System and method for automatically generating a base calling model

Abstract

The present application provides a system and method for automatically generating a model. The system for automatically generating a model comprises an input layer for receiving user input, a machine learning layer comprising a training data generation module and a training module, and an output layer for outputting the model. The training data generation module processes the user input to generate training data comprising pairs of associated features and labels, the features describing intensities corresponding to nucleic acid molecules at one or more locations, and the labels comprising one selected from the group consisting of bases A, T, C, and G. The training module trains a specified model based on at least a portion of the training data, including iteratively training the specified model using a training set and evaluating the performance of the specified model after each iteration of training using a validation set to obtain a base calling model. The system can automate and efficiently implement modeling, and can reduce the threshold for model development, so that users without machine learning expertise can develop models themselves.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence-based data processing, specifically to systems and methods for automatically generating base recognition models, and more specifically to systems and methods for generating base recognition models based on automatic training tools. Background Technology

[0002] The topics discussed in this section should not be considered prior art simply because they are mentioned here. Similarly, the technical problems mentioned in this section or related to the topics provided as background art should not be considered as having been previously recognized in the prior art. The topics in this section merely represent different methods, which themselves may correspond to specific embodiments of the technical solutions included in the claims.

[0003] In related technologies, sequencing generally involves detecting relevant signals to identify the type of nucleotides or bases linked to one or more nucleotide positions in a nucleic acid molecule; that is, determining at least a portion of the nucleotide sequence of the nucleic acid molecule through base calling. Changes in the signal and / or signal intensity corresponding to a specific position of a nucleotide linked to the nucleic acid molecule can indicate the type of base at that position. For example, different nucleotides can be labeled with different fluorescent molecules or optically detectable luminescent tags. In one round of sequencing, multiple nucleotides are contacted with the nucleic acid molecule to be tested, and then the luminescent signals are detected to distinguish the types of nucleotides or bases linked to the template in that round of sequencing. Typical examples include sequencing platforms from ILLUMINA or MGI. Alternatively, optical signal detection and nucleotide incorporation or linking can be separated into independent steps to achieve one round of sequencing, such as the sequencing platform from Element Bioscience; and others include Thermo Fisher's Ion... Torrent sequencing platforms do not rely on fluorescence or optical signals for detection. Instead, they utilize electrochemical sensors to detect pH changes caused by the release of H+ ions when nucleotides are incorporated into the nucleic acid molecule being tested, thus identifying the type of incorporated nucleotide or base. Other platforms, such as PacBio and Oxford Nanopore, are based on various nanopores for single-molecule sequencing. Compared to the sequencing wheels mentioned above, these platforms provide continuous sequencing. For example, they use zero-mode waveguide (ZMW) nanopores to observe the fluorescence signals emitted during the incorporation of phosphate-labeled modified nucleotides into a single-molecule template (the nucleic acid molecule being tested) to identify the type of incorporated base in real time. Alternatively, they may directly determine the base sequence of the nucleic acid molecule by physically detecting the electrical signals generated when different bases pass through the nanopore, without involving biochemical enzyme catalysis processes such as polymerization reactions. Currently, all commercially available sequencing platforms (also called sequencing systems or sequencers) include base recognition software adapted to the specific platform, model, or detection principle to determine the type of base in the nucleic acid molecule sequence based on the detection and analysis of relevant signals.

[0004] In related technologies, machine learning includes various algorithms, such as traditional machine learning algorithms like linear regression, decision trees, and support vector machines, as well as neural networks and deep learning. Applying machine learning to sequencing or gene detection fields, such as developing base recognition algorithms, base recognition result classification scoring, or quality score algorithms using machine learning, has been publicly disclosed in some cases. For example, Wei-Chun Kao, et al.: "naiveBayesCall: An Efficient Model-Based Base-Calling Algorithm for High-Throughput Sequencing", 2010-04-25, RESEARCH INCOMPUTATIONAL MOLECULAR BIOLOGY, SPRINGER BERLIN HEIDELBERG, BERLIN, HEIDELBERG, pp. 233–247, XP019141683, ISBN: 978-3-642-12682-6; CN118942549A, CN112789680, CN118429967A, CN118429965A, CN118116469A, CN117976042A and US10068053B, etc., are all incorporated herein by reference in their entirety.

[0005] Machine learning models used in common sequencing or related applications, including the model training, development, and usage processes outlined in the publicly available solutions, often involve significant manual intervention. For example, manual data collection and preprocessing, model selection, hyperparameter tuning, model training, and evaluation are frequently required in the early stages. This manual intervention inevitably leads to the need to configure and optimize machine learning models for different real-world scenarios, resulting in intensive manual operations. Intensive manual operations are generally prone to errors, inefficiency, and management difficulties. Furthermore, this process typically requires personnel with extensive relevant professional knowledge and experience, such as algorithm development experience or a professional background, especially when the machine learning model is complex and involves configuring and optimizing different types of algorithms. Summary of the Invention

[0006] The embodiments of this application are intended to at least solve one of the above-mentioned technical problems or at least provide a practical commercial option.

[0007] This application provides a system for automatically generating a base recognition model. The system includes: an input layer for receiving user input from a user, including sequencing data and optionally one or more tasks, wherein the sequencing data is obtained by detecting nucleic acid molecules linked to multiple locations on a surface using surface imaging sequencing technology; a machine learning layer including a training data generation module and a training module, wherein the training data generation module processes the user input from the input layer to generate training data, the training data including multiple pairs of associated features and labels, the features describing the intensity corresponding to nucleic acid molecules at one or more locations, the labels including a selection of bases A, T, C, and G; the training module training a specified model based on at least a portion of the training data, including iteratively training the specified model using a training set and evaluating the performance of the specified model after each iteration using a validation set, to obtain a base recognition model, wherein the training set and the validation set are each an independent, non-overlapping portion of the training data; and a result output layer for outputting the base recognition model. In related embodiments of this application, the system for automatically generating base recognition is also referred to as an automatic training tool.

[0008] This application also provides a method for automatically generating a base recognition model or developing a base recognition model. The method includes accessing user input using a system for automatically generating a base recognition model in any embodiment, and running the system using at least one hardware processor to present an optimal base recognition model to the user. The user input includes sequencing data and optionally one or more tasks. The sequencing data is obtained by detecting nucleic acid molecules linked to multiple locations on a surface using surface imaging sequencing technology.

[0009] This application also provides a base identification model generated by a system for automatically generating base identification models in any embodiment.

[0010] This application also provides a computer-readable storage medium storing a plurality of instructions for controlling a processor to execute a method for developing a base recognition model in any embodiment.

[0011] This application also provides a system for automatically generating a base recognition model. The system includes a memory storing a program and one or more processors coupled to the memory. The processors run the program to implement the method for developing a base recognition model in any embodiment.

[0012] This application also provides a sequencing system or sequencing platform that includes the system for automatically generating base recognition models as described in any of the above embodiments or examples.

[0013] The system, method, or related system for automatically generating base recognition models through any of the above embodiments automates the machine learning process, simplifies the modeling process, and reduces reliance on professionals such as algorithm developers, enabling non-machine learning experts, such as sequencing platform users without relevant algorithm development background or experience, to develop base recognition models independently.

[0014] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the embodiments of this application will become apparent and readily understood from the description of the embodiments in conjunction with the following drawings, wherein:

[0016] Figure 1 This is a schematic diagram of the system for automatically generating a base recognition model according to an embodiment of this application;

[0017] Figure 2 This is a schematic diagram illustrating how the entire image is converted into a set of blocks and the set of blocks is written as a matrix as input sequencing data in an embodiment of this application.

[0018] Figure 3 This is a schematic diagram of the system for automatically generating a base recognition model according to an embodiment of this application;

[0019] Figure 4 This is a schematic diagram of the system for automatically generating a base recognition model according to an embodiment of this application;

[0020] Figure 5 This is a schematic diagram of the system for automatically generating a base recognition model according to an embodiment of this application;

[0021] Figure 6 This is a schematic diagram of the system for automatically generating a base recognition model according to an embodiment of this application;

[0022] Figure 7 This is a schematic diagram of the system for automatically generating a base recognition model according to an embodiment of this application;

[0023] Figure 8 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0024] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0025] In this document, unless otherwise stated, the singular forms “a,” “an,” etc., include the plural referents (one or more). “A group” or “a plurality” refers to two or more.

[0026] In this document, terms such as "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number or order of the indicated technical features. In the description of this application, "multiple" means two or more, unless otherwise explicitly stated.

[0027] Unless otherwise stated, the terms "connected" and "linked" as used herein should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction between two components; they can refer to a connection through physical adsorption or other forces, or a chemical connection through chemical bonds, such as incorporation via polymerization reactions. Those skilled in the art will understand the specific meaning of these terms in the relevant examples based on the specific embodiments described, including the context and conventional understanding.

[0028] In this article, "sequencing" refers to nucleic acid sequencing, also known as "nucleic acid sequencing" or "gene sequencing," which involves determining the base sequence of the primary structure of a nucleic acid molecule. This can be achieved using sequencing by synthesis (SBS), sequencing by ligation (SBL), or sequencing by hybridization (SBH). Unless otherwise specified, sequencing by synthesis, as used in this article, includes not only the commonly understood SBS (typically ILLUMINA / Solexa technology), which uses polymerases to catalyze the incorporation of nucleotides into the nucleic acid molecule to be tested (polymerization reaction) and detects the corresponding reaction signals to identify the type of incorporated nucleotides, but also sequencing similar to SBS that uses polymerases or non-polymerases to controllably introduce or ligate nucleotides into the nucleic acid molecule to be tested, and directly or indirectly, simultaneously or sequentially detects the corresponding signals to determine the type of one or more ligated nucleotides. Examples include SBL, SBH, SBB (Sequencing by Binding), or SBE (Sequencing by Expansion), which use surface fluorescence imaging to achieve sequencing.

[0029] In the process of sequencing based on solid-phase surface signal detection, sequencing can also be described as measuring the intensity value corresponding to the location of one or more nucleic acid molecules. This intensity value, sometimes simply referred to as intensity, can be the intensity of any signal, such as electrical or electromagnetic radiation like visible light. A single base can have one intensity value, or multiple bases can have one or fewer intensity values. Furthermore, the intensity value can be specific to a location on the solid-phase surface or to multiple locations within a nucleic acid molecule. The intensity value can be converted or mapped to a predetermined numerical value (e.g., an integer in a binary or decimal number system), a continuous numerical value, or a range.

[0030] Sequencing can be performed using a sequencing platform. According to the embodiments of this application, the sequencing platforms that can be selected include, but are not limited to, Illumina's HiSeq, MiSeq, Nextseq and Novaseq series sequencing platforms, Thermo Fisher / Life Technologies' Ion Torrent platform, BGISEQ and MGISEQ / DNBSEQ platforms from BGI Genomics, Element Bioscience's AVITI and single-molecule sequencing platforms; the sequencing method can be selected as single-end sequencing, paired-end sequencing, or sequencing methods supported by the selected automated sequencing platform.

[0031] In this application's embodiments, "sequencing process" or "sequencing run" refers to the batch determination of intensity values ​​corresponding to the locations of one or more nucleic acids. For example, in a sequencing scenario involving multiple cycles (also called wheels or sequencing rounds) of imaging nucleic acid molecules on a substrate, such as a solid substrate surface, after a specified biochemical reaction, a series of intensity values ​​are obtained during the same sequencing run. Generally, nucleic acid molecules from the first sequencing run will not participate in a second sequencing run independent of the first sequencing run; for example, they will not be included in an image acquired at a certain time point (a certain round) of the second sequencing run.

[0032] In some examples, sequencing-by-synthesis (SBS) is used to perform multiple rounds of sequencing to obtain sequencing sequences or reads. For instance, the nucleic acid molecule to be tested is brought into contact with a polymerase and a modified nucleotide and placed under suitable conditions for polymerization. The modified nucleotide is controllably incorporated into the nucleic acid molecule to be tested, or in other words, single-base extension is controllably achieved. The corresponding reaction signal is detected, and the type of nucleotide incorporated into the nucleic acid molecule in that reaction is determined based on this signal. This process of controllable single-base extension and corresponding signal detection is repeated multiple times to detect the type of nucleotide or base incorporated into the nucleic acid molecule to be tested in multiple or multiple rounds of reactions, thereby reading a portion of the sequence of the nucleic acid molecule to be tested.

[0033] The nucleic acid molecule to be tested, also known as the nucleic acid template or template, can be an unamplified single molecule or an amplified molecular cluster or long chain containing multiple identical polynucleotide molecules, such as fluorescent clusters or DNA nanospheres (DNBs) formed by bridge amplification or rolling circle amplification used by current mainstream commercial sequencing platforms. The nucleic acid molecule to be tested can be single-stranded, double-stranded, and / or a triple-stranded or more-stranded complex hybridized with probes or primers.

[0034] The corresponding reaction signals can be, for example, fluorescence signals, which can be represented as image data (e.g., color or grayscale images) formed by collecting these fluorescence signals. These image data are then processed and analyzed to detect the nucleotides incorporated into the nucleic acid molecule to be tested in each or each round of reactions, so as to determine a portion of the base sequence of the nucleic acid molecule to be tested.

[0035] Specifically, in some examples, sequencing is achieved based on surface fluorescence imaging detection. The nucleic acid molecule to be tested is attached to a solid surface. For example, the nucleotide can be modified to have or be able to bind a fluorescent label, as well as a removable inhibitor group that prevents other nucleotides from polymerizing and attaching to the next position of the nucleic acid molecule to be tested (this type of modified nucleotide is also called a reversible terminator). After each polymerization reaction or single base extension reaction, the fluorescent label is excited to emit light, and these emission signals are collected to obtain an image of the nucleic acid molecule to be tested that has undergone a single base extension reaction at a specified surface position. Then, the inhibitor group and fluorescent label are removed to perform the next polymerization reaction and signal acquisition (photographing). This process of polymerization reaction-photographing-removal is repeated multiple times or multiple rounds to obtain image information related to the nucleotides attached to the nucleic acid molecule to be tested in each single base extension reaction.

[0036] Understandably, when a nucleic acid molecule to be tested undergoes a polymerization reaction at a designated location on the surface, it emits fluorescence. This fluorescence typically appears as a bright spot or bright patch with a higher intensity than the background signal at the corresponding location in the image acquired in that round of reflection. Therefore, based on the information contained in these image sets corresponding to specific chemical characteristics (the nucleic acid molecule to be tested undergoing a polymerization reaction), such as intensity and / or morphology, it is possible to determine whether a nucleic acid molecule to be tested exists at the designated location and whether it has undergone a polymerization reaction. Furthermore, by combining the pre-defined correspondence between the fluorescence emission signal and the nucleotide type, the type of nucleotide that has undergone a biochemical reaction and is linked to the nucleic acid molecule to be tested can be detected. This allows for the determination of at least a portion of the sequence of the nucleic acid molecule to be tested, obtaining what is known as a read.

[0037] The set of locations corresponding to chemical features (nucleic acid molecules to be tested) on a surface is determined based on an image, i.e., template acquisition. This can be determined by identifying features on the detection image that correspond to the chemical features, such as bright spots (locations), or by identifying other features on the image that have a specific spatial relationship with the chemical features. For example, for a regular array surface containing labeled regions (also often called tracer regions), the spatial relationship between the labeled regions and the reaction regions (the areas where the nucleic acid molecules to be tested are located) is generally known beforehand, and the signals from the labeled regions are clearly distinguishable from the signals from the reaction regions, or the signals from the labeled regions appear as detectable features on the image. By identifying the signals from the labeled regions and determining the location of the labeled regions on the image, the location of each amplicon (nucleic acid molecule to be tested) in the reaction region can be determined.

[0038] It should be noted that the term "nucleotide" as used herein includes ribonucleic acid or deoxyribonucleic acid, including natural nucleotides or their derivatives or modifications thereof (also called modified nucleotides or altered nucleotides, etc.). In this document, the term "nucleotide" is sometimes used to refer to the bases contained within it, which will be readily understood by those skilled in the art based on conventional understanding and / or context. Furthermore, the terms "nucleic acid sequence," "nucleotide sequence," "oligonucleotide sequence," or "base sequence" as used herein are interchangeable unless otherwise specified, and are sometimes simply referred to as "sequence." Additionally, sometimes one or more nucleotides or bases are also referred to as a sequence, which will be readily understood by those skilled in the art based on conventional understanding and / or context.

[0039] In some embodiments, a single round of sequencing may include one base extension reaction (one repeat). For example, four different nucleotides (dATP, dTTP, dGTP, and dCTP) may be placed in the same polymerization reaction system with multiple nucleic acid molecules to be tested, so that each nucleotide can be excited to emit a signal that is distinguishable from other types of nucleotides. Thus, the type of nucleotide incorporated or introduced at a position in any nucleic acid molecule to be tested can be determined by the information obtained from one repeated reaction. For example, the four nucleotides may be labeled with four different fluorescent bands for four-color or four-channel single-molecule sequencing or high-throughput sequencing; or, for example, the four nucleotides may be labeled with three different fluorescent bands and unlabeled (cold nucleotides) for two-color or three-color high-throughput sequencing. One round of sequencing includes one repeat, and one round of sequencing can detect the type of base at a position on any nucleic acid template.

[0040] The term "amplifier" refers to the amplified nucleic acid molecule to be tested. An amplifier is a cluster, chain, or group containing multiple identical polynucleotide sequences. The term "fluorescent cluster, group, or sphere" refers to an amplifier that can be excited to emit light, such as fluorescence, after undergoing a specified biochemical reaction. During sequencing, amplifiers can emit light and be detected by imaging. By processing the signals acquired during imaging and / or the corresponding image information, bases incorporated into or linked to the nucleic acid molecule to be tested after the specified biochemical reaction can be detected, thereby determining at least a portion of the sequence of the nucleic acid molecule to be tested.

[0041] In the embodiments of this application, unless otherwise stated, the terms amplicon, fluorescent cluster or sphere, nucleic acid molecule cluster, clone cluster, nucleic acid molecule to be tested, template, nucleic acid template, template spot, bright spot corresponding to chemical features or fluorescent clusters on the surface, etc., are generally used interchangeably. The terms signal intensity, intensity, brightness, and fluorescence brightness, etc., are also generally used interchangeably. The terms surface, chip, array, etc., are generally used interchangeably in the embodiments of this application.

[0042] The machine learning model referred to in the embodiments of this application, also simply as a model, refers to a determination technique that predicts the type of output bases or base combinations based on known results (training data). In some embodiments, it also includes a technique for predicting the type of output bases or base combinations with a quality score, whereby the quality score characterizes the accuracy, error rate, or reliability of the predicted bases or base combinations. The known results can be sequences or bases that are considered or assumed to be accurate, i.e., assumed to be correct. In some embodiments, the assumed accurate sequence or base is the result predicted by the expected model, sometimes also referred to as the corresponding benchmark true base. Supervised learning of the model can be performed using training data.

[0043] In the machine learning embodiments of this application, the concepts of label or target value are generally synonymous, referring to the result that the model aims to predict. The bases or base sequences in the label or target value in machine learning are assumed to be accurate sequences (bases or base combinations), considered accurate sequences, and sometimes referred to as benchmark true bases. The labels included in the training data may be inaccurate, but the model assumes they are accurate during training. The assumed sequence can be determined in various ways, as described in the embodiments of this application: using a reference sequence, such as the bases or base combinations corresponding to the positions of sequencing results aligned to the reference sequence; using sequences determined or verified by generally considered high-accuracy detection methods or means, such as Sanger sequencing and / or DNA synthesis; using consistent results from multiple detection methods or means; or using sequences or base combinations with a high probability of accuracy, such as bases with a quality fraction reaching or exceeding Q40, or Q40 reaching or exceeding 80, 85, or 90. Here, Q40 represents an error rate or probability of error of one in ten thousand. If the training model outputs highly accurate predictions, which is not the desired goal, the less accurate measurement results can be used as the hypothetical sequence and as training data. However, it is understandable that the accuracy of the prediction results of the model learned from this is generally not high.

[0044] The base identification referred to in the embodiments of this application, sometimes also called base category identification, base detection, base category detection, base readout, base determination, or base interpretation, refers to the determination of bases or base combinations at specified or unspecified positions in nucleic acid molecules. Base determination can be performed independently or as part of a specified base such as A or T, or a specified base combination such as ACT or GTCA. Base determination can be performed for the same genomic location (e.g., the scores or probabilities of detecting multiple bases at that location are similar) or for different locations. The base type at the specified or unspecified position can be determined using the score output from a machine learning model. For example, a score can be provided or assigned to each base or base combination, and base determination based on the score can be part of the model prediction. In addition, some models can provide scores, which can be used in subsequent processes. The score or value can reflect probability or likelihood. The probability score for each base or base combination is usually summed to a fixed value, such as 1; while the score reflecting likelihood generally does not need to be summed to a fixed value, although it can be summed to a fixed value, for example, by limiting the likelihood score for each to between 0 and 1, so that the likelihood score can also be summed to 1. In the embodiments of this application, the numerical values ​​or scores that reflect the probability or likelihood of each base or base combination are provided or assigned are sometimes collectively referred to as base recognition probability.

[0045] Furthermore, in the process of learning the quality scoring algorithm or quality scoring model or related model involved in certain embodiments of this application, the accuracy or reliability of the base determination results is scored based on the quality scoring model generated by training, so as to quantitatively assess or evaluate the accuracy or reliability of such base determination results. It can be used to compare the sequencing quality of different sequencing methods or sequencing platforms, and to determine whether the accuracy of the bases or sequences identified in a certain round or multiple rounds of sequencing or one or more sequencing runs meets expectations, and whether to continue the sequencing run, etc.

[0046] Please refer to Figure 1 Some embodiments provide a system 100 for automatically generating a base recognition model. The system 100 includes: an input layer 120 for receiving user input from a user, including sequencing data and optionally one or more tasks, wherein the sequencing data is obtained by detecting nucleic acid molecules linked at multiple locations on a surface using surface imaging sequencing technology; a machine learning layer 140 including a training data generation module 142 and a training module 144, wherein the training data generation module 142 processes the user input from the input layer 120 to generate training data, the training data including multiple pairs of associated features and labels, the features describing the intensity corresponding to nucleic acid molecules at one or more locations, the labels including a selection of bases A, T, C, and G; the training module 144 for training a specified model based on at least a portion of the training data, including iteratively training the specified model using a training set and evaluating the performance of the specified model after each iteration using a validation set to obtain a better base recognition model, wherein the training set and validation set are each independently non-overlapping portions of the training data; and an output layer 160 for outputting the better base recognition model. In related embodiments, the system 100 for automatically generating base recognition is also referred to as an automatic training tool.

[0047] The user input received by input layer 120 may include sequencing data, which can be one or more nucleic acid samples, data obtained through one or more library constructions, or data obtained through one or more sequencing runs on one or more sequencing platforms. Examples include electrical or electromagnetic radiation (e.g., visible light) detected during sequencing, and data that can be represented as images or image data or other forms or formats. Sequencing data can be obtained through self-sequencing, provided by others, publicly available data, or a mixture or combination of sequencing data from various sources. In some embodiments, the sequencing data provided by the user input may be in the form of an access link or path to reduce storage requirements and facilitate system 100's reception and connection to read the data. The model output by output layer 160 may also be in the form of an access path or link, for example, stored or saved in a specified path for easy retrieval and execution.

[0048] In some embodiments, the sequencing data originates from a sequencing platform that performs sequencing based on surface imaging detection. The sequencing data includes images of nucleic acid molecules at multiple locations on one or more surfaces, collected during one or more sequencing runs. The term "image" or "sequencing image" includes images directly acquired during sequencing, images acquired after preprocessing, or images reconstructed or recombined from images acquired during sequencing. In this embodiment, these images are sometimes referred to as "raw images." Subsequently, the training data generation module 142 processes this input image data, for example, extracting the position and intensity of each nucleic acid molecule in the image as feature values.

[0049] In some embodiments, sequencing data is a subset of data extracted from a sequencing image. The sequencing data includes the positions and intensity values ​​of multiple nucleic acid molecules on one or more surface regions. The sequencing image is an image of nucleic acid molecules at multiple locations on one or more surfaces obtained from one or more sequencing runs. In some examples, sequencing data is information obtained after image processing, containing information about the target nucleic acid molecules or amplicones in the image that represent or reflect the target signal, i.e., the nucleic acid molecules or amplicones that have undergone a specified biochemical reaction, such as the position and intensity values ​​of the amplicon. The size of the input sequencing data is smaller than the size of its corresponding sequencing image. The term "size," also referred to herein as file size, dimensions, or data volume, reflects the storage space required or occupied by the data. Thus, compared to the case where the input data is a sequencing image, this input can be directly used as feature values ​​or directly as part of the training data, giving the model generation using this system significant advantages in terms of computational cost and speed.

[0050] Specifically, the input sequencing data includes the position and intensity information of the amplicon on the image, and the size of the input data is smaller than the size of its corresponding image. Therefore, from a certain perspective, the input data can be considered compressed image information or data after extracting or filtering image information. It can be represented in a simplified image form or a non-image form, preferably in a data form that is easy for a computer or processor to process. In a preferred example, it is represented in a non-image data form, such as a matrix, multidimensional matrix, array, or multidimensional array, recording or presenting only the relative position and intensity information of the amplicon on the image. This significantly reduces the file size of the input data compared to the original input image, lowering the requirements for computing power and / or storage, and facilitating rapid computation or processing by the computer to quickly obtain prediction results.

[0051] In some specific embodiments, the size of the input sequencing data is less than or equal to half the size of its corresponding image. In some examples, particularly for regular surfaces, the file size of the input data is less than or equal to one-third the file size of its corresponding image. In some tests, particularly for high-density regular surfaces, the size of the input data is even less than one-quarter or one-fifth the file size of its corresponding image. Thus, compared to directly inputting image data, storage or space usage is significantly reduced, and running speed is significantly improved.

[0052] Specifically, in some examples, the input sequencing data is represented as multiple matrices. Each matrix comprises multiple submatrices, reflecting the position and intensity information of the amplicon on the images acquired in a round of sequencing. A single matrix contains multiple rows and columns, reflecting the information of the amplicon on all or a portion of a fluorescence channel image in a round of sequencing. Each element of the matrix corresponds to one amplicon, the row and column positions of the matrix elements reflect the position information of the corresponding amplicon on the corresponding image, and the values ​​of the matrix elements reflect the intensity information of the corresponding amplicon presented on the fluorescence channel image in that round of sequencing. In some embodiments, this matrix is ​​also referred to as a fluorescence intensity matrix. Thus, compared to the input image, the amount of information or size of the input sequencing data is greatly reduced, which helps to reduce the requirements for storage or computing power and improves the speed of computation. Moreover, such input sequencing data can be directly used as feature values ​​or part of feature values, reducing subsequent calculations.

[0053] In a more specific example, a single sequence of sequencing data can correspond to multiple amplicons in a region or field of view (FOV). A single sequence of data, such as one or a set of fluorescence intensity matrices, can correspond to the location and intensity information of amplicons that have undergone a specified reaction in an image acquired during a single round of sequencing of a field of view. Thus, compared to the original input image, the size of the input data is significantly reduced, lowering the requirements for computing power and / or storage.

[0054] Regarding the number and dimension of a set of matrices, it can be understood that a set of matrices corresponding to a set of images from a single FOV (Field of View) sequencing run of multi-fluorescence channels, such as the four fluorescence intensity matrices of amplicons from a single FOV sequencing run of four-color sequencing, the three fluorescence intensity matrices of amplicons from a single FOV sequencing run of three-color sequencing, and the two fluorescence intensity matrices of amplicons from a single FOV sequencing run of two-color high-throughput sequencing, can all be referred to as a three-dimensional matrix. Based on conventional understanding, those skilled in the art, according to the specific embodiments described, will understand the specific meaning of the descriptions involving matrices, the number of matrices, or their dimensions.

[0055] Furthermore, in some examples, the advantage of the input matrix compared to the original input image goes beyond just the difference in image and matrix size. For instance, in a single round of sequencing for four-color fluorescence imaging, assuming each image is 2000*2000 pixels, and the template point matrix (fluorescence brightness matrix) reflecting the position and intensity information of the amplicon in that image, written in matrix form, is 800*800 pixels, then the size is reduced to 0.4^2 times the original. Moreover, the brightness information corresponding to these 800*800 template points in the input matrix can be the brightness information corresponding to sub-pixel level positioning, such as brightness obtained through bilinear interpolation or other interpolation methods. Therefore, in terms of information content, this 800*800*4 template brightness matrix reflects the information of the four images: a 2000*2000*4 matrix plus an 800*800 position matrix. In other words, the information content of the input matrix reflects the position and intensity information of the amplicon, determined directly or indirectly based on image information, including information from the processed image. In this way, the input data not only contains target information and its size is significantly reduced, but it is also presented in a matrix form that is easy for machines to run and compute, further facilitating rapid prediction and obtaining accurate and reliable base identification results. Moreover, such input sequencing data can be directly used as feature values, or as part of the features or training data, to reduce subsequent calculations.

[0056] In other specific examples, the sequencing data is in matrix format. A matrix reflects the information of amplicon amplicons in a block of a fluorescence channel image during one round of sequencing. The input sequencing data also includes matrices reflecting the location information of each block within the image from which it originates. Thus, transforming the matrix corresponding to a single fluorescence channel image into a set of relatively small matrices, or in other words, transforming a fluorescence channel image into a set of blocks and extracting the amplicon information from simplified blocks into matrix form, allows for parallel processing of these relatively small matrices. Under the same hardware and software computing environment, this further improves the processing speed of input data and enables faster model training.

[0057] Please refer to Figure 2 In a specific example, an image is converted into a set of block matrices. Block matrices of different sizes are all zero-padded to the same size and then used as input. Therefore, the entire image is converted into a total of 3*3, or 9 blocks. Since the length and width of the blocks may be different, after splitting into 9 blocks, the edges of all blocks smaller than b*y can be padded with 0 values ​​to the size of b*y. Finally, an additional position encoding layer is added according to the position number of the block in the whole image. For example, the block number in the lower right corner is 9, so a Pos(P) position encoding layer is filled with all values ​​of 9 to obtain the final input matrix.

[0058] In some embodiments, the sequencing data can reflect at least 10 million sequence reads (also known as segment reads) and / or at least 1 billion base pairs (bp) of base reads. This generates sufficient training data to potentially generate a model with better predictive performance based on the system 100.

[0059] In some embodiments, in machine learning layer 140, training data generation module 142 processes user input from input layer 120, including sequencing data, to generate training data. The generated training data includes multiple pairs of associated features and labels. The features describe the intensity corresponding to one or more nucleic acid molecules at different locations. These features are, for example, sequencing data as described in any of the above embodiments, such as sequencing images, processed sequencing images, matrices corresponding to image-specific location information (amplifier location and intensity information), or block matrix groups corresponding to image-specific location information. The intensity, sometimes also called brightness, can be represented by the pixel / grayscale value, sub-pixel value, or sub-pixel value at the location of the nucleic acid molecule in the image. For example, if a feature of a nucleic acid molecule at a certain location occupies one or more pixels or multiple sub-pixels in the image, the intensity information of the nucleic acid molecule at that location can be represented by the intensity value of any pixel or sub-pixel location. Alternatively, the center or centroid of the feature can be determined, and the pixel or sub-pixel value at the center or centroid can represent the intensity information of the nucleic acid molecule at that location in the image. The intensity information of nucleic acid molecules at one or more locations in an image can be the original signal intensity value, the processed signal intensity value (e.g., after background removal and / or after at least one correction for chromatic aberration, crosstalk, and phase), or the intensity value after normalization or standardization. Crosstalk correction, phase correction (phasing or prephasing correction), etc., can be performed using publicly available methods such as those disclosed in EP3077943B1 and CN113012757B, the full contents of which are incorporated herein by reference. The term "tag" includes four bases selected from A, T, C, and G.

[0060] In some examples, the training data includes multiple pairs of associated features and labels. The features may be sequencing data or processed sequencing data formed by any relevant contextual embodiment or combination of relevant embodiments, or a portion thereof. The base or base combination category corresponding to the sequencing data (the corresponding benchmark true base) may be used as a label.

[0061] This application does not limit the method or approach for generating training data. Specifically, conventional base recognition methods, software, or tools, such as base recognition software typically carried or provided by various sequencing platforms / sequencers, can be used to identify the bases incorporated into the template nucleic acid in each round, thereby obtaining the base recognition results for each round or determining at least a portion of the sequence of the nucleic acid molecule to be tested. In related examples, the so-called benchmark true bases or reference answers are derived from user input or determined by processing the sequencing data and tasks input by the user. In some specific examples, the training data generated by the training data generation module 142 uses the bases on the reference sequence corresponding to the base recognition results of the amplicon as labels or target values. The so-called labels or target values ​​are sometimes also referred to as true results, correct answers, true labels, or benchmark true bases in this application embodiment.

[0062] In one embodiment, the user input also includes a reference sequence. The reference sequence is a known sequence, which may be a reference genome, a portion of a reference genome, or a known sequence formed by recombination and reorganization of one or more reference genomes. Please refer to... Figure 3 The training data generation module 142 also includes an alignment submodule 1422, which is used to align the sequence readout or base readout reflected by the sequencing data to the reference sequence in order to generate alignment results; and to determine the labels associated with the features based on the alignment results, for example, using the corresponding bases of the sequence or base readout corresponding to the sequencing data aligned to the reference sequence as the corresponding labels in order to generate training data.

[0063] In a specific example, the alignment submodule 1422 is connected to the sequencing system or sequencing platform, and to the base recognition software provided or carried by the sequencing system or sequencing platform itself.

[0064] Furthermore, during the training data generation module 142's construction of training data, sequences with poor sequencing quality or unreliable or low-reliability alignment positions can be removed to improve the reliability of the generated training data. Training a model based on highly reliable training data is beneficial for obtaining a more accurate and reliable prediction model. For example, in a certain example, for sequences that failed to align successfully to the reference genome or whose alignment results are unreliable or have low-reliability, these sequences or bases can be reassigned as "N" or marked as "N," and these bases or sequences marked as "N" can be excluded from training data or subsequent model training processes. This can improve the reliability of the generated training data or the model trained based on such training data.

[0065] In some embodiments, the so-called tag also includes combinations of any two, three, or four bases selected from A, T, C, and G. In one example, the so-called tag also includes combinations formed by the current round and the previous round of base identification, such as AA, AT, AC, AG, TT, TA, TC, TG, CC, CA, etc. In another example, the so-called tag also includes combinations formed by the current round and the base identification of the rounds before and after, such as AAA, ATG, ATT, ATC, ACG, ACT, ACA, TAT, CGA, GTA, etc. In yet another example, the tag also includes combinations formed by the current round, the previous round, and the two subsequent rounds of base identification, such as AATT, AATG, AATC, AAAT, CGTA, GTAC, TGAC, ACTG, etc. In a specific example, the sequence or base readout combination corresponding to the sequencing data is aligned to the corresponding base combination on the reference sequence as such a combination-type tag. Training a model based on this type of labeled training data typically requires high computing power and memory configuration.

[0066] The designated model can be any machine learning model, including deep learning neural networks. In some embodiments, the designated model trained in training module 144 is selected from at least one of Naive Bayes, logistic regression, support vector machines, tree models such as LightGBT, and neural networks. In some specific examples, the designated model includes neural networks such as models capable of semantic segmentation, such as U-net or its variations, and can be a single model or a cascaded model. In a specific example, the designated model is trained using the scheme disclosed in CN119479831A, the entire text of which is incorporated herein by reference.

[0067] Please refer to Figure 4 In some embodiments, the machine learning layer 140 further includes an evaluation module 146, which is connected to the training data generation module 142, the training module 144, and the output layer 160. The evaluation module 146 evaluates the performance of the base recognition module using a test set before the output layer 160 outputs the base recognition model, in order to determine whether the base recognition model meets preset requirements and can be output. The test set is a portion of the training data that does not overlap with the training set and validation set. Meeting the preset requirements means that at least one of the following metrics—accuracy, precision, recall, and speed—of the base recognition model's classification prediction results on the test set meets a preset standard, or that the model can perform the user-input task.

[0068] Please refer to Figure 5In some embodiments, the training data generation module 142 is a first training data generation module 142, the training data is first training data, the training module 144 is a first training module 144, the designated model is a first designated model, the so-called feature is a first feature, and the so-called label is a first label. The machine learning layer 140 also includes a second training data generation module 152 and a second training module 154. The second training data generation module 152 connects the first training data generation module 142 and the first training module 144, and is used to process at least a portion of the first training data through a base recognition model to generate a first prediction result, in order to generate second training data. The first prediction result includes the probability that the data is predicted to be identified as a base category in the first label and the base category. The second training data includes multiple pairs of associated second features and second labels. The second features describe the intensity corresponding to one or more nucleic acid molecules and / or describe the probability of at least one base category (at least one base recognition probability). The second labels include two categories: the first prediction result and the corresponding benchmark true base category are consistent or inconsistent. The second training module 154 is used to train a second specified model based on at least a portion of the second training data. This includes iteratively training the second specified model using the second training set and evaluating the performance of the second specified model after each iteration using a second validation set, in order to obtain a quality score model. The quality score output by the quality score model represents the accuracy, precision, or error rate of the base recognition results predicted by the base recognition model from the first training module 144 or the output layer 160; the second training set and the second validation set are each an independent, non-overlapping portion of the second training data. Thus, training data (second training data) for training the quality score model is generated using the generated base recognition model and its prediction results, and model learning is performed based on this training data to generate the quality score model.

[0069] In related technologies, the quality score or quality rating (Q-value) of base identification / classification results, regardless of the calculation method, rules, or custom rules used by different sequencing platforms from various companies, can generally be considered another quantitative representation of the probability of a correct or incorrect base detection result. It is an indicator of the accuracy or reliability of sequencing results. A higher Q-value indicates higher accuracy and reliability of the sequencing results. The Q-value is usually calculated based on the Phred algorithm, and the formula can be written as Q = -10 * log0 10The Q value (err) refers to the probability of an error in base identification, i.e., the error rate or probability of the base detection result. For example, common Q values ​​like Q20, Q30, and Q40 represent error rates of 1 percent (99% accuracy), 1 / 1000 (99.9% accuracy), and 1 / 10000 (99.99% accuracy), respectively. The Q value provides a direct understanding of the accuracy and reliability of sequencing data. Furthermore, in various practical applications that utilize sequencing data or results for detection, such as non-invasive prenatal screening for genetic variations, liquid biopsy for early tumor screening, and pathogen detection, Q values ​​are typically used to process sequencing data or set quality standards to ensure high-quality sequencing data or meet specific requirements, thereby guaranteeing the accuracy and reliability of the application's detection results.

[0070] In some examples, the features used to train the base recognition model (the first feature) can be used as the second feature. For example, the second feature also describes the intensity corresponding to one or more nucleic acid molecules at one or more positions. The specific value of the second feature can be determined by referring to the sequencing data or associated (first) feature values ​​in any of the above-mentioned related embodiments. Alternatively, the second feature value can be determined based on the prediction results of the base recognition model. For example, the second feature can describe the probability of at least one base class (base recognition probability), such as the probability values ​​of all base classes, the probability value of the base class with the highest probability, and the probability values ​​of the base classes with the highest and second-highest probabilities. The second feature can also describe both the intensity corresponding to one or more nucleic acid molecules at one or more positions and the probability of at least one base class. The base class referred to here is the same as the label or first label in the above or related embodiments. It can be a base class formed by a single base (A, T, C, G), or a base class formed by a single base combined with two or more bases, such as AT, ACG, etc.

[0071] Specifically, in some examples, the second feature describes the probability of at least one base category in the first prediction result, that is, the second feature describes the probability of identifying at least one base. For example, the second feature may be the probability value of identifying each base category, including base combination categories, in the first label; it may also be the probability value of the base category with the highest probability; the probability values ​​of the largest and second largest probability base categories; the probability values ​​of the top three base categories sorted in descending order of numerical value; the probability values ​​of the top four base categories sorted in descending order of numerical value; or any one or a combination of the above, or a processed or modified probability value, etc. The probability or probability value referred to here and in any related embodiment may be an absolute value, a relative value, a ratio, a modified value after normalization, or a value amplified or reduced based on some rule. In a specific example, the second feature describes the probability of multiple base categories in the first prediction result.

[0072] In some examples, the corresponding benchmark true base category is the base category that has been identified as correct. In a specific example, the corresponding benchmark true base category comes from sequences of human and non-human samples that have been consistently verified and fully characterized by multiple sequencing instruments, sequencing chemistry principles, and sequencing protocols. This can be considered the correct or true base category, or the accurate and reliable base category confirmed by the relevant detection methods to date. In another example, the corresponding benchmark true base category is the label or first label (target value or first target value) used in the process of training the base recognition model.

[0073] In some examples, the second specified model is selected from at least one of logistic regression, support vector machine, and tree model.

[0074] In some specific examples, training the second specified model in the second training module 154 further includes obtaining the classification probability of the base category of the first prediction result in the second training set; dividing the first prediction result into groups 'a' based on the classification probability and / or base category, and for each group, determining the error rate of the base category of the first prediction result of that group by comparing the base category in the first prediction result of that group with the corresponding benchmark true base category, where 'a' is a natural number and the value of 'a' is such that the logarithmic difference in the error rates of the base categories of the first prediction results of adjacent groups is less than 2; and associating a predetermined quality score with the error rate of the base category. Thus, a correspondence is established between the classification probability of each base identification result / base category and the predetermined quality score, which is the so-called quality score model.

[0075] Specifically, the first prediction result can be used to obtain the so-called classification probability P value, denoted as P0, through the quality score model of any of the above-mentioned related embodiments. P0 is a probability value between 0 and 1. Further, P0 can be divided into a segments, and the base categories in the first prediction result will be correspondingly divided into these a segments or groups, where a is a natural number not less than 5. In some specific examples, P0 is divided into a segments in an equal manner, for example, P0 is divided into 5 segments, then the ranges of these 5 segments of P0 are [0, 0.2), [0.2, 0.4), [0.4, 0.6), [0.6, 0.8), and [0.8, 1.0], respectively.

[0076] Furthermore, for each group, the base categories in the first prediction result of each group are compared with the corresponding benchmark true base categories to determine the error rate or accuracy rate of the base categories in the first prediction result of that group. This includes identifying or marking bases in the group that match the corresponding benchmark true base categories as correct, and identifying or marking those that do not match as incorrect. In this way, the accuracy rate or error rate of the base categories in that group can be calculated.

[0077] In the relevant examples, the Q value corresponding to the logarithm of the true error rate of a group or segment of bases is called the true Q value. For example, as mentioned above, the base categories in the first prediction result are divided into segments 'a' based on the P0 value. Then, the total error rate of the bases in each segment can be calculated, and the corresponding Q value can be calculated based on this total error rate. This total error rate is called the true error rate, and the corresponding Q value is called the true Q value. Specifically, following the example above, all bases (base categories) with P0 in the range [0, 0.2) are identified. Using the total number of bases in this segment as the denominator and the number of bases whose base identification results differ from or are inconsistent with the reference answer (benchmark true bases) as the numerator, the total error rate of the base category corresponding to P0 in the segment [0, 0.2) is calculated, and the logarithm of this total error rate is taken to convert it into a Q value. In this way, the true Q value of the base category corresponding to this segment is determined. This true Q value can be further assigned to each base in this segment, thus determining the true Q value or true error rate of each base category in this group or segment.

[0078] In some preferred examples, the setting or adjustment of the value of 'a' follows the principle that after dividing P0 into 'a' segments, the difference in the true quality value (true Q value) of the corresponding base categories of adjacent segments is not significant. Specifically, a difference of less than 2 in the true quality value of the base categories of adjacent segments can be considered as not significant. For example, a difference not greater than any value in the range [0.5, 2) can be considered as "not significant," such as a difference less than or not greater than 0.5, 0.8, 1, 1.1, 1.2, 1.5, etc. More preferably, a difference of less than 1 in the true Q value of base detection of adjacent segments is considered as not significant. If this preset limit is exceeded, P0 can be re-subdivided or further subdivided, i.e., a larger value of 'a' can be taken. After multiple tests following this principle, the value of 'a' is usually much greater than 10. The inventors have established these preferred principles or related settings or value rules by comprehensively considering the quality requirements of sequencing data for various detection applications and by testing and verifying sequencing data obtained from sequencing of various samples, including human and non-human samples, using sequencing platforms based on different sequencing principles or sequencing platforms from different vendors. This facilitates the rapid and automatic generation of better models or models that meet the user's input task. For example, the system 100 can determine a suitable 'a' or a model that conforms to certain principles through iterative calculations.

[0079] Compared to the methods described in the examples above, which directly segment the classification probability P0 to determine the bases (base categories) of the corresponding segments, some specific examples first classify the base categories in the first prediction result. This classification can be based on the current (round) base identification type, the previous n (round) base identification types, the next m (round) base identification types, or any two or all of these three combinations, where m and n are independent natural numbers. For example, classification can be based on the currently identified base categories such as A, C, G, and T; or based on the current (round) bases and the base categories identified in the previous or next round, such as CC, AA, AT, CG, etc.; or based on the current (round) bases and the base categories identified in the previous and next rounds, such as ATC, TAT, TTT, GCA, etc.; or based on any two or more of the aforementioned combinations, etc.

[0080] In examples where the base categories in the prediction results are first classified, the classified base categories are often directly referred to as base categories. However, if this might cause ambiguity or confusion, the classified base categories are sometimes referred to as new base categories. The classification probability of the classified base category can be the classification probability of the current (round) base category, for example, still P0 in the aforementioned examples, or a modified P0. For example, P0 can be modified or corrected based on the classification probability or base recognition probability of non-current bases in the new base category, such as preceding and / or following bases. The mapping relationship between P0 and the Q value can also be modified. Specifically, for example, if the classification probability of non-current bases in the base category is relatively low, for example, all less than 0.8, it can be considered to some extent that the current base category is "dragged down" by one or more preceding or following relatively unreliable bases, thus affecting the classification probability of the base category. If the reliability of a base class decreases, its P0 value or the mapping relationship between P0 and Q value can be adjusted. For example, multiplying it by a coefficient less than 1, such as 0.9, to update P0, or subtracting 1 or 2 from the Q value corresponding to P0, or assigning it to the adjacent lower-level segment Q value, can correct the classification probability of that base class or the mapping relationship between the classification probability of that base class and Q value. Alternatively, bases with relatively low classification probabilities (relatively unreliable bases) in that base class can be removed to update the base class, and the updated base class can replace the original base class in subsequent analysis processes. Or, the base class can even be directly removed, so that the base classification does not participate in subsequent processes. This helps to establish or generate a more accurate correspondence.

[0081] Furthermore, the subsequent processing of this method, which first classifies the base categories in the initial prediction results, can be similar to the methods or approaches described in the previous examples. For instance, it can involve segmenting the bases based on their classification probabilities P0, calculating the true error rate and / or true Q-value for each segment, and statistically analyzing or clarifying the number or distribution of base categories with each true error rate and / or true Q-value, in order to establish a correspondence between each classification probability and a predetermined quality score. The processes or methods and technical effects of the aforementioned related examples are also applicable to processing such new base categories and associated data, and will not be elaborated upon here.

[0082] The predetermined quality score refers to the predetermined Q-value segmentation requirements, such as parameters from user input, parameters that the user of system 100 can set and change, or parameters that system 100 automatically selects or adjusts based on user-input tasks. Currently, the Q-values ​​calculated based on the Phred algorithm or output by various sequencing platforms are commonly either continuous integers or non-contiguous integers. Specifically, the predetermined output Q-values ​​are continuous integers, such as all integers in the range [0,40], [0,45], or [0,50], etc.; the predetermined output Q-values ​​are non-contiguous integers, also known as Q-score binning, which divides continuous Q-values ​​into several intervals (bins), such as Q-values ​​forming an arithmetic sequence or consisting of non-contiguous integers without obvious rules, such as 0, 10, 20, 30, 40, and 50, or 0, 10, 17, 21, 28, 35, 40, and 50. Both continuous and discontinuous Q-values ​​are widely used in current commercially available sequencing platforms, with discontinuous Q-values ​​being more prevalent. Discontinuous Q-values ​​can simplify data processing, improve robustness, and facilitate visualization, significantly enhancing the efficiency and stability of data analysis. However, discontinuous Q-values ​​may also lead to the loss of some information. In specific applications, the appropriate or desired presentation of Q-values ​​generally requires a trade-off between the advantages and disadvantages.

[0083] In a certain example, the predetermined quality score matches the actual Q value in the relevant example above. That is, the actual Q value meets the predetermined quality score value requirement. Thus, the established correspondence between the classification probability of each base category and the predetermined quality score constitutes the establishment of the so-called quality score model.

[0084] In other examples, the predetermined quality score is a set of scores that differ from the true Q-values. Typically, it is necessary to further update these correspondences or mappings based on the established correspondence between the classification probabilities of bases or groups of bases and the true Q-values, combined with the quantity (e.g., percentage or distribution) of each base representing the true Q-value / true error rate, and the new Q-value requirements. This is to establish or update the correspondence, i.e., to establish or update a quality score model that conforms to the user's input task or requirements.

[0085] Specifically, in some specific examples where the predetermined quality score differs from the true Q value and the true Q value includes the value of the predetermined quality score, the so-called association of the predetermined quality score with the error rate of the base class includes, based on the error rate of the base class of the first prediction result of each group and the proportion or distribution of the number of base classes in each group, and the predetermined quality score, merging the groups into new groups with fewer groups and adjusting the grouping of the base classes of adjacent new groups, and calculating the error rate of the base class of each new group, so that the error rate of the base class of the final new group corresponds to the predetermined quality score, so as to establish a mapping relationship between the classification probability of each base class and the predetermined quality score.

[0086] Understandably, the so-called predetermined quality score, i.e., the predetermined Q-value requirement, is equivalent to predetermining the number of new segments and the error rate of each new segment. By merging one or more adjacent segments into fewer new segments, and adjusting the base assignments of adjacent new segments (equivalent to adjusting the proportion or distribution of base classes in each segment), and calculating the error rate of the base class in the new segment, the error rate of the segment calculated based on the P0 of the bases in each new segment is ultimately the predetermined error rate of that segment (the error rate of the new segment or the new Q-value). In this way, a mapping relationship is established between the classification probability of each base class and the predetermined quality score, thus establishing or generating the so-called quality score model.

[0087] Specifically, in one example, based on statistics of the second training data, including the true Q-value (true error rate) of each base (base class) in each group or segment, the distribution or trend of the number or proportion of bases in each group or segment, and the new quality score requirement (predetermined quality score), the bases of the original groups or segments are merged and regrouped or segmented, and the error rate of each segment after regrouping or segmentation (the true error rate of the bases in the resegmented segment) is made consistent with the predetermined quality score, so as to establish a new correspondence between each group of bases and the predetermined quality score, so as to update and generate the desired quality score model.

[0088] In System 100, the process of merging one or more adjacent segments into fewer new segments, allocating bases within these new segments, and calculating the error rate of the corresponding new segments can be achieved through iterative calculations. For example, if the intention is to divide segments of consecutive integers 0-50 (each integer representing a segment) into four new segments of 10, 20, 30, and 40, the initial division might result in a true Q value of 11 for the new segment 10 and 19 for the new segment 20. This indicates that the new segmentation is not reasonable or fails to meet the predetermined requirements. Therefore, the bases in the adjacent new segments are reallocated again. For instance, the lowest score (corresponding base) previously assigned to new segment 20 is reallocated to the previous new segment 10, thus lowering the true Q value of new segment 10 and raising the true Q value of new segment 20. This iterative update is repeated until the final true Q value / true error rate of the new segments matches the predetermined quality score.

[0089] Understandably, the calculation or determination of accuracy rate or precision rate in the relevant embodiments refers to the same meaning as the calculation or determination of error rate. Accuracy rate and error rate are corresponding and can be converted into each other. In addition, accuracy rate and correct probability or precision rate or accurate probability in the relevant embodiments generally refer to the same meaning unless otherwise stated; the same applies to the corresponding error rate, error probability, and probability of error, etc., which can be substituted unless otherwise stated.

[0090] In some specific examples, the process further includes merging and regrouping groups based on the error rate of the base categories in the first prediction results of each group and the proportion or distribution of the number of base categories in each group, so that the error rate of the base categories in the new group corresponds to the first quality score, thereby establishing a mapping relationship between the classification probability of each base category and the first quality score; and merging and regrouping new groups based on the error rate of the base categories in the new group and the proportion or distribution of the number of base categories in each new group, so that the error rate of the base categories in the final group corresponds to a predetermined quality score, thereby establishing a mapping relationship between the classification probability of each base category and the predetermined quality score, wherein the values ​​in the first quality score include all values ​​of the predetermined quality score. In this way, two mapping relationships can be directly generated, outputting two corresponding quality score models or a cascaded quality score model for users to compare and select.

[0091] The first quality score is the specified quality score, which can be called the standard quality score. In one example, the standard quality score comes from parameters that the user can optionally input, such as parameters that the user can set or adjust themselves. In another example, the standard quality score, or the corresponding mapping relationship, is an optional parameter built into the system 100, which can be mapped and is a parameter or relationship that the system 100 uses to determine whether to select or learn based on the task input by the user.

[0092] In a specific example, the first quality score is a continuous integer, such as a continuous integer Q-value that adapts to or covers the Q-values ​​of most current sequencing platforms, such as [0,45] or [0,50]. The predetermined quality score is a non-continuous integer, such as one that meets or adapts to the user-input task or specified application detection requirements. If the predetermined Q-values ​​are the same, but the upstream machine learning base recognition model has changed or it is unknown whether the upstream machine learning base recognition model is the same, then in most cases only the first mapping relationship needs to be updated and adjusted—that is, the mapping relationship between the classification probability of the preceding base category and the standard quality score—without needing to adjust subsequent mapping relationships. In this way, a quality score model can be quickly generated, facilitating the conversion and evaluation of the actual quality of sequencing data from different sequencing platforms that exhibit the same or different Q-values.

[0093] Furthermore, it can quickly and conveniently meet some special Q-value segmentation requirements. Specifically, the first mapping relationship (the preceding mapping relationship) maps all bases to standard segmentation intervals (standard quality scores) in any scenario. Therefore, such example schemes can provide standard intermediate results when verifying Q-value algorithm performance and troubleshooting algorithm problems, which is helpful for performance verification and problem troubleshooting. For example, the performance based on the preceding mapping relationship can evaluate whether the mapping relationship from P0 to consecutive integer Q values ​​of [0,50] is reasonable, while the performance based on the following mapping relationship can evaluate whether the segmentation mapping from consecutive integer Q values ​​of [0,50] to non-consecutive integer Q values ​​is reasonable, etc.

[0094] In some examples, output layer 160 is also used to output the quality score model in any of the related examples mentioned above.

[0095] In some examples, please refer to Figure 6 The evaluation module 146 is also connected to the second training data generation module 152, the second training module 154, and the output layer 160. It is used to evaluate the performance of the quality score model using the second test set before outputting the quality score model in any relevant example, so as to determine that the quality score model meets the preset requirements. The second test set is a part of the second training data that does not overlap with the second training set and the second validation set. Meeting the preset requirements means that at least one of the accuracy, precision, recall, and speed of the classification prediction results of the quality score model on the second test set meets the preset standard or is able to perform the user input task.

[0096] In some embodiments, the designated model or first designated model is selected from at least one of logistic regression, support vector machine, tree models such as LightGBT, and neural networks such as semantic segmentation models. The designated model or first designated model can be a single model, or a cascaded model or composite model containing multiple single models.

[0097] In some embodiments, please refer to Figure 7 The machine learning layer 140 also includes a selection module 148 for determining a specified model, a first specified model, or a second specified model based on user input. This can be done by the user directly selecting a specified model through user input, or by the task system 100 automatically selecting a specified model based on user input.

[0098] Additionally, certain embodiments of this application provide a method for developing a base recognition model, the method comprising accessing user input using a system 100 for automatically generating base recognition models as described in any of the above embodiments, and running the system 100 using at least one hardware processor to present an optimal base recognition model to the user. The user input includes sequencing data and optionally one or more tasks, the sequencing data being obtained by detecting nucleic acid molecules linked to multiple sites on a surface using surface imaging sequencing technology.

[0099] It is understood that the descriptions of the technical features, additional technical features, and related technical effects of the system 100 for automatically generating base recognition models in the above-described embodiments are also applicable to the method for developing base recognition models in this embodiment, and will not be repeated here.

[0100] For example, in some embodiments, sequencing data includes images of nucleic acid molecules at multiple locations on one or more surfaces obtained from one or more rounds of detection during one or more sequencing runs.

[0101] In other embodiments, the sequencing data is a subset of data extracted from the sequencing image, and the sequencing data includes the location and intensity values ​​of multiple nucleic acid molecules in one or more surface regions, and the sequencing image is an image of nucleic acid molecules at multiple locations on one or more surfaces obtained from one or more rounds of detection in one or more sequencing runs.

[0102] In some examples, sequencing data can reflect at least 10 million sequence reads and / or at least 1 billion base pairs reads.

[0103] In some embodiments, user input is received through a model generation framework or interface; and / or, the best base recognition model is presented to the user through a model generation framework or interface.

[0104] In some embodiments, system 100 is run using two or more operating systems, and at least one of file sharing, remote access, network communication, virtual machines, containers, and virtualization technologies or tools is used to enable interaction between the different operating systems, in order to quickly train and generate relevant models.

[0105] In some specific examples, a portion of system 100 is run using Windows and Linux respectively, and interaction between the different operating systems is achieved using at least one of WSL and / or SSH.

[0106] SSH (Secure Shell) is a type of "remote access" method (used to operate or access another operating system from one operating system). It is suitable for remote management of Linux and macOS and can be accessed from Windows via tools such as PuTTY and Termius. WSL (Windows Subsystem for Linux) is a type of "virtual file system and bridging tool" (bridging files or processes of different operating systems through virtualization technology). It enables running a Linux environment on Windows and allows for file and tool sharing.

[0107] In some embodiments, system 100 utilizes multi-threaded parallel computing to improve the speed of model generation. Specifically, in one example, system 100, such as training data generation modules 142 and / or 152 in its machine learning layer 140, runs interactively on multiple operating systems, with multi-threaded parallel mechanisms configured on each operating system.

[0108] Some embodiments of this application also provide a base recognition model generated using the system 100 for automatically generating base recognition models in any of the related embodiments described above.

[0109] Other embodiments of this application also provide a computer-readable storage medium storing one or more instructions for controlling a processor to perform the method for developing a base recognition model in any of the above embodiments. The computer-readable storage medium includes, but is not limited to, random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art.

[0110] A sequenced list of executable instructions for implementing logical functions may be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). In related embodiments, the computer-readable storage medium may be any apparatus that can contain, store, communicate, propagate, or transmit programs for instruction execution, or be used in conjunction with such a system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable storage media can even be paper or other suitable media on which the aforementioned program can be printed, because the aforementioned program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory. The various computer-readable storage media described in the related examples can represent one or more devices and / or other machine-readable storage media for storing information. The term "machine-readable storage medium" as used herein can include, but is not limited to, wireless channels and various other media capable of storing, containing, and / or carrying instructions and / or data.

[0111] Embodiments of this application also provide a system for automatically generating base recognition models, also referred to as a computer program product or computing device, the system including a memory storing a program and one or more processors coupled to the memory, the processors running the program in accordance with the method for developing base recognition models in any of the related embodiments.

[0112] When using a software implementation method, it can be implemented entirely or partially in the form of a computer program product. This computer program product includes one or more instructions or programs. When such computer programs or instructions are loaded and executed on a computer, the processes or functions in the relevant embodiments of this application are implemented, in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0113] The term "computer program product" or "computing device" is intended to refer to various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. A computing device can also refer to various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components, connections, associations, and functions of the components shown in the embodiments of this application are merely examples and are not intended to limit the description and / or related implementations of the embodiments of this application.

[0114] Please refer to Figure 8 The system 100 or electronic device 500 that automatically generates base recognition models includes a computing unit 501, which can perform various appropriate actions and processes based on a computer program stored in ROM (Read-Only Memory) 502 or a computer program loaded from storage unit 508 into RAM (Random Access Memory) 503. The RAM 503 may also store various programs and data required for the operation of the device 500. The computing unit 501, ROM 502, and RAM 503 are interconnected via a bus 504. An I / O (Input / Output) interface 505 is also connected to the bus 504.

[0115] Multiple components in electronic device 500 are connected to I / O interface 505, including: input unit 506, such as keyboard, mouse, etc.; output unit 507, such as various types of monitors, speakers, etc.; storage unit 508, such as disk, optical disk, etc.; and communication unit 509, such as network card, modem, wireless transceiver, etc. Communication unit 509 allows device 500 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0116] The processor or computing unit 501 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of computing unit 501 include, but are not limited to, CPUs (Central Processing Units), GPUs (Graphics Processing Units), various special-purpose AI (Artificial Intelligence) computing chips, various computing units running machine learning model algorithms, DSPs (Digital Signal Processors), and any suitable processor, controller, microcontroller, etc. Computing unit 501 executes the modules, methods, or processes described in the relevant embodiments, such as implementing the functions of machine learning layer 140, etc.

[0117] For example, in some examples, the method for developing a base recognition model can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program can be loaded and / or installed on device 500 via ROM 502 and / or communication unit 509. When the computer program is loaded into RAM 503 and executed by computing unit 501, one or more steps of the method in the aforementioned related embodiments can be performed. In other embodiments, computing unit 501 can be configured by any other suitable means (e.g., by means of firmware) to perform the method for developing a base recognition model using an automated training tool in any of the foregoing related embodiments.

[0118] It should be understood that the various parts of the system or method in the embodiments of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above-described related embodiments or implementations, multiple layers, modules, sub-modules, or steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, it can be implemented using any one or a combination of the following related prior art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0119] Those skilled in the art will understand that all or part of the steps of the system for automatically generating models in any of the above embodiments or the methods for implementing any of the above related embodiments can be performed by a program instructing related hardware. The program can be stored in a computer-readable storage medium and, when executed, includes one or a combination of steps for implementing the example methods.

[0120] Furthermore, the functional units in various implementation methods or embodiments can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0121] Publicly available schemes for constructing base recognition models based on machine learning include CN118429967A, CN118429965A, and CN118969086A, while examples of Q-value models based on machine learning include CN118471340A, CN117912550A, CN117976042A, and CN118116469A. The system 100 for automatically generating base recognition models in this application is applicable to the automated development of any publicly available machine learning-based base recognition model and / or Q-value algorithm. It automates the development process of machine learning models, reducing reliance on specialized knowledge and experience in algorithm development. The following description, in conjunction with one or more publicly available schemes or related optimizations and improvements for constructing base recognition models or Q-value algorithms based on machine learning, exemplarily describes the establishment, deployment, and use of the system (automatic development tool) for automatically generating base recognition models.

[0122] The following further examples describe the specific operating methods or implementation processes that may be included in some of the above embodiments or examples. However, this should not be construed as limiting the scope of the subject matter of this application to the following further examples or embodiments.

[0123] Establishment and Evaluation of Machine Learning-Based Base Recognition Model and Corresponding Q-Value Algorithm (Part 1) Establishment of Machine Learning Base Recognition Model

[0124] The main steps include:

[0125] 1. Determine the application scenario of the base recognition model to be built, and determine the corresponding platform, instrument, biochemical, and fluid dynamics versions of the model. Here, we take a multi-fluorescence channel based on surface imaging, such as the commonly known two-color or four-color fluorescence high-throughput sequencing platform, as an example.

[0126] 2. Under the defined operating environment and conditions in 1, sequence a standard library sample containing a known reference genome and save the original sequencing data, such as the original fluorescence image (sequencing image).

[0127] 3. Establishment of training data for the base recognition model (hereinafter referred to as training data I):

[0128] (1) Use existing basic base recognition software, such as the base recognition software built into various sequencing platforms (which may involve machine learning or not), to analyze the sequencing data and identify bases, and output the base recognition result file (fastq file). When using basic base recognition software for base recognition, retain the cluster location information during the base recognition process, as well as the fluorescence brightness information (intensity information) of each fluorescence channel in each round of each cluster (amplifier), for subsequent model feature value extraction.

[0129] (2) Align the fastq file with the reference genome to obtain the alignment result file (sam or bam file).

[0130] (3) Based on the brightness information of each cluster in each round, the brightness information of the first N rounds and the last M rounds (N and M are independent integers) and other optional information such as the brightest base in the first N rounds and the position information of the cluster are selected to create feature values.

[0131] (4) Create target values ​​based on the information after the sequence alignment of the cluster. For example, the target value is the reference base corresponding to each base after alignment, which is sometimes called the standard answer.

[0132] (5) Map the feature values ​​to the target values ​​one-to-one to complete the creation of training data I. When creating training data I, only sequences with high alignment confidence can be retained. Furthermore, it is preferable to exclude polymorphic sites from the scope of training data I. In short, when selecting training data, ensuring the accuracy of the "target value" as much as possible is beneficial for training a model with more accurate prediction results. In one example, for data with a high probability of error in the target value, either it should not be included in training data I, or the target value should be labeled as N, where N represents one of A, G, C, and T; in a scenario where deep learning, such as neural networks, is used for model training, all feature values ​​and target values ​​corresponding to clusters are written into training data I.

[0133] 4. Use a portion of the training data I to train the base recognition model. For example, in this example, the training data I is divided into three parts: one part is used for the development of the base recognition model (this part of the training data I is abbreviated as I.1), one part is used for the development of the corresponding Q-value model (this part of the training data I is abbreviated as I.2), and one part is used for validation (this part of the training data I is abbreviated as I.3).

[0134] 5. Establish a base recognition model: Randomly divide the training data I.1 into training and test sets, and establish a base recognition model (abbreviated as Model B0) based on the dataset. The "base recognition model" refers to the correspondence between "feature values" and "target values" mentioned earlier. Any machine learning or deep learning algorithm can be used to build the model. In this example, the model algorithm disclosed in CN119479831A is used, and its entirety is incorporated herein by reference.

[0135] In paired-end (PE) sequencing, in some best-case scenarios, the biochemical characteristics of R1 (read1) and R2 (read2) often differ. Therefore, it is preferable to extract features from them separately and train the model accordingly. Similarly, in scenarios involving bifacial microarray sequencing, the biochemical and optical characteristics of the two surfaces may differ. Therefore, it is generally beneficial to build models independently based on the detection or readout of each surface, which yields better predictive performance. This section uses the construction of one model as an example; the other model is similar and will not be repeated below.

[0136] (II) Establishment of the Q-value algorithm corresponding to the base recognition model B0

[0137] 1. Based on training data I.2, establish training data for the Q-value algorithm (hereinafter referred to as training data II).

[0138] (1) Using the base recognition model B0 established above, base recognition is completed based on the feature values ​​of training data I.2, and the base recognition probability corresponding to the model is obtained, that is, the normalized recognition probability of A, G, C and T output by each feature data after model prediction, and the sum of the four base recognition probabilities of each data is 1; and the base recognition result predicted by the model, which is usually the base corresponding to the maximum base recognition probability in each data.

[0139] (2) Compare the base identification results of the above model with the target value of training data I.2. If they match, mark them as 1 or correct; if they do not match, mark them as 0 or incorrect. This correct / incorrect information is used as the target value of training data II.

[0140] (3) The feature value of training data II can be exactly the same as that of training data I.2; it can also be the base recognition probability corresponding to each data point; it can also be one or more of the four base recognition probabilities, such as the feature value being the largest base recognition probability, or the largest and second largest base recognition probabilities, etc.; it can also be the derived calculation result based on the four base recognition probabilities, such as the ratio of the largest and second largest base recognition probabilities; or it can be any combination of any part of the above-mentioned parameters.

[0141] 2. A Q-value algorithm is established based on training data II (hereinafter referred to as the basic scheme).

[0142] a) Train the machine learning classification algorithm based on the training data II. Any suitable machine learning model can be selected, such as logistic regression, support vector machine, tree model, etc. The obtained model is denoted as model Q0.

[0143] b) Apply the Q0 model to training data II, that is, use the Q0 model to obtain the model prediction result based on the feature values ​​of training data II. Here, the final model classification result does not have to be output; only the classification probability P value given by the model during prediction can be output, denoted as P0.

[0144] c) Typically, P0 is a probability value between 0 and 1. P0 is divided into 'a' segments, and the bases in training data II are correspondingly divided into 'a' segments. The principle for choosing 'a' is that after dividing P0 into segments 'a', the difference in the true Q-values ​​of the bases corresponding to adjacent segments is not significant. In the actual operation of the example, it is generally required that the difference in the true Q-values ​​of adjacent segments is less than 1.

[0145] Here's some explanation of the "true Q-value". The Q-value is actually another way of representing the error rate. The relationship between Q and the error rate can be written as Q = -10 * log10(err). Here, Q refers to the sequencing quality score of the bases, i.e., the Q-value; err refers to the error probability or error rate of that base. The "true Q-value" refers to the Q-value calculated based on the true error rate (err) of a set of bases. For example, as mentioned above, if the bases are divided into segments 'a' based on the P0 value, then the total error rate of the bases in each segment can be calculated. For example, the error rate of that segment can be calculated by using the number of inconsistent bases as the numerator and the total number of bases in that segment as the denominator, and the corresponding Q-value can be calculated based on this error rate. This Q-value is the "true Q-value" defined here.

[0146] d) Write down the actual Q values ​​corresponding to the bases in segment a.

[0147] e) Define the segmentation requirements for the Q value based on the algorithm development needs. For example, some scenarios require Q to take the value of all integers in the range [0, 45], while others require Q to take the value of 5 specific values ​​in [0, 45], etc.

[0148] f) Based on the above requirements for Q values, segment the Q values. In the above steps, based on the statistics of training data II, the true Q value (true error rate) corresponding to each base in each of the a segments is known; at the same time, the distribution trend of the number of bases in different segments is also known. Therefore, based on the error rate and base percentage corresponding to each of these a segments, and the error rate requirement in the target Q value segment (i.e., the Q value mentioned above), the bases in the above a segments can be merged and segmented again. After merging the segments, the total error rate corresponding to the bases in each segment is consistent with the pre-required Q value.

[0149] g) Record the correspondence between each final P0 value and the Q value (segmented Q value), which is the Q-value algorithm corresponding to the machine learning algorithm.

[0150] 3. A Q-value algorithm is established based on training data II (hereinafter referred to as Supplementary Scheme 1).

[0151] a) The first few steps are consistent with a) and b) of the basic scheme.

[0152] b) Corresponding to step c) of the basic scheme, when segmenting bases according to the P0 value, the category of base identification is also considered to distinguish the bases. For example, in the basic scheme, all bases are placed together and divided into segment a according to the P0 value. In supplementary scheme 1, bases are first classified according to the final identification type. For example, they can be classified according to the current base identification category A, C, G, and T; or they can be classified according to the current base and the categories of the bases before and after it, such as ATC, TAT, etc. In short, bases can be classified according to the current base identification type, the identification types of the previous n bases, the identification types of the next m bases, or a combination of any two or three of the above, where n and m are independent integers.

[0153] c) After completing the above classification, perform the operations on each type of base according to d)-g) of the basic scheme.

[0154] d) Record the correspondence between the P0 values ​​and Q values ​​(segmented Q values) corresponding to each base classification. This correspondence is the so-called corresponding Q value algorithm.

[0155] Compared to the basic scheme, Supplementary Scheme 1 takes into account the potential P0 value preference between different base combinations and corrects for it when mapping Q values. The preference of P0 value relative to base combinations mainly comes from biochemical preference characteristics, that is, bases with the same brightness signal characteristics but different base combinations may have significantly different error rates.

[0156] 4. Establish a Q-value algorithm based on training data II (hereinafter referred to as Supplementary Scheme 2)

[0157] Both the basic scheme and supplementary scheme 1 can be combined with supplementary scheme 2. One of the characteristics of supplementary scheme 2 is that the segmentation of the Q value is divided into two steps.

[0158] Specifically, in steps e)-f) of the basic scheme or the corresponding steps of supplementary scheme 1, the bases segmented according to the P0 value are directly mapped to the final Q value segment. In supplementary scheme 2, however, all bases are first mapped to the Q value segment containing all integer values ​​in the range [0,50].

[0159] Specifically, the basic approach in the example is to directly map the bases segmented according to the P0 value to, for example, a Q-value segment of all integer values ​​in the range [0, 50]. Supplementary approach 1 first maps the bases in different groups to Q-value segments of all integer values ​​in the range [0, 50], and then merges all base categories according to the Q-value or segmentation requirements. Supplementary approach 2, in short, obtains the mapping relationship between the P0 values ​​corresponding to different categories of bases and, for example, a Q-value segment of all integer values ​​in the range [0, 50], which is referred to here as mapping relationship 2.1.

[0160] After obtaining mapping relation 2.1, the Q-value segments of all integer values ​​in the range [0,50] are then mapped to the final required Q-value segments, which is referred to here as mapping relation 2.2. Mapping relation 2.1 and mapping relation 2.2 together constitute the corresponding Q-value algorithm.

[0161] One advantage of Supplementary Scheme 2 is that mapping relationship 2.1 is implemented using the same logic in any scenario, which is beneficial for the implementation of automatic modeling software. If different scenarios have different Q-value segmentation requirements, then mapping relationship 2.2 can generally be adjusted directly. If the Q-value segmentation requirements are the same, but the upstream machine learning base recognition model has changed, then in most cases only mapping relationship 2.1 needs to be adjusted, and mapping relationship 2.2 usually does not need to be adjusted.

[0162] Since mapping relationship 2.1 maps all bases to standard segmented intervals in any scenario, scheme 2 can provide standard intermediate results when verifying the performance of the Q-value algorithm and troubleshooting algorithm problems, which is helpful for performance verification and problem troubleshooting. For example, the performance based on mapping relationship 2.1 can be used to evaluate whether the mapping relationship from P0 to integer Q values ​​of [0,50] is reasonable, while the performance based on mapping relationship 2.2 can be used to evaluate whether the segmented mapping from integer Q values ​​of [0,50] to large segmented Q values ​​is reasonable.

[0163] The advantage of Supplementary Scheme 2 is that it can easily meet some special Q-value segmentation requirements. The bases corresponding to the subdivided Q-value segments (e.g., the bases corresponding to the integer Q-values ​​[0, 50], or the mapping relationship between the bases directly segmented according to P0 in the basic scheme and the final target Q-value segmentation) are often not unique.

[0164] First, for example, during product development, the final target Q value is often given in the form of intervals. For instance, it might be required to "determine a value as the Q value from the intervals (0,15), (15,20), (20,30), (30,40), and (40,50)," but the specific value for each interval is often not predetermined. It can be determined during development based on the specific base classification, thus there can be many Q value segments that meet the requirements. Second, even if the specific final Q value is defined, there is more than one way to divide the bases to meet the requirements. In this case, the mapping relationship 2.1 of Supplementary Scheme 2 maps bases to standard Q value segments, which is more conducive to meeting special Q value segmentation requirements, such as "the Q35 segment cannot contain bases with a Q value lower than 30," and similar base division requirements.

[0165] (III) Model Performance Evaluation

[0166] a) Test the currently established base recognition model B0 and the corresponding Q-value algorithm on the validation set I.3.

[0167] b) Evaluation of base recognition model B0

[0168] Based on the brightness information of the current round in the feature values ​​of I.3, the base identification result under the traditional algorithm (not based on machine learning model prediction) can be obtained based on the brightest base. This result is denoted as base identification result 1 of I.3. Based on the feature values ​​of I.3, the base identification model B0 is used for prediction, and the result is denoted as base identification result 2. Base identification result 1 and base identification result 2 are compared with the target value of I.3 respectively, and the error rate is calculated for each (i.e., the number of base identification results that differ from the target value divided by the total number of entries in the I.3 data). The reduction in the error rate of base identification result 2 compared to base identification result 1 is the improvement in base identification accuracy brought about by base identification model B0.

[0169] c) Performance evaluation of the Q-value algorithm

[0170] Compare the base identification result 2 with the target value of I.3 to obtain the correctness result of each data entry under the base identification model B0, which is recorded as the base correctness result 2.

[0171] For the basic scheme, based on the base error result 2, the error rate corresponding to each Q-value segment is output, and the actual Q-value is calculated based on the error rate. The difference between the Q-value assigned by the algorithm and the actual Q-value is compared to determine whether the Q-value algorithm meets expectations.

[0172] For supplementary scheme 1, in addition to comparing the actual Q values ​​corresponding to the bases in different Q value segments, it is also necessary to compare the actual Q values ​​corresponding to the bases in different Q value segments for different base recognition types, so as to evaluate whether the total base Q value assignment is reasonable and whether the Q value assignment for specific types of bases is reasonable / meets expectations.

[0173] For supplementary scheme 2, in addition to comparing the actual Q values ​​corresponding to the bases in the final Q value segment, it is also necessary to compare the actual Q values ​​calculated for the bases corresponding to the [0,50] integer Q values ​​generated by mapping relationship 2.1.

[0174] In addition, before developing the base recognition model and Q-value algorithm, it is generally necessary to pre-set the pass indicators and qualified thresholds for the performance of the base recognition model and Q-value algorithm. For example, at least one of the indicators of accuracy, precision, recall and speed of the base recognition model for the classification prediction results of the validation set I.3 must meet the preset standard or be able to achieve the user input task.

[0175] When the performance of the base recognition model or Q-value algorithm fails to meet the threshold requirements, a mechanism for algorithm redevelopment can be triggered, or a manual intervention alert can be activated. For example, users can be prompted to add user input, such as sequencing data, to generate more training data for better model training.

[0176] Finally, the output model is a base recognition model and / or Q-value algorithm that has passed or met the preset indicators or expected requirements after performance evaluation, so as to complete machine learning modeling.

[0177] Model development process automation

[0178] 1. The above-described data processing method and model development logic can be implemented using any programming language on Linux or Windows systems.

[0179] 2. When using a Windows system to complete the main workflow of machine learning model training, there may be a need for system interaction between Windows and Linux. For example, since most mainstream bioinformatics analysis software is developed on Linux systems, even if there are Windows versions, their running speed is often much slower than the Linux versions. The workflow of building a base recognition model in the example above includes bioinformatics analysis steps such as alignment and error rate statistics. If the main workflow of model development is carried out on Windows, then completing this part of the bioinformatics process on Linux will help improve the efficiency of automated training tools.

[0180] Here are two examples of how to achieve interaction between Windows and Linux systems.

[0181] (1)WSL2

[0182] Installing a Linux subsystem on Windows via WSL (Windows Subsystem for Linux) enables interaction between Windows and Linux. Advantages: The Linux subsystem boots quickly, consumes few resources, and has almost no impact on Windows performance; WSL2 has a complete Linux kernel and supports the installation of most Linux software; WSL2 supports container technologies such as Docker / Singularity, making software deployment convenient and fast. Disadvantages: File access between Windows and Linux is via network output, resulting in slower read and write speeds compared to the native system.

[0183] (2) OpenSSH

[0184] Implementation: Install OpenSSH software on Windows and Linux, and interact via the SSH protocol. Advantages: Based on the SSH protocol, security is greatly improved, and the connection is stable; easy integration, by configuring passwordless SSH access, remote connections, task submissions, task status checks, and other operations can be performed in Python scripts; independent Windows and Linux nodes can maximize the computing power of each node. Disadvantages: Installation and deployment are relatively complex, requiring separate deployment of OpenSSH software on Windows and Linux, and configuration of passwordless access.

[0185] 3. During the model training phase, parallel computing can be used to improve the running speed of training tools.

[0186] For example, in the process of base identification using basic or traditional non-machine learning-based base identification software as described in (I) 3(1) above, multiple threads can be called simultaneously for analysis via code. When developing automatic training tools, it is generally necessary to analyze the computing power and memory occupied by the basic version of the base identification software in advance and write the calculation formula for the maximum number of threads corresponding to different computing power and memory. When using the automatic modeling tool (automatic training tool), the tool will automatically read the computing power and memory of the system used for modeling and automatically calculate the maximum number of threads that can run in parallel according to the maximum number of threads calculation formula.

[0187] Regarding the calculation of the maximum or optimal number of threads corresponding to specific computing power and memory, those skilled in the art can reasonably determine the number by comprehensively considering hardware specifications (such as the number of CPU cores, GPU hardware characteristics such as the number of multiprocessors, thread block size, etc., memory capacity and bandwidth) and task type (such as CPU-intensive or I / O-intensive). For example, the number of threads can be calculated based on processors such as CPU and / or GPU, memory and / or task type, or by comprehensively considering the number of threads, in order to reasonably calculate and optimize the number of threads.

[0188] (1) As mentioned in (a) 3 (2), whether it is performed on a Windows system or a Linux system, a multi-threaded parallel mechanism can be set. Similarly, the calculation formula for the optimal number of parallel threads is pre-entered, and the number of parallel threads is automatically set during actual operation.

[0189] (2) The training data production in (a) 3 (3)-(5) above can also be set up for multi-threaded parallel computing in the same way.

[0190] (3) The machine learning training steps of the aforementioned (I) 5-base recognition model can also be set up for parallel computation. Furthermore, in scenarios where models are trained separately using R1 and R2 data from paired-end (PE) sequencing, both R1 and R2 models can be trained in parallel if computing power and memory allow. Similarly, in scenarios where models are trained separately using multi-faceted data such as bifacial microarray sequencing, parallel computation is set up for each model training.

[0191] 4. The feature value form of the model can be determined before training to improve the data extraction speed. And when using the traditional or basic version algorithm for base identification in step (I) 3(1), the feature value can be directly calculated and output using the cached data in the base identification process, without retaining other unnecessary intermediate parameters, thereby improving the efficiency of data calculation and reducing the data writing time.

[0192] Use automated training tools

[0193] The following example illustrates how an automated training tool can output a model file based on user input, such as sequencing images, in a specific scenario.

[0194] Suppose a user needs to use a machine learning model in a specific application scenario. This scenario might require special sample features, specific biochemical and instrument configurations, and no suitable machine learning model has ever been built for this scenario. Under such circumstances, the user can use the automated model generation system (automated training process) to build the corresponding or suitable model themselves.

[0195] 1. Generally, users first define the sequencing conditions required for the model they need to build, such as biochemical parameters, instrument configuration, and library characteristics. Then, they perform sequencing under these defined conditions to obtain sequencing data for generating training data. For example, they might use a library with a known reference sequence. If the user doesn't specify a particular library for this sequencing scenario, a larger genome library, such as a human genome sample, can be used as the sequencing sample. Larger genome samples allow for more diverse sequence feature information in the training data, which typically helps improve the model's generalization ability. If the "special sequencing scenario" requires specific library features, such as for libraries with extreme base imbalances, then it's preferable to use a library with features identical or similar to the target sequencing library to help the machine learning model learn the sequencing preferences introduced by the library features.

[0196] 2. During the sequencing process described above, the user invokes the machine learning automatic training tool. Currently, mainstream commercially available sequencing platforms or common sequencing procedures often perform data processing, including base identification, during sequencing. The software included with the sequencer for this data processing is what is referred to as the traditional or basic version of the base identification algorithm. This traditional or basic version of the base identification algorithm or tool may be a machine learning model or software that does not involve machine learning. Here, the user can directly use the automatic training tool during the sequencing process. The automatic training tool will automatically invoke the traditional or basic version of the base identification algorithm during the sequencing steps. While invoking the traditional or basic version of the algorithm to complete base identification, it can simultaneously generate the feature values ​​needed by the machine learning model based on intermediate sequencing parameters such as the brightness of the original image and contextual base identification information, and store them in a specified path. Understandably, in a regular sequencing process (without invoking the automatic training tool), the intermediate sequencing parameters mentioned here are often parameter data that the sequencer needs or usually generates during base identification but generally does not need to be output or output to the user.

[0197] 3. The automated training process can automatically identify the sequencing progress. For example, after sequencing is complete, the automated training tool can automatically call the alignment tool, align the FASTQ file obtained from sequencing with a reference genome (e.g., user-inputted data), and output the alignment results to a specified path.

[0198] 4. After alignment is complete, the automatic training tool will begin automatically generating training data based on the stored feature values ​​and alignment results. The feature values ​​for the training data have already been output during the sequencing process; the target values ​​will be extracted based on the alignment results. See section (I) 3 of the aforementioned example for details. The training data will be automatically stored in the specified path.

[0199] 5. After the training dataset is created, the automatic training tool can automatically divide the training data, automatically complete the modeling of the machine learning model, and automatically complete the development of the Q-value algorithm, following the process described in (I) 4 and 5 and (II) above.

[0200] 6. After completing the machine learning model modeling and Q-value algorithm development, the automatic training tool will automatically save the model and algorithm files in the specified path. The machine learning and Q-value algorithm model files can be used as configuration files for the machine learning base recognition tool. As long as the configuration file is stored in the specific folder called by the base recognition tool, the next time the machine learning base recognition tool is used for sequencing, it will automatically use that configuration for base recognition and Q-value calculation. Therefore, after completing the modeling, users can directly copy the configuration file to the corresponding configuration file path of the base recognition tool on the sequencer, and the model can be used directly for the next sequencing run.

[0201] In this specification, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "certain examples," "specific examples," or "exemplary practice," etc., refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0202] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A system for automatically generating a base recognition model, characterized in that, include: The input layer is used to receive user input from the user, including sequencing data and optionally one or more tasks, wherein the sequencing data is obtained by detecting nucleic acid molecules connected to multiple locations on the surface using surface imaging sequencing technology; The machine learning layer includes a first training data generation module and a first training module, as well as a second training data generation module and a second training module. The first training data generation module processes user input from the input layer to generate first training data. The first training data includes multiple pairs of associated first features and first labels. The first features describe the intensity corresponding to one or more nucleic acid molecules at different locations. The first labels include those selected from bases A, T, C, and G. The first training module is used to train a first specified model based on at least a portion of the first training data, including performing multiple iterations of training on the first specified model using a first training set and evaluating the performance of the first specified model after each iteration using a first validation set, in order to obtain a base recognition model. The first training set and the first validation set are each independent, non-overlapping portions of the first training data. The second training data generation module connects the first training data generation module and the first training module, and is used to process at least a portion of the first training data through the base recognition model to generate a first prediction result, so as to generate second training data. The second training data includes multiple pairs of associated second features and second labels, wherein the second features are the same as or different from the first features, and the second labels are different from the first labels. The second training module is used to train a second specified model based on at least a portion of the second training data to obtain a quality score model. The quality score model outputs a quality score that reflects the accuracy of the first prediction result. The second specified model uses the same or a different algorithm as the first specified model. The output layer is used to output the base recognition model, or to output both the base recognition model and the quality score model.

2. The system according to claim 1, characterized in that, The sequencing data includes images of nucleic acid molecules at multiple locations on one or more surfaces obtained from one or more rounds of detection during one or more sequencing runs.

3. The system according to claim 1, characterized in that, The sequencing data is a subset of data extracted from the sequencing image. The sequencing data includes the position and intensity values ​​of multiple nucleic acid molecules in one or more surface regions. The sequencing image is an image of nucleic acid molecules at multiple locations on one or more surfaces obtained from one or more rounds of detection in one or more sequencing runs.

4. The system according to claim 2 or 3, characterized in that, The sequencing data can reflect at least 10 million sequence reads and / or at least 1 billion base pairs reads.

5. The system according to claim 1, characterized in that, The first tag also includes combinations of any two, three, or four bases selected from A, T, C, and G.

6. The system according to claim 4, characterized in that, The user input also includes a reference sequence, and the first training data generation module further includes an alignment submodule. The alignment submodule is used to align the sequence readout or base readout reflected by the sequencing data to the reference sequence in order to generate an alignment result; and to determine the first tag based on the alignment result in order to generate the first training data.

7. The system according to claim 5, characterized in that, The user input also includes a reference sequence, and the first training data generation module further includes an alignment submodule. The alignment submodule is used to align the sequence readout or base readout reflected by the sequencing data to the reference sequence in order to generate an alignment result; and to determine the first tag based on the alignment result in order to generate the first training data.

8. The system according to claim 1, characterized in that, The machine learning layer further includes an evaluation module, which connects the first training module and the result output layer. This evaluation module evaluates the performance of the base recognition model using a first test set before outputting the model, in order to determine whether the base recognition model meets preset requirements. The first test set is a portion of the first training data that does not overlap with the first training set and the first validation set. The condition of meeting the preset requirement is that at least one of the accuracy, precision, recall and speed of the base recognition model's classification prediction results for the first test set meets the preset standard or is able to perform the user input task.

9. The system according to claim 1, characterized in that, The first prediction result includes the probability and base category of the data being predicted and identified as various base categories in the first label. The second feature describes the intensity corresponding to one or more nucleic acid molecules and / or the probability of at least one base category in the first prediction result. The second label includes two categories: the first prediction result and the corresponding benchmark true base are consistent or inconsistent. The second training module is used to train a second specified model based on at least a portion of the second training data, including performing multiple iterations of training on the second specified model using the second training set and evaluating the performance of the second specified model after each iteration of training using the second validation set, so as to obtain a quality score model, wherein the second training set and the second validation set are each independent non-overlapping portions of the second training data.

10. The system according to claim 9, characterized in that, The second feature describes the probability of multiple base categories in the first prediction result.

11. The system according to claim 9, characterized in that, Training the second specified model in the second training module also includes, Obtain the classification probability of the base category of the first prediction result in the second training set; Based on the classification probability and / or the base category, the first prediction result is divided into group a. For each group, the error rate of the base category of the first prediction result of the group is determined by comparing the base category in the first prediction result of the group with the corresponding benchmark true base category. a is a natural number and the value of a can ensure that the logarithmic difference of the error rate of the base category of the first prediction result of the adjacent group is less than 2. as well as The predetermined quality score is correlated with the error rate of the base category.

12. The system according to claim 10, characterized in that, Training the second specified model in the second training module also includes, Obtain the classification probability of the base category of the first prediction result in the second training set; Based on the classification probability and / or the base category, the first prediction result is divided into group a. For each group, the error rate of the base category of the first prediction result of the group is determined by comparing the base category in the first prediction result of the group with the corresponding benchmark true base category. a is a natural number and the value of a can ensure that the logarithmic difference of the error rate of the base category of the first prediction result of the adjacent group is less than 2. as well as The predetermined quality score is correlated with the error rate of the base category.

13. The system according to claim 11, characterized in that, It also includes, Based on the error rate of the base class in the first prediction result of each group, the proportion or distribution of the number of base classes in each group, and the predetermined quality score, the groups are merged into new groups with fewer groups, and the attribution of the base classes of adjacent new groups is adjusted. The error rate of the base classes of each new group is calculated so that the error rate of the base classes of the final new group corresponds to the predetermined quality score, so as to establish a mapping relationship between the classification probability of each base class and the predetermined quality score.

14. The system according to claim 11, characterized in that, It also includes, Based on the error rate of the base category in the first prediction result of each group and the proportion or distribution of the number of base categories in each group, the groups are merged and then regrouped so that the error rate of the base category in the new group corresponds to the first quality score, so as to establish the mapping relationship between the classification probability of each base category and the first quality score. Based on the error rate of the base categories in the new group and the proportion or distribution of the number of base categories in each new group, the new groups are merged and then regrouped so that the error rate of the base categories in the final group corresponds to the predetermined quality score, so as to establish a mapping relationship between the classification probability of each base category and the predetermined quality score, wherein the value in the first quality score includes all values ​​of the predetermined quality score.

15. The system according to claim 14, characterized in that, The first quality score is a continuous integer, while the predetermined quality score is a non-continuous integer.

16. The system according to claim 14, characterized in that, The user input includes the predetermined quality score.

17. The system according to claim 11, characterized in that, The user input includes the corresponding benchmark true base category.

18. The system according to any one of claims 9-16, characterized in that, The result output layer is also used to output the quality score model.

19. The system according to claim 18, characterized in that, The machine learning layer further includes an evaluation module, which is connected to the second training module and the result output layer. This evaluation module is used to evaluate the performance of the quality score model using a second test set before outputting the quality score model, in order to determine whether the quality score model meets preset requirements. The second test set is a portion of the second training data that does not overlap with the second training set and the second validation set. The condition of meeting the preset requirement is that at least one of the following indicators of the classification prediction results of the quality score model on the second test set, namely accuracy, precision, recall and speed, meets the preset standard or is able to realize the user input task.

20. The system according to claim 1, characterized in that, The first or second specified model is selected from at least one of logistic regression, support vector machine, tree model and neural network.

21. The system according to claim 1, characterized in that, The machine learning layer also includes a selection module for determining the first specified model or the second specified model based on the user input.

22. A method for developing a base recognition model, characterized in that, The system includes accessing user input using a system for automatically generating a base recognition model as described in any one of claims 1-20, and running the system using at least one hardware processor to present an optimal base recognition model to the user, wherein the user input includes sequencing data and optionally one or more tasks, the sequencing data being obtained by detecting nucleic acid molecules linked to multiple locations on a surface using surface imaging sequencing technology.

23. The method according to claim 22, characterized in that, The sequencing data includes images of nucleic acid molecules at multiple locations on one or more surfaces obtained from one or more rounds of detection during one or more sequencing runs.

24. The method according to claim 22, characterized in that, The sequencing data is a subset of data extracted from the sequencing image. The sequencing data includes the position and intensity values ​​of multiple nucleic acid molecules in one or more surface regions. The sequencing image is an image of nucleic acid molecules at multiple locations on one or more surfaces obtained from one or more rounds of detection in one or more sequencing runs.

25. The method according to claim 23, characterized in that, The sequencing data can reflect at least 10 million sequence reads and / or at least 1 billion base pairs reads.

26. The method according to claim 24, characterized in that, The sequencing data can reflect at least 10 million sequence reads and / or at least 1 billion base pairs reads.

27. The method according to any one of claims 22-26, characterized in that, The user input is received through a model generation framework or interface; and / or the optimal base recognition model is presented to the user through the model generation framework or interface.

28. The method according to any one of claims 22-26, characterized in that, The system is run using two or more operating systems, and at least one of file sharing, remote access, network communication, virtual machines, containers, and virtualization technologies or tools is used to enable interaction between the different operating systems.

29. The method according to claim 28, characterized in that, The system is operated using Windows and Linux respectively, and interaction between the different operating systems is achieved using at least one of WSL and / or SSH.

30. The method according to claim 22, characterized in that, The system is run using multi-threaded parallel computing.

31. A computer-readable storage medium storing a base recognition model, characterized in that, The base recognition model is generated using the system for automatically generating base recognition models as described in any one of claims 1-21.

32. A computer-readable storage medium storing instructions for controlling a processor to perform the method of any one of claims 22-30.

33. A system for automatically generating a base recognition model, characterized in that, The system includes a memory storing a program and one or more processors coupled to the memory, the processors running the program to implement the method of any one of claims 22-30.

Images