Self-learning base code trained using oligo sequences

Abstract

A method for incrementally training a base chorer is disclosed. The method includes iteratively initially training the base chorer with exemplars comprising a single oligonucleotide sequence and generating labeled training data using the initially trained base chorer. In operation (i), the base chorer is further trained with exemplars comprising multiple oligonucleotide sequences and labeled training data is generated using the further trained base chorer. Operation (i) is iteratively repeated to further train the base chorer. In an embodiment, during at least one iteration, the complexity of the neural network configuration loaded into the base chorer is increased. In an embodiment, the labeled training data generated during an iteration is used to train the base chorer during a subsequent iteration.

Description

[Technical Field]

[0001] Priority application This application claims priority to U.S. Nonprovisional Patent Application No. 17 / 830,287, entitled "Self-Learned Base Caller, Trained Using Oligo Sequences," filed June 1, 2022 (Attorney Docket No. ILLM1038-3 / IP-2050-US), which claims the benefit of U.S. Provisional Patent Application No. 63 / 216,419, entitled "Self-Learned Base Caller, Trained Using Oligo Sequences," filed June 29, 2021 (Attorney Docket No. ILLM1038-1 / IP-2050-PRV), and U.S. Provisional Patent Application No. 63 / 216,404, entitled "Self-Learned Base Caller, Trained Using Organism Sequences," filed June 29, 2021 (Attorney Docket No. ILLM1038-2 / IP-2094-PRV). The priority application is incorporated herein by reference for all purposes.

[0002] This application claims priority to U.S. Non-Provisional Patent Application No. 17 / 830,316, entitled "Self-Learned Base Caller, Trained Using Organism Sequences," filed June 1, 2022 (Attorney Docket No. ILLM1038-5 / IP-2094-US), which claims the benefit of U.S. Provisional Patent Application No. 63 / 216,404, entitled "Self-Learned Base Caller, Trained Using Organism Sequences," filed June 29, 2021 (Attorney Docket No. ILLM1038-2 / IP-2094-PRV), and U.S. Provisional Patent Application No. 63 / 216,419, entitled "Self-Learned Base Caller, Trained Using Oligo Sequences," filed June 29, 2021 (Attorney Docket No. ILLM1038-1 / IP-2050-PRV). The priority application is incorporated herein by reference for all purposes.

[0003] The disclosed technology relates to artificial intelligence-based computers and digital data processing systems, and corresponding data processing methods and products for mimicking intelligence (i.e., knowledge-based systems, inference systems, and knowledge acquisition systems), including systems for reasoning with uncertainty (e.g., fuzzy logic systems), adaptive systems, machine learning systems, and artificial neural networks. Specifically, the disclosed technology relates to using deep neural networks, such as deep convolutional neural networks, to analyze data.

[0004] Built-in The following are incorporated by reference as if fully set forth herein: A concurrently filed PCT patent application entitled "SELF-LEARNED BASE CALLER, TRAINED USING ORGANISM SEQUENCES" (Attorney Docket No. ILLM ILLM1038-6 / IP-2094-PCT); U.S. Provisional Patent Application No. 62 / 979,384, entitled "ARTIFICIAL INTELLIGENCE-BASED BASE CALLING OF INDEX SEQUENCES," filed February 20, 2020 (Attorney Docket No. ILLM1015-1 / IP-1857-PRV); U.S. Provisional Patent Application No. 62 / 979,414, entitled "ARTIFICIAL INTELLIGENCE-BASED MANY-TO-MANY BASE CALLING," filed February 20, 2020 (Attorney Docket No. ILLM1016-1 / IP-1858-PRV); U.S. Nonprovisional Patent Application No. 16 / 825,987, entitled "TRAINING DATA GENERATION FOR ARTIFICIAL INTELLIGENCE-BASED SEQUENCING," filed March 20, 2020 (Attorney Docket No. ILLM1008-16 / IP-1693-US); U.S. Nonprovisional Patent Application No. 16 / 825,991, entitled "ARTIFICIAL INTELLIGENCE-BASED GENERATION OF SEQUENCING METADATA," filed March 20, 2020 (Attorney Docket No. ILLM1008-17 / IP-1741-US); U.S. Nonprovisional Patent Application No. 16 / 826,126, entitled "ARTIFICIAL INTELLIGENCE-BASED BASE CALLING," filed March 20, 2020 (Attorney Docket No. ILLM1008-18 / IP-1744-US); U.S. Nonprovisional Patent Application No. 16 / 826,134, entitled "ARTIFICIAL INTELLIGENCE-BASED QUALITY SCORING," filed March 20, 2020 (Attorney Docket No. ILLM1008-19 / IP-1747-US); and U.S. Patent Application Publication No. 16 / 826,168, entitled "ARTIFICIAL INTELLIGENCE-BASED SEQUENCING," filed March 21, 2020 (Attorney Docket No. ILLM 1008-20 / IP-1752-PRV-US). [Background technology]

[0005] The subject matter discussed in this section should not be assumed to be prior art merely as a result of its mention in this section. Similarly, it should not be assumed that the problems mentioned in this section, or problems associated with the subject matter provided as background, have been previously recognized in the prior art. The subject matter in this section merely represents different approaches, which themselves may also correspond to implementations of the claimed technology.

[0006] Rapid improvements in computing power have enabled deep convolutional neural networks (CNNs) to achieve great success in many computer vision tasks in recent years, with significantly improved accuracy. During the inference phase, many applications require low-latency processing of a single image with strict power consumption requirements, which reduces the efficiency of graphics processing units (GPUs) and other general-purpose platforms. This creates opportunities for specific acceleration hardware, such as field programmable gate arrays (FPGAs), by customizing digital circuits to be particularly effective for inferencing deep learning algorithms. However, deploying CNNs in portable and embedded systems remains challenging due to large data volumes, intensive computations, diverse algorithm structures, and frequent memory accesses.

[0007] Since convolution provides most of the operations in CNN, the convolution acceleration scheme significantly impacts the efficiency and performance of hardware CNN accelerators. Convolution involves multiply-and-accumulate (MAC) operations with four levels of loops that slide along kernels and feature maps. The first loop level calculates the MAC for pixels within a kernel window. The second loop level accumulates the sum of MAC products across various different input feature maps. After completing the first and second loop levels, adding a bias results in the final output element in the output feature map. The third loop level slides the kernel window within the input feature map. The fourth loop level generates various different output feature maps.

[0008] FPGAs, particularly for accelerating inference tasks, have attracted increasing interest and become more widely used due to their (1) high reconfigurability, (2) superior development time compared to application-specific integrated circuits (ASICs) in keeping up with the rapid evolution of CNNs, (3) good performance, and (4) energy efficiency compared to GPUs. The high performance and efficiency of FPGAs can be achieved by synthesizing circuits customized for specific computations and directly processing billions of operations with a customized memory system. For example, hundreds to thousands of digital signal processing (DSP) blocks in modern FPGAs support core convolution operations, such as multiply-and-accumulate operations with high parallelism. Dedicated data buffers between external on-chip memory and on-chip processing engines (PEs) can be designed to achieve prioritized data flow by configuring tens of megabytes of on-chip block random access memory (BRAM) on field-programmable gate array (FPGA) chips. Summary of the Invention [Problem to be solved by the invention]

[0009] Efficient data flow and hardware architecture for CNN acceleration is desired to minimize data communication while maximizing resource utilization to achieve high performance. This creates an opportunity to design methodologies and frameworks to accelerate the inference process of various CNN algorithms on acceleration hardware and achieve high performance, high efficiency, and high flexibility. [Brief explanation of the drawings]

[0010] In the drawings, like reference characters generally refer to like parts throughout the different views. Also, the drawings are not necessarily to scale, emphasis instead being placed upon illustrating the principles of the disclosed technology. In the following description, various embodiments of the disclosed technology are described with reference to the following drawings: [Figure 1] 1 shows a cross-sectional view of a biosensor that can be used in various embodiments. [Figure 2] 1 shows an implementation of a flow cell that includes clusters within its tiles. [Figure 3] An exemplary flow cell with eight lanes is shown, along with a zoom-in of one tile and its cluster and their surrounding background. [Figure 4] FIG. 1 is a simplified block diagram of a system for analysis of sensor data from a sequencing system, such as base call sensor output. [Figure 5] FIG. 1 is a simplified diagram illustrating aspects of base calling operations, including the functionality of a runtime program executed by a host processor. [Figure 6] 5 is a simplified diagram of a configuration of a configurable processor, such as the configurable processor of FIG. 4. [Figure 7] FIG. 1 is a diagram of a neural network architecture that can be implemented using a configurable or reconfigurable array configured as described herein. [Figure 8A] FIG. 8 is a simplified diagram of the organization of tiles of sensor data used by a neural network architecture such as that of FIG. 7. [Figure 8B] FIG. 8 is a simplified diagram of a patch of tiles of sensor data used by a neural network architecture such as that of FIG. 7. [Figure 9] 8 illustrates part of the configuration of a neural network such as that of FIG. 7 on a configurable or reconfigurable array such as a field programmable gate array (FPGA). [Figure 10]FIG. 10 is a diagram of another alternative neural network architecture that can be implemented using a configurable or reconfigurable array configured as described herein. [Figure 11] 1 shows one implementation of a dedicated architecture of a neural network-based base caller used to separate the processing of data in different sequencing cycles. [Figure 12] 1 illustrates one implementation of separated layers, each of which may contain convolutions. [Figure 13A] 1 illustrates one implementation of combination layers, each of which may include a convolution. [Figure 13B] 10 illustrates another implementation of combination layers, each of which may include convolutions. [Figure 14A] We illustrate a base-calling system that operates in a single oligo training stage to train a base-caller containing a neural network configuration using known synthetic oligo sequences. [Figure 14A1] 1 illustrates a comparison operation between a predicted base sequence and the corresponding ground truth base sequence. [Figure 14B] 14A illustrates further details of the base-calling system of FIG. 14A operating in a single oligo training stage to train a base-caller containing a neural network configuration using known synthetic oligo sequences. [Figure 15A] FIG. 14A illustrates the base calling system of FIG. 14A operating in the training data generation phase of a two-oligo training stage to generate labeled training data using two known synthetic sequences. [Figure 15B] 15A illustrates two corresponding exemplary selections of the two oligo sequences discussed with respect to FIG. 15A. [Figure 15C] 15A illustrates two corresponding exemplary selections of the two oligo sequences discussed with respect to FIG. 15A. [Figure 15D]Illustrated are exemplary mapping operations for either (i) mapping a predicted base call sequence to either the first oligo or the second oligo, or (ii) declaring uncertainty in mapping a predicted base call sequence to either of the two oligos. [Figure 15E] FIG. 15D illustrates labeled training data generated from the mapping, which training data is used by another neural network configuration illustrated in FIG. 16A. [Figure 16A] FIG. 14A illustrates the base-calling system of FIG. 14A operating in the training data consumption and training phases of a two-oligo training stage to train a base-caller with another neural network configuration (different from and more complex than the neural network configuration of FIG. 14A ) using two known synthetic oligo sequences. [Figure 16B] FIG. 14B illustrates the base calling system of FIG. 14A operating in the second iteration of the training data generation phase of the two-oligo training stage. [Figure 16C] FIG. 16B illustrates labeled training data generated from the illustrated mapping, which is used for further training. [Figure 16D] FIG. 14A illustrates the base-calling system of FIG. 14A operating in the second iteration of the "Training Data Consumption and Training Phase" of the "Two Oligo Training Stage" to train a base-caller with the neural network configuration of FIG. 16A using two known synthetic oligo sequences. [Figure 17A] 1 illustrates a flowchart depicting an exemplary method for iteratively training a neural network configuration for base calling using single-oligo and two-oligo sequences. [Figure 17B] 17 illustrates exemplary labeled training data generated by the Pth NN configuration at the end of method 1700 of FIG. 17A. [Figure 18A]14A illustrates the base-calling system of FIG. 14A operating in the first iteration of the "Training Data Consumption and Training Phase" of the "3-Oligo Training Stage" to train a base-caller with a 3-oligo neural network configuration. [Figure 18B] 14A is illustrated operating in the "training data generation phase" of the "three-oligo training stage" to train a base-caller comprising the three-oligo neural network configuration of FIG. 18A. [Figure 18C] Illustrates a mapping operation that either (i) maps the predicted base call sequence to any of the three oligos in FIG. 18B, or (ii) declares the mapping of the predicted base call sequence to be indeterminate. [Figure 18D] FIG. 18C shows labeled training data generated from the mapping, which is used to train another neural network configuration. [Figure 18E] 1 illustrates a flowchart depicting an exemplary method for iteratively training a neural network configuration for base calling using 3-oligo ground truth sequences. [Figure 19] 1 illustrates a flowchart depicting an exemplary method for iteratively training a neural network configuration for base calling using multiple oligo ground truth sequences. [Figure 20A] FIG. 14B illustrates the biological sequences used to train the base caller. [Figure 20B] 14A illustrates the base calling system of FIG. 14A operating in the training data generation phase of the first organism training stage to train a base calling system comprising a first organism-level neural network configuration using various subsequences of the first organism sequence of FIG. 20A. [Figure 20C] 1 illustrates an example of fading, where signal intensity decreases as a function of cycle number in a sequencing run of a base-calling operation. [Figure 20D]1 conceptually illustrates the decreasing signal-to-noise ratio as the sequencing progresses cycles. [Figure 20E] Illustrates base calling of the first L2 bases of the L1 bases of a subsequence, where the first L2 bases of the subsequence are used to map the subsequence to the biological sequence of Figure 20A. [Figure 20F] FIG. 20E illustrates labeled training data generated from the mapping, where the labeled training data includes a section of the biological sequence of FIG. 20A as ground truth. [Figure 20G] 14A illustrates the basecalling system of FIG. 14A operating in the "training data consumption and training phase" of the "organism-level training stage" to train a basecaller with a first organism-level neural network configuration. [Figure 21] 20A illustrates a flowchart depicting an exemplary method for iteratively training a neural network configuration for base calling using the simple biological sequence of FIG. 20A. [Figure 22] 14B illustrates the use of complex biological sequences for training the corresponding NN architecture for the base collaborators of FIG. 14A. [Figure 23A] 1 illustrates a flowchart depicting an exemplary method for iteratively training a neural network configuration for basecalling. [Figure 23B] 10 illustrates various charts illustrating the effectiveness of the base call training process discussed in this disclosure. [Figure 23C] 10 illustrates various charts illustrating the effectiveness of the base call training process discussed in this disclosure. [Figure 23D] 10 illustrates various charts illustrating the effectiveness of the base call training process discussed in this disclosure. [Figure 23E] 10 illustrates various charts illustrating the effectiveness of the base call training process discussed in this disclosure. [Figure 24] FIG. 1 is a block diagram of a base calling system according to one implementation. [Figure 25]FIG. 25 is a block diagram of a system controller that can be used in the system of FIG. 24. [Figure 26] FIG. 1 is a simplified block diagram of a computer system that can be used to implement the disclosed techniques. DETAILED DESCRIPTION OF THE INVENTION

[0011] As used herein, the terms "polynucleotide" or "nucleic acid" refer to deoxyribonucleic acid (DNA); however, where appropriate, those skilled in the art will recognize that the systems and devices herein can also be utilized with ribonucleic acid (RNA). These terms should be understood to include, as equivalents, analogs of either DNA or RNA made from nucleotide analogs. As used herein, these terms also encompass complementary cDNA or copy DNA produced from an RNA template, for example, by the action of reverse transcriptase.

[0012] Single-stranded polynucleotide molecules sequenced by the systems and devices herein can originate in single-stranded form as DNA or RNA, or in double-stranded DNA (dsDNA) form (e.g., genomic DNA fragments, PCR and amplification products, and the like). Thus, single-stranded polynucleotides can be the sense or antisense strand of a polynucleotide duplex. Methods for preparing single-stranded polynucleotide molecules suitable for use in the methods of the present disclosure using standard techniques are known in the art. The exact sequence of the primary polynucleotide molecule is generally not critical to the present disclosure and can be known or unknown. Single-stranded polynucleotide molecules can represent genomic DNA molecules (e.g., human genomic DNA), including both intron and exon sequences (coding sequences), as well as non-coding regulatory sequences such as promoter and enhancer sequences.

[0013] In certain embodiments, nucleic acids to be sequenced through use of the present disclosure are immobilized on a substrate (e.g., a substrate in a flow cell, or one or more beads on a substrate such as a flow cell, etc.). As used herein, the term "immobilized" is intended to encompass direct or indirect, covalent or non-covalent attachment, unless otherwise indicated explicitly or by context. While covalent attachment may be preferred in certain embodiments, what is generally required is that the molecule (e.g., nucleic acid) remain immobilized or associated with the support under conditions under which the support is intended to be used, e.g., in applications requiring nucleic acid sequencing.

[0014] The term "solid support" (or "substrate" in some usages) as used herein refers to any inert substrate or matrix to which nucleic acids can be attached, such as, for example, a glass surface, a plastic surface, latex, dextran, a polystyrene surface, a polypropylene surface, a polyacrylamide gel, a gold surface, a silicon wafer, etc. In many embodiments, the solid support is a glass surface (e.g., the flat surface of a flow cell channel). In certain embodiments, the solid support may be comprised of an inert substrate or matrix that has been "functionalized," for example, by the application of a layer or coating of an intermediate material that contains reactive groups that allow for covalent attachment to molecules such as polynucleotides. As a non-limiting example, such a support may comprise a polyacrylamide hydrogel supported on an inert substrate such as glass. In such embodiments, the molecule (polynucleotide) can be covalently attached directly to the intermediate material (e.g., the hydrogel), while the intermediate material can itself be non-covalently attached to the substrate or matrix (e.g., the glass substrate). Covalent attachment to a solid support should be interpreted accordingly to encompass this type of arrangement.

[0015] As described above, the present disclosure includes novel systems and devices for sequencing nucleic acids. As will be apparent to those skilled in the art, a reference herein to a specific nucleic acid sequence may, depending on the context, also refer to a nucleic acid molecule containing such a nucleic acid sequence. Sequencing a target fragment means establishing a chronological reading of bases. The bases read need not be consecutive, although this is preferred, and not all bases on the entire fragment need to be sequenced during sequencing. Sequencing can be performed using any suitable sequencing technique in which nucleotides or oligonucleotides are added sequentially to a free 3' hydroxyl group, resulting in the synthesis of a polynucleotide chain in a 5' to 3' direction. The identity of the added nucleotide is preferably determined after each nucleotide addition. Sequencing techniques using sequencing by ligation, in which not all consecutive bases are sequenced, and techniques such as massively parallel signature sequencing (MPSS), which removes bases rather than adding them to a surface strand, are also suitable for use with the systems and devices of the present disclosure.

[0016] In certain embodiments, the present disclosure discloses sequencing-by-synthesis (SBS), in which four fluorescently labeled, modified nucleotides are used to sequence high-density clusters (potentially millions of clusters) of amplified DNA present on the surface of a substrate (e.g., a flow cell). Various additional aspects of SBS procedures and methods that can be utilized with the systems and devices herein are disclosed, for example, in WO 04018497, WO 04018493, and U.S. Pat. No. 7,057,026 (nucleotides), WO 05024010 and WO 06120433 (polymerases), WO 05065814 (surface attachment techniques), and WO 9844151, WO 06064199, and WO 07010251, the contents of each of which are incorporated herein by reference in their entirety.

[0017] In a specific use of the system / device herein, a flow cell containing a nucleic acid sample for sequencing is placed in an appropriate flow cell holder. The sample for sequencing can take the form of a single molecule, amplified single molecules in the form of clusters, or beads containing molecules of nucleic acid. The nucleic acid is prepared to include oligonucleotide primers flanking the unknown target sequence. To initiate the first SBS sequencing cycle, one or more differently labeled nucleotides and a DNA polymerase, etc., are flowed into / through the flow cell by a fluid flow subsystem (various embodiments of which are described herein). A single nucleotide can be added at a time, or the nucleotides used in the sequencing procedure can be specifically designed to have reversible termination properties, thus allowing each cycle of the sequencing reaction to occur simultaneously in the presence of all four labeled nucleotides (A, C, T, G). When the four nucleotides are mixed together, the polymerase can select and incorporate the correct base, and each sequence is extended by a single base. In such methods using the system, the natural competition between the four alternatives results in greater accuracy than if only one nucleotide were present in the reaction mixture (thus exposing most of the sequence to the correct nucleotide). Sequences in which a particular base is repeated one after the other (e.g., homopolymers) are addressed with as much accuracy as any other sequence.

[0018] The fluid flow subsystem also flows appropriate reagents to remove blocked 3' ends (if appropriate) and fluorophores from each incorporated base. The substrate can then be exposed to either a second round of four blocked nucleotides, or, optionally, a second round using a different individual nucleotide. This cycle is then repeated, and the sequence of each cluster is read over multiple chemical cycles. Computerized aspects of the present disclosure can optionally align sequence data collected from each single molecule, cluster, or bead to determine the sequence of longer polymers, etc. Alternatively, image processing and alignment can be performed on separate computers.

[0019] The heating / cooling components of the system regulate reaction conditions within the flow cell channel and reagent storage area / reservoir (and optionally, camera, optics, and / or other components), while the fluid flow components allow the substrate surface to be exposed to appropriate reagents for incorporation (e.g., appropriate fluorescently labeled nucleotides to be incorporated) while unincorporated reagents are washed away. An optional movable stage on which the flow cell rests allows the flow cell to be properly oriented for laser (or other light) excitation of the substrate and optionally moves the flow cell relative to the objective to read different areas of the substrate. In addition, other components of the system are also optionally movable / adjustable (e.g., camera, objective, heater / cooler, etc.). During laser excitation, images / locations of fluorescence emitted from nucleic acids on the substrate are captured by the camera component, thereby recording the identity of the first base for each single molecule, cluster, or bead in the computer component.

[0020] The embodiments described herein may be used in a variety of biological or chemical processes and systems for academic or commercial analysis. More specifically, the embodiments described herein may be used in a variety of processes and systems in which it is desirable to detect an event, characteristic, quality, or property indicative of a desired response. For example, the embodiments described herein include cartridges, biosensors, and their components, as well as bioassay systems that operate with the cartridges and biosensors. In certain embodiments, the cartridges and biosensors include a flow cell and one or more sensors, pixels, photodetectors, or photodiodes coupled together in a substantially single structure.

[0021] The following detailed description of certain embodiments may be better understood when read in conjunction with the accompanying drawings. To the extent that the figures illustrate diagrams of functional blocks of various embodiments, the functional blocks are not necessarily indicative of a division between hardware circuitry. Thus, for example, one or more of the functional blocks (e.g., a processor or memory) may be implemented in a single piece of hardware (e.g., a general-purpose signal processor or random access memory, a hard disk, etc.). Similarly, a program may be a stand-alone program, may be incorporated as a subroutine within an operating system, may be a function within an installed software package, etc. It should be understood that the various embodiments are not limited to the arrangements and instrumentality shown in the figures.

[0022] As used herein, elements or steps described in the singular and followed by the word "a" or "an" should be understood as not excluding a plurality of those elements or steps, unless such exclusion is expressly stated. Furthermore, references to "one embodiment" are not intended to be interpreted as excluding the existence of additional embodiments that also incorporate the recited features. Furthermore, unless expressly stated to the contrary, embodiments "comprising" or "having" or "including" an element or elements having a particular characteristic may include the additional elements, whether or not they have that characteristic.

[0023] As used herein, a "desired reaction" includes a change in at least one of the chemical, electrical, physical, or optical properties (or qualities) of an analyte of interest. In certain embodiments, the desired reaction is a positive binding event (e.g., the incorporation of a fluorescently labeled biomolecule into the analyte of interest). More generally, the desired reaction may be a chemical conversion, chemical change, or chemical interaction. The desired reaction may also be a change in an electrical property. For example, the desired reaction may be a change in ion concentration in a solution. Exemplary reactions include, but are not limited to, chemical reactions such as reduction, oxidation, addition, elimination, rearrangement, esterification, amidation, etherification, cyclization, or substitution; binding interactions in which a first chemical binds to a second chemical; dissociation reactions in which two or more chemicals separate from one another; fluorescence, luminescence, bioluminescence, chemiluminescence; and biological reactions such as nucleic acid replication, nucleic acid amplification, nucleic acid hybridization, nucleic acid ligation, phosphorylation, enzyme catalysis, receptor binding, or ligand binding. The desired reaction may also be the addition or removal of a proton, which is detectable, for example, as a change in pH of the surrounding solution or environment. An additional desired response can be the detection of ion flow across a membrane (e.g., a natural or synthetic bilayer membrane), e.g., when ions flow through the membrane, the current is disrupted and this disruption can be detected.

[0024] In certain embodiments, the desired reaction involves the incorporation of a fluorescently labeled molecule into the analyte. The analyte may be an oligonucleotide, and the fluorescently labeled molecule may be a nucleotide. The desired reaction may be detected when excitation light is directed at the oligonucleotide bearing the labeled nucleotide and the fluorophore emits a detectable fluorescent signal. In alternative embodiments, the detected fluorescence is the result of chemiluminescence or bioluminescence. The desired reaction may also increase fluorophore (or Förster) resonance energy transfer (FRET) by bringing a donor fluorophore into close proximity with an acceptor fluorophore, decrease FRET by separating the donor and acceptor fluorophores, increase fluorescence by separating a quencher from the fluorophore, or decrease fluorophore by colocalizing the quencher and fluorophore.

[0025] As used herein, "reaction component" or "reactant" includes any substance that can be used to obtain a desired reaction. For example, reaction components include reagents, enzymes, samples, other biomolecules, and buffers. Reaction components are typically delivered to a reaction site in solution and / or immobilized at the reaction site. A reaction component can interact directly or indirectly with another substance, such as an analyte of interest.

[0026] As used herein, the term "reaction site" refers to a localized region where a desired reaction can occur. A reaction site may include a support surface of a substrate onto which a substance can be immobilized. For example, a reaction site may include a substantially planar surface within a channel of a flow cell having colonies of nucleic acids thereon. Typically, but not always, the nucleic acids in the colonies have the same sequence, e.g., clonal copies of a single-stranded or double-stranded template. However, in some embodiments, a reaction site may contain only a single nucleic acid molecule, e.g., in single-stranded or double-stranded form. Furthermore, multiple reaction sites may be distributed unevenly along the support surface or arranged in a predetermined manner (e.g., parallel in a matrix such as a microarray). A reaction site may also include a reaction chamber (or well) that at least partially defines a spatial region or volume configured to compartmentalize a desired reaction.

[0027] This application uses the terms "reaction chamber" and "well" interchangeably. As used herein, the term "reaction chamber" or "well" includes a spatial region that is in fluid communication with a flow channel. A reaction chamber may be at least partially isolated from the surrounding environment or another spatial region. For example, multiple reaction chambers may be separated from one another by a shared wall. As a more specific example, a reaction chamber may include a cavity defined by the inner surface of a well and have an opening or aperture such that the cavity is in fluid communication with the flow channel. A biosensor including such a reaction chamber is described in more detail in International Application PCT / US2011 / 057111, filed October 20, 2011, the entire contents of which are incorporated herein by reference.

[0028] In some embodiments, the reaction chamber is sized and shaped relative to a solid (including a semi-solid) so that the solid can be fully or partially inserted therein. For example, the reaction chamber is sized and shaped to accommodate only one capture bead. The capture bead may have clonally amplified DNA or other material thereon. Alternatively, the reaction chamber is sized and shaped to receive a number of beads or solid substrates. As another example, the reaction chamber may also be filled with a porous gel or material configured to control diffusion or filter fluids that may enter the reaction chamber.

[0029] In some embodiments, a sensor (e.g., a photodetector, photodiode) is associated with a corresponding pixel area on the sample surface of the biosensor. Thus, a pixel area is a geometric construct that represents the area on the sample surface of a biosensor of one sensor (or pixel). The sensor associated with a pixel area detects luminescence collected from the associated pixel area when a desired reaction occurs at a reaction site or reaction chamber above the associated pixel area. In flat surface embodiments, pixel areas can overlap. In some cases, multiple sensors can be associated with a single reaction site or a single reaction chamber. In other cases, a single sensor can be associated with a group of reaction sites or a group of reaction chambers.

[0030] As used herein, a "biosensor" includes a structure having multiple reaction sites and / or reaction chambers (or wells). The biosensor may include a solid-state imaging device (e.g., a CCD or CMOS imager) and, optionally, a flow cell attached thereto. The flow cell may include at least one flow channel in fluid communication with the reaction sites and / or reaction chambers. As one specific example, the biosensor is configured to fluidly and electrically couple to a bioassay system. The bioassay system may deliver reactants to the reaction sites and / or reaction chambers according to a predetermined protocol (e.g., sequencing-by-synthesis) and perform multiple imaging events. For example, the bioassay system may direct solutions to flow along the reaction sites and / or reaction chambers. At least one of the solutions may include four types of nucleotides with the same or different fluorescent labels. The nucleotides may bind to corresponding oligonucleotides located in the reaction sites and / or reaction chambers. The bioassay system can then illuminate the reaction sites and / or reaction chambers using an excitation light source (e.g., a solid-state light source such as a light-emitting diode or LED). The excitation light may have a predetermined wavelength or multiple wavelengths, including a range of wavelengths. The excited fluorescent label provides a luminescent signal that can be captured by a sensor.

[0031] In alternative embodiments, the biosensor may include electrodes or other types of sensors configured to detect other distinguishable characteristics. For example, the sensor may be configured to detect changes in ion concentration. In another example, the sensor may be configured to detect the flow of ionic current across a membrane.

[0032] As used herein, a "cluster" is a colony of similar or identical molecules or nucleotide sequences or DNA strands. For example, a cluster can be an amplification oligonucleotide or any other group of polynucleotides or polypeptides with the same or similar sequence. In other embodiments, a cluster can be any element or group of elements that occupy a physical region on the sample surface. In embodiments, the cluster is immobilized in a reaction site and / or reaction chamber during the base call cycle.

[0033] As used herein, the term "immobilized," when used in reference to a biomolecule or biological substance or chemical, includes substantially attaching the biomolecule or biological substance or chemical to a surface at the molecular level. For example, a biomolecule or biological substance or chemical may be immobilized on the surface of a substrate material using adsorption techniques, including non-covalent bonding (e.g., electrostatic forces, van der Waals, and hydrophobic interfacial dehydration), as well as covalent bonding techniques in which a functional group or linker facilitates attachment of the biomolecule to the surface. Immobilizing a biomolecule or biological substance or chemical to the surface of a substrate material may be based on the properties of the substrate surface, the liquid medium carrying the biomolecule or biological substance or chemical, and the properties of the biomolecule or biological substance or chemical itself. In some cases, the substrate surface may be functionalized (e.g., chemically or physically modified) to facilitate immobilization of the biomolecule (or biological substance or chemical) to the surface. The substrate surface may first be modified to have functional groups attached to the surface. The functional groups may then bind to the biomolecule or biological substance or chemical, immobilizing them thereon. Substances can be immobilized on a surface via a gel, for example, as described in US Patent Application Publication No. 2011 / 0059865(A1), which is incorporated herein by reference.

[0034] In some embodiments, nucleic acids can be attached to a surface and amplified using bridge amplification. Useful bridge amplification methods are described, for example, in U.S. Patent No. 5,641,658, International Publication No. 2007 / 010251, U.S. Patent No. 6,090,592, U.S. Patent Application Publication No. 2002 / 0055100 (A1), U.S. Patent No. 7,115,400, U.S. Patent Application Publication No. 2004 / 0096853 (A1), U.S. Patent Application Publication No. 2004 / 0002090 (A1), U.S. Patent Application Publication No. 2007 / 0128624 (A1), and U.S. Patent Application Publication No. 2008 / 0009420 (A1), each of which is incorporated herein in its entirety. Another useful method for amplifying nucleic acids on a surface is rolling circle amplification (RCA), for example, using methods described in more detail below. In some embodiments, a nucleic acid can be attached to a surface and amplified using one or more primer pairs. For example, one of the primers can be in solution, and the other primer can be immobilized (e.g., 5'-attached) on the surface. As an example, a nucleic acid molecule can hybridize to one of the primers on the surface, followed by extension of the immobilized primer to generate a first copy of the nucleic acid. The primer in solution then hybridizes to the first copy of the nucleic acid, which can be extended using the first copy of the nucleic acid as a template. Optionally, after the first copy of the nucleic acid is generated, the original nucleic acid molecule can hybridize to a second immobilized primer on the surface and be extended simultaneously or after the primer in solution is extended. In any embodiment, repeated rounds of extension (e.g., amplification) using the immobilized primer and the primer in solution provide multiple copies of the nucleic acid.

[0035] In certain embodiments, the assay protocols performed by the systems and methods described herein include the use of naturally occurring nucleotides and enzymes configured to interact with the naturally occurring nucleotides. Naturally occurring nucleotides include, for example, ribonucleotides (RNA) or deoxyribonucleotides (DNA). The naturally occurring nucleotides may be in monophosphate, diphosphate, or triphosphate form and may have a base selected from adenine (A), thymine (T), uracil (U), guanine (G), or cytosine (C). However, it will be understood that non-naturally occurring nucleotides, modified nucleotides, or analogs of the above nucleotides may be used. Some examples of useful non-naturally occurring nucleotides are described below with respect to reversible terminator-based sequencing by synthetic methods.

[0036] In embodiments including a reaction chamber, an article or solid substance (including a semi-solid substance) may be placed within the reaction chamber. When placed, the article or solid may be physically held or immobilized within the reaction chamber via interference fit, adhesion, or confinement. Exemplary articles or solids that may be placed within the reaction chamber include polymer beads, pellets, agarose gel, powders, quantum dots, or other solids that can be compressed and / or held within the reaction chamber. In certain embodiments, nucleic acid superstructures such as DNA balls can be placed within or within the reaction chamber, for example, by attaching them to the inner surface of the reaction chamber or by dwelling in a liquid within the reaction chamber. DNA balls or other nucleic acid superstructures can be preformed and then placed within or within the reaction chamber. Alternatively, DNA balls can be synthesized in situ in the reaction chamber. DNA balls can be synthesized by rolling circle amplification to generate concatemers of specific nucleic acid sequences, and the concatemers can be treated under conditions that form relatively compact balls. DNA balls and methods for their synthesis are described, for example, in U.S. Patent Application Publication Nos. 2008 / 0242560(A1) or 2008 / 0234136(A1), each of which is incorporated herein in its entirety. The material held or disposed within the reaction chamber can be in a solid, liquid, or gaseous state.

[0037] As used herein, "base calling" refers to identifying nucleotide bases in a nucleic acid sequence. Base calling refers to the process of determining the base call (A, C, G, T) of every cluster in a particular cycle. By way of example, base calling can be performed using the four-channel, two-channel, or one-channel methods and systems described in the incorporated materials of U.S. Patent Application Publication No. 2013 / 0079232. In certain embodiments, a base calling cycle is referred to as a "sampling event." In a one-dye, two-channel sequencing protocol, a sampling event includes two illumination steps in chronological order, such that a pixel signal is generated at each step. The first illumination step induces illumination from a given cluster representing nucleotide bases A and T in an AT pixel signal, and the second illumination step induces illumination from a given cluster representing nucleotide bases C and T in a CT pixel signal.

[0038] The disclosed technology, for example, the disclosed base code, can be implemented on processors such as central processing units (CPUs), graphics processing units (GPUs), field programmable gate arrays (FPGAs), coarse-grained reconfigurable architectures (CGRAs), application-specific integrated circuits (ASICs), application-specific instruction-set processors (ASIPs), and digital signal processors (DSPs).

[0039] Biosensors FIG. 1 shows a cross-sectional view of a biosensor 100 that can be used in various embodiments. The biosensor 100 has pixel areas 106′, 108′, 110′, 112′, and 114′, each capable of retaining two or more clusters (e.g., two clusters per pixel area) during a base call cycle. As shown, the biosensor 100 can include a flow cell 102 mounted on a sampling device 104. In the illustrated embodiment, the flow cell 102 is directly fixed to the sampling device 104. However, in alternative embodiments, the flow cell 102 can be removably coupled to the sampling device 104. The sampling device 104 has a sample surface 134 that can be functionalized (e.g., chemically or physically modified in a manner suitable for causing a desired reaction). For example, the sample surface 134 may be functionalized and may include multiple pixel regions 106′, 108′, 110′, 112′, and 114′, each capable of holding two or more clusters during a base calling cycle (e.g., having corresponding cluster pairs 106A, 106B, cluster pairs 108A, 108B, cluster pairs 110A, 110B, cluster pairs 112A, 112B, and cluster pairs 114A, 114B immobilized thereon). Each pixel region is associated with a corresponding sensor (or pixel or photodiode) 106, 108, 110, 112, and 114, such that light received by the pixel region is captured by the corresponding sensor. The pixel region 106′ may also be associated with a corresponding reaction site 106″ on the reaction surface 134 that holds the cluster pair, such that light emitted from the reaction site 106″ is received by the pixel region 106′ and captured by the corresponding sensor 106. As a result of this sensing structure, if two or more clusters are present in a particular sensor pixel area during a base call cycle (e.g., each with a corresponding cluster pair), the pixel signal in that base call cycle carries information based on all of the two or more clusters.As a result, the signal processing described herein is used to distinguish between clusters where there are more clusters than pixel signals at a given sampling event of a particular base call cycle.

[0040] In the illustrated embodiment, the flow cell 102 includes sidewalls 138, 125 and a flow cover 136 supported by the sidewalls 138, 125. The sidewalls 138, 125 are coupled to the sample surface 134 and extend between the flow cover 136 and the sidewalls 138, 125. In some embodiments, the sidewalls 138, 125 are formed from a curable adhesive layer that bonds the flow cover 136 to the sampling device 104.

[0041] The side walls 138, 125 are sized and shaped such that a flow channel 144 exists between the flow cover 136 and the sampling device 104. The flow cover 136 may comprise a material that is transparent to excitation light 101 propagating from outside the biosensor 100 into the flow channel 144. In one example, the excitation light 101 approaches the flow cover 136 at a non-orthogonal angle.

[0042] Also as shown, the flow cover 136 may include inlet and outlet ports 142, 146 configured to fluidly engage other ports (not shown). For example, these other ports may be from the cartridge or a workstation. The flow channel 144 is sized and shaped to direct fluid along the sample surface 134. The height H1 and other dimensions of the flow channel 144 may be configured to maintain a substantially uniform flow of fluid along the sample surface 134. The dimensions of the flow channel 144 may also be configured to control bubble formation.

[0043] By way of example, the flow cover 136 (or flow cell 102) may comprise a transparent material such as glass or plastic. The flow cover 136 may comprise a substantially rectangular block having a planar outer surface and a planar inner surface that defines the flow channel 144. The block may be attached to the sidewalls 138, 125. Alternatively, the flow cell 102 may be etched to define the flow cover 136 and sidewalls 138, 125. For example, a recess may be etched into the transparent material. When the etched material is attached to the sampling device 104, the recess may become the flow channel 144.

[0044] The sampling device 104 may be similar to an integrated circuit comprising, for example, multiple stacked substrate layers 120-126. The substrate layers 120-126 may include a base substrate 120, a solid-state imager 122 (e.g., a CMOS image sensor), a filter or light management layer 124, and a passivation layer 126. Note that the above is merely exemplary, and other embodiments may include fewer or additional layers. Furthermore, each of the substrate layers 120-126 may include multiple sublayers. The sampling device 104 may be fabricated using processes similar to those used in fabricating integrated circuits such as CMOS image sensors and CCDs. For example, the substrate layers 120-126, or portions thereof, may be grown, deposited, etched, etc. to form the sampling device 104.

[0045] The passivation layer 126 is configured to shield the filter layer 124 from the fluid environment of the flow channel 144. In some cases, the passivation layer 126 is also configured to provide a solid surface (i.e., the sample surface 134) onto which biomolecules or other analytes of interest can be immobilized. For example, each of the reaction sites can include a cluster of biomolecules immobilized on the sample surface 134. Thus, the passivation layer 126 can be formed from a material that allows the reaction sites to be immobilized thereon. The passivation layer 126 can also include a material that is at least transparent to the desired fluorescence. By way of example, the passivation layer 126 can include silicon nitride (SiN) and / or silica (SiO). However, other suitable materials can be used. In the illustrated embodiment, the passivation layer 126 can be substantially planar. However, in alternative embodiments, the passivation layer 126 can include recesses, such as pits, wells, or grooves. In the illustrated embodiment, the passivation layer 126 has a thickness of about 150-200 nm, more specifically about 170 nm.

[0046] The filter layer 124 may include various features that affect light transmission. In some embodiments, the filter layer 124 may perform multiple functions. For example, the filter layer 124 may be configured to (a) filter unwanted light signals, such as light signals from an excitation light source; (b) direct luminescence signals from reaction sites toward corresponding sensors 106, 108, 110, 112, and 114 configured to detect the luminescence signals from the reaction sites; or (c) block or prevent detection of unwanted luminescence signals from adjacent reaction sites. Thus, the filter layer 124 may also be referred to as a light management layer. In the illustrated embodiment, the filter layer 124 has a thickness of approximately 1-5 μm, more specifically, approximately 2-4 μm. In alternative embodiments, the filter layer 124 may include an array of microlenses or other optical components. Each of the microlenses may be configured to direct the luminescence signal from an associated reaction site to a sensor.

[0047] In some embodiments, the solid-state imager 122 and the base substrate 120 may be provided together as a previously constructed solid-state imaging device (e.g., a CMOS chip). For example, the base substrate 120 may be a silicon wafer, and the solid-state imager 122 may be mounted thereon. The solid-state imager 122 includes a layer of semiconductor material (e.g., silicon) and sensors 106, 108, 110, 112, and 114. In the illustrated embodiment, the sensors are photodiodes configured to detect light. In other embodiments, the sensors include photodetectors. The solid-state imager 122 may be fabricated as a single chip via a CMOS-based fabrication process.

[0048] The solid-state imager 122 may include a high-density array of sensors 106, 108, 110, 112, and 114 configured to detect activity indicative of a desired response from within or along the flow channel 144. In some embodiments, each sensor has an area of ​​approximately 1-2 square micrometers (μm 2 ) The array can include 500,000 sensors, 5 million sensors, 10 million sensors, or even 120 million sensors. Sensors 106, 108, 110, 112, and 114 can be configured to detect predetermined wavelengths of light that exhibit a desired response.

[0049] In some embodiments, the sampling device 104 includes a microcircuit arrangement, such as that described in U.S. Patent No. 7,595,882, which is incorporated herein by reference in its entirety. More specifically, the sampling device 104 may comprise an integrated circuit having a planar array of sensors 106, 108, 110, 112, and 114. The circuitry formed within the sampling device 104 may be configured for at least one of signal amplification, digitization, storage, and processing. The circuitry may collect and analyze the detected fluorescence and generate pixel signals (or detection signals) for communicating the detection data to a signal processor. The circuitry may also perform additional analog and / or digital signal processing within the sampling device 104. The sampling device 104 may include conductive vias 130 for signal routing (e.g., transmitting pixel signals to a signal processor). The pixel signals may also be transmitted through electrical contacts 132 of the sampling device 104.

[0050] The sampling device 104 is discussed in further detail with respect to U.S. Non-Provisional Patent Application No. 16 / 874,599, entitled "Systems and Devices for Characterization and Performance Analysis of Pixel-Based Sequencing," filed May 14, 2020 (Attorney Docket No. ILLM1011-4 / IP-1750-US), which is incorporated by reference as if fully set forth herein. The sampling device 104 is not limited to the above configuration or use as described above. In alternative embodiments, the sampling device 104 may take other forms. For example, the sampling device 104 may comprise a CCD device, such as a CCD camera, coupled to a flow cell or moved to interface with a flow cell having reaction sites therein.

[0051] Figure 2 shows one implementation of a flow cell 200 that includes clusters within its tiles. Flow cell 200 corresponds to flow cell 102 of Figure 1, e.g., without flow cover 136. Furthermore, the depiction of flow cell 200 is symbolic in nature, and flow cell 200 symbolically shows the various lanes and tiles therein without showing the various other components therein. Figure 2 shows a top view of flow cell 200.

[0052] In one embodiment, flow cell 200 is divided or segmented into multiple lanes, such as lanes 202a, 202b, ..., 202P, i.e., P lanes. In the example of Figure 2, flow cell 200 is shown as including eight lanes, i.e., P = 8 in this example, although the number of lanes in a flow cell is implementation specific.

[0053] In one embodiment, each lane 202 is further divided into non-overlapping regions called "tiles" 212. For example, Figure 2 shows an expanded view of a section 208 of an exemplary lane. The section 208 is shown to include multiple tiles 212.

[0054] In an embodiment, each lane 202 includes one or more tile columns. For example, in Figure 2, each lane 202 includes two corresponding tile columns 212, as shown in enlarged section 208. The number of tiles in each tile column in each lane is implementation specific, and in one example, there may be 50 tiles, 60 tiles, 100 tiles, or another suitable number of tiles in each tile column in each lane.

[0055] Each tile contains a corresponding number of clusters. During the sequencing procedure, the clusters on the tile and their surrounding background are imaged. For example, Figure 2 shows an example cluster 216 in an example tile.

[0056] Figure 3 shows an exemplary Illumina GA-IIx™ flow cell with eight lanes, including a zoomed-in view of one tile and its clusters and their surrounding background. For example, there are 100 tiles per lane on the Illumina Genome Analyzer II and 68 tiles per lane in the Illumina HiSeq2000. Tile 212 holds hundreds of thousands to millions of clusters. In Figure 3, an image generated from a tile with clusters shown as bright spots is shown in 308 (e.g., 308 is a magnified image view of the tile), and exemplary cluster 304 is labeled. Cluster 304 contains approximately 1,000 identical copies of the template molecule, although the clusters vary in size and shape. Clusters are grown from template molecules by bridge amplification of the input library prior to sequencing runs. The purpose of amplification and cluster growth is to increase the intensity of the emitted signal, as imaging devices cannot reliably sense a single fluorophore. However, the physical distance between the DNA fragments within the cluster 304 is small, so the imaging device perceives the cluster of fragments as a single spot 304 .

[0057] Clusters and tiles are discussed in further detail with respect to U.S. Non-Provisional Patent Application No. 16 / 825,987, entitled "TRAINING DATA GENERATION FOR ARTIFICIAL INTELLIGENCE-BASED SEQUENCING," filed March 20, 2020 (Attorney Docket No. ILLM1008-16 / IP-1693-US).

[0058] FIG. 4 is a simplified block diagram of a system for analyzing sensor data, such as base calling sensor output, from a sequencing system (see, e.g., FIG. 1). In the example of FIG. 4, the system includes a sequencing machine 400 and a configurable processor 450. The configurable processor 450 can execute a neural network-based base caller in cooperation with a runtime program executed by a host processor, such as a central processing unit (CPU) 402. The sequencing machine 400 includes a base call sensor (e.g., as discussed with reference to FIGS. 1-3) and a flow cell 401. The flow cell, as discussed with reference to FIGS. 1-3, can include one or more tiles in which clusters of genetic material are exposed to a sequence of analyte flow used to induce reactions within the clusters and identify bases in the genetic material. A sensor senses the reaction of each cycle of the sequence in each tile of the flow cell to provide tile data. An example of this technology is described in more detail below. Genetic sequencing is a data-intensive operation that converts base call sensor data into a sequence of base calls for each group of genetic material sensed during the base calling operation.

[0059] The system in this example includes a CPU 402 that executes a runtime program that coordinates the base calling operation, and memory 403 that stores the sequences of arrays of tile data, base call reads generated by the base calling operation, and other information used in the base calling operation. In this illustration, the system also includes memory 404 that stores a configuration file (or files), e.g., FPGA bit files, and neural network model parameters used to configure and reconfigure configurable processor 450 and to run the neural network. Sequencing machine 400 can include programs for configuring the configurable processor, and in some embodiments, can include a reconfigurable processor that runs the neural network.

[0060] Sequencing machine 400 is coupled to configurable processor 450 by bus 405. Bus 405 can be implemented using high-throughput technology, such as bus technology compatible with the PCIe (Peripheral Component Interconnect Express) standard currently maintained and developed by the PCI-SIG (Peripheral PCI Special Interest Group) standard. Also, in this embodiment, memory 460 is coupled to configurable processor 450 by bus 461. Memory 460 can be on-board memory located on a circuit board with configurable processor 450. Memory 460 is used for high-speed access by configurable processor 450 of working data used in base calling operations. Bus 461 can also be implemented using high-throughput technology, such as bus technology compatible with the PCIe standard.

[0061] Configurable processors, including field programmable gate arrays (FPGAs), coarse-grained reconfigurable arrays (CGRAs), and other configurable and reconfigurable devices, can be configured to implement various functions more efficiently or faster than can be achieved using a general-purpose processor executing a computer program. Configuring a configurable processor involves compiling a functional description to generate a configuration file, sometimes called a bitstream or bitfile, and distributing the configuration file to configurable elements on the processor.

[0062] The configuration file configures the circuit to set data flow patterns, including the use of distributed memory and other on-chip memory resources, lookup table contents, the operation of configurable logic blocks, and configurable execution units such as configurable interconnects and other elements of the configurable array. A configuration file is reconfigurable if it can be changed in the field by modifying a loaded configuration file. For example, the configuration file may be stored in a volatile SRAM element, in a non-volatile read-write memory element, or distributed among an array of configurable elements on a configurable or reconfigurable processor. Various commercially available configurable processors are suitable for use in base calling operations as described herein. Examples include commercially available products such as the Xilinx Alveo™ U200, Xilinx Alveo™ U250, Xilinx Alveo™ U280, Intel / Altera Stratix™ GX2800, Intel / Altera Stratix™ GX2800, and Intel Stratix™ GX10M. In some embodiments, the host CPU may be implemented on the same integrated circuit as the configurable processor.

[0063] The embodiments described herein use a configurable processor 450 to implement a multi-cycle neural network. The configuration file for the configurable processor can be implemented by specifying the logical functions to be performed using a high-level description language (HDL) or register transfer level (RTL) language specification. This specification can be compiled using resources designed by a selected configurable processor to generate the configuration file. The same or similar specifications can be compiled to generate the design of an application-specific integrated circuit, which may not be a configurable processor.

[0064] Thus, alternatives to the configurable processor in all embodiments described herein include a configured processor including an application specific ASIC or dedicated integrated circuit or set of integrated circuits, or a system-on-chip SOC device, configured to perform the neural network-based base calling operations described herein.

[0065] In general, the configurable and configured processors described herein that are configured to perform neural network operations are referred to herein as neural network processors.

[0066] Configurable processor 450 is configured, in this example, by a configuration file loaded using a program executed by CPU 402 or other source that configures an array of configurable elements on configurable processor 454 to perform base calling functions. In this example, the configuration includes data flow logic 451 coupled to buses 405 and 461 and that performs the function of distributing data and control parameters among elements used in base calling operations.

[0067] Configurable processor 450 is also configured with base call execution logic 452 to execute the multi-cycle neural network. Logic 452 includes a plurality of multi-cycle execution clusters (e.g., 453), which in this example include multi-cycle cluster 1 through multi-cycle cluster X. The number of multi-cycle clusters can be selected according to tradeoffs involving the desired throughput of operation and the available resources on the configurable processor.

[0068] The multi-cycle clusters are coupled to data flow logic 451 by data flow paths 454, implemented using configurable interconnect and memory resources on a configurable processor, and by control paths 455, implemented using configurable interconnect and memory resources on a configurable processor, for example, that provide control signals indicating available clusters, readiness to provide input units to available clusters for execution of neural network operations, readiness to provide trained parameters for the neural network, readiness to provide output patches of base call classification data, and other control data used in the execution of the neural network.

[0069] The configurable processor is configured to perform a multi-cycle neural network operation using the trained parameters to generate classification data for sensing cycles of the base calling operation. The neural network operation is performed to generate classification data for subject sensing cycles of the base calling operation. The neural network operation operates on an array including a number N of arrays of tile data from each sensing cycle of N sensing cycles, which, in the examples described herein, provide sensor data for different base calling operations for one base position per operation in the time series. Optionally, some of the N sensing cycles can be out of sequence as needed according to the particular neural network model being executed. The number N can be any number greater than 1. In some examples described herein, the sensing cycles of the N sensing cycles represent a set of sensing cycles for at least one sensing cycle preceding the subject sensing cycle and at least one sensing cycle following the subject sensing cycle in the time series. Examples described herein include an integer number N of 5 or greater.

[0070] The data flow logic 451 is configured to use an input unit for a given operation that includes tile data for an array of N spatially aligned patches to move the tile data and at least some trained parameters of the model from memory 460 to a configurable processor for operation of the neural network. The input unit can be moved by direct memory access operations in a single DMA operation, or in smaller units that move during available time slots in coordination with the execution of the deployed neural network.

[0071] The tile data of the sensing cycles described herein can include an array of sensor data having one or more features. For example, the sensor data can include two images analyzed to identify one of four bases at a base position in a genetic sequence of DNA, RNA, or other genetic material. The tile data can also include metadata about the images and sensors. For example, in a base calling embodiment, the tile data can include information about the alignment of the images with the clusters, such as distance from center information indicating the distance of each pixel in the array of sensor data from the center of the group of genetic material on the tile.

[0072] During execution of a multi-cycle neural network as described below, the tile data may also include data generated during execution of the multi-cycle neural network, called intermediate data, which may be reused rather than recomputed during execution of the multi-cycle neural network. For example, during execution of the multi-cycle neural network, the data flow logic may write the intermediate data to memory 460 in place of the sensor data for a given patch of the array of tile data. Such embodiments are described in more detail below.

[0073] As shown, a system for analyzing base calling sensor output is described that includes a memory (e.g., 460) accessible by a runtime program that stores tile data including sensor data for tiles from sensing cycles of a base calling operation. The system also includes a neural network processor, such as configurable processor 450, having access to the memory. The neural network processor is configured to perform neural network operations using trained parameters to generate classification data for the sensing cycles. As described herein, the neural network operations operate on an arrangement of N arrays of tile data from each sensing cycle of N sensing cycles comprising a subject cycle to generate classification data for the subject cycle. Data flow logic 451 is provided for moving the tile data and trained parameters from the memory to the neural network processor for execution of the neural network using input units including data for the N arrays of spatially aligned patches from each sensing cycle of the N sensing cycles.

[0074] Also described is a system in which the neural network processor has access to a memory and includes a plurality of execution clusters, and execution logic clusters in the plurality of execution clusters are configured to execute the neural network. The data flow logic accesses the memory and executes a cluster in the plurality of execution clusters to provide an input unit of tile data to an available execution cluster in the plurality of execution clusters, the input unit including an input unit including a number N of spatially aligned patches of the array of tile data from each sensing cycle, and causing the execution cluster to apply the N spatially aligned patches to the neural network to generate an output patch of classification data for the spatially aligned patches of the subject sensing cycle, where N is greater than 1.

[0075] Figure 5 is a simplified diagram illustrating aspects of base calling operations, including runtime program functionality executed by a host processor. In this diagram, image sensor output from a flow cell (such as those shown in Figures 1 and 2) is provided on line 500 to image processing thread 501, which can perform processes on the image, such as resampling, aligning, and arranging the array of sensor data for individual tiles, which can be used by a process to calculate a tile cluster mask for each tile in the flow cell, which can be used by a process to identify pixels in the array of sensor data that correspond to clusters of genetic material on the corresponding tile of the flow cell. To calculate the cluster mask, one exemplary algorithm is based on a process that uses a metric derived from the softmax output to detect unreliable clusters in early sequencing cycles; data from those wells / clusters is then discarded, and no output data is generated for those clusters. For example, the process can identify clusters with high confidence during the first N1 (e.g., 25) base calls and reject other clusters. Rejected clusters may be polyclonal or very weakly intense or unclear according to criteria. This procedure can be executed by the host CPU. Alternative implementations could potentially use this information to identify the necessary clusters that should be returned to the CPU, thereby limiting the storage required for intermediate data.

[0076] The output of image processing thread 501 is provided on line 502 to dispatch logic 510 in the CPU, which routes the array of tile data, depending on the status of the base calling operation, to data cache 504 over high-speed bus 503 or to multi-cluster neural network processor hardware 520, such as the configurable processor of FIG. 4, over high-speed bus 505. Hardware 520 returns classification data output by the neural network to dispatch logic 510, which passes the information to data cache 504 or over line 511 to thread 502, which can use the classification data to perform base calling and quality score calculations and arrange the data in a standard format for base called reads. The output of thread 502, which performs base calling and quality score calculations, is provided on line 512 to thread 503, which aggregates the base call reads, performs other operations such as data compression, and writes the resulting base call output to a specified destination for consumption by the customer.

[0077] In some embodiments, the host may include a thread (not shown) that performs final processing of the output of the hardware 520 supporting the neural network. For example, the hardware 520 may provide classification data output from the final layer of a multi-cluster neural network. The host processor may perform output activation functions, such as a softmax function, over the classification data to populate the data used by the base calling and quality scoring thread 502. The host processor may also perform input operations (not shown), such as resampling, batch normalization, or other adjustments to the tile data before inputting it to the hardware 520.

[0078] FIG. 6 is a simplified diagram of a configurable processor configuration, such as the configurable processor of FIG. 4 . In FIG. 6 , the configurable processor comprises an FPGA with multiple high-speed PCIe interfaces. The FPGA is configured with a wrapper 600 including the dataflow logic described with reference to FIG. 1 . The wrapper 600 manages interfacing and coordination with a runtime program in the CPU via a CPU communication link 609 and manages communication with an onboard DRAM 602 (e.g., memory 460) via a DRAM communication link 610. The dataflow logic in the wrapper 600 provides patch data obtained by traversing an array of tile data on the onboard DRAM 602 to a cluster 601 for a number N of cycles, and obtains process data 615 from the cluster 601 and delivers it to the onboard DRAM 602. The wrapper 600 also manages the transfer of data between the onboard DRAM 602 and host memory for both the input array of tile data and the output patch of classification data. The wrapper transfers the patch data to the assigned cluster 601 on line 613. The wrapper provides cluster 601 with trained parameters such as weights and biases obtained from on-board DRAM 602 on line 612. The wrapper provides cluster 601 with configuration and control data provided by or generated in response to a runtime program on the host on line 611 via CPU communication link 609. The cluster can also provide wrapper 600 with state signals on line 616 that are used in conjunction with control signals from the host to manage the traversal of the array of tile data to provide spatially aligned patch data and to run a multi-cycle neural network on the patch data using the resources of cluster 601.

[0079] As described above, multiple clusters may reside on a single configurable processor managed by wrapper 600 configured to run on corresponding ones of the multiple patches of tile data. Each cluster may be configured to provide classification data for base calls in a subject sensing cycle using the tile data of multiple sensing cycles described herein.

[0080] In an example system, model data, including kernel data such as filter weights and biases, can be sent from the host CPU to a configurable processor, so that the model can be updated as a function of cycle number. Base calling operations can typically involve hundreds of sensing cycles. In some embodiments, base calling operations can include paired-end reads. For example, model-trained parameters can be updated every 20 cycles (or other number of cycles) or according to an update pattern implemented in a particular system and neural network model. In some embodiments, where a sequence for a given string in a genetic cluster on a tile includes paired-end reads that include a first portion extending downward (or upward) from a first end of the string and a second portion extending upward (or downward) from a second end of the string, trained parameters can be updated at the transition from the first portion to the second portion.

[0081] In some embodiments, image data for multiple cycles of sensor data for a tile can be sent from the CPU to wrapper 600. Wrapper 600 can optionally perform some preprocessing and transformation of the sensor data and write the information to onboard DRAM 602. The input tile data for each sensing cycle can include an array of sensor data containing 4000 x 3000 pixels or more per tile per sensing cycle, with two features representing the colors of two images of the tile and including one or two bytes per pixel. In an embodiment where the number N is three sensing cycles used in each operation of the multi-cycle neural network, the array of tile data for each operation of the multi-cycle neural network can consume hundreds of megabytes per tile. In some embodiments of the system, the tile data also includes an array of DFC data stored once per tile, or other types of metadata about the sensor data and tile.

[0082] In operation, if a multi-cycle cluster is available, the wrapper assigns the patch to the cluster. The wrapper fetches the next patch of tile data for the cross section of the tile and sends it to the assigned cluster along with the appropriate control and configuration information. The cluster can be configured with enough memory on the configurable processor to have enough memory to hold the patch of data, including the patch, from multiple cycles in some systems being processed in place, and in various embodiments is processed using a ping-pong buffer technique or a raster scan technique.

[0083] When an assigned cluster completes its operation of the neural network for the current patch and generates an output patch, it signals the wrapper. The wrapper either reads the output patch from the assigned cluster, or the assigned cluster pushes data to the wrapper. The wrapper then assembles the output patch for the processed tile in DRAM 602. Once processing of the entire tile is complete and the output patch of data is transferred to DRAM, the wrapper sends the processed output array back to the host / CPU in a specific format. In some embodiments, the on-board DRAM 602 is managed by memory management logic within the wrapper 600. A runtime program can control the sequencing operations to complete analysis of all tile data arrays on every cycle, operating in a continuous flow to provide real-time analysis.

[0084] FIG. 7 is a diagram of a multi-cycle neural network model that can be implemented using the systems described herein. The example shown in FIG. 7 can be referred to as a 5-cycle input, 1-cycle output neural network. The input to the multi-cycle neural network model includes five spatially aligned patches (e.g., 700) from the tile data array for five sensing cycles of a given tile. The spatially aligned patches have the same aligned row and column dimensions (x, y) as other patches in the set, so that the information relates to the same cluster of genetic material on the tile in the sequence cycle. In this example, the subject patch is a patch from the array of tile data for cycle K. The set of five spatially aligned patches includes a patch from cycle K-2 that precedes the subject patch by two cycles, a patch from cycle K-1 that precedes the subject patch by one cycle, a patch from cycle K+1 that follows the patch from the subject cycle by one cycle, and a patch from cycle K+2 that follows the patch from the subject cycle by two cycles.

[0085] The model includes a separate stack 701 of neural network layers for each input patch. Thus, stack 701 receives patch tile data from cycle K+2 as input and is separate from stacks 702, 703, 704, and 705 so that they do not share input or intermediate data. In some embodiments, stacks 710-705 can all have the same model and the same trained parameters. In other embodiments, the models and trained parameters may be different in different stacks. Stack 702 receives patch tile data from cycle K+1 as input. Stack 703 receives patch tile data from cycle K as input. Stack 704 receives patch tile data from cycle K-1 as input. Stack 705 receives patch tile data from cycle K-2 as input. Each layer of the separate stacks performs a convolution operation of a kernel including multiple filters over the layer's input data. As in the example above, patch 700 may include three features: The output of layer 710 may include more features, such as 10-20 features. Similarly, the output of each of layers 711-716 may include any number of features suitable for a particular implementation. The filter parameters are the trained parameters of the neural network, such as weights and biases. The output feature sets (intermediate data) from each of stacks 701-705 are provided as input to an inverse layer 720 of temporal combination layers, in which the intermediate data from multiple cycles is combined. In the illustrated example, the inverse layer 720 includes a first layer including three combination layers 721, 722, and 723 that respectively receive intermediate data from three of the separated stacks, and a final layer including one combination layer 730 that receives intermediate data from the three temporal layers 721, 722, and 723.

[0086] The output of the final combination layer 730 is an output patch of classification data for the clusters located in the corresponding patch of the tile from cycle K. The output patches can be assembled into an output array of classification data for the tile in cycle K. In some embodiments, the output patches can have different sizes and dimensions than the input patches. In some embodiments, the output patches can include per-pixel data that can be filtered by the host to select cluster data.

[0087] The output classification data can then be applied to a softmax function 740 (or other output-driven function), optionally executed by the host or on a configurable processor, depending on the particular implementation. An output function different from softmax can be used (e.g., creating base call output parameters according to the maximum output, and then using a nonlinear mapping learned using the context / network output to give the base quality).

[0088] Finally, the output of the softmax function 740 is provided as the base call probability for cycle K (750) and may be stored in host memory for use in subsequent processing. Other systems may use different functions, e.g., different nonlinear models, for output probability calculation.

[0089] The neural network can be implemented using a configurable processor with multiple execution clusters to complete the evaluation of one tile cycle within or near the duration of the time interval of one sensing cycle, effectively outputting output data in real time. Dataflow logic can be configured to distribute input units of tile data and trained parameters to the execution clusters, and to distribute output patches for aggregation in memory.

[0090] A five-cycle input, one-cycle output neural network data input unit similar to that of FIG. 7 is described with reference to FIGS. 8A and 8B for a base calling operation using two-channel sensor data. For example, for a given base in a genetic sequence, the base calling operation can perform two streams of sample and two reactions, which generate two channels of signals, such as images, that can be processed to identify which one of four bases is located at the current position in the genetic sequence for each cluster of genetic material. In other systems, a different number of channels of sensor data can be utilized. For example, base calling can be performed using a one-channel method and system. The incorporated materials in U.S. Patent Application Publication No. 2013 / 0079232 discuss base calling using various numbers of channels, such as one, two, or four channels.

[0091] 8A shows an array of five cycles of tile data for a given tile, tile M, used to implement a five-cycle input, one-cycle output neural network. The five-cycle input tile data in this example is written to on-board DRAM or other memory in the system that can be accessed by the dataflow logic, and includes channel 1 array 801 and channel 2 array 811 for cycle K-2, channel 1 array 802 and channel 2 array 812 for cycle K-1, channel 1 array 803 and channel 2 array 813 for cycle K, channel 1 array 804 and channel 2 array 814 for cycle K+1, and channel 1 array 805 and channel 2 array 815 for cycle K+2. Also, array 820 of tile metadata can be written to memory once, with each cycle including the DFC file included for use as input to the neural network.

[0092] Although Figure 8A discusses a two-channel base calling operation, the use of two channels is merely an example, and base calling can be performed using any other suitable number of channels. For example, the incorporated materials in U.S. Patent Application Publication No. 2013 / 0079232 discuss base calling using various numbers of channels, such as one channel, two channels, or four channels, or another suitable number of channels.

[0093] The dataflow logic configures an input unit, which can be understood with reference to FIG. 8B, of tile data including spatially aligned patches of arrays of tile data for each execution cluster configured to perform a neural network execution on the input patches. The input unit of an assigned execution cluster is configured by the dataflow logic to read spatially aligned patches (e.g., 851, 852, 861, 862, 870) from each of arrays of tile data 801-805, 811, 815, 820 for five input cycles and deliver them via datapath (schematically 850) to memory on a configurable processor configured for use by the assigned execution cluster. The assigned execution cluster performs a five-cycle input / one-cycle output neural network execution and delivers subject cycle K output patches of classification data for the same patch of tiles for subject cycle K.

[0094] Figure 9 is a simplified representation of a neural network stack that can be used in a system like that of Figure 7 (e.g., 701 and 720). In this example, some functions of the neural network (e.g., 900, 902) run on the host, and other parts of the neural network (e.g., 901) run on a configurable processor.

[0095] In one example, the first function may be batch normalization (layer 910) formed on the CPU, however, in another example, batch normalization as a function may be fused into one or more layers, and there may not be a separate batch normalization layer.

[0096] Several spatially separated convolutional layers are implemented as the first set of convolutional layers of the neural network, as discussed above for the configurable processor. In this example, the first set of convolutional layers applies spatially 2D convolutions.

[0097] 9, for a number L / 2 of spatially separated neural network layers in each stack (where L was described with reference to FIG. 7), a first spatial convolution 921 is performed, followed by a second spatial convolution 922, followed by a third spatial convolution 923, etc. As shown in 923A, the number of spatial layers can be any practical number, which in context can range from a few to over 20 in different embodiments.

[0098] For SP_CONV_0, the kernel weights are stored in, for example, a (1, 6, 6, 3, L) structure because this layer has three input channels. In this example, the "6" in this structure comes from storing the coefficients in the transformed Winograd domain (the kernel size is 3x3 in the spatial domain, but expands in the transform domain).

[0099] For the other SP_CONV layers, the kernel weights are stored in a (1, 6, 6L) structure in this example since there are K(=L) inputs and outputs for each of these layers.

[0100] The output of the stack of spatial layers is provided to the temporal layers, including convolutional layers 924, 925, which run on an FPGA. Layers 924 and 925 may be convolutional layers that apply 1D convolutions over cycles. As shown in 924A, the number of temporal layers may be any practical number, which in context may range from a few to over 20 in different embodiments.

[0101] The first temporal layer, TEMP_CONV_0 layer 824, reduces the number of cycle channels from 5 to 3, as shown in Figure 7. The second temporal layer, layer 925, reduces the number of cycle channels from 3 to 1, as shown in Figure 7, reducing the number of feature maps to four outputs per pixel representing the confidence of each base call.

[0102] The outputs of the temporal layers are accumulated in output patches and delivered to the host CPU, where a softmax function 930, or other function, is applied to normalize the base call probabilities, for example.

[0103] Figure 10 shows an alternative implementation illustrating a 10-input, 6-output neural network that can be implemented for base calling operations. In this example, spatially aligned input patch tile data for cycles 0 through 9 are applied to separated stacks in the spatial layer, such as stack 1001 for cycle 9. The outputs of the separated stacks are applied to the inverse hierarchical arrangement of time stack 1020, and outputs 1035(2) through 1035(7) provide base call classification data for subject cycles 2 through 7.

[0104] Figure 11 shows one implementation of the dedicated architecture (e.g., Figure 7) of a neural network-based base caller used to separate the processing of data in different sequencing cycles. The motivation for using the dedicated architecture described above is first explained.

[0105] The neural network-based base caller processes data from the current sequencing cycle, one or more preceding sequencing cycles, and one or more subsequent sequencing cycles. Data from additional sequencing cycles provides sequence-specific context. The neural network-based base caller learns the sequence-specific context during training and base calls them. Additionally, data from pre- and post-sequencing cycles provide secondary contributions of pre-phasing and phasing signals to the current sequencing cycle.

[0106] Images captured in different sequencing cycles and in different image channels are misaligned and have residual registration errors with each other. To account for this misalignment, the specialized architecture includes a spatial convolution layer that does not mix information between sequencing cycles, but only mixes information within the same sequencing cycle.

[0107] The spatial convolutional layer uses so-called "decoupled convolutions" that operate on decoupling by processing the data for each of multiple sequencing cycles independently through a "dedicated, non-shared" array of convolutions. Decoupled convolutions convolve on the data and resulting feature maps only within a given sequencing cycle, i.e., the cycle, without convolving on the data and resulting feature maps of any other sequencing cycles.

[0108] For example, consider the input data as including (i) current data for the current (time t) sequencing cycle to be base-called, (ii) previous data for the previous (time t-1) sequencing cycle, and (iii) next data for the next (time t+1) sequencing cycle. The dedicated architecture then initiates three separate data processing pipelines (or convolution pipelines): a current data processing pipeline, a previous data processing pipeline, and a next data processing pipeline. The current data processing pipeline receives the current data for the current (time t) sequencing cycle as input and processes it independently through multiple spatial convolution layers to generate a so-called "current spatial convolutional representation" as the output of the final spatial convolutional layer. The previous data processing pipeline receives the previous data for the previous (time t-1) sequencing cycle as input and processes it independently through multiple spatial convolutional layers to generate a so-called "previous spatial convolutional representation" as the output of the final spatial convolutional layer. The next data processing pipeline receives the next data for the next (time t+1) sequencing cycle as input and processes it independently through multiple spatial convolutional layers to produce the so-called “next spatially convolved representation” as the output of the final spatial convolutional layer.

[0109] In some implementations, the current pipeline, one or more previous pipelines, and one or more next processing pipelines execute in parallel.

[0110] In some implementations, the spatial convolutional layer is part of a spatial convolutional network (or sub-network) within a dedicated architecture.

[0111] The neural network-based base caller further includes temporal convolutional layers that blend information between sequencing cycles, i.e., between cycles. The temporal convolutional layers receive their inputs from the spatial convolutional network and operate on the spatial convolutional representations produced by the final spatial convolutional layer for each data processing pipeline.

[0112] The inter-cycle operational freedom of the temporal convolutional layers arises from the fact that misalignment features present in the image data supplied as input to the spatial convolutional network are purged from the spatial convolutional representation by the stack or cascade of separated convolutions performed by the array of spatial convolutional layers.

[0113] The temporal convolutional layer uses so-called "combinatorial convolution," which convolves group-wise on input channels with subsequent inputs on a sliding window basis. In one implementation, the subsequent inputs are subsequent outputs generated by previous spatial or temporal convolutional layers.

[0114] In some implementations, the temporal convolutional layer is part of a temporal convolutional network (or sub-network) in a dedicated architecture. The temporal convolutional network receives its input from a spatial convolutional network. In one implementation, the first temporal convolutional layer of the temporal convolutional network combines the spatial convolutional representations between sequencing cycles by group. In another implementation, subsequent temporal convolutional layers of the temporal convolutional network combine successive outputs of previous temporal convolutional layers.

[0115] The output of the final temporal convolutional layer is fed into an output layer, which produces outputs that are used to base call one or more clusters in one or more sequencing cycles.

[0116] During forward propagation, the specialized architecture processes information from multiple inputs in two stages. In the first stage, decoupled convolutions are used to prevent mixing of information between inputs. In the second stage, combined convolutions are used to mix information between inputs. The results from the second stage are used to make a single inference on the multiple inputs.

[0117] This differs from batch-mode techniques, in which a convolutional layer processes multiple inputs in a batch simultaneously and makes a corresponding inference for each input in the batch. In contrast, dedicated architectures map multiple inputs to a single inference. A single inference may include two or more predictions, such as a classification score for each of the four bases (A, C, T, and G).

[0118] In one implementation, the inputs have a temporal ordering such that each input occurs at a different time step and has multiple input channels. For example, the multiple inputs may include three inputs: a current input generated by a current sequencing cycle at time step (t), a previous input generated by a previous sequencing cycle at time step (t-1), and a next input generated by a next sequencing cycle at time step (t+1). In another implementation, each input is derived from the current, previous, and next inputs by one or more previous convolutional layers, respectively, and includes k feature maps.

[0119] In one embodiment, each input may include five input channels: a red image channel (red), a red distance channel (yellow), a green image channel (green), a green distance channel (purple), and a scaling channel (blue). In another implementation, each input may include k feature maps generated by a previous convolutional layer, with each feature map treated as an input channel. In yet another example, each input may have just one channel, two channels, or another different number of channels. The incorporated materials in U.S. Patent Application Publication No. 2013 / 0079232 discuss base calling using various numbers of channels, such as one channel, two channels, or four channels.

[0120] FIG. 12 shows an implementation of separated layers, each of which may include a convolution. Separated convolution processes multiple inputs at once by applying a convolution filter to each input in parallel. In separated convolution, a convolution filter combines input channels within the same input and does not combine input channels within different inputs. In one implementation, the same convolution filter is applied to each input in parallel. In another implementation, a different convolution filter is applied to each input in parallel. In some implementations, each spatial convolution layer includes a bank of k convolution filters, each of which is applied to each input in parallel.

[0121] FIG. 13A shows one implementation of a combining layer, each of which may include a convolution. FIG. 13B shows another implementation of a combining layer, each of which may include a convolution. A combining convolution mixes information between different inputs by grouping corresponding input channels of the different inputs and applying a convolution filter to each group. The grouping of corresponding input channels and the application of the convolution filter occur on a sliding window basis. In this context, a window spans two or more consecutive input channels, for example, representing the output for two consecutive sequencing cycles. Because the window is a sliding window, most input channels are used in two or more windows.

[0122] In some implementations, the distinct inputs originate from an output array generated by a preceding spatial or temporal convolutional layer. In the output array, the distinct inputs are arranged as successive outputs and are therefore viewed as successive inputs by the next temporal convolutional layer. Then, in the next temporal convolutional layer, a combinatorial convolution applies a convolutional filter to groups of corresponding input channels in the successive inputs.

[0123] In one implementation, the successive inputs have a temporal ordering such that the current input is generated by the current sequencing cycle at time step (t), the previous input is generated by the previous sequencing cycle at time step (t-1), and the next input is generated by the next sequencing cycle at time step (t+1). In another implementation, each successive input is derived from the current, previous, and next inputs by one or more previous convolutional layers, respectively, and includes k feature maps.

[0124] In one embodiment, each input may include five input channels: a red image channel (red), a red distance channel (yellow), a green image channel (green), a green distance channel (purple), and a scaling channel (blue). In another implementation, each input may include k feature maps generated by a previous convolutional layer, with each feature map treated as an input channel.

[0125] The depth B of the convolution filter depends on the number of consecutive inputs whose corresponding input channels are convolved with the convolution filter on a sliding window basis for each group. In other words, the depth B is equal to the number of consecutive inputs in each sliding window and group size.

[0126] In Figure 13A, corresponding input channels from two consecutive inputs are combined within each sliding window, so B = 2. In Figure 13B, corresponding input channels from three consecutive inputs are combined within each sliding window, so B = 3.

[0127] In one implementation, the sliding windows share the same convolutional filter. In another implementation, a different convolutional filter is used for each sliding window. In some implementations, each temporal convolutional layer includes a bank of k convolutional filters, each of which is applied to successive inputs on a sliding window basis.

[0128] Further details of Figures 4-10 and variations thereof can be found in co-pending U.S. Non-Provisional Patent Application No. 17 / 176,147, filed February 15, 2021, entitled "HARDWARE EXECUTION AND ACCELERATION OF ARTIFICIAL INTELLIGENCE-BASED BASE CALLER" (Attorney Docket No. ILLM1020-2 / IP-1866-US), which is incorporated by reference as if fully set forth herein.

[0129] Training a base cola from scratch The base calling system is trained to predict base calls for unknown analytes that include a base sequence. For example, the base calling system may have a base caller that includes a neural network that predicts base calls for bases in the unknown analyte.

[0130] Training the neural network of a base calling system can be difficult. This is especially true when there is no labeled training data available to train the base calling system. In some embodiments, a real-time analysis (RTA) system can be used to generate labeled training data, which can be used to train the base calling system. An example of an RTA system is discussed in U.S. Pat. No. US 10,304,189 (B2), entitled "Data Processing System and Methods," issued May 28, 2019, which is incorporated by reference as if fully set forth herein. However, if the system lacks RTA or is unable to fully utilize the functionality of RTA, it can be difficult to initially generate labeled training data for training the neural network of the base calling system.

[0131] This disclosure discusses a self-learning base chore that generates initial labeled training data, trains itself using the labeled training data, generates further labeled training data using at least a partially trained base chore, trains itself using the further labeled training data, generates further, even more labeled training data, and iteratively repeats this process to fully train the base chore. This iterative training and labeled training data generation process involves different stages, such as a single-oligo stage, a multiple-oligo stage (two-oligo stage, three-oligo stage, etc.), followed by a single-organism stage, a complex-organism stage, and a further complex-organism stage. Thus, the complexity and / or length of the samples used for training and generation of labeled training data increases progressively and monotonically with each iteration, along with the complexity of the neural network configuration underlying the base chore, as discussed in more detail herein. Because the base chore is progressively self-trained, such a system obviates the need for an RTA to generate labeled training data. Thus, while the basecalling systems described herein may include RTAs, the iterative training processes discussed herein can be used in addition to or instead of RTAs to train the basecallers.

[0132] FIG. 14A shows a base-calling system 1400 operating in a single oligo training stage to train a base-caller 1414 containing a neural network (NN) configuration 1415 using a known synthetic sequence 1406.

[0133] In the example of Figure 14A, base calling system 1400 includes a sequencing machine 1404, such as sequencing machine 400 of Figure 4. In an embodiment, sequencing machine 1404 includes a biosensor (not illustrated in Figure 14A) that includes a flow cell 1405 similar to flow cell 102 of biosensor 100 of Figure 1.

[0134] As discussed with respect to Figures 2, 3, and 6, flow cell 1405 comprises a plurality of clusters 1407a, ..., 1407G. Specifically, in an embodiment, flow cell 1405 comprises a plurality of lanes of tiles, each tile including a corresponding plurality of clusters, as discussed with respect to Figure 2. In Figure 14A, flow cell 1405 is illustrated as including several such exemplary clusters 1407a, ..., 1407G. The base calling process predicts a base call (A, C, G, T) for each cluster in a particular cycle.

[0135] A typical flow cell 1405 can include multiple clusters 1407, such as thousands or millions of clusters. Merely by way of example, without limiting the scope of the present disclosure, and to illustrate some of the principles of the present disclosure, we will assume that there are 10,000 (or 10k) clusters 1407 in the flow cell 1405 (i.e., G=10,000), although a practical flow cell will likely have a much larger number of such clusters.

[0136] In an embodiment, known synthetic sequence 1406 is used as a sample for base-calling operations during the single oligo training phase. In an embodiment, known synthetic sequence 1406 comprises a synthetically generated oligomer. Oligonucleotides, short DNA or RNA molecules called oligomers or simply oligos, have wide applications in genetic testing, research, and forensics. These small amounts of nucleic acids, commonly produced in laboratories by solid-phase chemical synthesis, can be manufactured as single-stranded molecules with any user-specified sequence and are therefore extremely important for artificial gene synthesis, polymerase chain reaction (PCR), DNA sequencing, molecular cloning, and as molecular probes. The length of an oligonucleotide is typically referred to as a "mer." For example, a 6-nucleotide (nt) oligonucleotide is a hexamer, while a 25-nt oligonucleotide is typically referred to as a "25-mer." In an embodiment, the size of the oligomer or oligo comprising known synthetic sequence 1406 can have any suitable number of bases, such as 8, 10, 12, or more, and is implementation-specific. By way of example only, Figure 14A illustrates an oligo of known synthetic sequence 1406 containing 8 bases.

[0137] The oligo referred to in Figure 14A is labeled as oligo #1 (or oligo number 1). Because only one unique oligo is used in Figure 14A, the same oligo #1 is populated into each cluster 1407. Therefore, all 10k clusters 1407 are populated with the same oligo sequence; that is, copies of the same oligo are populated into all clusters 1407.

[0138] The sequencing machine 1404 generates sequence signals 1412a, ..., 1412G for corresponding clusters of the plurality of clusters 1407a, ..., 1407G. For example, for cluster 1407a, the sequencing machine 1404 generates a corresponding sequence signal 1412a indicating the base sequence populated into cluster 1407a for a series of sequencing cycles. Similarly, for cluster 1407b, the sequencing machine 1404 generates a corresponding sequence signal 1412b indicating the base sequence populated into cluster 1407b for a series of sequencing cycles, and so on. Base caller 1414 receives sequence signals 1412 and is responsible for calling (e.g., predicting) the corresponding bases. In an embodiment, base calling 1414, including NN configuration 1415 (and various other NN configurations described later herein), can be stored in memory 404, 403, and / or 406 and can execute on a host CPU (such as CPU 402 in FIG. 4) and / or configurable processor (such as configurable processor 450 in FIG. 4) that is local to sequencing machine 400. In another embodiment, base calling 1414 can be stored remotely from sequencing machine 400 (e.g., stored in the cloud) and executed by a remote processor (e.g., executed in the cloud). For example, in a remote version of base calling 1414, base calling 1414 receives sequence signals 1412 (e.g., over a network such as the Internet), performs base calling operations, and transmits the base calling results to sequencing machine 400 (e.g., over a network such as the Internet).

[0139] In examples, sequence signal 1412 includes an image captured by a sensor (e.g., a photodetector, a photodiode), as described previously herein. Accordingly, at least some of the examples and embodiments discussed herein relate to iteratively training a base caller (such as base caller 1414) to process sequence signals that include images. However, the principles of the present disclosure are not limited to training any particular type of base caller to receive any particular type of sequence signal. For example, the iterative training discussed herein in this disclosure is independent of the type of base caller trained or the type of sequence signal used. For example, the iterative training discussed herein in this disclosure can be used to train any other suitable type of base caller, such as a base caller configured to call bases based on sequence signals that do not include images. For example, sequence signals can include electrical signals (e.g., voltage signals, current signals), pH levels, and / or the like, and the iterative training methods discussed herein can be applied to training a base caller to receive any such type of sequence signal.

[0140] Neural network configuration 1415, as discussed in more detail herein, is a convolutional neural network (examples of which are illustrated in FIGS. 7, 9, 10, 11, and 12) that uses a relatively small number of layers and a relatively small number of parameters (e.g., compared to some other neural network configurations described later in this specification, such as neural network configuration 1615 of FIG. 16A).

[0141] An initially untrained base caller 1414 including a neural network configuration 1415 predicts base call sequences 1418a, ..., 1418G for corresponding clusters of a plurality of clusters 1407a, ..., 1407G based on corresponding sequence signals 1412a, ..., 1412G, respectively. For example, for cluster 1407a, base caller 1414 predicts corresponding base call sequence 1418a, comprising the base call for cluster 1407a for a set of sequencing cycles, based on corresponding sequence signal 1412a. Similarly, for cluster 1407b, base caller 1414 predicts corresponding base call sequence 1418b, comprising the base call for cluster 1407b for a set of sequencing cycles, based on corresponding sequence signal 1412b, and so on. Thus, G base call sequences 1418a, ..., 1418G are predicted by base caller 1414.

[0142] Assume that oligo#1 has eight bases generally labeled GA1, ..., GA8. Simply by way of example, and without limiting the scope of the present disclosure, assume that the eight bases of oligo#1 are A, C, T, T, G, C, A, C. Initially, base caller 1414 is untrained, and therefore errors in base calling are likely. For example, predicted base call sequence 1418a (generally labeled Sal, ..., Sa8) is C, A, T, C, G, C, A, G, as illustrated in FIG. 14A. Therefore, when comparing ground truth base sequence 1406 of oligo#1 (i.e., A, C, T, T, G, C, A, C) with predicted base sequence 1418a (i.e., C, A, T, C, G, C, A, G), there are errors in base calling for base numbers 1, 2, 4, and 8. Thus, in FIG. 14A , the ground truth base sequence 1406 and the predicted base sequence 1418a of oligo #1 are compared in operation 1413a, and the error between these two base sequences is used in a backward pass of the neural network configuration 1415 of the base call 1414 to update the gradients and weights of the neural network configuration 1415, etc., to train the neural network configuration 1415 (symbolically labeled as gradient update 1417 in FIG. 14A ).

[0143] FIG. 14A1 illustrates in more detail the comparison operation between the predicted base sequence 1418a and the ground truth base sequence 1406 of oligo #1. For example, referring to FIG. 14A and FIG. 14A1, the predicted base sequence 1418a is C, A, T, C, G, C, A, G, and the ground truth base sequence 1406 of oligo #1 is A, C, T, T, G, C, A, C. Therefore, when comparing the ground truth base sequence 1406 of oligo #1 (i.e., A, C, T, T, G, C, A, C) with the predicted base sequence 1418a (i.e., C, A, T, C, G, C, A, G), there are errors in the base calls for base numbers 1, 2, 4, and 8. For example, in FIG. 14A1, the error in the base call for base number 1 is given by "C should be A," i.e., the base call C should be the base call A. Similarly, the error in the base call for base number 2 is given by "A should be C," i.e., base call A should be base call B, etc. There are no errors in the base calls for base numbers 3, 5, 6, and 7 (illustrated as "Match (no error)" in FIG. 14A1). Thus, in FIG. 14A1, during comparison, each base call in predicted base call sequence 1418a is compared with the corresponding base call in the corresponding ground truth sequence (e.g., base sequence 1406 of oligo #1) to generate corresponding comparison results, as illustrated in FIG. 14A1.

[0144] Referring again to FIG. 14A , base calling system 1400 also includes mapping logic 1416, the functionality of which is described later in this specification. In an embodiment, mapping logic 1416 can be stored in memory 404, 403, and / or 406, and mapping logic 1416 can execute on a host CPU (such as CPU 402 in FIG. 4 ) and / or configurable processor (such as configurable processor 450 in FIG. 4 ) that is local to sequencing machine 400. In another embodiment, mapping logic 1416 can be stored remotely from sequencing machine 400 (e.g., stored in the cloud) and executed by a remote processor (e.g., executed in the cloud). For example, in a remote version of mapping logic 1416, the mapping logic receives data to be mapped from sequencing machine 400 (e.g., over a network such as the Internet), performs a mapping operation, and transmits the mapping results to sequencing machine 400 (e.g., over a network such as the Internet). Mapping operations are discussed in more detail later in this specification.

[0145] Figure 14A and various other figures, examples, and embodiments of the present disclosure refer to base callers that predict base call sequences. Various examples of such prediction of base call sequences are discussed herein. Further examples of base call prediction can be found in co-pending U.S. Provisional Patent Application No. 63 / 217,644 (Attorney Docket No. ILLM1046-1 / IP-2135-PRV), entitled "IMPROVED ARTIFICIAL INTELLIGENCE-BASED BASE CALLING OF INDEX SEQUENCES," filed July 1, 2021, which is incorporated by reference as if fully set forth herein.

[0146] Figure 14B illustrates further details of the base calling system 1400 of Figure 14A operating in a single oligo training phase to train a base caller 1414 comprising a neural network configuration 1415 using a known synthetic sequence 1406. For example, Figure 14B illustrates the use of predicted base call sequences 1418a, ..., 1418G to train the base caller 1414. For example, each of the predicted base call sequences 1418a, ..., 1418G is compared to the ground truth base sequence 1406 for oligo #1 (see comparison operations 1413a, ..., 1413G), and the resulting error is used for gradient updates and resulting parameter (e.g., weights and bias) updates (symbolically labeled in Figure 14A as gradient updates 1417) by the backpropagation section of the neural network configuration 1415.

[0147] Thus, neural network configuration 1415 has been trained using base call sequences 1418 predicted by neural network configuration 1415 and using the ground truth base sequence of oligo #1 1406. Because the training discussed with respect to Figures 14A and 14B uses a single oligo, this training phase is also referred to as the "single oligo training phase," and Figures 14A and 14B are labeled accordingly.

[0148] In an embodiment, the process of Figures 14A and 14B can be repeated iteratively. For example, in the first iteration of Figure 14A, NN configuration 1415 is at least partially trained. The at least partially trained NN configuration 1415 is again used during the second iteration to regenerate predicted base call sequences from sequence signal 1412 (e.g., as discussed with respect to Figure 14A), and the resulting predicted base call sequences are again compared to ground truth 1406 (i.e., oligo #1) to generate an error signal, which is used to further train NN configuration 1415. This process can be repeated iteratively multiple times until NN configuration 1415 is sufficiently trained. In an embodiment, this process can be repeated iteratively a certain number of times. In another embodiment, this process can be repeated iteratively until some error saturates (e.g., the error in successive iterations does not decrease significantly).

[0149] FIG. 15A illustrates the base calling system 1400 of FIG. 14A operating in the training data generation phase of a two-oligo training stage to generate labeled training data using two known synthetic sequences 1501A and 1501B.

[0150] The base calling system 1400 of Figure 15A is the same as the base calling system of Figure 14A, in that in both figures, the base calling system 1400 uses a neural network configuration 1415. Furthermore, two different, unique oligo sequences 1501A and 1501B are loaded onto various clusters of the flow cell 1405. By way of example only, and without limiting the scope of the present disclosure, assume that of the 10,000 clusters 1407, approximately 5,200 clusters are populated with oligo sequence 1501A and the remaining 4,800 clusters are populated with oligo sequence 1501B (although in other examples, the two oligos can be divided substantially equally among the 10,000 clusters).

[0151] The sequencing machine 1404 generates sequence signals 1512a, ..., 1512G for corresponding clusters of the plurality of clusters 1407a, ..., 1407G. For example, for cluster 1407a, the sequencing machine 1404 generates a corresponding sequence signal 1512a that indicates the bases of cluster 1407a for a series of sequencing cycles. Similarly, for cluster 1407b, the sequencing machine 1404 generates a corresponding sequence signal 1512b that indicates the bases for cluster 1407b for a series of sequencing cycles, and so on.

[0152] A base caller 1414, comprising at least a partially trained neural network configuration 1415 (e.g., trained by iteratively repeating the operations of Figures 14A and 14B), predicts base call sequences 1518a, ..., 1518G for corresponding ones of a plurality of clusters 1407a, ..., 1407G, based on corresponding sequence signals 1512a, ..., 1512G, respectively. For example, for cluster 1407a, base caller 1414 predicts corresponding base call sequence 1518a, comprising a base call for cluster 1407a for a series of sequencing cycles, based on corresponding sequence signal 1512a. Similarly, for cluster 1407b, base caller 1414 predicts corresponding base call sequence 1518b, comprising a base call for cluster 1407b for a series of sequencing cycles, based on corresponding sequence signal 1512b, and so on. Thus, G base call sequences 1518a, ..., 1518G are predicted by base caller 1414. Note that the neural network configuration 1415 in Figure 15A was trained earlier during the iterations of the single oligo training phase discussed with respect to Figures 14A and 14B. Thus, the predicted base call sequences 1518a, ..., 1518G are reasonably accurate, but not very accurate (because the base call 1414 is not fully trained).

[0153] In an embodiment, oligo sequences 1501A and 1501B are selected to have a sufficient edit distance between the bases of the two oligos. Figures 15B and 15C illustrate two corresponding exemplary selections of oligo sequences 1501A and 1501B of Figure 15A. For example, in Figure 15B, oligo 1501A is selected to have bases A, C, T, T, G, C, A, C, while oligo 1501B is selected to have bases C, C, T, A, G, C, A, C. Thus, the first and fourth bases in the two oligos 1510A and 1510B are different, resulting in an edit distance of 2 between the two oligos 1510A and 1510B.

[0154] In contrast, in Figure 15B, oligo 1501A is selected to have bases A, C, T, T, G, C, A, C, while oligo 1501B is selected to have bases C, A, T, G, A, T, A, G. Thus, in the example of Figure 15B, the first, second, fourth, fifth, sixth, and eighth bases in the two oligos 1510A and 1510B are different, resulting in an edit distance of 6 between the two oligos 1510A and 1510B.

[0155] In an embodiment, the two oligos 1501A and 1501B are selected such that the two oligos are separated by at least a threshold edit distance. By way of example only, the threshold edit distance could be 4 bases, 5 bases, 6 bases, 7 bases, or even 8 bases. Thus, the two oligos 1501A and 1501B are selected such that the two oligos are sufficiently different from each other.

[0156] Referring again to FIG. 15A , base caller 1414 does not know which oligo sequences are populated into which clusters. Therefore, base caller 1414 does not know the mapping between known oligo sequences 1501A, 1501B and the various clusters. In an embodiment, mapping logic 1416 receives predicted base call sequences 1518 and maps each predicted base call sequence 1518 to either oligo 1501A or oligo 1501B, or declares uncertainty when mapping a predicted base call sequence to either of the two oligos. FIG. 15D illustrates exemplary mapping operations for (i) mapping a predicted base call sequence to either oligo 1501A or oligo 1501B, or (ii) declaring uncertainty when mapping a predicted base call sequence to either of the two oligos.

[0157] In an example, the greater the edit distance between two oligos, the easier (or more accurate) it is to map an individual prediction to one of the two oligos. For example, referring to FIG. 15B, because the edit distance between the two oligos 1501A and 1501B is only 2, the two oligos are largely similar, and it may be relatively difficult to map a base call prediction to one of the two oligos. However, because the edit distance between the two oligos 1501A and 1501B in FIG. 15C is 6, the two oligos are very dissimilar, and it may be relatively easy to map a prediction to one of the two oligos. Thus, FIG. 15B, with an edit distance of 2, is labeled "less suitable for training," and FIG. 15C, with an edit distance of 6, is labeled "more suitable for training." Thus, in an example, oligos 1501A and 1501B according to FIG. 15C (but not according to FIG. 15B) are generated and used for training, as discussed in more detail in turn herein.

[0158] Referring again to Figure 15D, exemplary predicted base call sequences 1518a, 1518b, and 1518G are illustrated. Also illustrated are exemplary bases for two oligos 1501A and 1501B (the exemplary bases for the two oligos correspond to the bases illustrated in Figure 15C).

[0159] Because the neural network configuration 1415 is somewhat trained but not fully trained, the neural network configuration 1415 may be able to make base call predictions, but such base call predictions are prone to error.

[0160] Predicted base call sequence 1518a contains C, A, G, G, C, T, A, C. It is compared to the base call sequence of oligo 1501A, A, C, T, T, G, C, A, C, and to the base call sequence of oligo 1501B, C, A, T, G, A, T, A, G. Predicted base call sequence 1518a has its seventh and eighth bases matching the corresponding seventh and eighth bases of oligo 1501A, and its first, second, fourth, sixth, and seventh bases matching the corresponding bases of oligo 1501B. Thus, as illustrated in FIG. 15D , predicted base call sequence 1518a shares two bases of similarity with oligo 1501A, and predicted base call sequence 1518a shares five bases of similarity with oligo 1501B.

[0161] If predicted base call sequence 1518a is indeed for oligo 1501B (e.g., such that predicted base call sequence 1518a has five bases of similarity with oligo 1501B), this means that neural network configuration 1415 was able to correctly predict five bases of the eight-base sequence (i.e., correctly predicting the first, second, fourth, sixth, and seventh bases that match the corresponding bases in oligo 1501B). However, because neural network configuration 1415 was not fully trained, neural network configuration 1415 made errors in predicting the remaining three bases (i.e., the third, fifth, and eighth bases).

[0162] The mapping logic 1416 can map the predicted base call array to the corresponding oligo using appropriate logic. For example, assume that the predicted base call array has a similarity in the number of SAs with oligo 1501A and a similarity in the number of SBs with oligo 1501B. In an example, the mapping logic 1416 maps the predicted base call array to oligo 1501A when SA > ST and SB < ST, where ST is a threshold number. That is, the mapping logic 1416 maps the predicted base call array to oligo 1501A when the similarity level with oligo 1501A is higher than the threshold and the similarity level with oligo 1501B is lower than the threshold.

[0163] Similarly, in another example, the mapping logic 1416 maps the predicted base call array to oligo 1501B when SB > ST and SA < ST.

[0164] In yet another example, the mapping logic 1416 declares that the predicted base call array is uncertain when both SA and SB are less than the threshold ST or when both SA and SB are greater than the threshold ST.

[0165] The above discussion can be written in the form of equations as follows. For the predicted base call array, If SA > ST and SB < ST, map to oligo 1501A. (Equation 1) If SB > ST and SA < ST, map to oligo 1501B. (Equation 2) If both SA and SB are < ST, declare an uncertain mapping. Or (Equation 3) If both SA and SB are > ST, declare an uncertain mapping. (Equation 4)

[0166] The threshold ST depends on the number of bases in the oligo (which is 8 in the exemplary use case illustrated in the figure), the desired accuracy, and / or is implementation specific. By way of example only, the threshold ST is assumed to be 4 in the exemplary use case illustrated in FIG. 15D. Note that a threshold ST of 4 is merely an example, and the selection of the threshold ST may be implementation specific. By way of example only, during the first iteration of training, the threshold ST may have a relatively low value (e.g., 4). The threshold ST may have a relatively high value (e.g., 6 or 7) during later iterations of training (training iterations are described later in this specification). Thus, as the NN configuration becomes better trained during later training iterations, the threshold ST may be gradually increased. However, in another example, the threshold ST may have the same value throughout all iterations of training. While the threshold ST is selected as 4 in the example of FIG. 15D, in other exemplary implementations, the threshold ST may be, for example, 5, 6, or 7. In an example, the threshold ST may also be expressed as a percentage. For example, if the threshold ST is 4 and the total number of bases is 8, the threshold ST can be expressed as (4 / 8) × 100, i.e., 50%. The threshold ST can be a parameter that can be selected by the user, and in this embodiment, can be selected to be between 50% and 95%.

[0167] Referring again to Figure 15D, as noted above, predicted base call sequence 1518a has two bases of similarity with oligo 1501A, and predicted base call sequence 1518a has five bases of similarity with oligo 1501B. Therefore, SA = 2 and SB = 5. According to Equation 2, assuming a threshold ST of 4, predicted base call sequence 1518a maps to oligo 1501B.

[0168] Referring now to predicted base call sequence 1518b, predicted base call sequence 1518b has two bases of similarity with oligo 1501A, and predicted base call sequence 1518b has three bases of similarity with oligo 1501B. Therefore, SA=2 and SB=3. According to Equation 3, assuming a threshold ST of 4, predicted base call sequence 1518b is declared uncertain for mapping to any of the oligo sequences.

[0169] Referring now to predicted base call sequence 1518G, predicted base call sequence 1518G shares 6 bases of similarity with oligo 1501A, and predicted base call sequence 1518G shares 3 bases of similarity with oligo 1501B. Therefore, SA = 6 and SB = 3. According to Equation 2, assuming a threshold ST of 4, predicted base call sequence 1518G maps to oligo 1501A.

[0170] FIG. 15E illustrates labeled training data 1550 generated from the mapping of FIG. 15D, where the labeled training data 1550 is used by another neural network configuration 1615 (e.g., as illustrated in FIG. 16A, where the other neural network configuration 1615 is different from and more complex than the neural network configuration 1415 of FIGS. 14A, 14B, and 15A).

[0171] As illustrated in Figure 15E, some of the predicted base call sequences 1518 and corresponding sequence signals map to the base sequence of oligo 1501A (i.e., ground truth 1506a), some other predicted base call sequences 1518 and corresponding sequence signals map to the base sequence of oligo 1501B (i.e., ground truth 1506b), and the mapping of the remainder of the predicted base call sequences 1518 and corresponding sequence signals is uncertain.

[0172] For example, predicted base call sequences 1518c, 1518d, 1518G and corresponding sequence signals 1512c, 1512d, 1512G are mapped to the base sequence of oligo 1501A (i.e., ground truth 1506a), predicted base call sequences 1518a, 1518f and corresponding sequence signals 1512a, 1512f are mapped to the base sequence of oligo 1501B (i.e., ground truth 1506b), and the remaining mappings of predicted base call sequences 1518b, 1518e, 1518g and corresponding sequence signals 1512b, 1512e, 1512g are uncertain.

[0173] As an example only, assume that 2,600 base call sequences of training data 1550 are mapped to oligo 1501A and 3,000 base call sequences of training data 1550 are mapped to oligo 1501B. As illustrated in Figure 15E, the remaining 4,400 base call sequences are uncertain and do not map to either of the two oligos.

[0174] Note that Figures 15A, 15D, and 15E are referred to as the "training data generation phase" of the "two oligo training stage" because the labeling training data 1550 is generated using sequences from two oligos and using neural network configuration 1415.

[0175] FIG. 16A illustrates the base-calling system 1400 of FIG. 14A operating in the "training data consumption and training phase" of the "two-oligo training stage" to train a base-caller 1414 with another neural network configuration 1615 (different from and more complex than the neural network configuration 1415 of FIG. 14A) using two known synthetic sequences 1501A and 1501B.

[0176] The basecalling system 1400 of Figure 16A is the same as the basecalling system of Figure 14A. However, unlike Figure 14A (where neural network configuration 1415 was used in basecaller 1414), basecaller 1414 of Figure 16A uses a different neural network configuration 1615. Neural network configuration 1615 of Figure 16A differs from neural network configuration 1415 of Figure 14A. For example, neural network configuration 1615 is a convolutional neural network (examples of which are illustrated in Figures 7, 9, 10, 11, and 12) that uses a greater number of layers and parameters (e.g., weights and biases) than neural network configuration 1415. In another example, neural network configuration 1615 is a convolutional neural network that uses a greater number of convolutional filters than neural network configuration 1415. The configuration, topology, and number of layers and / or filters of the two neural network configurations 1415 and 1615 may differ in some examples.

[0177] In the "Training Data Consumption and Training Phase" of the "Two Oligo Training Stage" illustrated in Figure 16A, base caller 1414 with neural network configuration 1615 receives sequence signal 1512 previously generated during the "Training Data Generation Phase" of Figure 15A. That is, base caller 1414 with neural network configuration 1615 reuses previously generated sequence signal 1512. Therefore, because previously generated sequence signal 1512 is reused in the "Training Data Consumption and Training Phase" of the "Two Oligo Training Stage" illustrated in Figure 16A, sequencing machine 1404 and components therein play no role and are therefore illustrated using dotted lines. Similarly, mapping logic 1416 plays no role (because no mapping is performed in Figure 16A), and therefore mapping logic 1416 is also illustrated using dotted lines.

[0178] 16A, base caller 1414, comprising neural network configuration 1615, receives previously generated sequence signal 1512 and predicts base call sequence 1618 from sequence signal 1512. Predicted base call sequence 1618 includes predicted base call sequences 1618a, 1618b, ..., 1618G. For example, sequence signal 1512a is used to predict base call sequence 1618a, sequence signal 1512b is used to predict base call sequence 1618b, sequence signal 1512G is used to predict base call sequence 1618G, etc.

[0179] The neural network configuration 1615 has not yet been trained, and therefore the predicted base call sequences 1618a, 1618b, ..., 1618G may have many errors. The mapped training data 1550 of Figure 15E is now used to train the neural network configuration 1615. For example, from the training data 1550, the base call 1414 is (i) sequence signals 1512c, 1512d, 1512G are for the base sequence of oligo 1501A (i.e., ground truth 1506a); (ii) sequence signals 1512a, 1512f are for the base sequence of oligo 1501B (i.e., ground truth 1506b); and (iii) Know that the mapping of the constellation signals 1512b, 1512e, 1512g is uncertain.

[0180] Thus, sequence signals 1512 and predicted base call sequences 1518 are divided into two categories: (i) a first category including sequence signals 1512c, 1512d, 1512G (and corresponding predicted base call sequences 1518c, 1518d, 1518G) that can be mapped to the base sequence of oligo 1501A (i.e., ground truth 1506a); (iii) a third category containing sequence signals 1512b, 1512e, 1512g (and corresponding predicted base call sequences 1518b, 1518e, 1518g) that cannot be mapped to either the base sequence of oligos 1501A or 1501B.

[0181] Therefore, based on (iii) above, predicted base call sequences 1618b, 1618e, and 1618g (e.g., corresponding to sequence signals 1512b, 1512e, 1512g) are not used to train neural network configuration 1615. Accordingly, predicted base call sequences 1618b, 1618e, and 1618g are discarded during the training iterations and are not used for gradient update (symbolically illustrated in FIG. 16A using an "X" or "cross symbol" between predicted base call sequences 1618b, 1618e, and 1618g and gradient update box 1617).

[0182] Based on (i) above, base caller 1414 knows that predicted base call sequences 1618c, 1618d, and 1618G (e.g., corresponding to sequence signals 1512c, 1512d, and 1512G) are likely to be for oligo 1501A. That is, the base sequence of oligo 1501A is likely the ground truth for these predicted base call sequences 1618c, 1618d, and 1618G, but the untrained neural network configuration 1615 may have incorrectly predicted at least some bases in these predicted base call sequences. Thus, the neural network configuration uses comparison function 1613 to compare each of predicted base call sequences 1618c, 1618d, and 1618G to ground truth 1506a (which is the base sequence of oligo 1501A) and uses the generated errors for gradient update 1617 and the resulting training of neural network configuration 1615.

[0183] Similarly, based on (ii) above, the base caller knows that predicted base call sequences 1618a and 1618f (e.g., corresponding to sequence signals 1512a and 1512f, respectively) are likely to be for oligo 1501B. That is, the base sequence of oligo 1501B is likely the ground truth for these predicted base call sequences 1618a and 1618f, but the untrained neural network configuration 1615 may have incorrectly predicted at least some bases in these predicted base call sequences. Thus, the neural network configuration uses comparison function 1613 to compare each of predicted base call sequences 1618a and 1618f to ground truth 1506b (which is the base sequence of oligo 1501B) and uses the generated errors for gradient update 1617 and the resulting training of neural network configuration 1615.

[0184] At the end of the training data consumption and training phase of FIG. 16A, the NN configuration 1615 is at least partially trained.

[0185] Figure 16B illustrates the base calling system 1400 of Figure 14A operating in a second iteration of the training data generation phase of the two-oligo training stage. For example, in Figure 16A, neural network configuration 1615 was trained using training data 1550. In Figure 16B, some, or at least a partially trained, neural network configuration 1615 is used to generate further training data. For example, the at least a partially trained neural network configuration 1615 predicts base call sequences 1628 using previously generated sequence signals 1512. The predicted base call sequences 1628 in Figure 16B are likely to be relatively more accurate than the predicted base call sequences 1618 in Figure 16A because the predicted base call sequences 1618 in Figure 16A were generated using an untrained neural network configuration 1615, whereas the predicted base call sequences 1628 in Figure 16B were generated at least in part using the neural network configuration 1615.

[0186] Additionally, mapping logic 1416 maps each of predicted base call sequences 1628 to either oligo 1501A or oligo 1501B, or declares the mapping of predicted base call sequence 1628 to be uncertain (e.g., similar to the discussion with respect to Figure 15D).

[0187] FIG. 16C illustrates labeled training data 1650 generated from the mapping of FIG. 16B, which is used for further training.

[0188] As illustrated in FIG. 16C, some of the predicted base call sequences 1628 and corresponding sequence signals 1512 map to the base sequence of oligo 1501A (i.e., ground truth 1506a), some other predicted base call sequences 1628 and corresponding sequence signals 1512 map to the base sequence of oligo 1501B (i.e., ground truth 1506b), and the mapping of the remaining predicted base call sequences 1628 and corresponding sequence signals 1512 is uncertain.

[0189] For example, predicted base call sequences 1628 are sorted into three categories: (i) predicted base call sequences 1628c, 1628d, and 1628G and corresponding sequence signals 1512c, 1512d, and 1512G map to the base sequence of oligo 1501A (i.e., ground truth 1506a); (ii) predicted base call sequences 1628a, 1628b, and 1628f and corresponding sequence signals 1512a, 1512b, and 1512f map to the base sequence of oligo 1501B (i.e., ground truth 1506b); and (iii) the mapping of the remaining predicted base call sequences 1628e and 1628g and corresponding sequence signals 1512e and 1512g is uncertain.

[0190] As an example only, assume that 3,300 base call sequences in training data 1650 are mapped to oligo 1501A and 3,200 base call sequences in training data 1650 are mapped to oligo 1501B. As illustrated in Figure 16C, the remaining 3,500 base call sequences are uncertain and do not map to either of the two oligos.

[0191] Comparing the number of sequences with unmapped (or uncertain) base calls between the training data of Figures 15E and 16C, it is observed that this number is 4,400 in Figure 15E and 3,500 in Figure 16C. This is because the at least partially trained neural network configuration 1615 of Figure 16B (used to generate the mappings of training data 1650) is relatively more accurate and / or can be trained better than the at least partially trained neural network configuration 1415 of Figure 15A (used to generate the mappings of training data 1550). Thus, the number of sequences with uncertain base calls gradually decreases as the base calls become relatively more accurate (e.g., have fewer errors) and thus relatively more correctly mapped.

[0192] FIG. 16D illustrates the base-calling system 1400 of FIG. 14A operating in the second iteration of the "training data consumption and training phase" of the "2-oligo training stage" to train the base-caller 1414 comprising the neural network configuration 1615 of FIG. 16A using two known synthetic sequences 1501A and 1501B.

[0193] Figures 16A and 16D are at least partially similar. For example, Figures 16A and 16D are used to train neural network configuration 1615 using training data 1550 of Figure 15E and training data 1650 of Figure 16C, respectively. Note that in the initial stage of Figure 16A, neural network configuration 1615 is not trained at all, while in the initial stage of Figure 16D, neural network configuration 1615 is at least partially trained.

[0194] In Figure 16D, base caller 1414, including at least a partially trained neural network configuration 1615, receives sequence signal 1512 previously generated during the "training data generation phase" of Figure 15A and predicts base call sequence 1638 from sequence signal 1512. Predicted base call sequences 1638 include predicted base call sequences 1638a, 1638b, ..., 1638G. For example, sequence signal 1512a is used to predict base call sequence 1638a, sequence signal 1512b is used to predict base call sequence 1638b, sequence signal 1512G is used to predict base call sequence 1638G, etc.

[0195] Neural network configuration 1615 is not fully trained, and therefore predicted base call sequences 1638a, 1638b, ..., 1638G contain some errors, but the errors in predicted base call sequence 1638 of Figure 16D are likely to be fewer than the errors in predicted base call sequence 1618 of Figure 16A and predicted base call sequence 1628 of Figure 16B. Mapped training data 1650 of Figure 16C is now used to further train neural network configuration 1615. For example, from training data 1650, base call 1414 is (i) sequence signals 1512c, 1512d, 1512G are for the base sequence of oligo 1501A (i.e., ground truth 1506a); (ii) sequence signals 1512a, 1512b, 1512f are for the base sequence of oligo 1501B (i.e., ground truth 1506b); and (iii) It is known that the mapping of the constellation signals 1512e, 1512g is uncertain.

[0196] Thus, based on (iii) above, predicted base call sequences 1638e and 1638g in Figure 16D (e.g., corresponding to sequence signals 1512e and 1512g, respectively) are not used to train neural network configuration 1615. Accordingly, these predicted base call sequences 1638e and 1638g are discarded from the training data and are not used for gradient update (symbolically illustrated in Figure 16D using an "X" or "cross symbol" between predicted base call sequences 1618e, 1618g and gradient update box 1617).

[0197] Based on (i) above, base caller 1414 knows that predicted base call sequences 1638c, 1638d, and 1638G (e.g., corresponding to sequence signals 1512c, 1512d, and 1512G, respectively) are likely to be for oligo 1501A. That is, the base sequence of oligo 1501A is likely the ground truth for these predicted base call sequences 1638c, 1638d, and 1638G, but neural network configuration 1615 may have incorrectly predicted at least some bases in these predicted base call sequences. Thus, neural network configuration 1615 uses comparison function 1613 to compare each of predicted base call sequences 1638c, 1638d, and 1638G to ground truth 1506a (which is the base sequence of oligo 1501A) and uses the generated errors for gradient update 1617 and the resulting training of neural network configuration 1615. For example, during comparison, each base call in predicted base call sequence 1638c is compared with the corresponding base call in the corresponding ground truth sequence to generate a corresponding comparison result, for example, as discussed with respect to FIG. 14A1.

[0198] Similarly, based on (ii) above, base caller knows that predicted base call sequences 1638a, 1638b, and 1638f (e.g., corresponding to sequence signals 1512a, 1512b, and 1512f, respectively) are likely to be for oligo 1501B. That is, the base sequence of oligo 1501A is likely the ground truth for these predicted base call sequences 1638a, 1638b, and 1638f, but neural network configuration 1615 may have incorrectly predicted at least some bases in these predicted base call sequences. Thus, neural network configuration 1615 uses comparison function 1613 to compare each of predicted base call sequences 1638a, 1638b, and 1638f to ground truth 1506b (which is the base sequence of oligo 1501B) and uses the generated errors for gradient update 1617 and the resulting training of neural network configuration 1615.

[0199] FIG. 17A illustrates a flowchart depicting an exemplary method 1700 for iteratively training neural network configurations for base calling using single-oligo and two-oligo sequences. Method 1700 progressively trains NN configurations that are progressively and monotonically complex in nature. Increasing the complexity of the NN configuration may include increasing the number of layers of the NN configuration, increasing the number of filters of the NN configuration, increasing the topological complexity of the NN configuration, and / or the like. For example, method 1700 may reference a first NN configuration (which may be NN configuration 1415 discussed previously herein with reference to FIG. 14A and other figures), a second NN configuration (which may be NN configuration 1615 discussed previously herein with reference to FIG. 16A and other figures), a Pth NN configuration (not specifically discussed with reference to FIGS. 14A-16D), etc. In an example, as symbolically illustrated in box 1710 of Figure 17A, the complexity of the Pth NN configuration is higher than the complexity of the (P-1)th NN configuration, which is higher than the complexity of the (P-2)th NN configuration, etc. The complexity of the second NN configuration is therefore monotonically increasing (i.e., each subsequent stage of NN configuration has at least as much complexity as the previous stage of NN configuration).

[0200] Note that in method 1700, operation 1704a is for iteratively training a first NN configuration and generating labeled training data for a second NN configuration, operations 1704b1-1704bk are for training a second NN configuration and generating labeled training data for a third NN configuration, and operation 1704c is for training a third NN configuration and generating labeled training data for a fourth NN configuration. This process continues, with operation 1704P being for training a Pth NN configuration and generating labeled training data for the subsequent NN configuration. Thus, generally speaking, in method 1700, operation 1704i is for training an ith NN configuration and generating labeled training data for the (i+1)th NN configuration, where i=1, ..., P.

[0201] Method 1700 includes, at 1704a, (i) iteratively training a first NN configuration with a single oligo sequence and (ii) generating first two-oligo-labeled training data using the trained first NN configuration. As discussed, the first NN configuration is NN configuration 1415 of FIG. 14A, and the single oligo sequence includes oligo #1 discussed with respect to FIGS. 14A and 14B. The iterative training of the first NN configuration with the single oligo sequence is discussed with respect to FIGS. 14A and 14B. The generation of first two-oligo-labeled training data using the trained first NN configuration is discussed with respect to FIGS. 15A, 15D, and 15E, where the first two-oligo-labeled training data is training data 1550 of FIG. 15E.

[0202] Method 1700 then proceeds from 1704a to 1704b. Illustratively, operation 1704b is for training a second NN configuration (e.g., using the first two-oligo-labeled training data generated from operation 1704a) and using the trained second NN configuration to generate further two-oligo-labeled training data for training a third NN configuration. Operation 1704b includes sub-operations at blocks 1704b1 through 1704bk.

[0203] In block 1704b1, (i) a second NN configuration is trained using the first two-oligo-labeled training data generated in 1704a, and (ii) second two-oligo-labeled training data is generated using the at least partially trained second NN configuration. As discussed, the second NN configuration is NN configuration 1615 of FIG. 16A. Training of the second NN configuration using the first two-oligo-labeled training data is also illustrated in FIG. 16A. Generation of the second two-oligo-labeled training data (e.g., training data 1650 of FIG. 16C) using the at least partially trained second NN configuration is discussed with respect to FIGS. 16B and 16C.

[0204] Method 1700 then proceeds from block 1704b1 to block 1704b2. At block 1704b2, (i) a second NN configuration is further trained using the second two-oligo-labeled training data, and (ii) a third two-oligo-labeled training data is generated using the further trained second NN configuration. Training the second NN configuration using the second two-oligo-labeled training data is illustrated in Figure 16D. Generating the third two-oligo-labeled training data using the further trained second NN configuration is not illustrated, but is similar to the discussion regarding Figures 16B and 16C.

[0205] Note that block 1704b1 is the first iteration of training the second NN configuration, block 1704b2 is the second iteration of training the second NN configuration, and so on, and finally block 1704bk is the kth iteration of training the second NN configuration. As discussed, the operation of block 1704b1 is discussed in detail with respect to Figures 16A, 16B, and 16C. The operation of subsequent blocks 1704b2, ..., 1704bk may be similar to the discussion for block 1704b1.

[0206] Note that the same second NN configuration is used in all of the iterations 1704b1, ..., 1704bk, and thus these k iterations aim to repeatedly train the same second NN configuration without increasing the complexity of the second NN configuration.

[0207] Training of the second NN configuration progresses with each iteration of blocks 1704b1, 1704b2, ..., 1704bk. As the second neural network is gradually trained with each iteration 1704b1, ..., 1704bk, the second neural network makes progressively fewer errors in predicting base call sequences. For example, as shown in block 1704a and also illustrated in FIG. 15E, the first two-oligo-labeled training data (i.e., training data 1550) generated using the trained first NN configuration has 44% (i.e., 4,400 out of 10,000) uncertain mappings. As shown in block 1704b1 and also illustrated in FIG. 16C, the second two-oligo-labeled training data (i.e., training data 1650) generated using the partially trained second NN configuration has 35% (i.e., 3,500 out of 10,000) uncertain mappings. As shown in block 1704b2 and by way of example only, the third two-oligo-label training data generated using the further trained second NN configuration may have 32% (i.e., 3,200 out of 10,000) uncertain mappings. The percentage of uncertain mappings may gradually decrease with each iteration until it reaches, for example, about 20% at block 1704bk.

[0208] The number of iterations "k" for training the second NN configuration may be based on satisfying one or more convergence conditions. When the convergence conditions are satisfied, the iterations for training the second NN configuration may terminate. The convergence conditions are implementation-specific and dictate the number of iterations to undergo for training the second NN configuration. In an example, satisfying the convergence conditions indicates that further iterations may not significantly benefit further training of the second NN configuration, and thus, training iterations for the second NN configuration may terminate. Several examples of convergence conditions and satisfying them are discussed herein. For example, the second NN configuration may be iteratively trained until the percentage of uncertain mappings is less than a threshold percentage. Here, the convergence condition is satisfied when the percentage of uncertain mappings is less than a threshold percentage. For example, for the second NN configuration, this threshold may be approximately 20%, by way of example only. Thus, when the threshold is satisfied at iteration k, the convergence condition is satisfied and training of the second NN configuration terminates. Thus, the method proceeds to 1704c, where the Kth two-oligo-labeled training data generated in block 1704bk is used to train a third NN configuration that is more complex than the second NN configuration.

[0209] In another example, the iterations of the second NN configuration continue until the proportion of uncertain mappings reaches a certain saturation point (i.e., does not decrease significantly with successive iterations), satisfying the convergence condition. That is, in this example, saturation below a threshold indicates sufficient convergence of the iterative training (e.g., indicates satisfaction of the convergence condition), and the current iteration of the model can be terminated because further iterations will not significantly improve the model. For example, assume that at iteration (k-2) (e.g., at block 1704b(k-2)), the proportion of uncertain mappings is 21%, at iteration (k-1) (e.g., at block 1704b(k-2)), the proportion of uncertain mappings is 20.4%, and at iteration k (e.g., at block 1704bk), the proportion of uncertain mappings is 20%. Thus, in the last two iterations, the decrease in the proportion of uncertain mappings is relatively low (e.g., 0.6% and 0.4%, respectively), suggesting that training is nearly saturated and that further training will not significantly improve the second NN configuration. Here, saturation is measured as the difference between the proportion of uncertain mappings in two consecutive iterations. That is, if two consecutive iterations have approximately the same proportion of uncertain mappings, further iterations may not be helpful in further reducing this proportion, and therefore, training iterations can be terminated. Therefore, at this stage, the iterations for the second NN configuration are terminated, and method 1700 proceeds to 1704c for the third NN configuration.

[0210] In yet another embodiment, the number of iterations "k" is pre-specified, and completing k iterations satisfies the convergence condition, such that training for the current NN configuration can be terminated and the next NN configuration can be initiated.

[0211] Thus, at the end of the iterations for the second NN configuration (i.e., at the end of block 1704k), the method 1700 proceeds to block 1704c, where a third NN configuration is iteratively trained. The training of the third NN configuration also involves iterations similar to those discussed with respect to operations 1704b1, ..., 1704bk, and therefore will not be discussed in further detail.

[0212] This process of progressively training more complex NN configurations continues at 1704P of method 1700 until the Pth NN configuration has been trained and 2 oligo training data has been generated for training the next NN configuration.

[0213] Note that in an embodiment, as discussed herein, the same two oligo sequences may be used for all iterations of blocks 1704b1, ..., 1704bk, 1704c, ..., 1704P, however, in some other embodiments, not discussed herein, different two oligo sequences may be used for different iterations of method 1700 of Figure 17.

[0214] As discussed, the more complex the model, the better the model can be trained to predict base calls. For example, at the end of training a second NN configuration, the final labeled training data generated by the second NN configuration has 20% uncertain mappings. The percentage of uncertain mappings further decreases at the end of training a third NN configuration. For example, during the first training iteration of the third NN configuration, the percentage of uncertain mappings is 36% (e.g., because the third NN configuration is barely trained during the first iteration), and this percentage gradually decreases with subsequent training iterations of the third NN configuration. As illustrated in FIG. 17A, for example, assume that at the end of training a third NN configuration, the final labeled training data generated by the third NN configuration has 17% uncertain mappings. This percentage of uncertain mappings further decreases as the iterations of FIG. 17A progress, for example, at the end of training a Pth NN configuration, the final labeled training data generated by the Pth NN configuration has 12% uncertain mappings. Note that training ends with 12% uncertain mappings, for example, when the convergence condition (described above) for the Pth NN configuration is met. Thus, P NN configurations are trained in method 1700. The number "P" can be 3, 4, 5, or more, and is implementation-specific and can be based on meeting one or more corresponding convergence conditions. For example, if the (P-1)th NN configuration results in 12.05% uncertain mappings and the Pth NN configuration results in 12% uncertain mappings, there is a slight improvement of 0.05% uncertain mappings between the two NN configurations. This indicates that training of a new NN configuration with two oligo sequences is saturated. Here, saturation refers to the difference in the percentage of uncertain mappings between two successive NN configurations. If saturation is below a threshold (e.g., 0.1%), the two-oligo sequence training is terminated. In another example, the number "P" of NN configurations can be pre-specified by the user, for example, 3, 4, or more. As discussed below, once training with P NN configurations using two-oligo sequences is complete, more complex exemplars (such as three-oligo sequences) can be used for training.

[0215] Figure 17B illustrates exemplary final labeled training data generated by the Pth NN configuration at the end of method 1700 of Figure 17A. As discussed, at the end of training the Pth NN configuration, the final labeled training data generated by the Pth NN configuration has 12% (or 1,200 out of 10,000) uncertain mappings. The predicted base call sequences are sorted into three categories: (i) a first category containing predicted base call sequences that map to oligo 1501A, (ii) a second category containing predicted base call sequences that map to oligo 1501B, and (iii) a third category containing predicted base call sequences that do not map to either oligo 1501A or 1501B. The training data 1750 of Figure 17B will be apparent based on a discussion of the training data of Figures 15E and 16C.

[0216] FIG. 18A illustrates the basecalling system 1400 of FIG. 14A operating in the first iteration of the "training data consumption and training phase" of the "3-oligo training stage" to train a basecaller 1414 comprising a 3-oligo neural network configuration 1815. The reason for labeling the neural network configuration 1815 as a "3-oligo" neural network configuration 1815 will become clear later in this specification. FIG. 18A is at least partially similar to FIG. 16D. However, unlike FIG. 15D, labeled training data 1750 (see FIG. 17B) generated at the end of method 1700 (e.g., by the Pth NN configuration using 2-oligo base training) is used during training in FIG. 18A.

[0217] For example, in Figure 18A, base call 1414, which includes a three-oligo neural network configuration 1815, predicts base call sequences 1838a, 1838b, ..., 1838G. The mapped training data 1750 of Figure 17B is now used to further train the three-oligo neural network configuration 1815, similar to the training discussed with respect to Figure 16D.

[0218] FIG. 18B illustrates the base-calling system 1400 of FIG. 14A operating in the "training data generation phase" of the "three-oligo training stage" to train the base-caller 1414 comprising the three-oligo neural network configuration 1815 of FIG. 18A.

[0219] In Figure 18B, three different oligo sequences 1801A, 1801B, and 1801C are loaded into various clusters of flow cell 1405. By way of example only, and without limiting the scope of the present disclosure, assume that of the 10,000 clusters 1407, approximately 3,200 clusters contain oligo sequence 1801A, approximately 3,300 clusters contain oligo sequence 1801B, and the remaining 3,500 clusters contain oligo sequence 1801C (although in other examples, the three oligos can be divided substantially equally among the 10,000 clusters).

[0220] The sequencing machine 1404 generates sequence signals 1812a, ..., 1812G for corresponding clusters of the plurality of clusters 1407a, ..., 1407G. For example, for cluster 1407a, the sequencing machine 1404 generates a corresponding sequence signal 1812a that indicates the bases of cluster 1407a for a series of sequencing cycles. Similarly, for cluster 1407b, the sequencing machine 1404 generates a corresponding sequence signal 1812b that indicates the bases for cluster 1407b for a series of sequencing cycles, and so on.

[0221] The base caller 1414, including the neural network configuration 1815, predicts base call sequences 1818a, ..., 1818G for corresponding clusters of the plurality of clusters 1407a, ..., 1407G based on the corresponding sequence signals 1812a, ..., 1812G, respectively, as discussed, for example, with respect to FIG. 15A.

[0222] In embodiments, oligo sequences 1801A, 1801B, and 1801C are selected to have a sufficient edit distance between the bases of the three oligos, for example, as will become apparent based on a discussion of Figures 15B and 15C. For example, any of the three oligo sequences 1801A, 1801B, and 1801C is separated from another of the three oligo sequences 1801A, 1801B, and 1801C by at least a threshold edit distance. By way of example only, the threshold edit distance could be 4 bases, 5 bases, 6 bases, 7 bases, or even 8 bases. Thus, the three oligos are selected such that the three oligos are sufficiently different from one another.

[0223] Referring again to Figure 18B, in an embodiment, base caller 1414 does not know which oligo sequences are populated into which clusters. Therefore, base caller 1414 does not know the mapping between known oligo sequences 1801A, 1801B, and 1801C and the various clusters. Mapping logic 1416 receives predicted base call sequences 1818 and maps each predicted base call sequence 1818 to one of oligos 1801A, 1801B, or 1801C, or declares uncertainty when mapping a predicted base call sequence to any of the three oligos. Figure 18C illustrates the mapping operations for (i) mapping a predicted base call sequence to any of the three oligos 1801A, 1801B, or 1801C, or (ii) declaring uncertainty in the mapping of a predicted base call sequence to any of the three oligos.

[0224] As illustrated in Figure 18C, predicted base call sequence 1818a shares two bases of similarity with oligo 1801A, five bases of similarity with oligo 1801B, and one base of similarity with oligo 1801C. Assuming a threshold similarity ST of 4 (e.g., as discussed with respect to Equations 1-4), predicted base call sequence 1818a maps to oligo 1801B.

[0225] Similarly, in the example of Figure 18C, predicted base call sequence 1818b maps to oligo 1801C, and the mapping of predicted base call sequence 1818a is declared indeterminate by mapping logic 1416 of Figure 18B.

[0226] Figure 18D illustrates labeled training data 1850 generated from the mapping of Figure 18C, where the training data 1850 is used to train another neural network configuration. As illustrated in Figure 18D, some of the predicted base call sequences 1818 and corresponding sequence signals are mapped to the base sequence of oligo 1801A (i.e., ground truth 1806a), some of the predicted base call sequences 1818 and corresponding sequence signals are mapped to the base sequence of oligo 1801B (i.e., ground truth 1806b), some of the predicted base call sequences 1818 and corresponding sequence signals are mapped to the base sequence of oligo 1801C (i.e., ground truth 1806c), and the remaining mappings of the predicted base call sequences 1818 and corresponding sequence signals are uncertain. The training data 1850 of Figure 18D will be apparent based on the discussion of the training data 1550 of Figure 15E above.

[0227] FIG. 18E illustrates a flowchart depicting an example method 1880 for iteratively training neural network configurations for basecalling using three-oligo ground truth sequences. Method 1800 progressively trains a three-oligo NN configuration that is progressively and monotonically complex in nature. Increasing the complexity of the NN configuration may include increasing the number of layers of the NN configuration, increasing the number of filters of the NN configuration, increasing the topological complexity in the NN configuration, and / or the like, as also discussed with respect to FIG. 17A. For example, method 1880 references a first three-oligo NN configuration (which is the three-oligo NN configuration 1815 discussed earlier herein with respect to FIG. 18A), a second three-oligo NN configuration, a Qth NN configuration, etc. In an example, as symbolically illustrated in box 1890 of Figure 18E, the complexity of the Qth three-oligoNN configuration is higher than the complexity of the (Q-1)th three-oligoNN configuration, which is higher than the complexity of the (Q-2)th three-oligoNN configuration, and so on, with the complexity of the second three-oligoNN configuration being higher than the complexity of the first three-oligoNN configuration.

[0228] Note that in method 1880 of Figure 18E, operation 1704P is from the last block of method 1700 of Figure 17A, and operations 1888a1-1888am are for iteratively training a first three-oligoNN configuration and generating labeled training data for a second three-oligoNN configuration, operation 1888b is for iteratively training the second three-oligoNN configuration and generating labeled training data for a third three-oligoNN configuration, and so on. This process continues, and operation 1888Q is for training a Qth three-oligoNN configuration and generating labeled training data for training the subsequent NN configuration. Thus, generally speaking, in method 1880, operation 1888i is for training an ith three-oligoNN configuration and generating labeled training data for an (i+1)th three-oligoNN configuration, where i = 1, ..., Q.

[0229] Method 1880 includes, at 1704P, repeating operations 1704b1, ..., 1704bk to train the Pth NN configuration using the two-oligo ground truth data, and generating two-oligo labeled training data for training the next NN configuration, which is the last block of method 1700 of FIG. 17A.

[0230] Method 1880 then proceeds from 1704P to 1888a1. As illustrated, operation 1888a is for training a first three-oligo NN configuration (e.g., three-oligo neural network configuration 1815) using labeled training data (e.g., training data 1750 of FIG. 17B) generated from a previous block (e.g., block 1704P) and using the trained first three-oligo NN configuration to generate additional three-oligo labeled training data for subsequent training of a second three-oligo NN configuration. Operation 1888a includes sub-operations at blocks 1888a1-1888am.

[0231] In block 1888a1, (i) a first 3-oligo NN configuration (e.g., 3-oligo NN configuration 1815 of FIG. 18A) is trained using the labeled training data generated in 1704P, and (ii) 3-oligo labeled training data is generated using at least partially trained first 3-oligo NN configuration (e.g., training data 1850 of FIG. 18D).

[0232] Method 1880 then proceeds from 1888a1 to 1888a2, where (i) a first 3-oligo NN configuration is further trained using the 3-oligo labeled training data generated in a previous step (e.g., generated in block 1888a1), and (ii) new 3-oligo labeled training data is generated using the further trained first 3-oligo NN configuration.

[0233] The operations discussed with respect to block 1888a2 (and block 1888a2) are repeated iteratively at 1888a3, ..., 1888am. Note that blocks 1888a1, ..., 1888am are all for training the first three-oligoNN configuration. The number of iterations "m" may be implementation specific; example criteria used to select the number of iterations for training a particular NN model are discussed with respect to method 1700 of FIG. 17A (e.g., selection of the number of iterations "k" in this method).

[0234] After the first three-oligoNN configuration has been fully or satisfactorily trained in 1888am, method 1888 proceeds to block 1888b, where a second three-oligoNN configuration is iteratively trained. Training the second three-oligoNN configuration also involves iterations similar to those discussed with respect to operations 1888a1, ..., 1888am, and therefore will not be discussed in further detail.

[0235] This process of progressively training more complex NN configurations continues in method 1888 1888Q until the Qth 3-oligo NN configuration has been trained and corresponding 3-oligo training data has been generated for training the next NN configuration.

[0236] FIG. 19 illustrates a flowchart depicting an exemplary method 1900 for iteratively training a neural network configuration for base calling using multiple oligo ground truth sequences. FIG. 19 essentially summarizes the discussion related to FIGS. 14A-18E. For example, FIG. 19 illustrates an iterative training and labeled training data generation process using different oligo stages, such as a single oligo stage, a two-oligo stage, a three-oligo stage, etc. Thus, the complexity and / or length of the samples used for training and generation of labeled training data increases progressively and monotonically with each iteration, along with the complexity of the neural network configuration underlying the base calling.

[0237] Method 1900 includes, at 1904a, iteratively training one oligoNN configuration to generate labeled training data, e.g., as discussed with respect to block 1704a of method 1700 of Figures 14A and 14B and 17A.

[0238] Method 1900 further includes, at 1904b, iteratively training one or more two-oligo NN configurations using two-oligo sequences, e.g., as discussed with respect to blocks 1704b1-1704P of method 1700 of FIG. 17A, and generating labeled two-oligo training data.

[0239] Method 1900 further includes, at 1904c, iteratively training one or more 3-oligo NN configurations using the 3-oligo sequences and generating labeled 3-oligo training data, e.g., as discussed with respect to blocks 1888a1-1888Q of method 1880 of FIG. 18E.

[0240] This process continues, with progressively larger numbers of oligo sequences being used. Finally, in 1904N, one or more N oligo-NN configurations are trained using the N oligo sequences to generate corresponding N oligo-labeled training data, where N may be any suitable positive integer greater than or equal to 2. The operations in 1904N will become apparent based on a discussion of the operations in 1904b and 1904c.

[0241] 14A-19 are associated with training NN models using synthetically sequenced simple oligo sequences. For example, the oligo sequences used in these figures are likely to have fewer bases than sequences found in an organism's DNA. In embodiments, the oligo-based training discussed with respect to FIGS. 14A-19 is used to train increasingly complex NN models, generating increasingly rich labeled training datasets. For example, FIG. 19 uses an N-oligo NN configuration to output an N-oligo labeled training dataset, where the N-oligo labeled training dataset may have a much richer, more diverse, and larger labeled training dataset than a labeled training dataset associated with "less than N" number of oligos.

[0242] In practice, however, the sequencing machine 1404 and base caller 1414 will base call sequences that are much more complex than simple oligo sequences. For example, in practice, the sequencing machine 1404 and base caller 1414 will base call biological sequences that are much more complex than simple oligo sequences. Therefore, the base caller 1414 must be trained on base sequences found in biological DNA and RNA that are more complex than oligo sequences.

[0243] FIG. 20A illustrates an organism sequence 2000 used to train the base chorus 1414 of FIG. 14A . The organism sequence can be the sequence of an organism with a relatively small number of bases, such as phix (also called phiX). The phix bacteriophage is a single-stranded DNA (ssDNA) virus. The phix174 bacteriophage is an ssDNA virus that infects Escherichia coli and was the first DNA-based genome sequenced in 1977. phix (such as phiX174) virus particles have also been successfully assembled in vitro. In embodiments, after training the base chorus 1414 with oligo sequences (as discussed with respect to FIGS. 14A-19 ), the base chorus 1414 can be further trained with simple organism DNA, such as phix DNA, although this does not limit the scope of the present disclosure. For example, a more complex organism, such as a bacterium (such as E. coli or E. coli), can be used in place of phix. Thus, biological sequence 2000 can be phix or another relatively simple biological DNA. Biological sequence 2000 is pre-sequenced, i.e., the base sequence of biological sequence 2000 is known a priori (e.g., sequenced by a sequencing machine and a previously trained base caller different from that illustrated in FIG. 14A).

[0244] As illustrated in FIG. 20A , when biological sequence 2000 is loaded into sequencing machine 1404 of FIG. 14A , biological sequence 2000 is divided or partitioned into a plurality of subsequences 2004 a, 2004 b, ..., 2004 N. Each subsequence is loaded into one or more corresponding clusters. Thus, each cluster 1407 is populated with the corresponding subsequence 2004 and its synthetic copies. Any suitable criteria can be used to partition biological sequence 2000, such as the maximum size of a subsequence that can be populated into a cluster. For example, if each cluster of a flow cell can be populated with subsequences having a maximum of approximately 150 bases, then each of subsequences 2004 can be partitioned accordingly so that each has a maximum of 150 bases. In some embodiments, each subsequence 2004 can have a substantially equal number of bases, while in other embodiments, each subsequence 2004 can have a different number of bases. Subsequence 2004b is used as an example to discuss the teachings of the present disclosure and is assumed to have L1 bases. By way of example only, the number L1 can be between 100 and 200, but can have any other suitable value and is implementation specific.

[0245] FIG. 20B illustrates the base calling system 1400 of FIG. 14A operating in the training data generation phase of the first organism training stage to train a base calling system 1414 comprising a first organism-level neural network configuration 2015 using subsequences 2004a, ..., 2004S of the first organism sequence 2000 of FIG. 20A.

[0246] Although not illustrated in Figure 20B, note that the first organism-level NN configuration 2015 is initially trained using the N-oligo-labeled training data from method 1904 of Figure 19. Thus, the first organism-level NN configuration 2015 is at least partially pre-trained. The base calling system 1400 of Figure 20B is the same as the base calling system of Figure 14A, except that in the two figures, the base calling systems 1400 use different neural network configurations and different analytes.

[0247] As described above, subsequences 2004a, ..., 2004S are loaded into corresponding clusters 1407. For example, subsequence 2004a is loaded into cluster 1407a, subsequence 2004b is loaded into cluster 1407b, etc. Note that each cluster 1407 contains multiple sequenced copies of the same subsequence 2004. For example, the subsequences loaded into a cluster are synthetically replicated such that the cluster has multiple copies of the same subsequence, which serves to generate a corresponding sequence signal 2012 for the cluster.

[0248] Note that base caller 1414 does not know which subsequences are populated into which clusters. For example, if subsequence 2004a and a composite copy of it are loaded into a particular cluster, base caller 1414 does not know which cluster subsequence 2004a is populated into. As described later in this specification, mapping logic 1416 is intended to map each subsequence 2004 to a corresponding cluster 1407 to facilitate the training process.

[0249] The sequencing machine 1404 generates sequence signals 2012a, ..., 2012G for corresponding clusters of the plurality of clusters 1407a, ..., 1407G. For example, for cluster 1407a, the sequencing machine 1404 generates a corresponding sequence signal 2012a that indicates the bases of cluster 1407a for a series of sequencing cycles. Similarly, for cluster 1407b, the sequencing machine 1404 generates a corresponding sequence signal 2012b that indicates the bases for cluster 1407b for a series of sequencing cycles, and so on.

[0250] In an embodiment, each subsequence 2004 is loaded into a corresponding cluster 1407, but the base caller 1414 does not know which subsequence is loaded into which cluster. Therefore, the base caller 1414 does not know the mapping between the subsequences 2004 and the clusters 1407. Because each cluster 1407 generates a corresponding sequence signal 2012, the base caller 1414 does not know the mapping between the subsequences 2004 and the sequence signal 2012.

[0251] Base caller 1414, including neural network configuration 2015, predicts base call sequences 2018a, ..., 2018G for corresponding clusters of a plurality of clusters 1407a, ..., 1407G based on corresponding sequence signals 2012a, ..., 2012G, respectively. For example, for cluster 1407a, base caller 1414 predicts corresponding base call sequence 2018a, including base calls for cluster 1407a for a set of sequencing cycles, based on corresponding sequence signal 2012a. Similarly, for cluster 1407b, base caller 1414 predicts corresponding base call sequence 2018b, including base calls for cluster 1407b for a set of sequencing cycles, based on corresponding sequence signal 2012b, and so on.

[0252] Note that neural network configuration 2015 is only partially trained, not fully trained, and therefore, neural network configuration 2015 may not accurately predict some or most of the bases of individual subsequences.

[0253] Furthermore, as base calling progresses in a partial sequence, it becomes increasingly difficult to call bases due to, for example, phasing or pre-phasing fading and / or noise. Figure 20C illustrates an example of fading, in which signal intensity decreases as a function of cycle number in a sequencing run of a base-calling operation. Fading is the exponential decay of the fluorescent signal intensity of a cluster as a function of cycle number. As a sequencing run progresses, the specimen strands are washed extensively, exposed to laser emissions that create reactive species, and subjected to harsh environmental conditions. All of this results in the gradual loss of fragments in each specimen, reducing its fluorescent signal intensity. Fading is also referred to as extinction or signal decay. Figure 20C illustrates an example of fading. In Figure 20C, the intensity values ​​of specimen fragments with AC microsatellites exhibit exponential decay.

[0254] Figure 20D conceptually illustrates the decreasing signal-to-noise ratio as the sequencing cycle progresses. For example, as sequencing progresses, signal intensity decreases and noise increases, resulting in a substantial decrease in the signal-to-noise ratio, making accurate base calling increasingly difficult. Physically, it has been observed that later synthesis steps attach tags to the sensor at different positions than earlier synthesis steps. When the sensor is below the sequence being synthesized, signal decay results from tags being attached to strands further away from the sensor in later sequencing steps than in earlier steps. This causes signal decay as the sequencing cycle progresses. In some designs, when the sensor is above the substrate holding the cluster, the signal may increase as sequencing progresses, instead of decreasing.

[0255] In the investigated flow cell designs, noise increases while the signal decays. Physically, phasing and prephasing increase noise as sequencing progresses. Phasing refers to a sequencing step in which the tag cannot advance along the sequence. Prephasing refers to a sequencing step in which the tag jumps forward two positions instead of one during a sequencing cycle. Both phasing and prephasing are relatively infrequent, occurring approximately once every 500–1000 cycles. Phasing is slightly more frequent than prephasing. Because phasing and prephasing affect individual strands within a cluster that generate intensity data, the intensity noise distribution from the cluster accumulates as sequencing progresses, with binomial, trinomial, and quaternary unfolding.

[0256] Further details of fading, signal attenuation, and signal-to-noise ratio reduction, as well as Figures 20C and 20D, can be found in U.S. Non-Provisional Patent Application No. 16 / 874,599, entitled "Systems and Devices for Characterization and Performance Analysis of Pixel-Based Sequencing," filed May 14, 2020 (Attorney Docket No. ILLM1011-4 / IP-1750-US), which is incorporated by reference as if fully set forth herein.

[0257] Thus, during base calling, the reliability or predictability of the base calling decreases as the sequencing cycles progress. For example, with reference to a particular subsequence, such as subsequence 2004b in Figure 20A, calling bases 1-10 of subsequence 2004b may generally be more reliable than calling bases 10-20 or calling bases 50-60. In other words, the first few bases of the L1 bases of subsequence 2004b are relatively more likely to be accurately predicted than the remaining bases of the L1 bases of subsequence 2004b.

[0258] Figure 20E illustrates base calling of the first L2 bases of the L1 bases of subsequence, where the first L2 bases of subsequence 2004b are used to map subsequence 2004b to sequence 2000.

[0259] For example, referring to Figures 20A, 20B, and 20E, sequencing machine 1404 generates sequence signal 2012b corresponding to subsequence 2004b (i.e., assume subsequence 2004b has been populated into cluster 1407b). However, base caller 1414 does not know where the subsequence corresponding to sequence signal 2012b fits into sequence 2000. That is, base caller 1414 does not know, in particular, that subsequence 2004b has been loaded into cluster 1407b.

[0260] As illustrated in Figure 20E, a partially trained NN configuration 2015 (e.g., trained using the N-oligo-labeled training data from method 1904 of Figure 19) receives sequence signal 2012b and predicts the L1 bases indicated by sequence signal 2012b. The L1 base predictions include predictions of the first L2 bases, and the predictions of the first L2 bases of subsequence 2004b are used to map subsequence 2004b to sequence 2000.

[0261] In an embodiment, the number L2 is 10. The number L2 can be any suitable number, such as 8, 10, 12, 13, or the like, so long as L2 is relatively smaller than L1. For example, L2 is less than 10% of L1, less than 25% of L1, or the like.

[0262] For example, the first L2 bases of subsequence 2004b predicted by NN configuration 2015 are A, C, C, T, G, A, G, C, G, A, as illustrated in Figure 20E. The remining (L1-L2) base predictions are generally illustrated as B1, ..., B1 in Figure 20E.

[0263] Here, the NN configuration 2015 may have correctly predicted the first L2 bases, or it may have made one or more errors in these L2 base predictions. The mapping logic 1416 attempts to map the predictions of the first L2 bases to the corresponding L2 consecutive bases in the biological sequence 2000. In other words, the mapping logic 1416 attempts to match the predictions of the first L2 bases to the L2 consecutive bases in the biological sequence 2000 so that the subsequence 2004b within the biological sequence 2000 can be identified.

[0264] As illustrated in FIG. 20E, the mapping logic 1416 can find a "substantial" and "unique" match between the first L2 bases predicted for the subsequence 2004b and the consecutive L2 bases in the biological sequence 2000. Note that a "substantial" match means that the match may not be 100% and that there may be one or more errors in the match. For example, the first L2 bases of the subsequence 2004b predicted by the NN configuration 2015 are A, C, C, T, G, A, G, C, G, A, while the corresponding substantially matching consecutive L2 bases in the biological sequence 2000 are A, G, C, T, G, A, G, C, G, A. Thus, the second base of these two L2 base sequences does not match, but the remaining bases match. As long as the number of such mismatches is below a threshold percentage, the mapping logic 1416 declares the two L2 base fragments to match. The threshold percentage of mismatch can be 10%, or 20%, or a similar percentage of the number L2. Thus, in an embodiment, L2 is 10, and the matching logic 1416 can tolerate up to two mismatches (or 20% mismatches). Thus, the mapping logic 1416 aims to map the first L2 bases predicted for the subsequence 2004b, or a slight variation thereof (e.g., the variation represents the error tolerance during matching), to the L2 consecutive bases of the biological sequence 2000. The value of the threshold percentage can be implementation-specific and user-configurable. Simply by way of example, during the first iteration of training, the threshold percentage can be relatively high (e.g., 20%), and the threshold percentage can have a relatively low value (e.g., 10%) during later iterations of training. Thus, during the early stages of training iterations, the threshold percentage can be relatively high because the likelihood of error in base-calling predictions is relatively high. As the NN configuration becomes better trained, it becomes more likely to make better base-calling predictions, and so the threshold percentage can be gradually lowered, although in another embodiment, the threshold percentage can remain the same throughout all iterations of training.

[0265] Also, in an embodiment, a match between two L2 bases must be unique for proper mapping; a non-unique match may result in the match and mapping being declared indeterminate. Thus, the first L2 bases predicted for subsequence 2004b (or a small variation thereof) can only occur once in organism sequence 2000 for the match and mapping to be valid. Typically, in practical base sequences of simpler organisms, the consecutive L2 bases (or small variations thereof) are likely to occur only once in organism sequence 2000.

[0266] For example, referring to the example of Figure 20E, if one section of biological sequence 2000 has an occurrence of the consecutive bases A, G, C, T, G, A, G, C, G, A, and another section of biological sequence 2000 has another occurrence of the consecutive bases A, C, A, T, G, A, G, C, G, A, then both sections of biological sequence 2000 could potentially match the first L bases (A, C, C, T, G, A, G, C, G, A) of subsequence 2004b predicted by NN configuration 2015. Thus, in this example, the match is not unique, and mapping logic 1416 does not know which of the two sections of biological sequence 2000 maps to the L bases on subsequence 2004b. In such a scenario, mapping logic 1416 declares there is no reliable match (i.e., declares an uncertain mapping).

[0267] Referring to the example of Figure 20E, as illustrated, the first L2 bases of subsequence 2004b predicted by NN configuration 2015 "substantially" and "uniquely" match the corresponding contiguous L2 bases of biological sequence 2000. Also, given section 2000B (having L1 bases) of biological sequence 2000, the first L2 prediction of subsequence 2004b "substantially" and "uniquely" matches the first L2 bases of section B of biological sequence 2000. Therefore, subsequence 2004b is most likely actually section 2000B of biological sequence 2000. In other words, section 2000B of biological sequence 2000 was most likely partitioned in Figure 20A to form subsequence 2004b.

[0268] Thus, section 2000B of biological sequence 2000 serves as ground truth for sequence signal 2012b corresponding to subsequence 2004b. Figure 20F illustrates labeled training data 2050 generated from the mapping of Figure 20E, where labeled training data 2050 includes a section of biological sequence 2000 of Figure 20A as ground truth.

[0269] In the labeled training data 2050 of Figure 20F, by way of example only, subsequences 2004a, 2004d do not map to any section of biological sequence 2000 due to uncertain mapping. For example, as discussed with respect to Figure 20E, for mapping logic 1416 to declare a final mapping, there must be a substantial and unique match between the first L2 bases of the subsequence and the corresponding section of biological sequence 2000. NN configuration 2015 may generate a relatively large number of errors in the first L2 bases of each of subsequences 2004a, 2004d, resulting in these subsequences not being able to map to any corresponding section of biological sequence 2000.

[0270] In the labeled training data 2050 of Figure 20F, subsequence 2004b (and therefore sequence signal 2012b) is mapped to section 2000B of biological sequence 2000, as discussed with respect to Figure 20E. Similarly, subsequence 2004c is mapped to section 2000C of biological sequence 2000, and subsequence 2004S is mapped to section 2000S of biological sequence 2000. For example, subsequence 2004c is mapped to section 2000C of biological sequence 2000 (e.g., having the same number of bases as subsequence 2004c) such that the first L2 base predictions of subsequence 2004c "substantially" and "uniquely" match the first L2 bases of section 2000C.

[0271] Figure 20G illustrates the basecalling system 1400 of Figure 14A operating in the "training data consumption and training phase" of the "organism-level training stage" to train a basecaller 1414 comprising a first organism-level neural network configuration 2015. For example, the labeled training data 2050 of Figure 20F is used in the training of Figure 20G.

[0272] For example, L1 bases of subsequence 2004b predicted by base caller 1414 are compared to section 2000B of biological sequence 2000. Note that L1 bases of subsequence 2004b predicted by base caller 1414 have the first L2 bases compared to biological sequence 2000 to generate the mapping of Figure 20F. Because the remaining (L1-L2) bases are likely to contain many errors, the remaining (L1-L2) bases were not compared while generating the mapping of Figure 20F. This is because, as discussed with respect to Figures 20C and 20D, bases occurring later in the subsequence are more likely to be incorrectly predicted due to fading, phasing, and / or pre-phasing. In Figure 20G, all L1 bases of subsequence 2004b predicted by base caller 1414 are compared to the corresponding L1 bases on section 2000B of biological sequence 2000.

[0273] Thus, the mapping in Figure 20F identifies the portion of biological sequence 2000 (i.e., section 2000B) to which subsequence 2004b is compared in Figure 20G. Once the mapping is complete and labeled training data 2050 is generated, labeled training data 2050 is used in Figure 20G to compare and generate an error signal, which is used for gradient update 2017 in the backward pass of NN configuration 2015 and for training the resulting NN configuration 2015.

[0274] Note that some of the subsequences (such as subsequences 2004a and 2004d, see Figure 20F) do not ultimately match the corresponding sections of biological sequence 2000, and therefore the base call predictions corresponding to these subsequences are not used in the training of Figure 20G.

[0275] FIG. 21 illustrates a flowchart depicting an exemplary method 2100 for iteratively training neural network configurations for basecalling using the simple organism sequence 2000 of FIG. 20A. Method 2100 progressively trains NN configurations that are monotonically complex in nature. As previously described herein, increasing the complexity of the NN configuration may include increasing the number of layers of the NN configuration, increasing the number of filters of the NN configuration, increasing the topological complexity of the NN configuration, and / or the like. For example, method 2100 references a first organism-level NN configuration (the NN configuration 2015 discussed previously herein with respect to FIGS. 20B, 20G, and other figures), a second organism-level NN configuration, an Rth organism-level NN configuration, etc. In an embodiment, the complexity of the Rth organism-level NN configuration is higher than the complexity of the (R-1)th organism-level NN configuration, which is higher than the complexity of the (R-2)th organism-level NN configuration, etc., and the complexity of the second organism-level NN configuration is higher than the complexity of the first organism-level NN configuration.

[0276] Note that in method 2100, operation 2104a (including blocks 2104a1, ..., 2104am) is for training a first organism-level NN configuration and generating labeled training data for a second organism-level NN configuration, operation 2104b is for training a second organism-level NN configuration and generating labeled training data for a third organism-level NN configuration, etc. This process continues, and finally operation 2104R is for training an Rth organism-level NN configuration and generating labeled training data for the next stage of NN configuration. Thus, generally speaking, in method 2100, operation 2104i is for training an ith organism-level NN configuration and generating labeled training data for the (i+1)th organism-level NN configuration, where i=1, ..., R.

[0277] Method 2100 includes, at 2104a1, (i) training a first organism-level NN configuration (e.g., organism-level NN configuration 2015 of FIG. 20B, although training of this NN configuration is not illustrated in FIG. 20B) using the N-oligo-labeled training data from 1904N of method 1900 of FIG. 19, and (ii) generating labeled training data using at least partially trained first organism-level NN configuration 2015. The labeled training data is illustrated in FIG. 20F, and its generation is discussed with respect to FIGS. 20E and 20F.

[0278] Method 2100 then proceeds from 2104a1 to 2104a2, during which a second iteration is performed to train a first organism-level NN configuration 2015. For example, in 2104a2, (i) the first organism-level NN configuration 2015 is further trained using labeled training data from the previous stage, e.g., as discussed with respect to Figure 20G, and (ii) further labeled training data is generated using the at least partially trained first organism-level NN configuration 2015 (e.g., similar to the discussion with respect to Figures 20E and 20F).

[0279] The training and generation operations are repeated iteratively, eventually completing the training of the first organism-level NN configuration 2015 at 2104am. Note that block 2014a1 is the first iteration of training the first organism-level NN configuration 2015, block 2104a2 is the second iteration of training the first organism-level NN configuration 2015, and so on, until finally block 2104am is the mth iteration of training the first organism-level NN configuration 2015. The number of iterations can be based on one or more factors, such as those previously discussed herein with respect to method 1700 of FIG. 17A (e.g., when the criteria for selecting the number of iterations "k" were discussed). The complexity of the first organism-level NN configuration 2015 does not change during the iterations of 2104a1, ..., 2104am.

[0280] At the end of the iterations for the first organism-level NN configuration 2015 (i.e., at the end of block 2104am), the method 2100 proceeds to block 2104b, where a second organism-level NN configuration is iteratively trained. The training of the second organism-level NN configuration and the generation of associated training-labeled data will also involve iterations similar to those discussed with respect to operations 2104a1, ..., 2104am, and therefore will not be discussed in further detail.

[0281] This process of progressively training more complex NN configurations associated with the generation of training labeled data continues until, at 2104R of method 2100, the Rth organism-level NN configuration has been trained and corresponding labeled training data has been generated for training the next NN configuration.

[0282] Figure 22 illustrates the use of complex biological sequences for training corresponding NN configurations for the base call 1414 of Figure 14A. For example, as discussed with respect to Figures 20A-21, relatively simple biological sequences 2000 containing approximately L1 bases per subsequence are used to iteratively train R simple organism-level NN configurations to generate corresponding labeled training data. For example, method 2100 of Figure 21 illustrates such iterative training and generation of labeled training data using simple biological sequences 2000. As discussed, simple biological sequences 2000 can be Phix or another organism with a relatively simple (or relatively small) gene sequence.

[0283] 22 also illustrates the use of a relatively complex biological sequence 2200a. The biological sequence 2200a is more complex than the biological sequence 2000 because, for example, the number of bases in the composite biological sequence 2200a is greater than the number of bases in the biological sequence 2000. By way of example only, the biological sequence 2000 may have approximately 1 million bases, and the composite biological sequence 2200a may have 4 million bases. In another example, each subsequence partitioned from the composite biological sequence 2200a has a greater number of bases than each subsequence partitioned from the biological sequence 2000. In yet another example, the number of subsequences partitioned from the composite biological sequence 2200a is greater than the number of subsequences partitioned from the biological sequence 2000. For example, when segmenting composite biological sequence 2200a and biological sequence 2000, the number of subsequences segmented from composite biological sequence 2200a will be greater than the number of subsequences segmented from biological sequence 2000 because (i) composite biological sequence 2200a has more bases than biological sequence 2000 and (ii) each subsequence may have at most a threshold number of bases. In an embodiment, composite biological sequence 2200a includes genetic material from bacteria, such as E. coli, or other suitable biological sequence that is more complex than biological sequence 2000.

[0284] As illustrated in Figure 22, the composite biological sequence 2200a is used to iteratively train Ra composite biological-level NN configurations and generate labeled training data. The training and generation of labeled training data is similar to that discussed with respect to method 2100 of Figure 21 (the difference is that method 2100 is specifically directed to biological sequence 2000, whereas here a composite biological sequence 2200a is used).

[0285] This iterative process continues, and finally, a relatively more complex biological sequence 2200T is used. The more complex biological sequence 2200T is more complex than biological sequences 2000 and 2200a. For example, the number of bases in the more complex biological sequence 2200T is greater than the number of bases in each of biological sequences 2000 and 2200a. In another embodiment, each subsequence partitioned from the more complex biological sequence 2200T has a greater number of bases than each subsequence partitioned from biological sequences 2000 or 2200a. In yet another embodiment, the number of subsequences partitioned from the more complex biological sequence 2200T is greater than the number of subsequences partitioned from biological sequences 2000 or 2200a. In an embodiment, the more complex biological sequence 2200T includes genetic material from a complex species, such as genetic material from a human or other mammal.

[0286] As illustrated in Figure 22, biological sequence 2200T is used to iteratively train R T more complex biological-level NN configurations and generate labeled training data. The training and generation of labeled training data is similar to that discussed with respect to method 2100 of Figure 21 (the difference is that method 2100 is specifically directed to biological sequence 2000, whereas here biological sequence 2000T is used).

[0287] FIG. 23A illustrates a flowchart depicting an exemplary method 2300 for iteratively training a neural network configuration for base calling. Method 2300 summarizes at least some of the embodiments and examples discussed herein with respect to FIGS. 14A-22. Method 2300 incrementally trains a NN configuration that is monotonically complex in nature, as discussed herein. Method 2300 also uses monotonically complex gene sequences as exemplars. Method 2300 is used to train the base callers 1414 of various figures discussed herein.

[0288] Method 2300 begins at 2304, where a base chore 1414 including a NN configuration 1415 (see, e.g., FIG. 14A) is iteratively trained using single-oligo ground truth data, as discussed with respect to block 1704 of method 1700 of FIG. 17A. The at least partially trained NN configuration 1415 of FIG. 14A is used to generate labeled training data, as also discussed with respect to block 1704 of method 1700 of FIG. 17A.

[0289] Next, method 2300 proceeds from 2304 to 2308, where one or more NN configurations are iteratively trained using the two oligo sequences and corresponding labeled training data is generated, for example, as discussed with respect to method 1700 of FIG. 17A.

[0290] Next, method 2300 proceeds from 2308 to 2312, where one or more NN configurations are iteratively trained using the three oligo sequences and corresponding labeled training data is generated, for example, as discussed with respect to method 1900 of FIG. 19.

[0291] In 2316, one or more NN configurations are iteratively trained using N oligo sequences, for example, as discussed with respect to method 1900 of FIG. 19, and this process of training NN configurations using increasing numbers of oligos continues until corresponding labeled training data is generated.

[0292] Method 2300 then proceeds to 2320, where training and generation of labeled training data involves an organism. At 2320, a simple organism sequence, such as simple organism sequence 2000 of FIG. 20A, is used. One or more NN configurations are trained using the simple organism sequence (see, e.g., method 2100 of FIG. 21), and labeled training data is generated.

[0293] As method 2300 proceeds from 2320, increasingly complex biological sequences are used, for example, as discussed with respect to Figure 22. Finally, at 2328, one or more NN configurations are iteratively trained using complex biological sequences (e.g., the further complex biological sequences 2200T of Figure 22) to generate corresponding labeled training data.

[0294] Thus, method 2300 continues until base caller 1414 is "sufficiently trained." "Sufficiently trained" may imply that base caller 1414 can now make base calls with an error rate less than the target error rate. As discussed, the training process can continue iteratively until sufficient training and a target base calling error rate are achieved (see, e.g., the "Error Rate" chart in Figure 23E). At the end of method 2300, base caller 1414 comprising the final NN configuration of method 2300 is now fully trained. Thus, trained base caller 1414 comprising the final NN configuration of method 2300 can now be used for inference, e.g., to sequence unknown gene sequences.

[0295] 23B-23E illustrate various charts illustrating the effectiveness of the base-corer training process discussed in this disclosure. Referring to FIG. 23B, illustrated is a chart 2360 depicting the mapping percentage of training data generated by (i) a first two-oligo NN configuration, such as NN configuration 1615, trained using the neural network-based training data generation technique discussed herein, and (ii) a NN configuration trained using a conventional two-oligo training data generation technique. The white bars in chart 2360 illustrate mapping data from the first two-oligo NN configuration trained using training data generated using the neural network-based model discussed herein. Thus, the white bars in chart 2360 illustrate mapping data generated using the various techniques discussed herein. The gray bars in chart 2360 illustrate data associated with a NN configuration trained with training data generated by a conventional non-neural network-based model, such as a Real Time Analysis (RTA) model. An example of an RTA model is discussed in U.S. Pat. No. 10,304,189 (B2), issued May 28, 2019, entitled "Data Processing System and Methods," which is incorporated by reference as if fully set forth herein. Accordingly, the gray bars in chart 2360 illustrate mapping data generated using conventional techniques. In an example, the white bars in chart 2360 may be generated in operation 1704b1 of method 1700 of FIG. 17A . Chart 2360 illustrates the percentage of base call predictions that mapped to oligo 1, the percentage of base call predictions that mapped to oligo 2, and the percentage of base call predictions that cannot be conclusively mapped to either oligo 1 or 2 (i.e., the uncertain percentage). As can be seen, the uncertain percentage of training data generated using the techniques discussed herein is slightly higher than the uncertain percentage of training data generated using conventional techniques.Thus, initially (eg, at the beginning of the training iterations), the conventional techniques slightly outperform the training data generation techniques discussed herein.

[0296] 23C, illustrated is a chart 2365 depicting mapping percentages in training data generated using (i) a first two-oligo NN configuration (such as NN configuration 1615) trained using the neural network-based training data generation techniques discussed herein (white bars), (ii) a second two-oligo NN configuration trained using the neural network-based training data generation techniques discussed herein (dotted bars), and (iii) an NN configuration trained using a conventional two-oligo training data generation technique, such as an RTA-based conventional training data generation technique (gray bars). In an example, the first two-oligo NN configuration (white bars) and the second two-oligo NN configuration (dotted bars) correspond to operations 1704b and 1704c, respectively, of method 1700 of FIG. 17A. Chart 2365 illustrates the percentage of base call predictions that mapped to Oligo 1, the percentage of base call predictions that mapped to Oligo 2, and the percentage of base call predictions that could not be conclusively mapped to either Oligo 1 or Oligo 2 (i.e., the uncertain percentage). As can be seen, the uncertain percentage for the training data generated using the first two-oligo NN configuration is higher than each of (i) the training data generated using the second two-oligo NN configuration and (ii) the training data generated using the conventional technique. Furthermore, the uncertain percentage for the training data generated using the second two-oligo NN configuration is approximately equivalent to the training data generated using the conventional technique. Thus, with iterations and more complex NN configurations, the training data generated using the NN-based configuration is approximately equivalent to the training data generated using the conventional technique.

[0297] Referring now to Figure 23D, a chart 2370 is illustrated depicting the mapping percentages of training data generated by (i) a first four-oligo NN configuration trained using the neural network-based training data generation technique discussed herein (white bars), and (ii) a NN configuration trained using a conventional four-oligo training data generation technique, e.g., an RTA-based technique (gray bars). As can be seen, the uncertain percentages of training data generated using the techniques discussed herein are comparable to the uncertain percentages of training data generated using conventional techniques. Thus, when training is shifted to four-oligo sequences, the conventional techniques and the training data generation techniques discussed herein produce comparable results.

[0298] Referring now to Figure 23E, there is illustrated a chart 2375 depicting the error rate of data generated by (i) a NN configuration trained using the complex biological sequences discussed herein, e.g., with respect to operation 2328 of method 2300 of Figure 23A (solid line), and (ii) a NN configuration trained using a conventional complex biological training data generation technique, e.g., an RTA-based technique (dashed line). As can be seen, the error rate of the data generated using the techniques discussed herein is comparable to that of data generated using the conventional techniques. Thus, the conventional techniques and the training data generation techniques discussed herein produce comparable results. As discussed, the training data generation techniques discussed herein can be used in place of conventional techniques, for example, when the conventional techniques are not available or ready for training data generation.

[0299] 24 is a block diagram of a base calling system 2400 according to one implementation. The base calling system 2400 can operate to obtain any information or data related to at least one of biological or chemical substances. In some implementations, the base calling system 2400 is a workstation, which can be similar to a benchtop device or desktop computer. For example, most (or all) of the systems and components for performing the desired reactions can be within a common housing 2416.

[0300] In certain implementations, base calling system 2400 is a nucleic acid sequencing system (or sequencer) configured for various applications, including, but not limited to, de novo sequencing, resequencing of whole genomes or targeted genomic regions, and metagenomics. Sequencers may also be used for DNA or RNA analysis. In some implementations, base calling system 2400 may also be configured to generate reaction sites within a biosensor. For example, base calling system 2400 may be configured to receive a sample and generate surface-attached clusters of clonally amplified nucleic acids from the sample. Each cluster may constitute or be part of a reaction site within a biosensor.

[0301] Exemplary base calling system 2400 may include a system receptacle or interface 2412 configured to interact with biosensor 2402 to effect a desired reaction within biosensor 2402. In the description that follows with respect to Figure 24, biosensor 2402 is loaded into system receptacle 2412. However, it is understood that a cartridge containing biosensor 2402 may be inserted into system receptacle 2412, and that in some conditions, the cartridge may be temporarily or permanently removed. As noted above, the cartridge may include, among other things, fluid control and fluid storage components.

[0302] In certain implementations, base calling system 2400 is configured to perform multiple parallel reactions within biosensor 2402. Biosensor 2402 includes one or more reaction sites where desired reactions can occur. The reaction sites may be immobilized, for example, on a solid surface of the biosensor or on beads (or other movable substrates) located within corresponding reaction chambers of the biosensor. The reaction sites may include, for example, clusters of clonally amplified nucleic acids. Biosensor 2402 may include a solid-state imaging device (e.g., a CCD or CMOS imager) and a flow cell attached thereto. The flow cell may include one or more flow channels that receive solutions from base calling system 2400 and direct the solutions toward the reaction sites. Optionally, biosensor 2402 may be configured to engage a thermal element for transferring thermal energy into and out of the flow channels.

[0303] Base calling system 2400 may include various components, assemblies, and systems (or subsystems) that interact with each other to perform a predetermined method or assay protocol for biological or chemical analysis. For example, base calling system 2400 includes a system controller 2404, which may be in communication with the various components, assemblies, and subsystems of base calling system 2400, and also includes biosensor 2402. For example, in addition to system receptacle 2412, base calling system 2400 may also include a fluid control system 2406 for controlling fluid flow throughout the fluidic network of base calling system 2400 and biosensor 2402, a fluid reservoir system 2408 configured to hold any fluids (e.g., fluids, gases, or liquids) that may be used by the bioassay system, a temperature control system 2410 that may regulate the temperature of the fluids in the fluidic network, fluid reservoir system 2408, and / or biosensor 2402, and an illumination system 2409 configured to illuminate biosensor 2402. As described above, when a cartridge having biosensor 2402 is loaded into system receptacle 2412, the cartridge may also include fluid control and fluid storage components.

[0304] Base calling system 2400 may also include a user interface 2414 for interacting with a user. For example, user interface 2414 may include a display 2413 for displaying or requesting information from a user and a user input device 2415 for receiving user input. In some implementations, display 2413 and user input device 2415 are the same device. For example, user interface 2414 may include a touch-sensitive display configured to detect the presence of individual touches and identify the location of the touches on the display. However, other user input devices 2415, such as a mouse, touchpad, keyboard, keypad, handheld scanner, voice recognition system, motion recognition system, etc., may also be used. As described in more detail below, base calling system 2400 may communicate with various components, including biosensor 2402 (e.g., in the form of a cartridge), to perform desired reactions. Base calling system 2400 may also be configured to analyze data obtained from the biosensor to provide desired information to the user.

[0305] System controller 2404 may include any processor- or microprocessor-based system, including systems using microcontrollers, reduced instruction set computers (RISC), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), logic circuits, and any other circuits or processors capable of performing the functions described herein. The above examples are merely exemplary and, thus, are not intended to limit the definition and / or meaning of the term system controller. In an exemplary implementation, system controller 2404 executes sets of instructions stored in one or more storage elements, memories, or modules to at least one of acquire and analyze detection data. The detection data may include multiple sequences of pixel signals, such that sequences of pixel signals from each of millions of sensors (or pixels) can be detected over many base call cycles. The storage elements may be in the form of information sources or physical memory elements within base calling system 2400.

[0306] The set of instructions may include various commands that instruct the base call system 2400 or biosensor 2402 to perform specific operations, such as the methods and processes of various implementations described herein. The set of instructions may be in the form of a software program, which may form part of a tangible, non-transitory computer-readable medium or media. As used herein, the terms "software" and "firmware" are used interchangeably and include any computer program stored in memory executed by a computer, including RAM memory, ROM memory, EPROM memory, EEPROM memory, and non-volatile RAM (NVRAM) memory. The above memory types are exemplary only and thus not limiting of the types of memory that may be used to store a computer program.

[0307] The software may be in various forms, such as system software or application software. Furthermore, the software may be in the form of a collection of separate programs, or a program module or portion of a program module within a larger program. The software may also include modular programming in the form of object-oriented programming. After acquiring the detection data, the detection data may be processed automatically by the processed base calling system 2400 in response to user input, or may be processed in response to a request made by another processing machine (e.g., a remote request via a communications link). In another implementation shown, the system controller 2404 includes an analysis module 2538 (shown in FIG. 25 ). In other implementations, the system controller 2404 does not include the analysis module 2538, but instead has access to the analysis module 2538 (e.g., the analysis module 2538 may be separately hosted on the cloud).

[0308] The system controller 2404 may be connected to the biosensor 2402 and other components of the base calling system 2400 via a communication link. The system controller 2404 may also be communicatively connected to an off-site system or server. The communication link may be a wire, a cord, or wireless. The system controller 2404 may receive user input or commands from a user interface 2414 and user input devices 2415.

[0309] The fluid control system 2406 includes a fluid network and is configured to direct the flow of one or more fluids through the fluid network. The fluid network may be in fluid communication with the biosensor 2402 and the fluid reservoir system 2408. For example, fluid may be selected from the fluid reservoir system 2408 and directed to the biosensor 2402 in a controlled manner, or fluid may be drawn from the biosensor 2402 and directed to, for example, a waste reservoir within the fluid reservoir system 2408. Although not shown, the fluid control system 2406 may include a flow sensor that detects the flow rate or pressure of the fluid within the fluid network. The sensor may be in communication with the system controller 2404.

[0310] Temperature control system 2410 is configured to regulate the temperature of fluids in different regions of the fluid network, fluid reservoir system 2408, and / or biosensor 2402. For example, temperature control system 2410 may include a thermal circulator that interacts with biosensor 2402 and controls the temperature of fluids flowing along reaction sites within biosensor 2402. Temperature control system 2410 may also regulate the temperature of solid elements or components of base calling system 2400 or biosensor 2402. Although not shown, temperature control system 2410 may include sensors for detecting the temperature of fluids or other components. The sensors may be in communication with system controller 2404.

[0311] The fluid storage system 2408 is in fluid communication with the biosensor 2402 and may store various reaction components or reactants used to carry out a desired reaction. The fluid storage system 2408 may also store fluids for washing or cleaning the fluidic network and the biosensor 2402 and for diluting the reactants. For example, the fluid storage system 2408 may include various reservoirs for storing samples, reagents, enzymes, other biomolecules, buffers, aqueous, and non-polar solutions, etc. Additionally, the fluid storage system 2408 may also include a waste reservoir for receiving waste from the biosensor 2402. In implementations that include a cartridge, the cartridge may include one or more of a fluid storage system, a fluid control system, or a temperature control system. Accordingly, one or more of the components described herein for these systems may be contained within the cartridge housing. For example, the cartridge may have various reservoirs for storing samples, reagents, enzymes, other biomolecules, buffers, aqueous, and non-polar solutions, waste, etc. Thus, one or more of the fluid reservoir system, fluid control system, or temperature control system may be removably engaged with the bioassay system via a cartridge or other biosensor.

[0312] The illumination system 2409 may include a light source (e.g., one or more LEDs) and multiple optical components for illuminating the biosensor. Examples of light sources include lasers, arc lamps, LEDs, or laser diodes. The optical components may be, for example, reflectors, polarizers, beam splitters, collimators, lenses, filters, wedges, prisms, mirrors, detectors, etc. In implementations using an illumination system, the illumination system 2409 may be configured to direct excitation light to the reaction sites. As an example, a fluorophore may be excited by a green wavelength of light, so the wavelength of the excitation light may be approximately 532 nm. In one implementation, the illumination system 2409 is configured to generate illumination parallel to a surface normal of the surface of the biosensor 2402. In another implementation, the illumination system 2409 is configured to generate illumination that is off-angled relative to the surface normal of the surface of the biosensor 2402. In yet another implementation, the illumination system 2409 is configured to generate illumination having multiple angles, including some parallel illumination and some off-angle illumination.

[0313] System receptacle or interface 2412 is configured to engage biosensor 2402 in at least one of mechanical, electrical, and fluidic manners. System receptacle 2412 can hold biosensor 2402 in a desired orientation to facilitate fluid flow through biosensor 2402. System receptacle 2412 can also include electrical contacts configured to engage biosensor 2402 so that base calling system 2400 can communicate with and / or provide power to biosensor 2402. Additionally, system receptacle 2412 can include a fluid port (e.g., a nozzle) configured to engage biosensor 2402. In some implementations, biosensor 2402 is removably coupled to system receptacle 2412 both electrically and fluidically.

[0314] Additionally, the base calling system 2400 may communicate remotely with other systems or networks, or with other bioassay systems 2400. Detection data obtained by the bioassay system 2400 may be stored in a remote database.

[0315] FIG. 25 is a block diagram of a system controller 2404 that can be used in the system of FIG. 24. In one implementation, the system controller 2404 includes one or more processors or modules that can communicate with each other. Each of the processors or modules may include algorithms (e.g., instructions stored on a tangible and / or non-transitory computer-readable storage medium) or sub-algorithms for performing a particular process. The system controller 2404 is conceptually illustrated as a collection of modules, but may also be implemented using any combination of dedicated hardware boards, DSPs, processors, etc. Alternatively, the system controller 2404 may be implemented using a single processor or an off-the-shelf PC with multiple processors, with functional operations distributed among the processors. As a further option, the modules described below may be implemented using a hybrid configuration in which certain modular functions are performed using dedicated hardware, while remaining modular functions are performed using an off-the-shelf PC, etc. The modules may also be implemented as software modules within a processing unit.

[0316] During operation, the communication port 2520 may transmit information (e.g., commands) to the biosensor 2402 ( FIG. 24 ) and / or the subsystems 2406, 2408, 2410 ( FIG. 24 ). In implementations, the communication port 2520 may output multiple arrays of pixel signals. The communication port 2520 may receive user input from the user interface 2414 ( FIG. 24 ) and transmit data or information to the user interface 2414. Data from the biosensor 2402 or the subsystems 2406, 2408, 2410 may be processed in real time by the system controller 2404 during a bioassay session. Additionally or alternatively, the data may be temporarily stored in system memory during a bioassay session and processed in slower than real time or offline operation.

[0317] As shown in FIG. 25, the system controller 2404 may include multiple modules 2531-2539 in communication with a main control module 2530. The main control module 2530 may be in communication with a user interface 2414 (FIG. 24). While the modules 2531-2539 are shown in direct communication with the main control module 2530, the modules 2531-2539 may also be in direct communication with each other, the user interface 2414, and the biosensor 2402. The modules 2531-2539 may also be in communication with the main control module 2530 through other modules.

[0318] The plurality of modules 2531-2539 include system modules 2531-2533, 2539 that communicate with subsystems 2406, 2408, 2410, and 2409, respectively. Fluid control module 2531 may communicate with fluid control system 2406 to control valves and flow sensors in the fluid network to control the flow of one or more fluids through the fluid network. Fluid storage module 2532 can notify a user when fluid is low or when a waste reservoir is at or near full capacity. Fluid storage module 2532 may also communicate with temperature control module 2533 so that fluid can be stored at a desired temperature. Illumination module 2539 may communicate with illumination system 2409 to illuminate reaction sites at specified times during a protocol, such as after a desired reaction (e.g., a binding event) has occurred. In some implementations, illumination module 2539 can communicate with illumination system 2409 to illuminate reaction sites at a specified angle.

[0319] The plurality of modules 2531-2539 may also include a device module 2534 that communicates with the biosensor 2402 and an identification module 2535 that determines identification information associated with the biosensor 2402. The device module 2534 may, for example, communicate with the system receptacle 2412 to confirm that the biosensor has established electrical and fluidic connection with the base calling system 2400. The identification module 2535 may receive a signal that identifies the biosensor 2402. The identification module 2535 may use the identification information of the biosensor 2402 to provide other information to the user. For example, the identification module 2535 may determine and subsequently display the lot number, manufacturing date, or recommended protocol for operating the biosensor 2402.

[0320] The plurality of modules 2531-2539 also includes an analysis module 2538 (also referred to as a signal processing module or signal processor) that receives and analyzes signal data (e.g., image data) from the biosensor 2402. The analysis module 2538 includes memory (e.g., RAM or flash) for storing the detection data. The detection data can include multiple sequences of pixel signals, such that sequences of pixel signals from each of millions of sensors (or pixels) can be detected over many base call cycles. The signal data can be stored for subsequent analysis or transmitted to the user interface 2414 to display desired information to the user. In some implementations, the signal data can be processed by a solid-state imager (e.g., a CMOS image sensor) before the analysis module 2538 receives the signal data.

[0321] Analysis module 2538 is configured to acquire image data from the photodetector during each of a plurality of sequencing cycles, the image data being derived from the luminescence signals detected by the photodetector, and process the image data for each of the plurality of sequencing cycles through a neural network (e.g., neural network-based template generator 2548, neural network-based base caller 2558 (see, e.g., Figures 7, 9, and 10), and / or neural network-based quality scorer 2568) to generate base calls for at least some of the analytes during each of the plurality of sequencing cycles.

[0322] Protocol modules 2536 and 2537 communicate with main control module 2530 to control the operation of subsystems 2406, 2408, and 2410 in implementing a predetermined assay protocol. Protocol modules 2536 and 2537 may include instruction sets for instructing base calling system 2400 to perform specific operations according to a predetermined protocol. As shown, a protocol module may be a sequencing-by-synthesis (SBS) module 2536 configured to issue various commands to execute a sequencing-by-synthesis process. In SBS, the extension of nucleic acid primers along a nucleic acid template is monitored to determine the sequence of nucleotides in the template. The underlying chemical process may be polymerization (e.g., catalyzed by a polymerase enzyme) or ligation (e.g., catalyzed by a ligase enzyme). In certain polymer-based SBS implementations, fluorescently labeled nucleotides are added to primers (thereby extending the primers) in a template-dependent manner, such that detection of the order and type of nucleotides added to the primers can be used to determine the sequence of the template. For example, to initiate the first SBS cycle, one or more labeled nucleotides, DNA polymerase, etc. can be delivered into / through a flow cell containing an array of nucleic acid templates. The nucleic acid templates may be located at corresponding reaction sites. Primer extension can detect incorporated labeled nucleotides through an imaging event, and these reaction sites can be detected. During the imaging event, an illumination system 2409 can provide excitation light to the reaction sites. Optionally, the nucleotides can further include a reversible termination feature that terminates further primer extension once the nucleotide is added to the primer. For example, a nucleotide analog with a reversible terminator moiety can be added to the primer such that further extension does not occur until a deblocking agent is delivered to remove the moiety. Thus, in another implementation using reversible termination, a command can be given to deliver a deblocking reagent to the flow cell (before or after detection occurs).One or more commands can be given to effect washing between the various delivery steps. The cycle is then repeated n times to extend the primer by n nucleotides, thereby detecting a sequence of length n. Exemplary sequencing techniques are described, for example, in Bentley et al., Nature 456:53-59 (2008), WO 04 / 018497, U.S. Patent No. 7,057,026, WO 91 / 06678, WO 07 / 123744, U.S. Patent No. 7,329,492, U.S. Patent No. 7,211,414, U.S. Patent No. 7,315,019, and U.S. Patent No. 7,405,281, each of which is incorporated herein by reference.

[0323] In the nucleotide delivery step of the SBS cycle, any single type of nucleotide can be delivered at a time, or multiple different nucleotide types (e.g., A, C, T, and G together) can be delivered. In nucleotide delivery configurations where only a single type of nucleotide is present at a time, different nucleotides do not need to have distinct labels because they can be distinguished based on the temporal separation inherent in individualized delivery. Thus, a sequencing method or apparatus can use single-color detection. For example, the excitation source only needs to provide excitation at a single wavelength or a single wavelength range. In nucleotide delivery configurations where delivery results in multiple different nucleotides being present in the flow cell at a given time, the sites at which different nucleotide types incorporate can be distinguished based on the different fluorescent labels attached to each nucleotide type in the mixture. For example, four different nucleotides, each bearing one of four different fluorophores, can be used. In one implementation, the four different fluorophores can be distinguished using excitation in four different regions of the spectrum. For example, four different excitation radiation sources can be used. Alternatively, fewer than four different excitation sources can be used, but optical filtering of the excitation radiation from a single source can be used to generate different excitation radiation ranges in the flow cell.

[0324] In some implementations, fewer than four different colors can be detected in a mixture having four different nucleotides. For example, pairs of nucleotides can be detected at the same wavelength but can be distinguished based on differences in intensity for one member of the pair, or based on a change to one member of the pair (e.g., through chemical modification, photochemical modification, or physical modification) that causes a distinct signal to appear or disappear compared to the signal detected for the other member of the pair. Exemplary devices and methods for distinguishing four different nucleotides using detection of fewer than four colors are described, for example, in U.S. Patent Application Nos. 61 / 538,294 and 61 / 619,878, which are incorporated herein by reference in their entireties. U.S. Patent Application No. 13 / 624,200, filed September 21, 2012, is incorporated herein by reference in its entirety.

[0325] The multiple protocol modules may also include a sample preparation (or generation) module 2537 configured to issue commands to the fluidic control system 2406 and the temperature control system 2410 to amplify the product in the biosensor 2402. For example, the biosensor 2402 may be coupled to the base calling system 2400. The amplification module 2537 can issue instructions to the fluidic control system 2406 to deliver the necessary amplification components to a reaction chamber in the biosensor 2402. In other implementations, the reaction site may already contain some components for amplification, such as template DNA and / or primers. After delivering the amplification components to the reaction chamber, the amplification module 2537 can instruct the temperature control system 2410 to cycle through different temperature steps according to a known amplification protocol. In some implementations, amplification and / or nucleotide incorporation is performed isothermally.

[0326] The SBS module 2536 can issue commands to perform bridge PCR, in which clusters of clonal amplicons are formed over localized regions within the flow cell channel. After generating amplicons via bridge PCR, the amplicons may be "linearized" to create single-stranded template DNA, and sstDNA and sequencing primers may be hybridized to universal sequences flanking the region of interest. For example, reversible terminator-based sequencing by synthesis methods can be used, as described above or as follows.

[0327] Each base calling or sequencing cycle can extend the sstDNA by a single base, which can be achieved, for example, by using a modified DNA polymerase and a mixture of four types of nucleotides. Different types of nucleotides can have unique fluorescent labels, and each nucleotide can further have a reversible terminator that allows only a single base to be incorporated in each cycle. After a single base is added to the sstDNA, excitation light can be incident on the reaction site and fluorescence emission can be detected. After detection, the fluorescent label and terminator can be chemically cleaved from the sstDNA. Another similar base calling or sequencing cycle can be as follows: In such a sequencing protocol, the SBS module 2536 can instruct the fluid control system 2406 to direct the flow of reagent and enzyme solutions through the biosensor 2402. Exemplary reversible terminator-based SBS methods that can be utilized with the devices and methods described herein are described in U.S. Patent Application Publication No. 2007 / 0166705(A1), U.S. Patent Application Publication No. 2006 / 0188901(A1), U.S. Patent No. 7,057,026, U.S. Patent Application Publication No. 2006 / 0240439(A1), U.S. Patent Application Publication No. 2006 / 02814714709(A1), WO 05 / 065814, WO 06 / 064199, each of which is incorporated herein by reference in its entirety. Exemplary reagents for reversible terminator-based SBS are described in U.S. Pat. No. 7,541,444, U.S. Pat. No. 7,057,026, U.S. Pat. No. 7,427,673, U.S. Pat. No. 7,566,537, and U.S. Pat. No. 7,592,435, each of which is incorporated herein by reference in its entirety.

[0328] In some implementations, the amplification and SBS modules may operate in a single assay protocol, for example, template nucleic acids are amplified and subsequently sequenced within the same cartridge.

[0329] Base calling system 2400 may also allow the user to reconfigure the assay protocol. For example, base calling system 2400 may provide the user with options through user interface 2414 to modify the determined protocol. For example, if it is determined that biosensor 2402 will be used for amplification, base calling system 2400 may request the temperature of the annealing cycle. Additionally, base calling system 2400 may issue a warning to the user if the user provides user input that is not generally accepted for the selected assay protocol.

[0330] In an implementation, biosensor 2402 includes millions of sensors (or pixels), each of which generates a sequence of pixel signals over successive base call cycles. Analysis module 2538 detects the sequences of pixel signals and attributes them to corresponding sensors (or pixels) according to the row-wise and / or column-wise positions of the sensors on the array of sensors.

[0331] Each sensor in the array of sensors can generate sensor data for a tile of the flow cell, where the tile is within an area on the flow cell where a cluster of genetic material is placed during base calling. The sensor data can include image data within an array of pixels. For a given cycle, the sensor data can include two or more images, generating multiple features per pixel as tile data.

[0332] 26 is a simplified block diagram of a computer 2600 system that can be used to implement the disclosed techniques. The computer system 2600 includes at least one central processing unit (CPU) 2672 that communicates with several peripheral devices via a bus subsystem 2655. These peripheral devices may include, for example, a storage subsystem 2610, including memory devices and a file storage subsystem 2636, a user interface input device 2638, a user interface output device 2676, and a network interface subsystem 2674. The input and output devices allow user interaction with the computer system 2600. The network interface subsystem 2674 provides an interface to external networks, including interfaces to corresponding interface devices in other computer systems.

[0333] User interface input devices 2638 can include pointing devices such as keyboards, mice, trackballs, touchpads, or graphics tablets, scanners, touchscreens integrated into displays, audio input devices such as voice recognition systems and microphones, and other types of input devices. In general, use of the term "input device" is intended to encompass all possible types of devices and ways of inputting information into computer system 2600.

[0334] The user interface output devices 2676 may include a display subsystem, a printer, a fax machine, or a non-visual display such as an audio output device. The display subsystem may include a flat panel device such as an LED display, a cathode ray tube (CRT), a liquid crystal display (LCD), a projection device, or some other mechanism for producing a visible image. The display subsystem may also provide a non-visual display such as an audio output device. In general, use of the term "output device" is intended to encompass all possible types of devices and ways for outputting information from the computer system 2600 to a user or to another machine or computer system.

[0335] The storage subsystem 2610 stores programming and data constructs that provide the functionality of some or all of the modules and methods described herein. These software modules are generally executed by the deep learning processor 2678.

[0336] In one implementation, the neural network is implemented using a deep learning processor 2678, which may be a configurable and reconfigurable processor, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), and / or a coarse-grained reconfigurable architecture (CGRA) and a graphics processing unit (GPU) or other configured device. The deep learning processor 2678 may be hosted by a deep learning cloud platform such as Google Cloud Platform™, Xilinx™, and Cirrascale™. Examples of deep learning processors 14978 include Google's Tensor Processing Unit (TPU)™, rackmount solutions such as the GX4 Rackmount Series™, GX149 Rackmount Series™, NVIDIA DGX-1™, Microsoft's Stratix V FPGA™, Graphcore's Intelligent Processor Unit (IPU)™, Qualcomm's Zeroth Platform™ with Snapdragon processors™, NVIDIA's Volta™, NVIDIA's DRIVE PX™, NVIDIA's JETSON TX1 / TX2 MODULE™, Intel's Nirvana™, Movidius VPU™, Fujitsu's DPI™, ARM's DynamicIQ™, and IBM's TrueNorth™.

[0337] The memory subsystem 2622 used in the storage subsystem 2610 may include several memories, including a main random access memory (RAM) 2634 for storing instructions and data during program execution, and a read only memory (ROM) 2632 in which fixed instructions are stored. The file storage subsystem 2636 may provide persistent storage for program and data files and may include a hard disk drive, associated removable media, a CD-ROM drive, an optical drive, or a removable media cartridge. Modules that implement the functionality of an embodiment may be stored by the file storage subsystem 2636 in the storage subsystem 2610 or in another machine accessible by the processor.

[0338] Bus subsystem 2655 provides a mechanism for allowing the various components and subsystems of computer system 2600 to communicate with each other as intended. Although bus subsystem 2655 is shown schematically as a single bus, alternative implementations of the bus subsystem may use multiple buses.

[0339] The computer system 2600 itself can be of various types, including a personal computer, a portable computer, a workstation, a computer terminal, a network computer, a television, a mainframe, a server farm, a loosely distributed set of loosely networked computers, or any other data processing system or user device. Due to the ever-changing nature of computers and networks, the description of computer system 2600 shown in Figure 26 is intended only as a specific example for purposes of illustrating a preferred implementation of the present invention. Many other configurations of computer system 2600 can have more or fewer components than the computer system shown in Figure 26.

[0340] The present inventors disclose the following items.

[0341] section Node set #1 (self-learning base cola trained using oligo sequences). 1. A computer-implemented method for progressively training a base classifier, comprising: Iteratively initially training a base caller with samples containing a single oligonucleotide sequence and generating labeled training data using the initially trained base caller; (i) further training the base caller using samples containing the multi-oligonucleotide sequence, and generating labeled training data using the further trained base caller; and further training the base collaborator by repeating step (i) during at least one iteration while increasing the complexity of the neural network configuration loaded into the base collaborator, wherein the labeled training data generated during an iteration is used to train the base collaborator during an immediately subsequent iteration. 1a. Further training the base call with an exemplar comprising the multi-oligonucleotide sequence, during at least one iteration, increasing the number of unique oligonucleotide sequences of the multi-oligonucleotide sequence within the exemplar; Section 1 Method. 2. First, iteratively train the base classifier with samples containing single oligonucleotide sequences. During the first iteration of the first training of the base cola, Injecting a single known oligonucleotide sequence into multiple clusters of flow cells; generating a plurality of sequence signals corresponding to a plurality of clusters, each sequence signal of the plurality of sequence signals representing a base sequence loaded into a corresponding cluster of the plurality of clusters; predicting a corresponding base call for the known single oligonucleotide sequence based on each sequence signal of the plurality of sequence signals, thereby generating a plurality of predicted base calls; For each sequence signal of the plurality of sequence signals, generating a corresponding error signal based on a comparison of (i) the corresponding predicted base call and (ii) the bases of the known single oligonucleotide sequence, thereby generating a plurality of error signals corresponding to the plurality of sequence signals; and initially training a base coder during a first iteration based on the plurality of error signals. 2a. First train the base colleague during the first iteration. 3. The method of clause 2, including updating weights and / or biases of the neural network configuration based on the plurality of error signals using a backpropagation path of the neural network configuration loaded into the base call. 3. First, iteratively train the base classifier with samples containing single oligonucleotide sequences. During the second repetition of the first training of the base cola, which is performed after the first repetition of the first training, predicting corresponding further base calls for the known single oligonucleotide sequence based on each sequence signal of the plurality of sequence signals using the base callers partially trained during the first iteration of the initial training, thereby generating a plurality of further predicted base calls; For each sequence signal of the plurality of sequence signals, generating a corresponding further error signal based on a comparison of (i) the corresponding further predicted base call and (ii) the bases of the known single oligonucleotide sequence, thereby generating a plurality of further error signals corresponding to the plurality of sequence signals; The method of clause 2, further comprising: further initially training the base coder during a second iteration based on a plurality of further error signals. 4. First, iteratively train the base classifier with samples containing a single oligonucleotide sequence. 4. The method of clause 3, comprising: repeating a second iteration of the initial training of the base caller with exemplars comprising a single oligonucleotide sequence for multiple instances until a convergence condition is met. 5. The method of clause 4, wherein the convergence condition is met if, during two successive iterations of the second iteration of the initial training of the base collaborator, the decrease in the error signal is less than a threshold. 6. The method of Section 4, where the convergence condition is met if the second iteration of the initial training of the base cola is repeated for at least a threshold number of instances. 7. The plurality of sequence signals corresponding to the plurality of clusters generated during the first iteration of the initial training of the base collaborator are reused for the second iteration of the initial training of the base collaborator; Section 3 Method. 8. Comparing (i) the corresponding predicted base calls with (ii) the bases of a known single oligo sequence; 3. The method of clause 2, comprising, for a first predicted base call, (i) comparing a first base of the first predicted base call to a first base of the known single oligo sequence, and (ii) comparing a second base of the first predicted base call to a second base of the known single oligo sequence to generate a corresponding first error signal. 9. Further training of the base chore repeatedly further training the base caller for N1 iterations using an exemplar containing two known unique oligonucleotide sequences; further training the base caller for N2 iterations using an exemplar comprising three known unique oligonucleotide sequences; The method of clause 1, in which N1 iterations are performed before N2 iterations. 10. A first neural network configuration is loaded into the base collaborator to iteratively initially train the base collaborator using exemplars containing a single oligonucleotide sequence, and the base collaborator is further iteratively trained; further training the base caller for N1 iterations using exemplars containing two known unique oligonucleotide sequences, thereby (i) for a first subset of N1 iterations, a second neural network configuration is loaded into the base collaborator; (ii) the method of clause 1, wherein for a second subset of N1 iterations occurring after the first subset of N1 iterations, a third neural network configuration is loaded into the base collaborator, and the first, second, and third neural network configurations are different from one another. 11. The method of clause 10, wherein the second neural network configuration is more complex than the first neural network configuration, and the third neural network configuration is more complex than the second neural network configuration. 12. The method of clause 10, wherein the second neural network configuration has a greater number of layers than the first neural network configuration. 13. The method of clause 10, wherein the second neural network configuration has a greater number of weights than the first neural network configuration. 14. The method of clause 10, wherein the second neural network configuration has a greater number of parameters than the first neural network configuration. 15. The method of clause 10, wherein the third neural network configuration has a greater number of layers than the second neural network configuration. 16. The method of clause 10, wherein the third neural network configuration has a greater number of weights than the second neural network configuration. 17. The method of clause 10, wherein the third neural network configuration has a greater number of parameters than the second neural network configuration. 18. Further training the base caller for N1 iterations using an exemplar containing two known unique oligonucleotide sequences, during one of the N1 iterations: (i) dispensing a first known oligonucleotide sequence of two known unique oligonucleotide sequences into a first plurality of clusters of a flow cell, and (ii) dispensing a second known oligonucleotide sequence of the two known unique oligonucleotide sequences into a second plurality of clusters of a flow cell; predicting a corresponding base call for each cluster of the first and second plurality of clusters, such that a plurality of predicted base calls is generated; (i) mapping a first predicted base call of the plurality of predicted base calls to a first known oligonucleotide sequence, and (ii) mapping a second predicted base call of the plurality of predicted base calls to a second known oligonucleotide sequence, while refraining from mapping a third predicted base call of the plurality of predicted base calls to either the first or second known oligonucleotide sequence; generating (i) a first error signal based on comparing the first predicted base call to the first known oligonucleotide sequence, and (ii) a second error signal based on comparing the second predicted base call to the second known oligonucleotide sequence; and further training the base coder based on the first and second error signals. 19. Mapping the first predicted base call to a first known oligonucleotide sequence of two known unique oligonucleotide sequences, comparing each base of the first predicted base call with the corresponding base of the first and second known oligonucleotide sequences; determining that the first predicted base call has similarity to the first known oligonucleotide sequence by at least a threshold number of bases and similarity to the second known oligonucleotide sequence by less than a threshold number of bases; and mapping the first predicted base call to the first known oligobase sequence based on determining that the first predicted base call has similarity to the first known oligobase sequence by at least a threshold number of bases. 20. Refraining from mapping the third predicted base call to either the first or second known oligonucleotide sequence; comparing each base of the first predicted base call with the corresponding base of the first and second known oligonucleotide sequences; determining that the first predicted base call has similarity to each of the first and second known oligonucleotide sequences by less than a threshold number of bases; and refraining from mapping the third predicted base call to either the first or second known oligobase sequences based on determining that the first predicted base call has similarity with each of the first and second known oligobase sequences by less than a threshold number of bases. 21. Refraining from mapping the third predicted base call to either the first or second known oligonucleotide sequence; comparing each base of the first predicted base call with the corresponding base of the first and second known oligonucleotide sequences; determining that the first predicted base call has a similarity of more than a threshold number of bases with each of the first and second known oligonucleotide sequences; and refraining from mapping the third predicted base call to either the first or second known oligobase sequences based on determining that the first predicted base call has similarity to each of the first and second known oligobase sequences by more than a threshold number of bases. 22. Generating labeled training data using a further trained base colleague for one of the N1 iterations; re-predicting the corresponding base calls after further training the base caller during one of the N1 iterations, such that for each cluster of the first and second plurality of clusters, another plurality of predicted base calls is generated; (i) remapping a first subset of the other plurality of predicted base calls to a first known oligobase sequence, and (ii) remapping a second subset of the other plurality of predicted base calls to a second known oligobase sequence, while refraining from mapping a third subset of the other plurality of predicted base calls to either the first or second known oligobase sequences; and generating labeled training data based on the remapping, such that the labeled training data includes (i) a first subset of the other plurality of predicted base calls, where the first known oligobase sequence forms ground truth data for the first subset of the other plurality of predicted base calls, and (ii) a second subset of the other plurality of predicted base calls, where the second known oligobase sequence forms ground truth data for the second subset of the other plurality of predicted base calls. 23. The labeled training data generated during one of the N1 iterations is used to train the base classifier during the immediately following iteration of the N1 iterations. Section 22 method. 24. The neural network configuration of the base cola is the same during one iteration of the N1 iterations as during the immediately following iteration of the N1 iterations. Section 23 method. 25. The neural network configuration of the base collaborator during an iteration immediately following the N1 iterations is different and more complex than the neural network configuration of the base collaborator during one of the N1 iterations. Section 23 method. 26. Further training of the base chore repeatedly 2. The method of claim 1, comprising monotonically increasing the number of unique oligonucleotide sequences in the multi-oligonucleotide-containing sample with each iteration during the iterative further training. 27. Using base caller to predict base call sequences for unknown samples sequenced with known sequences of oligos; labeling each unknown analyte with a ground truth sequence that matches a known sequence; training a base classifier using labeled unknown analytes; Computer-implemented methods. 28. The computer-implemented method of clause 27, further comprising repeating the using, labeling, and training until convergence is satisfied. 29. Using base caller to predict base call sequences for a population of unknown specimens that have been sequenced to have two or more known sequences of two or more oligos; Sorting unknown samples from the population of unknown samples based on classification of the base call sequences of the selected unknown samples into known sequences; labeling each subset of the selected unknown specimens with a respective ground truth sequence that matches each known sequence based on the classification; training a base classifier using each labeled subset of the selected unknown analytes; Computer-implemented methods. 30. The computer-implemented method of clause 29, further comprising repeating the using, filtering, labeling, and training until convergence is satisfied. 31. A non-transitory computer-readable storage medium having stored thereon computer program instructions for progressively training a base caller, the instructions, when executed on a processor, performing: Iteratively initially training a base caller with samples containing a single oligonucleotide sequence and generating labeled training data using the initially trained base caller; (i) further training the base caller using samples containing the multi-oligonucleotide sequence, and generating labeled training data using the further trained base caller; and further training the base collaborator by repeating step (i) during at least one iteration while increasing the complexity of the neural network configuration loaded into the base collaborator, wherein the labeled training data generated during an iteration is used to train the base collaborator during an immediately subsequent iteration. 31a. The command is 32. The computer-readable storage medium of claim 31, further comprising increasing the number of unique oligonucleotide sequences of the multi-oligonucleotide sequence within the exemplar during at least one iteration of further training the base call with an exemplar including the multi-oligonucleotide sequence. 32. First, iteratively train the base classifier using samples containing a single oligonucleotide sequence. During the first iteration of the first training of the base cola, Injecting a single known oligonucleotide sequence into multiple clusters of flow cells; generating a plurality of sequence signals corresponding to a plurality of clusters, each sequence signal of the plurality of sequence signals representing a base sequence loaded into a corresponding cluster of the plurality of clusters; predicting a corresponding base call for the known single oligonucleotide sequence based on each sequence signal of the plurality of sequence signals, thereby generating a plurality of predicted base calls; For each sequence signal of the plurality of sequence signals, generating a corresponding error signal based on a comparison of (i) the corresponding predicted base call and (ii) the bases of the known single oligonucleotide sequence, thereby generating a plurality of error signals corresponding to the plurality of sequence signals; and initially training a base coder during a first iteration based on the plurality of error signals. 32a. Training the base colleague first during the first iteration 33. The computer-readable storage medium of claim 32, comprising: updating weights and / or biases of the neural network configuration based on the plurality of error signals using a backpropagation path of the neural network configuration loaded into the base code. 33. First, iteratively train the base classifier using samples containing a single oligonucleotide sequence. During the second repetition of the first training of the base cola, which is performed after the first repetition of the first training, predicting corresponding further base calls for the known single oligonucleotide sequence based on each sequence signal of the plurality of sequence signals using the base callers partially trained during the first iteration of the initial training, thereby generating a plurality of further predicted base calls; For each sequence signal of the plurality of sequence signals, generating a corresponding further error signal based on a comparison of (i) the corresponding further predicted base call and (ii) the bases of the known single oligonucleotide sequence, thereby generating a plurality of further error signals corresponding to the plurality of sequence signals; and further initially training the base coder during a second iteration based on a plurality of further error signals. 34. First, iteratively train the base classifier using samples containing a single oligonucleotide sequence. and repeating a second iteration of the initial training of the base caller using exemplars including a single oligonucleotide sequence for the plurality of instances until a convergence condition is met. 35. The computer-readable storage medium of clause 34, wherein the convergence condition is met if, during two successive iterations of the second iteration of the initial training of the base collaborator, a decrease in the plurality of further error signals is less than a threshold. 36. The computer-readable storage medium of clause 34, wherein the convergence condition is met if the second iteration of the initial training of the base collaborator is repeated for at least a threshold number of instances. 37. A plurality of sequence signals corresponding to a plurality of clusters generated during a first iteration of the initial training of the base collaborator are reused for a second iteration of the initial training of the base collaborator; The computer-readable storage medium of clause 33. 38. Comparing (i) the corresponding predicted base calls with (ii) the bases of a known single oligo sequence 33. The computer-readable storage medium of clause 32, comprising, for the first predicted base call, (i) comparing a first base of the first predicted base call to a first base of the known single oligo sequence, and (ii) comparing a second base of the first predicted base call to a second base of the known single oligo sequence to generate a corresponding first error signal. 39. Further training of the base chore repeatedly further training the base caller for N1 iterations using an exemplar containing two known unique oligonucleotide sequences; further training the base caller for N2 iterations using an exemplar comprising three known unique oligonucleotide sequences; 32. The computer-readable storage medium of claim 31, wherein N1 iterations occur before N2 iterations. 40. During initial iterative training of a base collaborator using exemplars containing a single oligonucleotide sequence, a first neural network configuration is loaded into the base collaborator, and the base collaborator is further iteratively trained; further training the base caller for N1 iterations using exemplars containing two known unique oligonucleotide sequences, thereby (i) for a first subset of N1 iterations, a second neural network configuration is loaded into the base collaborator; (ii) for a second subset of N1 iterations occurring after the first subset of N1 iterations, a third neural network configuration is loaded into the base collaborator, and the first, second, and third neural network configurations are different from one another. 41. The computer-readable storage medium of clause 40, wherein the second neural network configuration is more complex than the first neural network configuration, and the third neural network configuration is more complex than the second neural network configuration. 42. The computer-readable storage medium of clause 40, wherein the second neural network configuration has a greater number of layers than the first neural network configuration. 43. The computer-readable storage medium of clause 40, wherein the second neural network configuration has a greater number of weights than the first neural network configuration. 44. The computer-readable storage medium of clause 40, wherein the second neural network configuration has a greater number of parameters than the first neural network configuration. 45. The computer-readable storage medium of clause 40, wherein the third neural network configuration has a greater number of layers than the second neural network configuration. 46. ​​The computer-readable storage medium of clause 40, wherein the third neural network configuration has a greater number of weights than the second neural network configuration. 47. The computer-readable storage medium of clause 40, wherein the third neural network configuration has a greater number of parameters than the second neural network configuration. 48. Further training the base caller for N1 iterations using samples containing two known unique oligonucleotide sequences, during one of the N1 iterations: (i) dispensing a first known oligonucleotide sequence of two known unique oligonucleotide sequences into a first plurality of clusters of a flow cell, and (ii) dispensing a second known oligonucleotide sequence of the two known unique oligonucleotide sequences into a second plurality of clusters of a flow cell; predicting a corresponding base call for each cluster of the first and second plurality of clusters, such that a plurality of predicted base calls is generated; (i) mapping a first predicted base call of the plurality of predicted base calls to a first known oligonucleotide sequence, and (ii) mapping a second predicted base call of the plurality of predicted base calls to a second known oligonucleotide sequence, while refraining from mapping a third predicted base call of the plurality of predicted base calls to either the first or second known oligonucleotide sequence; generating (i) a first error signal based on comparing the first predicted base call to the first known oligonucleotide sequence, and (ii) a second error signal based on comparing the second predicted base call to the second known oligonucleotide sequence; and further training the base coder based on the first and second error signals. 49. Mapping a first predicted base call to a first known oligonucleotide sequence of two known unique oligonucleotide sequences, comparing each base of the first predicted base call with the corresponding base of the first and second known oligonucleotide sequences; determining that the first predicted base call has similarity to the first known oligonucleotide sequence by at least a threshold number of bases and similarity to the second known oligonucleotide sequence by less than a threshold number of bases; and mapping the first predicted base call to the first known oligobase sequence based on determining that the first predicted base call has similarity to the first known oligobase sequence by at least a threshold number of bases. 50. Refraining from mapping the third predicted base call to either the first or second known oligonucleotide sequence comparing each base of the first predicted base call with the corresponding base of the first and second known oligonucleotide sequences; determining that the first predicted base call has similarity to each of the first and second known oligonucleotide sequences by less than a threshold number of bases; and refraining from mapping the third predicted base call to either the first or second known oligobase sequences based on determining that the first predicted base call has similarity with each of the first and second known oligobase sequences by less than a threshold number of bases. 51. Refraining from mapping the third predicted base call to either the first or second known oligonucleotide sequence; comparing each base of the first predicted base call with the corresponding base of the first and second known oligonucleotide sequences; determining that the first predicted base call has a similarity of more than a threshold number of bases with each of the first and second known oligonucleotide sequences; and refraining from mapping the third predicted base call to either the first or second known oligobase sequences based on determining that the first predicted base call has similarity with each of the first and second known oligobase sequences by more than a threshold number of bases. 52. Generating labeled training data using a base colleague further trained for one of the N1 iterations; re-predicting the corresponding base calls after further training the base caller during one of the N1 iterations, such that for each cluster of the first and second plurality of clusters, another plurality of predicted base calls is generated; (i) remapping a first subset of the other plurality of predicted base calls to a first known oligobase sequence, and (ii) remapping a second subset of the other plurality of predicted base calls to a second known oligobase sequence, while refraining from mapping a third subset of the other plurality of predicted base calls to either the first or second known oligobase sequences; and generating the labeled training data based on the remapping, such that the labeled training data includes (i) a first subset of the other plurality of predicted base calls, where the first known oligobase sequence forms ground truth data for the first subset of the other plurality of predicted base calls, and (ii) a second subset of the other plurality of predicted base calls, where the second known oligobase sequence forms ground truth data for the second subset of the other plurality of predicted base calls. 53. The labeled training data generated during one of the N1 iterations is used to train the base classifier during the immediately following iteration of the N1 iterations. The computer-readable storage medium of clause 52. 54. The neural network configuration of the base cola is the same during one iteration of the N1 iterations as during the immediately following iteration of the N1 iterations. The computer-readable storage medium of clause 53. 55. The neural network configuration of the base collaborator during an iteration immediately following the N1 iterations is different and more complex than the neural network configuration of the base collaborator during one of the N1 iterations. The computer-readable storage medium of clause 53. 56. Further training of the base chore repeatedly 32. The computer-readable storage medium of claim 31, comprising monotonically increasing the number of unique oligonucleotide sequences in the sample containing the multi-oligonucleotide sequence as the iterations progress during the iterative further training.

[0342] Node set #2 (self-learning base cola trained using biological sequences) A1. A computer-implemented method for progressively training a base classifier, comprising: first training a base collaborator and generating labeled training data using the first trained base collaborator; (i) further training the base coder using samples containing biological base sequences and generating labeled training data using the further trained base coder; iteratively further training the base classifier by repeating step (i) for N iterations; Further training the base caller for N1 iterations among the N iterations using samples including the first biological base sequence selected into the first plurality of base subsequences; and and iteratively further training the base collaborator for N2 iterations out of the N iterations using samples including a second biological base sequence selected into a second plurality of base subsequences; The complexity of the neural network configuration loaded into the base collaborator increases monotonically with N iterations, The indicator generated during one of N iterations The computer-implemented method, wherein the training data is used to train a base collaborator during an iteration immediately following the N iterations. A1a. Training the base cola first is the best way to The method of clause A1, including initially training a base collaborator with exemplars including one or more oligonucleotide sequences, and generating labeled training data using the initially trained base collaborator. A2. The method of paragraph A1, wherein N1 iterations are performed before N2 iterations, and the second biosequence has a greater number of bases than the first biosequence. A3. Further training the base chorus for N1 repetitions is performed during one of the N1 repetitions. (i) inputting a first base partial sequence of the first plurality of base partial sequences of the first organism into a first cluster of the plurality of clusters of the flow cell, (ii) inputting a second base partial sequence of the first plurality of base partial sequences of the first organism into a second cluster of the plurality of clusters of the flow cell, and (iii) inputting a third base partial sequence of the first plurality of base partial sequences of the first organism into a third cluster of the plurality of clusters of the flow cell; (i) receiving a first sequence signal from a first cluster indicating a base partial sequence input to the first cluster, (ii) a second sequence signal from a second cluster indicating a base partial sequence input to the second cluster, and (iii) a third sequence signal from a third cluster indicating a base partial sequence input to the third cluster; (i) generating a first predicted base subsequence based on the first sequence signal, (ii) generating a second predicted base subsequence based on the second sequence signal, and (iii) generating a third predicted base subsequence based on the third sequence signal; (i) mapping the first predicted base subsequence to a first section of the first organism base sequence, and (ii) mapping the second predicted base subsequence to a second section of the first organism base sequence, while not mapping the third predicted base subsequence to any section of the first organism base sequence; The method of clause A1, comprising: generating labeled training data comprising: (i) a first predicted base subsequence mapped to a first section of a first biosequence, wherein the first section of the first biosequence is a ground truth for the first predicted base subsequence; and (ii) a second predicted base subsequence mapped to a second section of the first biosequence, wherein the second section of the first biosequence is a ground truth for the second predicted base subsequence. A3a. Further training of the base chorus for N1 repetitions is performed during one of the N1 repetitions. The method of clause A3, including training the base caller using the labeled training data generated during an initial training of the base caller prior to generating the first, second, and third predicted base subsequences. A4. The first predicted base subsequence has L1 bases, one or more of the L1 bases of the first predicted base subsequence do not match corresponding bases in the first section of the first biological base sequence due to an error in base calling prediction by the base caller; Method of Section A3. A5. A first predicted base subsequence has L1 bases, and the L1 bases of the first predicted base subsequence include the first L2 bases followed by the subsequent L3 bases, and mapping the first predicted base subsequence to a first section of a first biological base sequence is substantially and uniquely matching the first L2 bases of the first predicted sequence with the L2 consecutive bases of the first biological sequence; identifying a first section of the first biological sequence such that the first section (i) includes L2 consecutive bases as an initial base, and (ii) includes L1 bases; and mapping the first predicted base subsequence to the identified first section of the first organism base sequence. A6.The method is The method of A5, further comprising substantially and uniquely matching the first L2 bases of the first predicted base sequence, while refraining from aiming to match the subsequent L3 bases of the first predicted base sequence with any bases of the first biological base sequence. A7. The method of A5, wherein the first L2 bases of the first predicted base sequence substantially match L2 consecutive bases of the first biological base sequence, whereby at least a threshold number of bases of the first L2 bases of the first predicted base sequence match L2 consecutive bases of the first biological base sequence. A8. The method of A5, wherein the first L2 bases of the first predicted base sequence uniquely match the L2 consecutive bases of the first biological base sequence, thereby substantially matching only the L2 consecutive bases of the first biological base sequence and not matching any other L2 consecutive bases of the first biological base sequence. A9. The third predicted base subsequence has L1 bases, and the third predicted base subsequence does not map to any of the base subsequences of the first plurality of base subsequences, (i) the method of clause A3, comprising substantially and uniquely not matching the first L2 bases of the L1 bases of the third predicted sequence to the first L2 consecutive bases of the first biological sequence. A10. One iteration of the N1 iterations is a first iteration of the N1 iterations, and further training the base collaborator during a second iteration of the N1 iterations; training a base classifier using the labeled training data generated during a first iteration of the N1 iterations; using a base caller trained with the labeled training data generated during the first of the N1 iterations to generate (i) a first further predicted base subsequence based on the first sequence signal, (ii) a second further predicted base subsequence based on the second sequence signal, and (iii) a third further predicted base subsequence based on the third sequence signal; (i) mapping the first additional predicted base subsequence to a first section of the first organism base sequence, (ii) mapping the second additional predicted base subsequence to a second section of the first organism base sequence, and (iii) mapping the third additional predicted base subsequence to a third section of the first organism base sequence; and generating further labeled training data comprising: (i) a first further predicted base subsequence mapped to a first section of the first biosequence, wherein the first section of the first biosequence is a ground truth for the first further predicted base subsequence; (ii) a second further predicted base subsequence mapped to a second section of the first biosequence, wherein the second further section of the first biosequence is a ground truth for the second further predicted base subsequence; and (iii) a third further predicted base subsequence mapped to a third section of the first biosequence, wherein the third further section of the first biosequence is a ground truth for the third further predicted base subsequence. A11. Generating a first error between (i) a first predicted base subsequence generated during a first iteration of the N1 iterations and (ii) a first section of the first biological base sequence; (i) generating a second error between a further first predicted base subsequence generated during a second iteration of the N1 iterations and (ii) a first section of the first biological base sequence; The second error is less than the first error because the base classifier is better trained during the second iteration compared to the first iteration. Method of Section A10. A12. The first, second, and third sequence signals generated during the first iteration are reused in a second iteration to generate a first further predicted base subsequence, a second further predicted base subsequence, and a third further predicted base subsequence, respectively. Method of Section A10. A13. The neural network configuration of the base code is the same between the first iteration of the N1 iterations and the second iteration of the N1 iterations. Method of Section A10. A13a. The neural network configuration of the base class is reused for multiple iterations until a convergence condition is met. Method of Section A13. A14. The neural network configuration of the base collaborator during the first of the N1 iterations is different and more complex than the neural network configuration of the base collaborator during the second of the N1 iterations. Method of Section A10. A15. Further training the base caller for N1 iterations of the N iterations using a sample including the first biological base sequence; further training the base collaborator using the first neural network configuration loaded into the base collaborator for a first subset of N1 iterations; The method of clause A1, including further training the base collaborator using a second neural network configuration loaded into the base collaborator for a second subset of N1 iterations, the second neural network configuration being different from the first neural network configuration. A16. The method of clause A15, wherein the second neural network configuration has a greater number of layers than the first neural network configuration. A17. The method of clause A15, wherein the second neural network configuration has a greater number of weights than the first neural network configuration. A18. The method of clause A15, wherein the second neural network configuration has a greater number of parameters than the first neural network configuration. A19. Further training of the base chore repeatedly For one or more of the N1 iterations using a sample including a first biological base sequence, loading a first neural network configuration into the base collaborator; For one or more of the N2 iterations using a specimen containing a second biological base sequence, loading a second neural network configuration into the base cola, the second neural network configuration being different from the first neural network configuration. A20. The method of clause A19, wherein the second neural network configuration has a greater number of layers than the first neural network configuration. A21. The method of clause A19, wherein the second neural network configuration has a greater number of weights than the first neural network configuration. A22. The method of clause A19, wherein the second neural network configuration has a greater number of parameters than the first neural network configuration. A23. Further training the base caller for N1 iterations of the N iterations using a sample including the first biological base sequence; The method of clause A1, including repeating further training with the first biological sequence until a convergence condition is met after N1 iterations. A24. The method of clause A23, wherein the convergence condition is met when the decrease in the generated error signal between two successive iterations of the N1 iterations is less than a threshold. A25. The method of clause A23, wherein the convergence condition is met after completing N1 iterations.

[0343] B1. A non-transitory computer-readable storage medium having stored thereon computer program instructions for progressively training a base caller, the instructions, when executed on a processor, first training a base collaborator and generating labeled training data using the first trained base collaborator; (i) further training the base coder using samples containing biological base sequences and generating labeled training data using the further trained base coder; iteratively further training the base classifier by repeating step (i) for N iterations; Further training the base caller for N1 iterations among the N iterations using samples including the first biological base sequence selected into the first plurality of base subsequences; and and iteratively further training the base collaborator for N2 iterations out of the N iterations using samples including a second biological base sequence selected into a second plurality of base subsequences; The complexity of the neural network configuration loaded into the base collaborator increases monotonically with N iterations, A non-transitory computer-readable storage medium, wherein labeled training data generated during an iteration of the N iterations is used to train a base colleague during an iteration immediately following the N iterations. B1a. Further training of the base chore repeatedly The computer-readable storage medium of clause B1, comprising initially training a base collaborator with exemplars comprising one or more oligonucleotide sequences, and generating labeled training data using the initially trained base collaborator. B2. The computer-readable storage medium of clause B1, wherein N1 iterations are performed before N2 iterations, and the second biosequence has a greater number of bases than the first biosequence. B3. Further training the base chorus for N1 repetitions is performed during one of the N1 repetitions. (i) inputting a first base partial sequence of the first plurality of base partial sequences of the first organism into a first cluster of the plurality of clusters of the flow cell, (ii) inputting a second base partial sequence of the first plurality of base partial sequences of the first organism into a second cluster of the plurality of clusters of the flow cell, and (iii) inputting a third base partial sequence of the first plurality of base partial sequences of the first organism into a third cluster of the plurality of clusters of the flow cell; (i) receiving a first sequence signal from a first cluster indicating a base partial sequence input to the first cluster, (ii) a second sequence signal from a second cluster indicating a base partial sequence input to the second cluster, and (iii) a third sequence signal from a third cluster indicating a base partial sequence input to the third cluster; (i) generating a first predicted base subsequence based on the first sequence signal, (ii) generating a second predicted base subsequence based on the second sequence signal, and (iii) generating a third predicted base subsequence based on the third sequence signal; (i) mapping the first predicted base subsequence to a first section of the first organism base sequence, and (ii) mapping the second predicted base subsequence to a second section of the first organism base sequence, while not mapping the third predicted base subsequence to any section of the first organism base sequence; and generating labeled training data including (i) a first predicted base subsequence mapped to a first section of a first biosequence, the first section of the first biosequence being a ground truth for the first predicted base subsequence, and (ii) a second predicted base subsequence mapped to a second section of the first biosequence, the second section of the first biosequence being a ground truth for the second predicted base subsequence. B3a. Further training of the base chorus for N1 repetitions is performed during one of the N1 repetitions. The computer-readable storage medium of clause B3, including training the base caller using the labeled training data generated during an initial training of the base caller prior to generating the first, second, and third predicted base subsequences. B4. The first predicted base subsequence has L1 bases; one or more of the L1 bases of the first predicted base subsequence do not match corresponding bases in the first section of the first biological base sequence due to an error in base calling prediction by the base caller; The computer-readable storage medium of clause B3. B5. The first predicted base subsequence has L1 bases, and the L1 bases of the first predicted base subsequence include the first L2 bases followed by the subsequent L3 bases, and mapping the first predicted base subsequence to a first section of the first organism base sequence is substantially and uniquely matching the first L2 bases of the first predicted sequence with the L2 consecutive bases of the first biological sequence; identifying a first section of the first biological sequence such that the first section (i) includes L2 consecutive bases as an initial base, and (ii) includes L1 bases; and mapping the first predicted base subsequence to the identified first section of the first organism base sequence. B6. substantially and uniquely matching the first L2 bases of the first predicted base sequence, while refraining from seeking to match the subsequent L3 bases of the first predicted base sequence with any bases of the first biological base sequence; B5 computer readable storage medium. B7. The computer-readable storage medium of B5, wherein the first L2 bases of the first predicted base sequence substantially match L2 consecutive bases of the first biological base sequence, whereby at least a threshold number of bases of the first L2 bases of the first predicted base sequence match L2 consecutive bases of the first biological base sequence. B8. The computer-readable storage medium of B5, wherein the first L2 bases of the first predicted base sequence uniquely match the L2 consecutive bases of the first biological base sequence, thereby substantially matching only the L2 consecutive bases of the first biological base sequence and not matching any other L2 consecutive bases of the first biological base sequence. B9. The third predicted base subsequence has L1 bases, and the third predicted base subsequence does not map to any of the base subsequences of the first plurality of base subsequences; (i) not substantially and uniquely matching the first L2 bases of the L1 bases of the third predicted base sequence with the first L2 consecutive bases of the first biological base sequence. B10. One iteration of the N1 iterations is a first iteration of the N1 iterations, and further training the base collaborator during a second iteration of the N1 iterations; training a base classifier using the labeled training data generated during a first iteration of the N1 iterations; using a base caller trained with the labeled training data generated during the first of the N1 iterations to generate (i) a first further predicted base subsequence based on the first sequence signal, (ii) a second further predicted base subsequence based on the second sequence signal, and (iii) a third further predicted base subsequence based on the third sequence signal; (i) mapping the first additional predicted base subsequence to a first section of the first organism base sequence, (ii) mapping the second additional predicted base subsequence to a second section of the first organism base sequence, and (iii) mapping the third additional predicted base subsequence to a third section of the first organism base sequence; and generating further labeled training data including: (i) a first further predicted base subsequence mapped to a first section of the first biosequence, wherein the first section of the first biosequence is a ground truth for the first further predicted base subsequence; (ii) a second further predicted base subsequence mapped to a second section of the first biosequence, wherein the second further section of the first biosequence is a ground truth for the second further predicted base subsequence; and (iii) a third further predicted base subsequence mapped to a third section of the first biosequence, wherein the third further section of the first biosequence is a ground truth for the third further predicted base subsequence. B11. Generating a first error between (i) a first predicted base subsequence generated during a first iteration of the N1 iterations and (ii) a first section of the first biological base sequence; (i) generating a second error between the additional first predicted base subsequence generated during the second iteration of the N1 iterations and (ii) the first section of the first biological base sequence; The second error is less than the first error because the base classifier is better trained during the second iteration compared to the first iteration. The computer-readable storage medium of clause B10. B12. The first, second, and third sequence signals generated during the first iteration are reused in a second iteration to generate a first additional predicted base subsequence, a second additional predicted base subsequence, and a third additional predicted base subsequence, respectively. The computer-readable storage medium of clause B10. B13. The neural network configuration of the base code is the same between the first iteration of N1 iterations and the second iteration of N1 iterations. The computer-readable storage medium of clause B10. B13a. The base class neural network configuration is reused for multiple iterations until a convergence condition is met. The computer-readable storage medium of clause B13. B14. The neural network configuration of the base collaborator during the first of the N1 iterations is different and more complex than the neural network configuration of the base collaborator during the second of the N1 iterations. The computer-readable storage medium of clause B10. B15. Further training the base caller for N1 iterations of the N iterations using a sample containing the first biological base sequence; further training the base collaborator using the first neural network configuration loaded into the base collaborator for a first subset of N1 iterations; and for a second subset of the N1 iterations, further training the base collaborator using a second neural network configuration loaded into the base collaborator, the second neural network configuration being different from the first neural network configuration. B16. The computer-readable storage medium of clause B15, wherein the second neural network configuration has a greater number of layers than the first neural network configuration. B17. The computer-readable storage medium of clause B15, wherein the second neural network configuration has a greater number of weights than the first neural network configuration. B18. The computer-readable storage medium of clause B15, wherein the second neural network configuration has a greater number of parameters than the first neural network configuration. B19. Further training of the base chore repeatedly For one or more of the N1 iterations using a sample including a first biological base sequence, loading a first neural network configuration into the base collaborator; and for one or more of the N2 iterations using a specimen containing a second biological base sequence, loading a second neural network configuration into the base collaborator, the second neural network configuration being different from the first neural network configuration. B20. The computer-readable storage medium of clause B19, wherein the second neural network configuration has a greater number of layers than the first neural network configuration. B21. The computer-readable storage medium of clause B19, wherein the second neural network configuration has a greater number of weights than the first neural network configuration. B22. The computer-readable storage medium of clause B19, wherein the second neural network configuration has a greater number of parameters than the first neural network configuration. B23. Further training the base caller for N1 iterations of the N iterations using a sample containing the first biological base sequence; The computer-readable storage medium of clause B1, including repeating further training using the first biological sequence until a convergence condition is met after N1 iterations. B24. The computer-readable storage medium of clause B23, wherein the convergence condition is met when a decrease in the generated error signal between two consecutive iterations of the N1 iterations is less than a threshold value. B25. The computer-readable storage medium of clause B23, wherein the convergence condition is met after completing N1 iterations.

[0344] 1. A computer-implemented method for progressively training a base classifier, comprising: (i) using a base caller to predict single oligo base call sequences for a population of single oligo unknown analytes (i.e., unknown target sequences) that have been sequenced to have known sequences of oligos; (ii) labeling each single oligo unknown analyte in the population of single oligo unknown analytes with a single oligo ground truth sequence that matches the known sequence; and (iii) starting with a single oligo training phase in which the labeled population of single oligo unknown analytes is used to train the base caller; (i) using the base caller to predict multi-oligo base call sequences for a population of multi-oligo unknown analytes that have been sequenced to have two or more known sequences of two or more oligos; (ii) sorting multi-oligo unknown analytes from the population of multi-oligo unknown analytes based on classification of the multi-oligo base call sequences of the sorted multi-oligo unknown analytes into known sequences; (iii) labeling each subset of the sorted multi-oligo unknown analytes with a respective multi-oligo ground truth sequence that matches a respective known sequence based on the classification; and (iv) further training the base caller using each labeled subset of the sorted multi-oligo unknown analytes; A computer-implemented method comprising: (i) using a base caller to predict organism-specific base call sequences for a population of organism-specific unknown analytes that have been sequenced to have one or more known subsequences of a reference sequence for the organism; (ii) sorting organism-specific unknown analytes from the population of organism-specific unknown analytes based on mapping the organism-specific base call sequences of the sorted organism-specific unknown analytes to sections of the reference sequence containing the known subsequences; (iii) labeling each subset of the sorted organism-specific unknown analytes with a respective organism-specific ground truth sequence that matches each known subsequence based on the mapping; and (iv) continuing with one or more organism-specific training phases to further train the base caller using each labeled subset of the sorted organism-specific unknown analytes. 2. The computer-implemented method of clause 1, further comprising performing multiple iterations of the single-oligo training stage before proceeding to the multi-oligo training stage until a convergence condition is met. 3. The computer-implemented method of clause 1, further comprising performing multiple iterations of each of the multi-oligo training stages before proceeding to the organism-specific training stage until a convergence condition is met. 4. The computer-implemented method of clause 3, wherein in each iteration of the subject multi-oligo training stage, selecte...

Claims

1. 1. A computer-implemented method for progressively training a base classifier, comprising: Iteratively initially training a base coder with a single oligonucleotide sequence exemplar having a single known oligonucleotide sequence, and generating labeled training data using the initially trained base coder, the base coder having an initial neural network configuration including several layers and parameters; (i) further training the base caller with an exemplar comprising a multi-oligobase sequence comprising at least two known oligobase sequences that differ from each other by a threshold edit distance that quantifies the number of nucleotide positions at which each nucleotide base in the at least two known oligobase sequences differs, and generating labeled training data using the further trained base caller; iteratively further training the base caller by repeating step (i) while, during at least one iteration, increasing the complexity of the initial neural network configuration of the base caller by increasing the number of layers and parameters relative to the initial neural network configuration to adjust the initial neural network configuration for the multi-oligobase sequence, wherein labeled training data generated during the iteration is used to train the base caller during an immediately subsequent iteration; 11. A computer-implemented method comprising:

2. During at least one iteration of further training the base call with the sample comprising a multi-oligonucleotide sequence, Increasing the number of unique oligonucleotide sequences of the multi-oligonucleotide sequence within the sample; further increasing the number of layers and parameters to further tune the initial neural network configuration for the increased number of unique oligobase sequences; The computer-implemented method of claim 1 further comprising:

3. first repeatedly training the base caller with the sample comprising the single oligonucleotide sequence; During the first iteration of the initial training with the base collaborator, Injecting the single oligonucleotide sequence into multiple clusters of a flow cell; generating a plurality of sequence signals corresponding to the plurality of clusters, each sequence signal of the plurality of sequence signals representing a base sequence loaded into a corresponding cluster of the plurality of clusters; predicting a corresponding base call for the single known oligonucleotide sequence based on each sequence signal of the plurality of sequence signals, thereby generating a plurality of predicted base calls; For each sequence signal of the plurality of sequence signals, generating a corresponding error signal based on a comparison of (i) a corresponding predicted base call and (ii) a base of the single known oligonucleotide sequence, thereby generating a plurality of error signals corresponding to the plurality of sequence signals; initially training the base classifier during the first iteration based on the plurality of error signals; 3. The computer-implemented method of claim 1, comprising:

4. first repeatedly training the base caller with the sample comprising the single oligonucleotide sequence; During a second iteration of the initial training with the repeated base collaborators that occurs after the first iteration of the initial training with the repeated base collaborators, predicting corresponding further base calls for the single known oligonucleotide sequence based on each sequence signal of the plurality of sequence signals using the base callers partially trained during the first iteration, thereby generating a plurality of further predicted base calls; For each sequence signal of the plurality of sequence signals, generating a corresponding further error signal based on a comparison of (i) a corresponding further predicted base call and (ii) a base of the single known oligonucleotide sequence, thereby generating a plurality of further error signals corresponding to the plurality of sequence signals; further initially training the base coder during the second iteration based on the plurality of further error signals; The computer-implemented method of claim 3 further comprising:

5. 5. The computer-implemented method of claim 4, wherein the plurality of sequence signals corresponding to the plurality of clusters generated during the first iteration of iteratively initially training the base collaborator are reused for the second iteration of iteratively initially training the base collaborator.

6. Comparing (i) the corresponding predicted base calls with (ii) the bases of the single known oligo sequence, 5. The computer-implemented method of claim 4, comprising, for a first predicted base call, (i) comparing a first base of the first predicted base call to a first base of the single known oligo sequence, and (ii) comparing a second base of the first predicted base call to a second base of the single known oligo sequence to generate a corresponding first error signal.

7. Repeating the base chore for further training; further training the base caller for N1 iterations using an exemplar comprising two known unique oligonucleotide sequences; further training the base caller for N2 iterations using an exemplar comprising three known unique oligonucleotide sequences; Including, The computer-implemented method of claim 1 , wherein the N1 iterations occur before the N2 iterations.

8. 1. A system comprising: at least one processor; a non-transitory computer-readable storage medium having stored thereon computer program instructions, the computer program instructions, when executed on the at least one processor, Iteratively initially training a base coder with a single oligonucleotide sequence exemplar having a single known oligonucleotide sequence, and generating labeled training data using the initially trained base coder, the base coder having an initial neural network configuration including several layers and parameters; (i) further training the base caller with an exemplar comprising a multi-oligobase sequence comprising at least two known oligobase sequences that differ from each other by a threshold edit distance that quantifies the number of nucleotide positions at which each nucleotide base in the at least two known oligobase sequences differs, and generating labeled training data using the further trained base caller; iteratively further training the base caller by repeating step (i) while, during at least one iteration, increasing the complexity of the initial neural network configuration of the base caller by increasing the number of layers and parameters relative to the initial neural network configuration to adjust the initial neural network configuration for the multi-oligobase sequence, wherein labeled training data generated during the iteration is used to train the base caller during an immediately subsequent iteration; A system including a non-transitory computer-readable storage medium that performs the actions, including:

9. During an initial iterative training of the base collaborator using the sample containing the single oligonucleotide sequence, a first neural network configuration is loaded into the base collaborator, and the base collaborator is further iteratively trained; further training the base caller for N1 iterations using an exemplar comprising two known unique oligonucleotide sequences, thereby (i) during a first subset of the N1 iterations, a second neural network configuration is loaded into the base core; 9. The system of claim 8, wherein (ii) during a second subset of the N1 iterations that occurs after the first subset of the N1 iterations, a third neural network configuration is loaded into the base core, and the first, second, and third neural network configurations are different from one another.

10. 10. The system of claim 9, wherein the second neural network configuration is more complex than the first neural network configuration, and the third neural network configuration is more complex than the second neural network configuration.

11. 11. The system of claim 9 or 10, wherein the second neural network configuration has a greater number of layers, a greater number of weights, or a greater number of parameters than the first neural network configuration.

12. 10. The system of claim 9, wherein the third neural network configuration has a greater number of layers, a greater number of weights, or a greater number of parameters than the second neural network configuration.

13. Further training the base caller for the N1 iterations with the exemplar comprising two known unique oligonucleotide sequences, during one of the N1 iterations; (i) dispensing a first known oligobase sequence of the two known unique oligobase sequences into a first plurality of clusters of a flow cell, and (ii) dispensing a second known oligobase sequence of the two known unique oligobase sequences into a second plurality of clusters of the flow cell; predicting a corresponding base call for each cluster of the first and second plurality of clusters, such that a plurality of predicted base calls is generated; (i) mapping a first predicted base call of the plurality of predicted base calls to the first known oligobase sequence, and (ii) mapping a second predicted base call of the plurality of predicted base calls to the second known oligobase sequence, while refraining from mapping a third predicted base call of the plurality of predicted base calls to either the first or second known oligobase sequence; (i) generating a first error signal based on comparing the first predicted base call to the first known oligobase sequence, and (ii) generating a second error signal based on comparing the second predicted base call to the second known oligobase sequence; further training the base caller based on the first and second error signals; The system of claim 9 , comprising:

14. mapping the first predicted base call to the first known oligobase sequence of the two known unique oligobase sequences; comparing each base of the first predicted base call with the corresponding base of the first and second known oligobase sequences; determining that the first predicted base call has similarity to the first known oligonucleotide sequence by at least a threshold number of bases and similarity to the second known oligonucleotide sequence by less than the threshold number of bases; mapping the first predicted base call to the first known oligobase sequence based on determining that the first predicted base call has similarity to the first known oligobase sequence by at least the threshold number of bases; The system of claim 13 , comprising:

15. refraining from mapping the third predicted base call to either the first or second known oligobase sequences; comparing each base of the first predicted base call with the corresponding base of the first and second known oligobase sequences; determining that the first predicted base call has similarity to each of the first and second known oligobase sequences by less than a threshold number of bases; refraining from mapping the third predicted base call to either the first or second known oligobase sequences based on determining that the first predicted base call has similarity with each of the first and second known oligobase sequences by less than the threshold number of bases; The system of claim 13 , comprising:

16. refraining from mapping the third predicted base call to either the first or second known oligobase sequences; comparing each base of the first predicted base call with the corresponding base of the first and second known oligobase sequences; determining that the first predicted base call has a similarity of more than a threshold number of bases with each of the first and second known oligonucleotide sequences; refraining from mapping the third predicted base call to either the first or second known oligobase sequences based on determining that the first predicted base call has similarity to each of the first and second known oligobase sequences by more than the threshold number of bases; The system of claim 13 , comprising:

17. generating labeled training data using the further trained base coder for the one iteration of the N1 iterations; after further training the base caller during the one iteration of the N1 iterations, re-predicting corresponding base calls such that an additional plurality of predicted base calls is generated for each cluster of the first and second plurality of clusters; (i) remapping a first subset of said additional plurality of predicted base calls to said first known oligobase sequences, and (ii) remapping a second subset of said additional plurality of predicted base calls to said second known oligobase sequences, while refraining from mapping a third subset of said additional plurality of predicted base calls to either said first or second known oligobase sequences; generating labeled training data based on the remapping, such that the labeled training data includes: (i) the first subset of the additional plurality of predicted base calls, where the first known oligobase sequence forms ground truth data for the first subset of the additional plurality of predicted base calls; and (ii) the second subset of the additional plurality of predicted base calls, where the second known oligobase sequence forms the ground truth data for the second subset of the additional plurality of predicted base calls; The system of claim 13 , comprising:

18. the labeled training data generated during the one iteration of the N1 iterations is used to train the base classifier during an immediately subsequent iteration of the N1 iterations; the neural network configuration of the base call remains the same during the one iteration of the N1 iterations and the immediately subsequent iteration of the N1 iterations; or the neural network configuration of the base caller during the immediately subsequent iteration of the N1 iterations is different and more complex than the neural network configuration of the base caller during the one iteration of the N1 iterations; 20. The system of claim 17.

19. 1. A non-transitory computer-readable storage medium having stored thereon computer program instructions for progressively training a base caller, the computer program instructions, when executed on a processor, comprising: Iteratively initially training a base coder with a single oligonucleotide sequence exemplar having a single known oligonucleotide sequence, and generating labeled training data using the initially trained base coder, the base coder having an initial neural network configuration including several layers and parameters; (i) further training the base caller with an exemplar comprising a multi-oligobase sequence comprising at least two known oligobase sequences that differ from each other by a threshold edit distance that quantifies the number of nucleotide positions at which each nucleotide base in the at least two known oligobase sequences differs, and generating labeled training data using the further trained base caller; iteratively further training the base caller by repeating step (i) while, during at least one iteration, increasing the complexity of the initial neural network configuration of the base caller by increasing the number of layers and parameters relative to the initial neural network configuration to adjust the initial neural network configuration for the multi-oligobase sequence, wherein labeled training data generated during the iteration is used to train the base caller during an immediately subsequent iteration; 10. A non-transitory computer-readable storage medium for performing actions, comprising:

20. performing actions that, when executed on the processor, include iteratively further training the base classifier; 20. The non-transitory computer-readable storage medium of claim 19, further storing computer program instructions for: responsive to increasing the number of layers and parameters relative to the initial neural network configuration to adjust the initial neural network configuration for the multi-oligobase sequence prior to further training the base coder with exemplars comprising the multi-oligobase sequence, further training the base coder using labeled training data generated during a final iteration of iteratively initially training the base coder.

Images