
[0042] (a-iv) Libraries prepared using next-generation sequencers; and
[0043] (av) Obtain the sequence information (reads) of nucleic acids from the next-generation sequencer.
[0044] In this invention, the next-generation sequencer can be used with any sequencing method known in the art. Sequencing of nucleic acids isolated using selection methods is typically performed using next-generation sequencing (NGS). Next-generation sequencing includes any sequencing method (e.g., 10) that determines the nucleotide sequence of each nucleic acid molecule or the nucleotide sequence of a surrogate cloned from each nucleic acid molecule to make it highly similar. 5 (One or more molecules are sequenced simultaneously). In one embodiment, the relative abundance of nucleic acid species in the library can be estimated by counting the relative number of homologous sequences appearing in the data generated by the sequencing experiment. Next-generation sequencing is known in the art and is described, for example, in Metzker, M. (2010), Nature Biotechnology Reviews 11:31-46, which is incorporated herein by reference.
[0045] In one implementation, next-generation sequencing is performed to determine the nucleotide sequence of each nucleic acid molecule (e.g., using the HelioScope gene sequencing system from Helicos Biosciences or the PacBio RS system from Pacific Biosciences). In other implementations, massively parallel short-read sequencing produces more base sequences per sequencing unit than other sequencing methods (e.g., other sequencing methods that produce fewer but longer reads), determining the nucleotide sequence of a surrogate cloned from each nucleic acid molecule (e.g., using the Solexa sequencing analyzer from Illumina, CA, San Diego; 454Life Sciences (Branford, Connecticut) and Ion Torrent). Other methods or apparatus for next-generation sequencing may be available from, but are not limited to, 454Life Sciences (Branford, Connecticut), Applied Biosystems (CA, Foster City; SOLiD sequencers), Helicos Biosciences Corporation (MA, Cambridge), and emulsion and microfluidic sequencing nanodroplets (e.g., GnuBIO droplets).
[0046] Platforms used for next-generation sequencing include, but are not limited to, the Roche / 454 FLX System genome sequencer (GS), the Illumina / Solexa genome analyzer (GA), the Life / APG Support Oligonucleotide Ligation Detection (SOLiD) system, the Polonator G.007 system, the HelioScope gene sequencing system from Helical Biosciences, and the PacBio RS system from Pacific Biosciences.
[0047] For example, NGS technology may include one or more template preparation, sequencing, imaging, and data analysis steps.
[0048] Template preparation methods involve randomly breaking nucleic acids (e.g., genomic DNA or cDNA) into small sizes and preparing sequencing templates (e.g., fragment templates or paired templates). Spatially separated templates can be attached to or immobilized on solid surfaces or vectors, allowing for simultaneous large-scale sequencing reactions. Examples of template types that can be used in NGS reactions include templates amplified from clones derived from single DNA molecules, as well as single DNA molecule templates.
[0049] Methods for preparing templates from clone amplification include, for example, emulsion PCR (emPCR) and solid-phase amplification.
[0050] EmPCR can be used to prepare templates for NGS. Typically, a library of nucleic acid fragments is prepared, and adapters containing universal primer promoter sites are ligated to the ends of the fragments. The fragments are then denatured into single strands and captured using beads. Each bead captures a single nucleic acid molecule. After emPCR bead amplification and enrichment, large quantities of template can be attached, immobilized on a polyacrylamide gel (e.g., from Polonator) on a standard microscope slide, chemically cross-linked to an amino-coated glass surface (e.g., Life / APG; Polonator), or deposited in a single PicoTiterPlate (PTP) well (e.g., Roche / 454). At this point, the NGS reaction can be performed.
[0051] Solid-phase amplification (SPA) can also be used to prepare templates for NGS. Typically, the front and back primers are covalently linked to a solid support. The surface density of the amplified fragment is defined as the ratio of primers to template on the support. SPA can prepare millions of spatially separated template clusters (e.g., Illumina / Solexa). The ends of these template clusters can hybridize with universal primers used for NGS reactions.
[0052] Other methods for preparing clonal amplification templates include, for example, multiple substitution amplification (MDA) (Lasken RS; Curr. Opin. Microbiol. 2007; 10(5):510-6). MDA is a non-PCR-based DNA amplification method. The reaction involves annealing random hexameric primers to a template and synthesizing DNA at isothermal temperature using a high-fidelity enzyme (typically Ф29). MDA can produce large-scale products with a low error rate.
[0053] Template amplification methods such as PCR can bind NGS platforms to specific regions of the target or enriched genome (e.g., exons). Representative template enrichment methods include, for example, droplet PCR (Tewhey R. et al., Nature Biotech. 2009, 27: 1025-1031), custom-designed oligonucleotide microarrays (e.g., Roche / NimbleGen oligonucleotide microarrays), liquid-phase hybridization (e.g., molecular inverted probes, MIPs) (Porreca GJ et al., Nature Methods, 2007, 4: 931-936; Krishnakumar S. et al., Proc. Natl. Acad. Sci. USA, 2008, 105: 9296-9310; Turner EH et al., Nature Methods, 2009, 6: 315-316), and biotinylated RNA capture sequences (Gnirke A. et al., Nat. Biotechnol. 2009; 27(2): 182-9).
[0054] Single-molecule templates are another type of template that can be used in NGS reactions. Spatially separated single-molecule templates can be immobilized on a solid support using various methods. In one method, each primer molecule is covalently ligated to the solid support. A adapter is added to the template, and then the template hybridizes with the immobilized primers. In another method, single-molecule templates are covalently ligated to the solid support by initiation and extension of single-stranded single-molecule templates from the immobilized primers. Universal primers then hybridize with the template. In yet another method, a single polymerase molecule is attached to the solid support, wherein the primer template binds to the solid support.
[0055] Representative NGS sequencing and imaging methods include, but are not limited to, cyclic resectable sequencing (CRT), ligation sequencing (SBL), pyrosequencing, and real-time sequencing.
[0056] CRT uses an excisable terminator in a cyclic approach that includes at least nucleotide incorporation, fluorescence imaging, and a cleavage step. Typically, DNA polymerase incorporates a single fluorescently modified nucleotide that is complementary to the template base in the primer. DNA synthesis terminates upon incorporation of the single nucleotide, and the unincorporated nucleotide is washed away. Imaging is performed to determine the homology of the incorporated labeled nucleotide. Then, in the cleavage step, the terminator / inhibitor and fluorescent dye are removed. Representative NGS platforms using the CRT method include, but are not limited to, the Illumina / Solexa Genomics Analyzer (GA), which combines a clonal amplification template method with a 4-color CRT method involving detection using total internal reflection fluorescence (TIRF); and the Helicos Biosciences / HelioScope, which combines a single-molecule template method with a 1-color CRT method involving detection using TIRF.
[0057] SBL uses DNA ligase and either a 1-base-coding probe or a 2-base-coding probe for sequencing.
[0058] Typically, fluorescently labeled probes hybridize to complementary sequences of adjacent primer templates. DNA ligase is used to ligate the dye-labeled probes to primers. After washing away unligated probes, fluorescence imaging is performed to identify the ligated probes. A resectable probe can be used to remove the fluorescent dye, which regenerates the 5'-PO4 group for subsequent ligation cycles. Alternatively, new primers can hybridize to the template after the old primers have been removed. Representative SBL platforms include, but are not limited to, Life / APG / SOLiD (which supports oligonucleotide ligation detection) and use dibase-encoded probes.
[0059] Pyrosequencing is based on the detection of DNA polymerase activity using a different chemiluminescent enzyme. Typically, this method involves sequencing single-stranded DNA by synthesizing complementary strands one base pair at a time and detecting the actual bases added at each step. The template DNA is immobilized, and during the reaction, solutions of A, C, G, and T nucleotides are sequentially added and removed. Light is generated only when the nucleotide solution replenishes the unpaired bases of the template. The sequence of the solution that produces the chemiluminescent signal is used to determine the sequence of the template. Representative pyrosequencing platforms include, but are not limited to, those from Roche / 454, which use DNA templates generated via emPCR from 1-2 million beads deposited in PTP wells.
[0060] Real-time sequencing involves imaging the sequential incorporation of dye-labeled nucleotides during DNA synthesis. Representative real-time sequencing platforms include, but are not limited to, those from Pacific Biosciences, which use DNA polymerase molecules attached to the surface of their respective zero-mode waveguide (ZMW) detectors to obtain sequence information as phosphate-linked nucleotides are incorporated into the growing primer strands.
[0061] The Life / VisiGen platform uses genetically engineered DNA polymerase and attached fluorescent dyes to generate enhanced signals via fluorescence resonance energy transfer (FRET) after incorporation of nucleotides; and the platform from LI-COR Biosciences uses dye quencher nucleotides in sequencing reactions.
[0062] Other NGS methods include, but are not limited to, nanopore sequencing, hybridization sequencing, sequencing based on nanotransistor arrays, polymerase cloning sequencing (Polony sequencing), sequencing based on scanning tunneling microscopy (STM), and sequencing based on nanowire molecular sensors.
[0063] Nanopore sequencing involves the electrophoresis of nucleic acid molecules in solution through nanoscale pores, which provide highly enclosed regions for the analysis of single nucleic acid polymers. Representative nanopore sequencing methods are described in Branton D. et al., Nat. Biotechnol. 2008; 26(10):1146-53 and elsewhere.
[0064] Hybridization sequencing is a non-enzymatic method using DNA microarrays. Typically, individual DNA pools are fluorescently labeled and hybridized into an array containing a known sequence. The hybridization signal from a given point on the array can be used to identify the DNA sequence. The binding of one DNA strand to its complementary strand in a DNA double helix can be sensitive even to single-base mismatches, when the hybridization region is short or when specific mismatch detection proteins are present. Representative hybridization sequencing methods are described in, for example, Hanna GJ et al., J. Clin. Microbiol. 2000; 38(7):2715-21; and Edwards JR et al., Mut. Res. 2005; 573(1-2):3-12.
[0065] Polymerase cloning sequencing is based on polymerase cloning (Polony) amplification and multiple single-base extension (FISSEQ). Polymerase cloning amplification is a method for in situ amplification of DNA on a polyacrylamide membrane. Representative polymerase cloning sequencing methods are described, for example, in U.S. Patent Application Publication No. 2007 / 0087362.
[0066] Devices based on nanotransistor arrays, such as carbon nanotube field-effect transistors (CNTFETs), can also be used in NGS. For example, DNA molecules are extended and driven through nanotubes via microfabricated electrodes. The DNA molecules sequentially contact the surface of the carbon nanotubes, and due to charge transfer between the DNA molecules and the nanotubes, a difference in current from each base is generated. This difference is recorded to sequence the DNA. Representative nanotransistor array-based sequencing methods are described, for example, in U.S. Patent Publication No. 2006 / 0246497.
[0067] Scanning tunneling microscopy (STM) can also be used in next-generation sequencing (NGS). STM uses a piezoelectrically controlled probe to perform a raster scan of a sample, forming an image on its surface. STM can be used to image the physical properties of individual DNA molecules, for example, by integrating a flexible actuator-driven gap with a scanning tunneling microscope to produce coherent electron tunneling imaging and spectroscopy. Representative sequencing methods using STM are described, for example, in U.S. Patent Application Publication No. 2007 / 0194225.
[0068] Molecular analysis devices composed of nanowire molecular sensors can also be used in NGS. Such devices can detect interactions between nucleic acid molecules and nitrogen-containing substances on nanowires such as DNA. Molecular guides are configured to direct molecules near the molecular sensor to allow for interactions and subsequent detection. Representative sequencing methods using nanowire molecular sensors are described, for example, in U.S. Patent Application Publication No. 2006 / 0275779.
[0069] Double-stranded sequencing can be used in next-generation sequencing (NGS). Double-stranded sequencing uses blocking and unblocking primers to sequence the sense and antisense strands of DNA. Typically, the method includes: annealing the unblocking primer to the first strand of the nucleic acid; annealing the second blocking primer to the second strand of the nucleic acid; extending the nucleic acid along the first strand with a polymerase; terminating the first sequencing primer; unblocking the second primer; and extending the nucleic acid along the second strand. A representative double-stranded sequencing method is described, for example, in U.S. Patent No. 7,244,567.
[0070] Data analysis phase.
[0071] After NGS reads are formed, they are compared with known reference sequences or spliced from scratch into known reference sequences.
[0072] For example, by aligning NGS reads with a reference sequence (e.g., a wild-type sequence), genetic modifications in a sample (e.g., a tumor sample), such as single nucleotide polymorphisms and structural variations, can be identified. Methods for aligning NGS reads with sequences are described, for example, in Trapnell C. and Salzberg SL, Nature Biotech., 2009, 27:455-457.
[0073] Examples of de novo splicing are described, for example, Warren R. et al., Bioinformatics, 2007, 23:500-501; Butler J. et al., Genome Res., 2008, 18:810-820; Zerbino DR and Birney E., Genome Res., 2008, 18:821-829.
[0074] Sequence alignment or sequence assembly can be performed using read data from one or more NGS platforms (e.g., by mixing Roche / 454 and Illumina / Solexa read data). In this invention, alignment can be performed using the BWA algorithm and hg19 sequences, but is not limited to these methods.
[0075] In this invention, the sequence alignment in step (b) includes using computational methods or approaches of computer algorithms to determine whether sequences may originate from the genome (e.g., short read sequences obtained through next-generation sequencing) or to determine whether there is identity between read sequences and reference sequences by evaluating the similarity between them. Various algorithms can be applied to the sequence alignment problem. Some algorithms are relatively slow but capable of achieving relatively high specificity. These include, for example, algorithms based on dynamic programming. Dynamic programming is a method of solving complex problems by breaking them down into simple steps. Other methods are more efficient but generally not exhaustive and include, for example, heuristic algorithms and probabilistic methods designed for searching large-scale databases.
[0076] Typically, the alignment process can involve two steps: candidate selection and sequence alignment. Candidate selection reduces the search space for sequence alignment across the entire genome to obtain a shorter list of possible alignment locations. As the term literally means, sequence alignment involves aligning sequences including those obtained during the candidate selection process. This can be done using broad alignment (e.g., Needleman-Wunsch alignment) or local alignment (e.g., Smith-Waterman alignment).
[0077] Most attribute sorting algorithms likely fall into one of three categories based on indexing methods: hash table-based algorithms (e.g., BLAST, ELAND, SOAP), suffix tree algorithms (e.g., Bowtie, BWA), and merge sort (e.g., slider). Short read sequences are typically used for alignment. Examples of sequence alignment algorithms / programs for short read sequences include, but are not limited to, BFAST (Homer N. et al., PLoS One. 2009; 4(11):e7767), BLASTN (from blast.ncbi.nlm.nih.gov.com.uk), BLAT (Kent WJ Genome Res. 2002; 12(4):656-64), Bowtie (Langmead B. et al., GenomeBiol. 2009; 10(3):R25), BWA (Li H. and Durbin R., Bioinformatics, 2009, 25:1754-60), BWA-SW (Li H. and Durbin R., Bioinformatics, 2010; 26(5):589-95), CloudBurst (Schatz M.C., Bioinformatics, 2009; 25(11):1363-9), Corona Lite (Applied Biosystems, Carlsbad, California, USA), CASHX (Fahlgren N. et al., RNA, 2009; 15, 992-1002), CUDA-EC (Shi H. et al., J. Comput. Biol. 2010; 17(4): 603-15), ELAND (bioit.dbi.udel.edu / howto / eland on the World Wide Web), GNUMAP (Clement NL et al., Bioinformatics. 2010; 26(1): 38-45), GMAP (Wu TD and Watanabe CK, Bioinformatics, 2005; 21(9): 1859-75), GSNAP (Wu TD and Nacu S., Bioinformatics, 2010; 26(7): 873-81), Geneious Compiler (Geneious) Assembler (Biomatters Ltd., Auckland, New Zealand), LAST, MAQ (Li H. et al., Genome Res. 2008; 18(11): 1851-8), Mega-BLAST (at ncbi.nlm.nih.gov / blast / megablast on the World Wide Web).shtml), MOM (Eaves HL and Gao Y. Bioinformatics. 2009; 25(7): 969-70), MOSAIK (on the World Wide Web bioinformatics.bc.edu / marthlab / Mosaik), NovoAlign (on the World Wide Web novocraft.com / main / index.php), PALMapper (on the World Wide Web fml.tuebingen.mpg.de / raetsch / suppl / palmapper), PASS (Campagna D. et al., Bioinformatics, 2009; 25(7): 967-8), PatMaN (Prufer K. et al., Bioinformatics, 2008; 24(13): 1530-1), PerM (Chen Y. et al., Bioinformatics, 2009, 25(19): 2514-2521), ProbeMatch (Kim YJ et al., Bioinformatics. 2009; 25(11):1424-5), QPalma (de Bona F et al., Bioinformatics, 2008, 24(16):i174), RazerS (Weese D et al., Genome Research, 2009, 19:1646-1654), RMAP (Smith AD et al., Bioinformatics, 2009; 25(21):2841-2), SeqMap (Jiang H et al., Bioinformatics, 2008; 24:2395-2396), Shrec (Salmela L., Bioinformatics, 2010; 26(10):1284-90), SHRiMP (Rumble SM et al., PLoS Comput.Biol., 2009, 5(5):e1000386), SLIDER (Malhis N. et al., Bioinformatics, 2009, 25(1):6-13), SLIM search (Muller T. et al., Bioinformatics, 2001; 17Suppl 1:S182-9), SOAP (Li R. et al., Bioinformatics, 2008; 24(5):713-4), SOAP2 (Li R. et al., Bioinformatics, 2009; 25(15):1966-7), SOCS (Ondov BD)Ning et al., Bioinformatics, 2008; 24(23):2776-7), SSAHA (Ning Z. et al., Genome Res. 2001; 11(10):1725-9), SSAHA2 (Ning Z. et al., Genome Res. 2001; 11(10):1725-9), Stampy (Lunter G. and Goodson M., Genome Res. 2010, preprint (epubahead of print)), Taipan (taipan.sourceforge.net on the World Wide Web), UGENE (ugene.unipro.ru on the World Wide Web), XpressAlign (bcgsc.ca / platform / bioinfo / software / XpressAlign on the World Wide Web), and ZOOM (Bioinformatics Solutions Inc., Waterloo, Ontario, Canada).
[0078] Sequence alignment algorithms can be selected based on many factors, including, for example, sequencing technology, read length, number of reads, available computational resources, and sensitivity / scoring requirements. Different sequence alignment algorithms can achieve different levels of speed, alignment sensitivity, and alignment specificity. Alignment specificity refers to the percentage of target sequence residues that are correctly aligned compared to the predicted alignment, as typically shown in the submission. Alignment sensitivity also refers to the percentage of target sequence residues that are aligned in the predicted alignment, as typically shown in the submission.
[0079] When speed is the primary consideration, alignment algorithms such as ELAND or SOAP can be used to align short reads (e.g., from Illumina / Solexa sequencers) with a reference genome. When specificity is considered the most important factor, alignment algorithms such as BLAST or Mega-BLAST are used to determine similarity using even shorter reads (e.g., Roche FLX), although these methods are slower. When quality scores are important and therefore accuracy is crucial (e.g., in rapid large-scale SNP searches), alignment algorithms such as MAQ or NovoAlign can be used for single-end or paired-end data. Alignment algorithms such as Bowtie or BWA use the Burrows-Wheeler transform (BWT), thus requiring a relatively small memory footprint. Alignment algorithms such as BFAST, PerM, SHRiMP, SOCS, or ZOOM map color space reads and can therefore be used with ABI's SOLiD platform. In some applications, results from two or more classification algorithms can be combined.
[0080] In this invention, the length of the sequence information (read segment) in step (b) is 5 to 5000 bp, and the number of sequence information (read segments) used can be 5000 to 5 million, but this invention is not limited thereto.
[0081] In this invention, any vectorized data that can be generated by using nucleic acid fragments based on aligned reads can be used as the vectorized data in step (c) without limitation, but the vectorized data is preferably a Grand Canyon map (GC map), but is not limited thereto.
[0082] In this invention, the vectorized data is preferably an image, but is not limited thereto. An image is essentially composed of pixels. If an image composed of pixels is vectorized, it may be represented as a monochrome 2D vector (black and white), a three-channel 2D vector (RGB color), or a four-channel 2D vector (CMYK color), depending on the type of image.
[0083] The vectorized data of the present invention is not limited to image data, and, for example, can be input data for an artificial intelligence model using an n-channel 2D vector (multichannel vector) created by stacking n black and white images.
[0084] In this invention, the GC plot is created by taking a specific portion (a constant bin or bins of different sizes) on the X-axis and taking values on the Y-axis that can be represented by nucleic acid fragments (such as the distance between nucleic acid fragments or the count of nucleic acid fragments). In this invention, the bin can be from 1 kbp to 10 Mbp, but is not limited to this.
[0085] In this invention, the method may further include, prior to step (c), separating nucleic acid fragments that satisfy the mapping quality score from the aligned nucleic acid fragments.
[0086] In this invention, the mapping quality score can vary according to a desired standard, but is preferably 15 to 70, more preferably 50 to 70, and most preferably 60.
[0087] In this invention, the GC map in step (c) is characterized by generating vectorized data by calculating the distribution of aligned nucleic acid fragments in each chromosome bin based on the count of nucleic acid fragments in each bin or the distance between nucleic acid fragments.
[0088] In this invention, the calculated count of nucleic acid fragments or the vectorization of the calculated distance between nucleic acid fragments can be performed without restriction using any known method for vectorizing calculated values.
[0089] In this invention, calculating the distribution of aligned sequence information in each chromosome bin based on the counting of nucleic acid fragments can be performed using a method including the following steps:
[0090] i) Divide the chromosomes into predetermined bins;
[0091] ii) Determine the count of the aligned nucleic acid fragments in each bin;
[0092] iii) Normalize the count of nucleic acid fragments in each identified bin by dividing it by the total number of nucleic acid fragments in the sample; and
[0093] iv) Create a GC graph where the X-axis represents the order of the corresponding bins and the Y-axis represents the normalized values calculated in step iii).
[0094] In this invention, calculating the distribution of aligned sequence information in each chromosome bin based on the distance between nucleic acid fragments can be performed using a method including the following steps:
[0095] i) Divide the chromosomes into predetermined bins;
[0096] ii) Determine the distance between the aligned nucleic acid fragments in each bin (fragment distance, FD);
[0097] iii) Based on the fragment distance calculated for each bin, determine the representative distance (RepFD) between fragments in each bin;
[0098] iv) Normalize the representative distance (RepFD) between fragments calculated in step iii) by the representative total nucleic acid fragment distance; and
[0099] v) Create a GC graph, where the X-axis represents the order of the corresponding bins and the Y-axis represents the normalized values calculated in step iv).
[0100] In this invention, the predetermined bin can be from 1Kbp to 3Gbp, but is not limited to this.
[0101] In this invention, the method may further include grouping nucleic acid fragments. Grouping can be based on the adapter sequences of the aligned nucleic acid fragments. For forward-aligned and reverse-aligned nucleic acid fragments, the distance between nucleic acid fragments with selected sequence information can be calculated separately.
[0102] In this invention, FD is defined as the distance between the reference value of the i-th nucleic acid fragment and the reference value of at least one nucleic acid fragment selected from the (i+1)-n-th nucleic acid fragments among the obtained n nucleic acid fragments.
[0103] In this invention, FD can be one or more values selected from the group consisting of: the sum, difference, product, mean, logarithm of the product, logarithm of the sum, median, quantile, minimum, maximum, variance, standard deviation, absolute deviation of the median and coefficient of variation, and / or one or more reciprocals thereof, weighted values, and statistical values, but the invention is not limited thereto.
[0104] As used in this article, the expression “one or more values…and / or one or more reciprocals thereof” is intended to indicate that one or more of the above values or a combination of two or more of them may be used.
[0105] As used in this article, the term "reference value for a nucleic acid fragment" can be a value obtained by adding any value to or subtracting any value from the median of the nucleic acid fragment.
[0106] The FD of the obtained n nucleic acid fragments can be defined as follows.
[0107] FD=Dist(Ri~Rj)(1 <i<j<n),
[0108] The Dist function calculates one or more values selected from the following groups: the sum, difference, product, mean, logarithm of the product, logarithm of the sum, median, quantile, minimum, maximum, variance, standard deviation, absolute deviation of the median, and coefficient of variation, and / or one or more reciprocals thereof, weighted values, and statistical values, but not limited to these.
[0109] In other words, as used in this article, FD (fragment distance) refers to the distance between aligned nucleic acid fragments. Here, the number of cases where nucleic acid fragments are selected for distance calculation can be defined as follows. When there are a total of N nucleic acid fragments, the number of combinations of distances between nucleic acid fragments is... That is, when i is 1, i+1 is 2, and FD can be defined as the distance between the first nucleic acid fragment and one or more nucleic acid fragments selected from the second to the nth nucleic acid fragments.
[0110] In this invention, the FD can be obtained by calculating the distance between a specific position within the i-th nucleic acid fragment and a specific position within at least one of the (i+1)-nth nucleic acid fragments.
[0111] For example, if a nucleic acid fragment is 50 bp in length and is aligned to position 4183 on chromosome 1, the genetic position values for calculating the distance between that nucleic acid fragment and another nucleic acid fragment are 4183 and 4232 on chromosome 1.
[0112] If a 50bp nucleic acid fragment adjacent to the nucleic acid fragment aligns to position 4232 on chromosome 1, the genetic position values for calculating the distance between the nucleic acid fragment and another nucleic acid fragment are 4232 and 4281 on chromosome 1, and the FD between the two nucleic acid fragments is 1 to 99.
[0113] If another adjacent 50bp nucleic acid fragment aligns to position 4123 on chromosome 1, the genetic position values used to calculate the distance between this nucleic acid fragment and the other nucleic acid fragment are positions 4123 and 4172 on chromosome 1. The FD between the two nucleic acid fragments is 61-159, and the FD between this nucleic acid fragment and the first exemplary nucleic acid fragment is 12-110. The FD can be one or more of the following groups: sum, difference, product, mean, logarithm of product, logarithm of sum, median, quantile, minimum, maximum, variance, standard deviation, absolute deviation of median, and coefficient of variation, and / or one or more reciprocals thereof, weighted values, and statistical values, but not limited thereto, and preferably the reciprocal of one of the two FD values, but not limited thereto, and preferably the reciprocal of one of the two FD values, but not limited thereto.
[0114] Preferably, in this invention, FD can be a value obtained by adding any value to or subtracting any value from the median of the nucleic acid fragment.
[0115] In this invention, the median of a function FD (Function Determination) is the value located at the very center when the calculated FDs are compared in ascending order. For example, when there are three values, namely 1, 2, and 100, the median is 2, which is located at the center. If there are an even number of FDs, the median is determined to be the average of two medians. For example, if the FDs are 1, 10, 90, and 200, the median is 50, which is the average of 10 and 90.
[0116] In this invention, any value can be set without limitation, as long as it can be used to indicate the position of nucleic acid fragments, but preferably 0 to 5 kbp or 0 to 300% of the length of nucleic acid fragments, 0 to 3 kbp or 0 to 200% of the length of nucleic acid fragments, or 0 to 1 kbp or 0 to 100% of the length of nucleic acid fragments, more preferably 0 to 500 bp or 0 to 50% of the length of nucleic acid fragments, but not limited thereto.
[0117] In this invention, during paired-end sequencing, FD can be derived based on the position values of the forward and reverse reads.
[0118] For example, if in a pair of 50bp paired reads, the forward read aligns to position 4183 on chromosome 1 and the reverse read aligns to position 4349, then the two ends of the nucleic acid fragment are located at positions 4183 and 4349, and the reference values for calculating the distance between the nucleic acid fragments are 4183 and 4349. In this case, if in another pair of paired reads adjacent to this nucleic acid fragment, the forward read aligns to position 4349 on chromosome 1 and the reverse logarithmic read aligns to position 4515, then the position values of this nucleic acid fragment are 4349 and 4515. The distance between the two nucleic acid fragments can be from 0 to 333, and most preferably 166, which corresponds to the median distance between the individual nucleic acid fragments.
[0119] In this invention, when sequence information is obtained through paired-end sequencing, the method may further include excluding nucleic acid fragments with mapping quality scores lower than reference values from the calculation process.
[0120] In this invention, during single-end sequencing, FD can be derived based on a position value of one type of forward read or reverse read.
[0121] In this invention, during single-end sequencing, if a position value is derived based on sequence information from forward alignment, an arbitrary value is added to it, and if a position value is derived based on sequence information from reverse alignment, an arbitrary value is subtracted. Any value can be set without limitation, as long as FD clearly indicates the position of the nucleic acid fragment, but is preferably 0 to 5 kbp or 0 to 300% of the nucleic acid fragment length, 0 to 3 kbp or 0 to 200% of the nucleic acid fragment length, or 0 to 1 kbp or 0 to 100% of the nucleic acid fragment length, more preferably 0 to 500 bp or 0 to 50% of the nucleic acid fragment length, but is not limited thereto.
[0122] In this invention, the nucleic acid to be analyzed can be sequenced and expressed in units called "reads". Depending on the sequencing method, reads can be divided into single-end sequencing reads (SE) and paired-end sequencing reads (PE). SE reads are obtained by sequencing one of the 5' and 3' ends of the nucleic acid molecule to a predetermined length in a random orientation, while PE reads are obtained by sequencing both of the 5' and 3' ends of the nucleic acid molecule to a predetermined length. As is known to those skilled in the art, due to this difference, when sequencing in SE mode, one read is generated from a nucleic acid fragment, while in PE mode, a pair of two reads are generated from a nucleic acid fragment.
[0123] The ideal method for accurately calculating the distance between nucleic acid fragments involves sequencing the nucleic acid molecule from start to finish, aligning the reads, and using the median (center) of the aligned read position values. However, this method faces technical limitations due to the constraints of sequencing technology and its high cost. Therefore, sequencing methods such as SE or PE are used. In PE mode, since the start and end positions of the nucleic acid molecule can be identified, the accurate position (median) of the nucleic acid fragment can be determined by combining these values. In SE mode, the accuracy of the position (median) calculation is limited because only information from one end of the nucleic acid fragment can be used.
[0124] Furthermore, when calculating the distance between nucleic acid molecules using the end information of all reads sequenced (aligned) in both forward and reverse directions, inaccurate values may be obtained due to the sequencing direction factor.
[0125] Therefore, due to technical reasons related to sequencing methods, the 5' end of the forward read has a smaller position value compared to the center position value of the nucleic acid molecule, while the 3' end of the reverse read has a larger position value. This characteristic can be used to estimate the value near the center of the nucleic acid molecule by adding any value (the extended bp) to the forward read and subtracting it from the reverse read.
[0126] In other words, the arbitrary value (extended bp) can vary depending on the sample used, and since cell-free nucleic acids are known to have an average nucleic acid length of approximately 166 bp, its arbitrary value (extended bp) is therefore set at approximately 80 bp. If experiments are conducted using a fragmentation device (e.g., sonication), approximately half of the target length set during the fragmentation process can be set as the extended bp.
[0127] In this invention, the representative FD (RepFD) includes at least one selected from the group consisting of: the sum, difference, product, mean, median, quantile, minimum, maximum, variance, standard deviation, absolute deviation of the median, and coefficient of variation and / or their reciprocals, and is preferably the median or mean of the FD or their reciprocals, but is not limited thereto.
[0128] In this invention, vectorized data may include multiple chromosome-specific maps in an image.
[0129] In this invention, any artificial intelligence model can be used in step (d) without restriction, as long as it is a model capable of learning to distinguish cancer types from images, and preferably a deep learning model.
[0130] In this invention, any artificial intelligence model can be used without restriction, as long as it is an artificial neural network algorithm capable of analyzing vectorized data based on artificial neural networks, and preferably selected from the group consisting of: convolutional neural networks (CNN), deep neural networks (DNN), recurrent neural networks (RNN), and autoencoders, but not limited thereto.
[0131] In this invention, the recurrent neural network is selected from the group consisting of: Long Short-Term Memory (LSTM) neural networks, Gated Recurrent Unit (GRU) neural networks, vanilla recurrent neural networks, and attentive recurrent neural networks.
[0132] In this invention, when the artificial intelligence model is a CNN, the loss function for performing binary classification is represented by Equation 1 below, and the loss function for performing multi-class classification is represented by Equation 2 below.
[0133] Equation 1: Binary Classification
[0134]
[0135] Model (x) i = The output of the artificial intelligence model in response to the i-th input
[0136] y = actual label value
[0137] n = the number of input data
[0138] Equation 2: Multiclass Classification
[0139]
[0140] Model (x) i ) j = The j-th AI model output in response to the i-th input
[0141] y = actual label value
[0142] n = the number of input data
[0143] c = Number of classes
[0144] In this invention, binary classification means that the artificial intelligence model learns to identify the presence or absence of cancer, while multi-class classification means that the artificial intelligence model learns to distinguish between two or more types of cancer.
[0145] In this invention, when the artificial intelligence model is a CNN, the learning process includes the following steps:
[0146] i) Classify the generated GC graph into training data, validation data, and test data.
[0147] The training data is used to train the CNN model, the validation data is used for hyperparameter tuning and validation, and the test data is used for performance evaluation after the optimal model is built; and
[0148] ii) Constructing the optimal CNN model through hyperparameter tuning and training; and
[0149] iii) Use validation data to compare the performance of multiple models obtained through hyperparameter tuning, and determine the model with the best validation data as the best model.
[0150] In this invention, hyperparameter tuning is the process of optimizing the values of various parameters (number of convolutional layers, number of dense layers, number of convolutional filters, etc.) that constitute the CNN model. Bayesian optimization and grid search methods are used for hyperparameter tuning.
[0151] In this invention, predetermined hyperparameters are used to optimize the intrinsic parameters (weights) of the CNN model, and the model is determined to be overfitting when the validation loss begins to increase compared to the training loss. Training is stopped before this determination.
[0152] In this invention, any value obtained by the artificial intelligence model analyzing the input vectorized data in step (d) can be used without limitation, as long as it is a specific fraction or real number, and the value is preferably a depth probability index (DPI), but is not limited thereto.
[0153] In this invention, the "depth probability index" refers to the value expressed as a probability value by adjusting the output of the artificial intelligence to a scale of 0 to 1, using the sigmoid function in the case of binary classification and the SoftMax function in the case of multi-class classification for the last layer of the artificial intelligence model.
[0154] In binary classification, a sigmoid function is used for training, adjusting the DPI to 1 to assume cancer has occurred. For example, when given breast cancer samples and normal samples, training is performed so that the DPI for breast cancer samples is close to 1.
[0155] In multi-class classification, the SoftMax function is used to extract as many DPIs as there are classes. The sum of the DPIs is adjusted to 1, and training is performed so that the DPI for the cancer type is effectively adjusted to 1. For example, assuming there are three classes—breast cancer, liver cancer, and normal—when a breast cancer sample is input, training is performed to adjust the DPI for the breast cancer class to approximately 1.
[0156] In this invention, the output value obtained in step (d) is obtained for each cancer type.
[0157] In this invention, the artificial intelligence model is trained to adjust the output value to approximately 1 if cancer is present, and to approximately 0 if cancer is absent. Therefore, performance (training, validation, and testing accuracy) is measured based on a cutoff value of 0.5. In other words, if the output value is 0.5 or greater, cancer is determined to be present, and if it is less than 0.5, cancer is determined not to be present.
[0158] Here, it will be obvious to those skilled in the art that the cutoff value of 0.5 can be arbitrarily changed. For example, to reduce false positives, the cutoff value can be set higher than 0.5 as a more stringent standard for determining the presence of cancer, and to reduce false negatives, the cutoff value can be set lower than 0.5 as a weaker standard for determining the presence of cancer.
[0159] Most preferably, the cutoff value can be set by using a trained artificial intelligence model to determine the probability of DPI by applying unseen data (including data containing solutions different from those obtained during training).
[0160] In this invention, predicting the cancer type by comparing the output results (e) includes identifying the cancer type that shows the highest value among the output results as the cancer of the sample.
[0161] On the other hand, the present invention relates to a device for diagnosing cancer and predicting cancer type based on artificial intelligence, the device comprising:
[0162] A decoder configured to extract nucleic acids from biological samples and decode sequence information;
[0163] The aligner is configured to compare the decoded sequence with a reference genome database;
[0164] The data generator is configured to generate vectorized data using nucleic acid fragments based on aligned sequence information (reads);
[0165] A cancer diagnosis unit is configured to input generated vectorized data into a trained artificial intelligence model, analyze the data, and compare the resulting values with cutoff values to determine the presence of cancer; and
[0166] A cancer type predictor is used to analyze the output and thus predict the type of cancer.
[0167] In this invention, the decoder may include a nucleic acid syringe for injecting nucleic acids extracted from a separate device and a sequence information analyzer for analyzing the sequence information of the injected nucleic acids, preferably an NGS analyzer, but not limited thereto.
[0168] In this invention, the decoder can receive and decode sequence information data generated in a separate device.
[0169] In this invention, the vectorized data of the data generator can be a Grand Canyon graph (GC graph).
[0170] In this invention, a GC graph is a graph in which a specific portion (bins of constant size or bins of different sizes) is set as the X-axis, and values expressed based on nucleic acid fragments, such as the distance or number between nucleic acid fragments, are set as the Y-axis. In this invention, a bin can be from 1 kbp to 10 Mbp, but is not limited to this.
[0171] In this invention, the data generator may further include a nucleic acid fragment classifier configured to separately classify nucleic acid fragments that meet the mapping quality scores of the aligned nucleic acid fragments before generating vectorized data.
[0172] In this invention, the mapping quality score can vary according to a desired standard, but is preferably 15 to 70, more preferably 50 to 70, and most preferably 60.
[0173] In this invention, a GC graph of the data generator is generated using data vectorized by calculating the count of nucleic acid fragments in each bin or the distance between nucleic acid fragments.
[0174] In this invention, the calculated count of nucleic acid fragments or the vectorization of the calculated distance between nucleic acid fragments can be performed without restriction using any known method for vectorizing calculated values.
[0175] In this invention, calculating the distribution of aligned sequence information in each chromosome bin based on the counting of nucleic acid fragments can be performed using a method including the following steps:
[0176] i) Divide the chromosomes into predetermined bins;
[0177] ii) Determine the count of the aligned nucleic acid fragments in each bin;
[0178] iii) Normalize the count of nucleic acid fragments in each identified bin by dividing it by the total number of nucleic acid fragments in the sample; and
[0179] iv) Create a GC graph where the X-axis represents the order of the corresponding bins and the Y-axis represents the normalized values calculated in step iii).
[0180] In this invention, calculating the distribution of aligned sequence information in each chromosome bin based on the distance between nucleic acid fragments can be performed using a method including the following steps:
[0181] i) Divide the chromosomes into predetermined bins;
[0182] ii) Determine the distance between the aligned nucleic acid fragments in each bin (fragment distance, FD);
[0183] iii) Based on the fragment distance calculated for each bin, determine the representative distance (RepFD) between fragments in each bin;
[0184] iv) Normalize the representative distance (RepFD) between fragments calculated in step iii) by the representative total nucleic acid fragment distance; and
[0185] v) Create a GC graph, where the X-axis represents the order of the corresponding bins and the Y-axis represents the normalized values calculated in step iv).
[0186] In this invention, the predetermined bin can be from 1Kbp to 3Gbp, but is not limited to this.
[0187] In this invention, the method may further include grouping nucleic acid fragments. Grouping can be based on the adapter sequences of the aligned nucleic acid fragments. For forward-aligned and reverse-aligned nucleic acid fragments, the distance between nucleic acid fragments with selected sequence information can be calculated separately.
[0188] In this invention, FD is defined as the distance between the reference value of the i-th nucleic acid fragment and the reference value of at least one nucleic acid fragment selected from the (i+1)-n-th nucleic acid fragments among the obtained n nucleic acid fragments.
[0189] In this invention, FD can be one or more values selected from the group consisting of: the sum, difference, product, mean, logarithm of the product, logarithm of the sum, median, quantile, minimum, maximum, variance, standard deviation, absolute deviation of the median, and coefficient of variation, and / or one or more reciprocals thereof, weighted values, and statistical values, but the invention is not limited thereto.
[0190] As used in this article, the expression “one or more values…and / or one or more reciprocals thereof” is intended to indicate that one of the above values or a combination of two or more thereof may be used.
[0191] As used in this article, the term "reference value for a nucleic acid fragment" can be a value obtained by adding any value to or subtracting any value from the median of the nucleic acid fragment.
[0192] In this invention, any model can be used without restriction as an artificial intelligence model for cancer diagnosis units, as long as it is a model capable of learning to distinguish cancer types from images, and preferably a deep learning model.
[0193] In this invention, any artificial intelligence model can be used without restriction, as long as it is an artificial neural network algorithm capable of analyzing vectorized data based on artificial neural networks, and preferably selected from the group consisting of: convolutional neural networks (CNN), deep neural networks (DNN), recurrent neural networks (RNN), and autoencoders, but not limited thereto.
[0194] In this invention, the recurrent neural network is selected from the group consisting of: Long Short-Term Memory (LSTM) neural network, Gated Recurrent Unit (GRU) neural network, ordinary recurrent neural network, and attention recurrent neural network.
[0195] In this invention, when the artificial intelligence model is a CNN, the loss function for performing binary classification is represented by Equation 1 below, and the loss function for performing multi-class classification is represented by Equation 2 below.
[0196] Equation 1: Binary Classification
[0197]
[0198] Model (x) i = The output of the artificial intelligence model in response to the i-th input
[0199] y = actual label value
[0200] n = the number of input data
[0201] Equation 2: Multiclass Classification
[0202]
[0203] Model (x) i ) j = The j-th AI model output in response to the i-th input
[0204] y = actual label value
[0205] n = the number of input data
[0206] c = Number of classes
[0207] In this invention, binary classification means that the artificial intelligence model learns to identify the presence or absence of cancer, and multi-class classification means that the artificial intelligence model learns to distinguish between two or more types of cancer.
[0208] In this invention, when the artificial intelligence model is a CNN, the learning process includes the following steps:
[0209] i) Classify the generated GC graphs into training, validation, and test data.
[0210] Training data is used when training a CNN model, validation data is used for hyperparameter tuning and validation, and test data is used for performance evaluation after the optimal model is built.
[0211] ii) Constructing the optimal CNN model through hyperparameter tuning and training; and
[0212] iii) Use validation data to compare the performance of multiple models obtained through hyperparameter tuning, and determine the model with the best validation data as the best model.
[0213] In this invention, hyperparameter tuning is the process of optimizing the values of various parameters (number of convolutional layers, number of dense layers, number of convolutional filters, etc.) that constitute the CNN model. Bayesian optimization and grid search methods are used for hyperparameter tuning.
[0214] In this invention, during training, predetermined hyperparameters are used to optimize the intrinsic parameters (weights) of the CNN model, and when the validation loss begins to increase compared to the training loss, the model is determined to be overfitting. Training is stopped before this determination.
[0215] In this invention, any value obtained by analyzing the input vectorized data by an artificial intelligence model in the cancer diagnosis unit can be used without limitation, as long as it is a specific fraction or real number, and the value is preferably a depth probability index (DPI), but not limited thereto.
[0216] In this invention, the "depth probability index" refers to the value expressed as a probability value by adjusting the output of the artificial intelligence to a scale of 0 to 1, using the sigmoid function in the case of binary classification and the SoftMax function in the case of multi-class classification for the last layer of the artificial intelligence model.
[0217] In binary classification, a sigmoid function is used for training, adjusting the DPI to 1 to assume cancer has occurred. For example, when given breast cancer samples and normal samples, training is performed so that the DPI for breast cancer samples is close to 1.
[0218] In multi-class classification, the SoftMax function is used to extract as many DPIs as there are classes. The sum of the DPIs is adjusted to 1, and training is performed so that the DPI for the cancer type is effectively adjusted to 1. For example, assuming there are three classes—breast cancer, liver cancer, and normal—when a breast cancer sample is input, training is performed to adjust the DPI for the breast cancer class to approximately 1.
[0219] In this invention, the output value obtained by the cancer diagnosis unit is obtained for each type of cancer.
[0220] In this invention, the artificial intelligence model is trained to adjust the output value to approximately 1 if cancer is present, and to approximately 0 if cancer is absent. Therefore, performance (training, validation, and testing accuracy) is measured based on a cutoff value of 0.5. In other words, if the output value is 0.5 or greater, cancer is determined to be present, and if it is less than 0.5, cancer is determined not to be present.
[0221] Here, it will be obvious to those skilled in the art that the cutoff value of 0.5 can be arbitrarily changed. For example, to reduce false positives, the cutoff value can be set higher than 0.5 as a more stringent standard for determining the presence of cancer, and to reduce false negatives, the cutoff value can be set lower than 0.5 as a weaker standard for determining the presence of cancer.
[0222] Most preferably, the probability of DPI can be determined by using a trained artificial intelligence model to apply unseen data (including data containing solutions different from those obtained during training), thereby setting a cutoff value.
[0223] In this invention, the cancer type predictor predicts the cancer type by comparing the output results, and identifies the cancer type with the highest value in the output results as the cancer of the sample.
[0224] In another aspect, the present invention relates to a computer-readable storage medium comprising instructions configured to be executed by a processor for diagnosing cancer and predicting cancer type by means of the following steps:
[0225] (a) Extracting nucleic acids from biological samples to obtain sequence information;
[0226] (b) Align the obtained sequence information (reads) with the reference genome database;
[0227] (c) Vectorized data is generated using nucleic acid fragments based on the aligned sequence information (reads);
[0228] (d) Input the generated vectorized data into the trained artificial intelligence model, analyze the output values, and compare the output values with the cutoff values to determine the presence of cancer; and
[0229] (e) Predict cancer type by analyzing output values.
[0230] In this invention, step (a) may include obtaining previously generated sequence information, wherein the previously generated sequence information is obtained by extracting nucleic acids from a biological sample using an NGS device or the like.
[0231] In another aspect, the method according to this disclosure can be implemented using a computer. In one embodiment, the computer includes one or more processors coupled to a chipset. Furthermore, memory, storage devices, a keyboard, a graphics adapter, a clicking device, a network adapter, etc., are coupled to the chipset. In one embodiment, the performance of the chipset is obtained by a memory controller hub and an I / O controller hub. In another embodiment, the memory may be directly coupled to the processor, rather than the chipset. The storage device is any device capable of maintaining data, including hard disk drives, optical disc read-only memories (CD-ROMs), DVDs, or other storage devices. The storage device relates to data and instructions used by the processor. The clicking device may be a mouse, trackball, or other type of clicking device and is used in conjunction with a keyboard to transmit input data to the computer system. The graphics adapter displays images and other information on a monitor. The network adapter is connected to the computer system via a local area network or a remote communication network. However, the computer used herein is not limited to the above configurations, may not have some of the above configurations, may further include additional configurations, and may also be part of a storage area network (SAN), and the computer of the present invention may be configured to execute modules in a program for implementing the method according to the present invention.
[0232] As used herein, a module can refer to a functional and structural combination of hardware and software that implement the technical concept of the present invention. For example, it will be apparent to those skilled in the art that a module can refer to a logical unit of predetermined code and hardware resources that execute said predetermined code, and not necessarily to physically connected code or a type of hardware.
[0233] The method according to the invention can be implemented in hardware, firmware, or software, or a combination thereof. When the method is implemented in software, the storage medium includes any medium that stores or transmits data in a form readable by a device such as a computer. For example, a computer-readable medium may include read-only memory (ROM), random access memory (RAM), magnetic disk storage media, optical storage media, flash memory devices, and other electrical, optical, or acoustic signal transmission media.
[0234] In this respect, the present invention relates to a computer-readable medium comprising an execution module for executing a processor to perform operations including the steps described above according to the invention.
[0235] Example
[0236] The invention will be described in more detail below with reference to embodiments. However, it will be apparent to those skilled in the art that these embodiments are merely illustrative and should not be construed as limiting the scope of the invention.
[0237] Example 1. Next-generation sequencing of DNA extracted from blood.
[0238] Blood samples of 10 mL were collected from each of 184 healthy subjects and 580 cancer patients and stored in EDTA tubes. Within 2 hours of blood collection, the plasma was initially centrifuged at 1200 g and 4°C for 15 minutes, followed by a second centrifugation at 16000 g and 4°C for 10 minutes to separate the plasma supernatant (excluding precipitate). Cell-free DNA was extracted from the separated plasma using the Tiangen micro DNA kit, and library preparation was performed using the TruSeq NanoDNA HT library preparation kit (Illumina). Sequencing was then performed using a DNBseq G400 instrument (MGI) in 100-base paired-end mode. Approximately 170 million reads were generated per sample.
[0239] Example 2. Generation of GC map based on nucleic acid fragment distance
[0240] A GC map (vectorized) was generated using the NGS data produced in Example 1 above. The hg19 reference chromosome was divided into bins of 100 k bases, and the generated NGS reads were assigned to each bin. Then, the reciprocal of the median FD (fragment distance) for each bin was calculated, and an image representing the reciprocal of the previously calculated median FD for each bin location on the X-axis and the Y-axis was created. Figure 2 ).
[0241] Example 3. Construction and learning process of CNN model
[0242] The basic configuration of a CNN model is as follows: Figure 3 As shown in the diagram. The activation function uses ReLU (Rectified Linear Unit), and each convolutional layer uses 20 10×10 patches. The pooling mode used in this paper is the largest, using 2x2 patches. Five fully connected layers are used, and each layer includes 175 hidden nodes. Finally, the final DPI value is calculated using the sigmoid function value. The hyperparameter values used in the CNN model are obtained through Bayesian optimization, and the model configuration can be varied depending on the data used and the model optimization.
[0243] Example 4. Construction and performance testing of a deep learning model for cancer diagnosis using GC graphs based on nucleic acid fragment distance.
[0244] The performance of DPI values output from a deep learning model built using a GC graph based on distances between nucleic acid fragments was evaluated using the reads obtained in Example 1. All samples were divided into training, validation, and test groups. A model was built using the training samples, and then the performance of the model built using the training samples was evaluated using the validation and test group samples.
[0245] [Table 1]
[0246]
[0247] [Table 2]
[0248]
[0249] As a result, from Table 2 and Figure 4 It can be seen that the accuracy was 100%, 99.7%, and 90% in the training, validation, and test groups, respectively, and the AUC values indicating the ROC analysis results were 1.00, 0.95, and 0.938 in the training, validation, and test groups, respectively.
[0250] Figure 4 (A) shows the results of the analysis using receiver operating characteristic (ROC) curves to measure accuracy. Accuracy is determined to increase with increasing area under the curve (AUC). The AUC value is between 0 and 1, with an expected (baseline) AUC of 0.5 when randomly predicting label values and an expected (baseline) AUC of 1 when correctly predicting label values.
[0251] Figure 4 (B) is a box plot showing the cancer probability values (DPI values) calculated in the artificial intelligence model of the present invention for the normal sample group and the cancer patient sample group, where the red line represents the DPI cutoff of 0.5.
[0252] Example 5. Construction and performance testing of a deep learning model for cancer diagnosis using GC graphs based on nucleic acid fragment distance.
[0253] The performance of DPI values output from a deep learning model built using a GC graph based on distances between nucleic acid fragments was tested using the reads obtained in Example 1. All samples were divided into training, validation, and test groups. A model was built using the training samples, and then the performance of the model built using the training samples was evaluated using the validation and test group samples.
[0254] [Table 3]
[0255]
[0256] [Table 4]
[0257]
[0258] As a result, from Table 4 and Figure 5It can be seen that the accuracy was 100%, 91%, and 86.8% in the training, validation, and test groups, respectively, and the AUC values indicating the ROC analysis results were 1.0, 0.968, and 0.936 in the training, validation, and test groups, respectively.
[0259] Figure 5 (A) shows the results of the analysis using receiver operating characteristic (ROC) curves to measure accuracy. Accuracy is determined to increase with increasing area under the curve (AUC). The AUC value is between 0 and 1, with an expected (baseline) AUC of 0.5 when randomly predicting label values and an expected (baseline) AUC of 1 when correctly predicting label values.
[0260] Figure 5 (B) is a box plot showing the cancer probability values (DPI values) calculated in the artificial intelligence model of the present invention for the normal sample group and the cancer patient sample group, where the red line represents the DPI cutoff of 0.5.
[0261] Although specific configurations of the invention have been described in detail, those skilled in the art will understand that this description is provided for illustrative purposes and to set forth preferred embodiments, and should not be construed as limiting the scope of the invention. Therefore, the essential scope of the invention is defined by the appended claims and their equivalents.
[0262] Industrial applicability
[0263] The method for diagnosing cancer and predicting cancer type based on artificial intelligence according to the present invention includes generating vectorized data and analyzing it using an AI algorithm. Compared with the method of determining chromosome number based on read count (using read correlation values as standardized values one by one), it can still show similar effects despite low read coverage and is therefore useful.
Claims
1. A method for providing information based on artificial intelligence to diagnose cancer and predict cancer type, the method comprising: (a) Extracting nucleic acids from biological samples to obtain sequence information; (b) Align the sequence information reads with a reference genome database; (c) Vectorized data is generated from alignment-based sequence information reads using nucleic acid fragments. The vectorized data mentioned above is the Grand Canyon map, or GC map, which is generated by a method including the following steps: i) Divide the chromosomes into predetermined bins; ii) Calculate the distance between the aligned nucleic acid fragments in each bin, where the distance is the fragment distance FD; iii) Based on the fragment distances calculated in each bin, determine the representative distance between fragments in each bin, i.e., RepFD; iv) Normalize the representative distance between fragments calculated in step iii) by RepFD, which is the representative total nucleic acid fragment distance; as well as v) Create a GC graph, where the X-axis represents the order of the corresponding bins and the Y-axis represents the normalized values obtained in step iv). (d) Input the generated vectorized data into the trained artificial intelligence model, analyze the obtained output value, and compare the obtained output value with the cutoff value to determine whether cancer exists; as well as (e) Predict the cancer type by identifying the cancer type that shows the highest value in the output values as the cancer of the sample.
2. The method according to claim 1, wherein, Step (a) includes: (ai) Obtaining nucleic acids from blood, semen, vaginal cells, hair, saliva, urine, oral cells, amniotic fluid containing placental cells or fetal cells, tissue cells or mixtures thereof; (a-ii) Use salting out, column chromatography or beading to remove proteins, fats and other residues from the obtained nucleic acids to obtain purified nucleic acids; (a-iii) Prepare single-end or double-end sequencing libraries from the purified nucleic acid or nucleic acid randomly fragmented by enzyme digestion, pulverization or hydraulic shearing. (a-iv) Libraries prepared using next-generation sequencers; and (av) Obtain the sequence information reads of the nucleic acid described in the next-generation sequencer.
3. The method according to claim 1, wherein, The representative FD, or RepFD, includes at least one of the following groups selected from: the sum, difference, product, mean, median, quantile, minimum, maximum, variance, standard deviation, absolute deviation of the median, coefficient of variation, its reciprocal, and combinations thereof.
4. The method according to claim 1, wherein, The artificial intelligence model in step (d) is trained to distinguish between vectorized data of normal chromosomes and vectorized data of abnormal chromosomes.
5. The method according to claim 4, wherein, The artificial intelligence model is selected from the following group: Convolutional Neural Network (CNN), Deep Neural Network (DNN), Recurrent Neural Network (RNN), and Autoencoder.
6. The method according to claim 5, wherein, When the AI model is a CNN and learns binary classification, the loss function is represented by Equation 1 below; and when the AI model is a CNN and learns multi-class classification, the loss function is represented by Equation 2 below. Equation 1: Binary Classification Model (x) i = Output of the artificial intelligence model in response to the i-th input y = actual label value n = the number of input data Equation 2: Multiclass Classification Model (x) i ) j =Response to the j-th AI model output of the i-th input y = actual label value n = the number of input data c = the number of classes.
7. The method according to claim 1, wherein, In step (d), the output value obtained by the artificial intelligence model through analyzing the input vectorized data is the Depth Probability Index (DPI). The depth probability index is expressed as a probability value by the following: for the last layer of the AI model, the sigmoid function is used in the case of binary classification, and the SoftMax function is used in the case of multi-class classification, by adjusting the AI output to a scale of 0 to 1.
8. The method according to claim 1, wherein, The cutoff value in step (d) is 0.5, and cancer is determined to be present when the resulting output value is 0.5 or greater.
9. A device for diagnosing cancer and predicting cancer type based on artificial intelligence, the device comprising: A decoder configured to extract nucleic acids from biological samples and decode sequence information; The aligner is configured to compare the decoded sequence with a reference genome database; The data generator is configured to generate vectorized data from aligned sequence information reads using nucleic acid fragments. The vectorized data mentioned above is the Grand Canyon map, or GC map, which is generated through a method including the following steps: i) Divide the chromosomes into predetermined bins; ii) Calculate the distance between the aligned nucleic acid fragments in each bin, where the distance is the fragment distance FD; iii) Based on the fragment distances calculated in each bin, determine the representative distance between fragments in each bin, i.e., RepFD; iv) Normalize the representative distance between fragments calculated in step iii) by RepFD, which is the representative total nucleic acid fragment distance; as well as v) Create a GC graph, where the X-axis represents the order of the corresponding bins and the Y-axis represents the normalized values obtained in step iv). A cancer diagnosis unit is configured to input generated vectorized data into a trained artificial intelligence model, analyze the data, and compare the resulting output value with a cutoff value to determine whether cancer is present. as well as A cancer type predictor is used to analyze the obtained output values and thereby predict the cancer type by identifying the cancer type that shows the highest value in the output values as the cancer of the sample.
10. A computer-readable storage medium comprising instructions configured to be executed by a processor for diagnosing cancer and predicting cancer type by including the following steps: (a) Extracting nucleic acids from biological samples to obtain sequence information; (b) Align the obtained sequence information reads with the reference genome database; (c) Vectorized data is generated from alignment-based sequence information reads using nucleic acid fragments. The vectorized data mentioned above is the Grand Canyon map, or GC map, which is generated through a method including the following steps: i) Divide the chromosomes into predetermined bins; ii) Calculate the distance between the aligned nucleic acid fragments in each bin, where the distance is the fragment distance FD; iii) Based on the fragment distances calculated in each bin, determine the representative distance between fragments in each bin, i.e., RepFD; iv) Normalize the representative distance between fragments calculated in step iii) by RepFD, which is the representative total nucleic acid fragment distance; as well as v) Create a GC graph, where the X-axis represents the order of the corresponding bins and the Y-axis represents the normalized values obtained in step iv). (d) Input the generated vectorized data into the trained artificial intelligence model, analyze the output values, and compare the output values with the cutoff values to determine whether cancer exists; as well as (e) Analyze the obtained output values and thereby predict the cancer type by identifying the cancer type that shows the highest value in the output values as the cancer of the sample.
Citation Information
Patent Citations
Apparatus and method for recommending schedule to user using asscociation pattern learning
KR1020180124550A
Ultra-rapid DNA sequencing method with nano-transistors array based devices
US20060246497A1
Method and apparatus for molecular analysis using nanowires
US20060275779A1
Polony fluorescent in situ sequencing beads
US20070087362A1
Coherent electron junction scanning probe interference microscope, nanomanipulator and spectrometer with assembler and DNA sequencing applications
US20070194225A1
