Nanorobot system for high-throughput single-cell DNA sequencing

Abstract

A nanoscale robotic system for single-cell DNA sequencing of a DNA strand placed on a glass slide, the system utilizes an atomic force microscope (AFM) whose end effector is in the form of a probe with a tip. The AFM causes its probe tip to scan the base pairs of the DNA strand. A pair of spaced electrodes at the tip contact the two ends of the DNA strand, and the current between the bases of the DNA strand is measured by a current measurement system connected to the electrodes. An artificial intelligence-based data analysis system determines the DNA sequence based on the current from the current measurement system. The AFM tip is guided to the DNA strand by comparing a compressed desired intensity local scan image with a compressed actual intensity local scan image and using the difference to control the position of the tip.

Description

Technical Field

[0001] The present invention relates to DNA sequencing, and in particular to high-throughput single-cell DNA sequencing using an atomic force microscope (AFM)-based nanorobotic system and artificial intelligence. Background Art

[0002] DNA sequencing has become a key technology in many fields, including biology, medicine, and anthropology [JS2017]. In particular, single-cell DNA sequencing is the basis for new diagnostic methods, personal medicine, and tissue engineering. The development of DNA sequencing technology plays a vital role in dissecting the basic mechanisms of biological processes. With the need to determine the molecular mechanisms that control cell behavior at the single-cell level, the development of high-throughput single-cell sequencing technology has shown that it is possible to obtain the nuclear information of the genome or transcriptome of thousands of single cells in a short time. This opens the black box of single-cell identification, processing, and association. For example, high-throughput single-cell sequencing is applied to embryonic development, stem cell biology, cancer, disease progression, etc. As a comprehensive record of subpopulations and cell state changes, high-throughput single-cell sequencing can provide a deep understanding of molecular mechanisms and the order of events. The application of high-throughput single-cell sequencing technology in stem cell biology has made great progress, especially in stem cell differentiation, cell regeneration, reprogramming, and embryonic development.

[0003] The development of DNA sequencing technology has gone through more than 40 years of history, including electrophoresis methods (first generation, such as Sanger sequencing method), massively parallel methods (second generation, such as Illumina's Hi-Seq genome sequencer), and ongoing real-time, long-reading methods (third generation, such as Oxford's nanopore-based sequencer). Each breakthrough in DNA sequencing technology has brought major technological leaps in the development of new diagnostics and treatments, and has had a significant social and economic impact.

[0004] The first and second generation DNA sequencing technologies are mainly based on the polymerase chain reaction (PCR) method to produce a large number of copies of target DNA fragments by chemical methods to achieve parallel identification of base pairs. The accuracy of base pair detection is over 99.9%. In recent years, derivatives of PCR-based DNA sequencing methods (such as double-stranded sequencing protocols) have been developed to further improve sequencing accuracy. Products currently on the market include the bottleneck sequencing system (BotSeqS) [ML2016] and nano-rate sequencing (NanoSeq) [FA2021].

[0005] However, PCR-based DNA sequencing methods are often costly, time-consuming, and require large numbers of samples [SJ2020]. In addition, they can only sequence short reads of a few hundred base pairs [PP2018, MT2019] and rely on post-processing to splice short sequences together. In addition, PCR sequencing is difficult to perform single-cell sequencing, which requires the extraction of available single cells from dead cells, which may result in the loss of key information of some rare cell subpopulations [TL2016][SW2019][LC2020].

[0006] To overcome these problems, the third generation of DNA sequencing technology is emerging. Among them, the most common representative technology is the nanopore-based method [PP2018, MT2019, RR2019]. They use natural or synthetic nanopores or gaps to let a single DNA chain pass through, while recording the changes in the electrical signal of the part inside the hole or gap for DNA base decoding. Therefore, long DNA chains such as chains directly from cells can be sequenced without PCR amplification. Compared with the first and second generation sequencing technologies, sample preparation is relatively simple and the sequencing cost is low. At present, the nanopore sequencing technology based on Oxford Nanopore on the market is based on biological nanopores, which utilize transmembrane proteins that have the natural ability to attract small charged particles through their ion channels [JS2017, MT2019, CG2016]. The biological nanopore is inserted into a synthetic polymer membrane immersed in an electrolyte solution. Since the membrane usually exhibits a fairly high resistance, ions pass through the nanopore only when a suitable potential is applied to the membrane. When a single DNA chain enters the nanopore, it causes a unique disruption in the ion flow, which is called a "change in conductance." This phenomenon is mainly due to the fact that the four bases adenine (A), thymine (T), cytosine (C) and guanine (G) show different electrostatic potentials [TO2006, LA2015, TO2012]. In addition, some researchers also believe that the conductivity of different bases is related to their geometric properties, rather than their electronic properties [XG2006].

[0007] Biological nanopores usually have natural and fixed geometric dimensions and unstable morphology. They are sensitive to the environment and difficult to engineer / integrate into devices. Over the years, technologies based on solid-state nanopores / nanogaps [PP2018] have also been considered and attempted. They have high robustness, low sensitivity to environmental conditions, and good pore size adjustment capabilities and integration capabilities in arrays [JL2001, FT2013]. However, due to the limitations of manufacturing technology, it is still difficult to manufacture solid-state nanopores with market-acceptable costs and benefits.

[0008] Biological nanopore sequencing technology generally faces several major challenges that lead to measurement errors: First, the speed at which the DNA chain passes through the nanopore is relatively high (about 0.5 cm / s) [PP2018] and cannot be controlled. This often leads to missing bases and measurement noise. In addition, the DNA is in a liquid when the measurement is performed. This makes electrical measurements more difficult. In addition, the DNA chain can only pass through the nanopore once, and the measurement cannot be repeated. Therefore, statistical and artificial intelligence-based methods cannot be used to improve measurement accuracy.

[0009] In addition, methods based on scanning probe microscopy (SPM) have also been applied to DNA sequencing. However, due to the limitations of the precision of SPM motion control, it is difficult to accurately position the probe on the DNA base. Motion control requires sub-nanometer precision, which was impossible to achieve in the past. In addition, the limitations of available nanofabrication technologies make it impossible to manufacture the nanostructures required for DNA sequencing on the SPM probe.

[0010] The current inventor and his team have been working on the development and implementation of atomic force microscopy (AFM)-based nanorobots for over 20 years [XN2011]. This includes the development of basic technologies such as online visualization [XN2013], improving nanomanipulation efficiency [LG2014, LG2015], increasing scanning speed [SB2014], and improving AFM positioning accuracy in the task space [XN2013]. In addition, AFM-based nanorobots have been successfully applied to the fabrication of CNT-based infrared sensors [FC2009, CH2010] and biomedical research [YR2015].

[0011] However, the motion control of AFMs currently available on the market relies on position sensors. Such systems cannot directly measure the tip position at the sub-nanometer level due to system drift and probe deformation. In addition, the visual tracking methods required to obtain the position introduce errors in the vector space. Therefore, features need to be extracted from the image to determine and correct the errors. This process is time-consuming if the features of the image are not clear.

[0012] Although featureless methods have been proposed, two key issues remain. First, since the problem is still formulated in vector space, these methods require a complex calibration process that is extremely difficult to perform when high precision is required, such as nanometer-scale motion control. Second, all featureless methods require successful feedback control of the entire image, which is time-consuming for systems with slow sensor feedback, such as AFM, which can only acquire images by scanning pixel by pixel. Therefore, the challenge is to develop new AFM image-based motion control methods to achieve sub-nanometer motion accuracy. Summary of the invention

[0013] The nanorobot system of the present invention overcomes the difficulties of the third generation DNA sequencing technology based on the latest breakthroughs in nanorobot motion control, nanomanufacturing and artificial intelligence-based data analysis. Therefore, the robot system of the present invention can achieve sub-nanometer position measurement without the need for a liquid environment, controllable measurement speed and multiple measurements. Therefore, the measurement accuracy is significantly improved and the sequencing cost is reduced by several percent. Single-cell sequencing is faster and more convenient.

[0014] The system of the present invention includes a nanorobot manipulator based on an atomic force microscope (AFM), a high-precision tunnel current measurement system, a nanorobot end effector, and a data analysis system based on artificial intelligence (AI). The movement of the robot is controlled by a new semantic compression feedback control system, which can achieve sub-nanometer motion tracking accuracy. The nanorobot manipulator is equipped with an end effector that can measure the current between two bases. The single-stranded DNA molecule is fixed on a glass slide or a substrate (such as mica), and the nanorobot moves the end effector along the DNA molecule and measures the current between the two base contact points. The measurement data is analyzed by a new Generative Adversarial Tri-model (GAT) machine learning scheme to obtain the DNA sequence. The measurement can be repeated multiple times until sufficient measurement data is obtained to achieve the desired confidence level of the result.

[0015] The newly developed semantic compressed motion control overcomes the main problems of traditional image-based control by eliminating feature extraction, system calibration for establishing coordinate transformations, and the high computational requirements of image processing and pattern recognition. In addition, the system uses a non-vector space control approach. Considering the local scanned images as a set, the motion control problem is formulated in the space of the local scanned image set, rather than in the traditional vector space. In this way, all the intensity information in the image can be used for control. Therefore, feature extraction is not required. Since the problem is formulated in the set space, no complex calibration process is required. Secondly, non-vector space control can also be used with compressed feedback (i.e., feedback of compressed information of the compressed image, for example). Therefore, both sampling time and computation time can be significantly reduced, and the system feedback speed is faster, which is crucial for high-precision motion control. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The patent or application file contains at least one drawing executed in color.

[0017] The above and other objects and advantages of the present invention will become more apparent when considered in conjunction with the following detailed description and accompanying drawings, in which like reference numerals represent similar elements in the different views, and in which:

[0018] Figure 1A nanorobot system for DNA sequencing based on atomic force microscopy (AFM) according to the present invention;

[0019] Figure 2 is a local scanning control based on non-vector space motion control according to the present invention;

[0020] Figure 3 It is a non-vector space motion control based on compressed local scanned images;

[0021] Figure 4 is an embodiment of an AFM-based nanorobotic system with semantic compression feedback according to the present invention;

[0022] Figure 5 Explains the use of Figure 4 The experimental results of the embodiment of the nanorobot end effector tracking DNA chain, wherein Figure 5 (a) shows the robot end effector on a DNA strand. Figure 5 (b) shows the DNA strand tracking error;

[0023] Fig. 6A illustrates the structure of the AFM cantilever, Figure 6B Target parameters for sub-10 nm separation electrodes and AFM probes are described;

[0024] Figure 7 is a series of images showing the sequence of steps to fabricate a sub-10 nm gap dual-tip AFM probe prototype.

[0025] Figures 8A-8G Shows Figure 7 An image of the preliminary results of the fabrication, where Fig. 8A shows the welding of carbon nanotubes on an AFM cantilever tip using EBID, Figure 8B shows the formation of fingers on a polystyrene dimer, Figure 8C shows high aspect ratio (>20) nanowires in a 4×4 array, Fig.8D The invention describes a FIB for fabricating a substrate array with a diameter of 100 nm on a Si+SiO2 substrate coated with a gold film, with a spacing of 500 nm between the substrate arrays and on-chip leads with a width of 50 nm. Fig. 8E Shows: Using Figure 7 The process shown successfully achieves a planar AFM cantilever tip using focused ion beam chemical vapor deposition (FIBCVD) steps 1-3 and dicing, Fig.8F Nanotubes bridging between planes are shown and sectioned using the TEM's electron beam. Figure 8G It shows that an initial gap of 7.6nm has been achieved;

[0026] Fig. 9It is a schematic diagram of the design of a dual-tip AFM probe with a sub-10 nm gap;

[0027] Fig.10 The steps of the fabrication process for a sub-10 nm gap dual-tip AFM probe are described.

[0028] Fig.11A is an image of a pentamer disk optical antenna array, Fig. 11B yes Fig.11A Magnified image with gap dimensions, Fig. 11C Here is a further zoomed-in view of the disk: Fig.11D A specially engineered optical antenna with a sub-10nm gap is shown. Fig.11E A probe chip having an optical antenna thereon is described, Fig.11F A series of images illustrating laser stamping are shown;

[0029] Fig.12 is a series of schematic diagrams of the design of a sub-10 nm gap dual-tip AFM probe based on strain engineering;

[0030] Fig.13 A series of steps for the fabrication process of a sub-10 nm gap dual-tip AFM probe based on strain engineering are shown.

[0031] FIG. 14A to FIG. 14I The curly design of the helical structure is shown, Figures 14J to 14K The bistable curled claw that can be used as a probe is shown;

[0032] Fig.15 It is the manufacturing of nanorobot end effectors;

[0033] Fig.16 is an actual measurement scenario using nanorobots;

[0034] Fig.17 It is the equivalent battery chain model of the DNA chain;

[0035] Fig.18 is the point contact approximation between the electrode and the DNA strand;

[0036] Fig.19 It is the connection of the T nucleobases at the beginning and end of the DNA chain;

[0037] Fig. 20 It is an ideal equation for DNA sequence based on measured values;

[0038] Fig.21 is a schematic diagram of the nanorobot's motion error;

[0039] Fig. 22 is the actual measurement matrix with β noise;

[0040] Fig.23 is the actual equation for the DNA sequence based on noisy measurements;

[0041] Fig.24 It is a neural network structure used to solve overdetermined equations;

[0042] Fig.25 It is a flowchart of the DNA iterative sequencing framework;

[0043] Fig.26 This is the DNA sequencing accuracy graph under pure Gaussian contact noise;

[0044] Fig. 27 This is a plot of DNA sequencing accuracy under pure beta motion error from the nanorobot motion;

[0045] Fig.28 is a plot of DNA sequencing accuracy under joint contact noise and motion error;

[0046] Fig.29 is a schematic diagram of the Generative Adversarial Triad (GAT) model;

[0047] Fig.30 A schematic diagram of the neural network structure used in the GAT model (left) and an explanation of the method of reproducing arbitrary functions (right); and

[0048] Fig.31 is a schematic diagram of the GAT model used for noise calibration. DETAILED DESCRIPTION

[0049] like Figure 1 As shown, the system of the present invention includes a nanorobot scanner or manipulator 12 based on an atomic force microscope (AFM), a nanorobot end effector 14 located above the AFM platform and moved by it, a high-precision tunnel current measurement system 16, and a data analysis system 18 based on artificial intelligence. The AFM system includes an atomic force microscope that provides various types of surface measurements that meet the needs of scientists and engineers. It is powerful because the AFM can generate atomic resolution images with angstrom-level resolution height information with minimal sample preparation. The AFM uses an end effector in the form of a cantilever with a very sharp tip 17 to scan the sample surface. When the tip approaches the surface, the close-range attraction between the surface and the tip causes the cantilever to deflect toward the surface. However, when the cantilever is closer to the surface, so that the tip contacts it, the increasingly large repulsive force plays a major role and causes the cantilever to deflect away from the surface.

[0050] The laser beam emitted by the laser 13 is used to detect the deflection of the cantilever towards or away from the surface. By reflecting the incident beam off the cantilever flat top, any cantilever deflection will cause a slight change in the direction of the reflected beam. The reflected beam can be received by the position sensitive photodiode (PSPD) 15, which can be used to track changes in the cantilever deflection (Z axis). Therefore, if the AFM tip passes over a raised surface feature, the resulting cantilever deflection (and subsequent change in the reflected beam direction) can be recorded by the PSPD.

[0051] The AFM images the topography of the sample surface by scanning the cantilever over the region of interest (X and Y axis). The lifting features of the sample surface affect the deflection of the cantilever, which is monitored by the PSPD. By using a feedback loop through the detector electronics 11 and the controller electronics 19, the height of the tip above the surface can be controlled - thereby maintaining a constant laser position - the AFM can generate accurate topographic maps of surface features.

[0052] In the present invention, the robot motion control is a new type of semantic compression feedback control system. Semantic compression is a process by which the representation of information (e.g., sensory data) is compressed / simplified, i.e., the amount of data representing the information is reduced based on the semantic content (i.e., the meaning of the information). Semantic feedback control is to feed back the semantically compressed sensory data instead of the original sensory data.

[0053] The semantic compression feedback control of the present invention not only has sub-nanometer motion tracking accuracy in the vertical direction, but also can scan the DNA sample chain 20 on the microscope sample platform 21. The end effector 16 can measure the current between the bases of the DNA chain. The single-stranded DNA molecule 20 is fixed on a substrate of the sample platform, such as mica. The nanorobot or scanner 12 moves the end effector along the DNA molecule and measures the current between two contact points on both sides of the DNA molecule. The measurement data is analyzed by the new generative adversarial three-model (GAT) machine learning scheme of the analysis system 18 to obtain the DNA sequence.

[0054] It is important to note that the measurement does not need to be performed in a liquid environment, and the movement speed of the nanorobot manipulator can be optimized and controlled to achieve reliable current measurement. More importantly, the measurement can be repeated multiple times until sufficient measurement data is obtained to achieve the required confidence level in the measurement result.

[0055] Positioning of the end effector along the DNA strand is achieved by the newly developed semantically compressed motion control of the present invention. This motion control system has successfully overcome the main problems of conventional image-based control by eliminating feature extraction, system calibration for establishing coordinate transformations, and the high computational requirements of image processing and pattern recognition.

[0056] In order to solve the shortcomings of the prior art, the present invention relies on a newly developed non-vector space control method, such as Figure 2 As shown. Considering the local scanned images as a set, the motion control problem can be expressed in the space of the local scanned image set instead of the traditional vector space. In this way, all the intensity information in the image can be used for control. Therefore, feature extraction is not required. Since the problem is expressed in the set space, a complex calibration process is not required. Secondly, non-vector space control can also be used with compressed feedback (i.e., feedback of compressed information of compressed images, for example). Therefore, both sampling and calculation time can be significantly reduced. Therefore, the system has a faster feedback rate, which is crucial for high-precision motion control.

[0057] The overall schematic diagram of non-vector space control under compressed feedback is as follows: Figure 3 As shown. Unlike the full local scan image feedback, both the target and the feedback of the local scan image are compressed. Sensing in compressed feedback is a new signal acquisition paradigm. Instead of sampling the signal directly, it samples the signal in a compressed form. The system then relies on advanced optimization algorithms or greedy algorithms to recover the original signal [XN2007]. During the AFM scanning process, a compressed scanning mode is used instead of the traditional zigzag scanning to obtain a compressed local scan image, which has much less data and is much faster than the full scan [XN2013][SB2014], [LG2014][LG2015]. In addition, in the local scan image, prior knowledge of the image content is usually available, such as the DNA chain to be measured. Semantic compressed sensing has been developed to exploit the knowledge of the content to improve sampling [LC2017][LC2018].

[0058] Non-vector space control methods based on compressed feedback such as Figure 3As shown. The term "non-vector space" means that control is not performed in the traditional vector space but in the space of sets. Zhao et al., "Compressive feedback based non-vector space control," American Control Conference (June 2012). Considering an initial set of compressed actual local scanned images 30 and a target or desired set 32, a stable non-vector space controller 34 is designed to make the set dynamics converge to the target set. The difference between the actual compressed image 30 and the desired or target image 32 is determined in a subtractor or comparator 33. This difference signal is applied to the controller 34, which in turn drives the AFM-based robot 35 so that the actual local image is close to the target image due to the feedback loop. Due to the compressed feedback, the controller can work even if only some elements of the feedback set are available; that is, the same controller can still stabilize the set dynamics around the target set using compressed feedback. By treating the image as a set, the controller is applied to visual servoing (vision-based robot control). This eliminates the need for image processing for feature extraction, which is an essential process in traditional servo methods. In addition, compressed feedback can reduce the size of the feedback image. This is important when the sampling process is time-consuming, such as using atomic force microscopy (AFM) imaging. The visual servo characterization of the controller is further applied to motion control in nanomanipulation. Simulation results show that the controller has good performance. The proposed framework can be extended to other systems where the signal can be represented as a set.

[0059] Inspired by recent advances in neuroscience, in the primate brain, object recognition is represented in the inferior temporal cortex and is invariant to different transformations that do not affect object recognition [BS2004]. Recently, Chang et al. [CL2017] proposed an axial model to study how primates encode and decode facial features. They found that in the primate brain, human faces are represented by very simple neural codes that correspond to facial content along specific axes in face space. Facial images can be linearly restored by the responses of about 200 cells, with different cells carrying complementary facial information for reconstruction.

[0060] Based on the axis model in [CL2017], in the brains of primates, objects of the same kind have similar or even identical features, and the images of objects are linearly expressed through many content-related axes. For example, faces of different people have a common visual structure, namely eyes, mouths, and noses, and their relative positions are similar. Therefore, images can be reasonably represented by some common high-level parts [MN2015]. This is the basis of semantic compression.

[0061] For local images in AFM motion, there is a certain common semantic content. This fact leads to a low-rank representation of the image [LC2017], and therefore, the common semantic content is extracted and emphasized during the compressive sampling process. The semantic compressive scanning method utilizes the sensing matrix learned from a set of DNA chain images. The sensing matrix training process can be performed by various methods, such as the popular deep learning strategy [LC2017]. In order to achieve an efficient low-rank representation of the local scanning AFM images of DNA chains, the non-negative matrix factorization (NMF) [LC2018] representation is used to learn the semantic content of the AFM images of DNA chains.

[0062] By applying semantic compression feedback, high-speed feedback control using AFM local scan images is achieved. It enables the nanorobot end effector to achieve high-precision tracking of DNA strands for measurement. The inventors implemented and experimentally tested semantic compression feedback control in an AFM-based nanorobot system, such as Figure 4 As shown. Figure 5 The experimental results shown in the paper demonstrate that the DNA chain can be tracked with sub-nanometer precision, which lays the foundation for using nanorobot systems for DNA sequencing.

[0063] Figure 4 (a) and Figure 4 (b) is related to Figure 3 Equivalent hardware and block diagrams / images. Figure 3As shown, in the prior art, the positioning of the tip along the DNA chain is done in vector space, that is, based on the position of the tip in X, Y, and Z coordinates. However, the distance required by the present invention is too small to be effective. On the contrary, the present invention adopts a new technology using non-vector space. Through this new control process, only an image of the local area close to the tip is formed. Therefore, the amount of data is greatly reduced compared with the system using the entire image. The local image is then semantically compressed to further reduce the amount of data. The image required for semantic compression of the local part of the chain is also formed from repeated previous scans. The actual local compressed image is then compared with the expected or ideal local compressed image, and the difference is the control error signal fed back to the AFM to control the tip to move to the expected position represented by the compressed ideal local image. By using image comparison, there is no need to use vector space coordinates. By using local images, the amount of data and the time required to process it are greatly reduced. In addition, the amount of data can be further reduced by using semantic compression on local images (including actual local images and expected local images).

[0064] exist Figure 4 In the arrangement of (a), the AFM includes a probe 43, a microscope platform 44, its controller 49 and a processor 48. A signal access module (SAM) 46 provides access to signals in the AFM controller 49, such as voltages on actuators, laser deflection signals from photodetectors, etc., and also allows commands to be sent to the controller 49 and the microscope platform 44. Device 45 is an I / O or interface board with a digital-to-analog converter. The tip or probe motion control system includes a control system 40, a tactile joystick 41 and a program stored in a nanorobot PC 42. System 40 is the software interface of the system, which has a display for images, system information and the ability to diagnose problems. As shown by the Ethernet lines, these processors can share information and work in parallel with each other. It is also possible to run the functions of PC 42 in PC 48, which is a processor provided by the AFM vendor.

[0065] The required (target) local scan image is created from the AFM local images of DNA obtained in the previous scan. These images are stored in the nanorobot PC 42. The images are transmitted from the I / O device 45 to the control system 40 through the nanorobot PC 42.

[0066] When the probe 43 moves on the DNA strand on the substrate under the control of the joystick 41 or by autonomous scanning, the AFM 44 takes actual local images. These images can be accessed by the nanorobot PC42 through the SAM46 and the D / A converter 45. In this processor, the images are semantically compressed. The nanorobot PC42 and the AFM PC48 compare the semantically compressed expected image previously stored in the processor 42 with the actual local compressed image. The processor 42 sends the difference signal back to the controller 49 (which is a non-vector space controller 34') through the converter 45 and the SAM 46 to correct the motion error.

[0067] The control system 40 and / or the joystick 41 constitute Figure 4 (b) of the motion planner 32′. The ideal image 31′ is compressed in the planner 32′ to form a compressed image 37′. Its local part 38′ is passed to the comparator 33′. The actual local compressed image comes from the AFM 35′, and its compressed local part 39′ is also passed to the comparator 33′. The result of the comparison (error) 36′ is communicated to the non-vector space controller 34". The output of the controller 34′ is applied to the motion control of the AFM 35′ to change the position of the probe in a feedback loop that stabilizes its position along the DNA chain.

[0068] To carry out the present invention, it is necessary to fabricate an AFM cantilever with dual cylindrical nanoelectrodes on its cone / cone tip, e.g. Fig. 6A and Figure 6B The cylindrical metal tips are specified to have sub-10 nm spacing, a diameter of 50 nm or less, and a height of 45 nm. These tips were achieved by developing and optimizing three technologies: (1) prototyping using focused ion beam chemical vapor deposition (FIB-CVD), electron beam induced deposition (EBID), and electron beam cutting using the beam of a transmission electron microscope (TEM); (2) batch manufacturing using electron beam lithography (EBL), laser stamping, and chemical etching; and (3) an alternative approach and risk mitigation program for batch manufacturing called "prestressed thin film strain engineering."

[0069] In-situ prototype of a sub-10 nm gap dual-tip AFM cantilever based on metal layer coating, FIB chemical vapor deposition (FIB-CVD), EBID, and electron beam cutting inside a TEM Figure 7 The process is as follows:

[0070] Step 1: First, a metal (gold) layer is pre-coated on both sides (top and bottom) of a commercial AFM probe, and the cantilever and its supporting base are cut using FIB to form microelectrodes and leads with widths of approximately 2 μm to 5 μm and 20 μm to 50 μm, respectively. In the case of an AFM cantilever without metal coating, the Au layer is sputtered on both sides.

[0071] Step 2 The cone or conical tip may or may not have a pre-coated metal layer. In the former case, the pre-coating layer will be separated into two parts using FIB, whereas the electrode will be directly written on the side surface using FIBCVD, with a line width of about 1 μm and a thickness of about 50 to 100 nm.

[0072] Step 3: Use FIB to grind away the cone or conical tip to form two small facets with side lengths of 1 μm to 3 μm. Figure 7 As shown, a hollow cone or pyramid is preferred. This leaves room for the passage of electrons, as described in step 6.

[0073] Step 4 FIBCVD or EBID is applied to form a nanotip from the flat surface generated in the previous step. The precursor is Pt for FIBCVD or CpPtMe3 for EBID. The formation direction is 30° with both sides tilted inward. The diameter of the cylindrical nanotip is between 30 and 50 nm and the length is between 1 and 3 μm.

[0074] Step 5 Using the same method as Step 4, form a lateral bridge on top of the cylindrical tip.

[0075] Step 6 then uses a focused high-energy (200keV to 300keV) electron beam to generate a sub-10 nm gap between the cylindrical nanotips, i.e., by cutting the bridge into two parts at the center. To achieve the high-precision sub-10 nm gap, a confocal scanning transmission electron microscope (CsTEM) was used. Attempts on nanowires of similar size showed that the key to success was the control of internal stress, which can be caused by the fabrication process of steps 4 and 5 or the electrode bombardment of step 6.

[0076] Although other steps involving FIB and EBID have been described by one of the inventors and his team, Figures 8A to 8C Step 6 is the main challenge in creating this structure, as shown in [DL2002][DL2004][DL2007] and has been used routinely for 20 years by other research laboratories around the world [MS2000][KR2004]. Figure 7 The FIBCVD and cutting for the first three steps shown, and Fig.8D The prefabricated AFM cantilever with a flat tip shown in Figure 6 was successfully implemented. The carbon nanotubes were then bridged between the planes and cut using the electron beam of the TEM ( Fig.8F). An initial gap of 7.6 nm was achieved, which is encouraging and shows the feasibility of the proposed step 6. However, the gap could not be maintained within 10 nm and the gap increased (29.4 nm) after 5 min. This is due to the internal stress caused by electron beam bombardment during the assembly process and cutting. To avoid this, an annealing process was added between step 5 and step 6. Since it is caused by electron beam bombardment, the physics behind the expansion is complicated. To avoid the charging effect, the electrodes were shortened before step 6 or cut off after step 6 instead of in step 3. In any case, the feasibility and scalability of this method are no worse than the nanopore-based method that has been successfully marketed in 2015 [GS2010][HS2016][WA2018].

[0077] It has been noted that this prototyping process is a time-consuming process, typically requiring 5-7 hours for a single cantilever. To shorten the time, the bridge (step 5) is made as thin as possible. For proof of concept, demonstration of new designs or new parameters, or small batch manufacturing, this method is an effective and efficient way to form cantilevers simply and flexibly. In addition, electrode direct writing based on FIBCVD is also an irreplaceable process based on the EBL method.

[0078] Moderate-volume fabrication of dual-tip AFM cantilevers with sub-10 nm gaps can be achieved using EBL patterning and laser stamping gap reduction. Fig. 9 As shown in the figure, the design of the sub-10 nm gap dual-tip AFM cantilever is different from the conventional one in two aspects: two conductive nanotips with diameters of about 50 nm and a very challenging requirement for the spacing (sub-10 nm). To achieve it, the conventional AFM probe fabrication process is combined with two specially designed processes: (1) fabrication of 50 nm cylindrical tips and (2) reduction of the gap. In addition, the FIBCVD-based method is suitable for nanoelectrode writing. Fig.10 As shown, the preparation process is as follows:

[0079] Step 1: Insulation layer preparation. Plasma enhanced chemical vapor deposition (PECVD) is used to form a 50nm thick silicon dioxide (SiO2) film. Since silicon wafers are doped semiconductors, making electrodes and antennas directly on silicon wafers may cause the device to short-circuit, so an insulating layer is introduced to isolate the electrodes and conductive tips from the Si substrate. Since SiO2 has good corrosion resistance, it is also used as a mask material for wet etching.

[0080] Step 2-4 Tip Fabrication. Several 50 nm crossed cylindrical tips with a minimum gap of about 15 nm were prepared by using an EBL system to pattern the tips on a spin-on photoresist. The pattern was then developed using 45 nm thick Au electron beam evaporation.

[0081] Step 5 uses laser punching to compress the small tip gap. This is a new design technology (such as Fig.11E and 11F ), used to achieve a smaller gap between cylindrical tips by pressing the prepared metal disk from the top using a ns YAG laser. To achieve this, three layers of materials were used, including a 4 μm thick aluminum layer as a compressor, a graphite layer as an ablative coating, and a CaF2 / ZeSe glass layer for transparent confinement.

[0082] Step 6-10 Fabrication of support base and cantilever. A pre-designed patterned photolithography mask is used with UV irradiation for top and bottom side patterning of the profile of the support base and cantilever, which is achieved using spin coating of photoresist, development, and dry etching of the SiO2 layer using inductively coupled plasma (ICP). The resulting window of Si wet etching is achieved by using a KOH solution. A tapered tip is formed after etching, while the cantilever maintains a thickness greater than the final size.

[0083] Steps 11-15 fabricate extraction electrodes on the support base and microelectrodes on the cantilever. A 50 nm thick SiO2 layer is formed on the top surface using PECVD as an insulating layer for fabricating extraction electrodes on the support base and microelectrodes on the cantilever. Then, ICP dry etching based on ultraviolet irradiation is used and developed on a spin-coated photoresist layer to selectively remove SiO2 on the tip top. Then, a 45 nm thick gold film is deposited as an electrode using an electron beam evaporator.

[0084] Step 16 uses FIBCVD to connect the nano tip to the micro electrode on the cantilever. FIBCVD is used to directly write the nano electrode that connects the micro electrode and the tip on the cantilever. The nano electrode follows a path along the side and top surfaces of the cone to achieve a three-dimensional electrical connection.

[0085] Steps 17-18 End of cantilever etching. To complete the probe fabrication, the SiO2 is etched away using buffered oxide etch (BOE), followed by Si etching using KOH. Note that the sacrificial layers on both the front and back sides are etched away from both the top and bottom sides, and therefore faster than areas of the cantilever that are only etched from the bottom side.

[0086] Typically, 375 probes can be patterned from a single 4” wafer, where individual 3.4mm x 1.6mm probes are used to support the base and can be several hundred microns in length. The probes are then tested for conductivity / insulation before being used for DNA sequencing.

[0087] One of the inventors of this patent and his team have been working on plasmonic devices such as optical antennas for more than 10 years. This work demonstrated spheres on pillars [CX2010a], matched [CX2010b] and pentamer optical antennas and arrays thereof [HC2018]. Fig.11A The spacing between the disks (shown) is key to improving the performance of these devices. In this work, an average gap of ca 15 nm ( Fig. 11B ), which is also the best solution obtained from commercially available equipment. Sometimes, sub-10 nm gaps are achieved by tilting the sample stage or performing shadow lithography using disk pillars as masks on each other ( Fig. 11C ). However, this approach still lacks full controllability. Currently, the GRF project is supporting the research of "Plasmonic Nanosensors and Antennas for Signal Transduction and Energy Harvesting in Biomedical Microrobots" (PI: Lixin Dong, Period: 01Jan 21-31Dec 23). One of the goals of this project is the design, optimization and fabrication of molecular plasmonic detectors. Fig.11D A cantilever chip with an optical antenna integrated on it is shown.

[0088] A newly developed process called "laser stamping / shock" [GH2014][HY2016] was used to try to reduce the prefabricated gap between the nanodisk tips. In this project, laser stamping was used to further reduce the gap to sub-10 nm, based on the metal nanodisks with an initial gap of 15 to 20 nm, which is typically obtained from EBL and electron beam irradiation. Previously reported nanodisks have demonstrated the feasibility of a 5 nm gap ( Fig.11E and 11F ). In addition, the laser stamping technique can achieve gap reduction in a batch manufacturing manner, which is very compatible with other batch manufacturing processes of AFM tips.

[0089] The key bottleneck for the scalability of the second approach (i.e., EBL-based patterning, laser stamping-based nanogap reduction, and chemical etching) is the single-beam nature of the process, i.e., electron beam or focused ion beam. Although the remaining processes are scalable compared to the mainstream technologies in the semiconductor industry's existing technologies, the high-energy beam-based technologies still face huge challenges shared by the entire semiconductor industry. Although AFM cantilever manufacturing is not cost-effective, extreme ultraviolet lithography (EUV) has become an excellent alternative for mobile phone chip manufacturing. With the maturity of 2 to 7nm processes, EUV processes are very promising in the application of preparing AFM tips with sub-10nm gaps.

[0090] Large-scale batch fabrication of sub-10 nm gap dual-tip AFM cantilevers using strain engineering. Fig.12An alternative design for high-volume manufacturing is shown, in which a three-layer metal-insulator-metal (MIM) sandwich is patterned on a prestressed film on a substrate plane and curled after etching to form a vertical tip at the end of the cantilever beam. In this design, the thickness of the insulator layer determines the spacing of the metal layers, i.e., the gap between the tips.

[0091] To combine the nanogap electrode structure with the AFM probe, the design relies on basic semiconductor manufacturing processes that utilize standard AFM probe manufacturing processes and combine them with a 3D self-curling manufacturing process. The design uses a cross section at the end of the self-curling 3D structure as a tip for DNA sequencing. The nanogap on the cross section of the curled structure relies primarily on atomic layer deposition (ALD) technology, which can be precisely controlled to the sub-nanometer level. In addition, the design takes full account of compatibility with established AFM probe manufacturing processes, enabling reliable mass production.

[0092] This method ( Fig.13 ) is manufactured as follows:

[0093] Step 1: Insulating Layer Vapor Deposition SiO2 vapor deposition is performed on both sides of the silicon wafer using PECVD.

[0094] Step 2: Evaporation of sacrificial layer from the rolled mesa. A 20nm thick aluminum film layer is used as the sacrificial material. The aluminum layer is deposited and then patterned under ultraviolet irradiation. The patterned film is evaporated with an electron beam and then subjected to a lift-off process.

[0095] Steps 3 and 4 are the evaporation and patterning of the self-curling mesa layer. The self-curling layer is Si with intrinsic stress obtained by PECVD. x N y The film has a thickness of 30nm. Then, the irradiation mask is designed using UV lithography, and the mesa is fabricated using ICP etching technology.

[0096] Step 5 Bottom electrode fabrication: A 50 nm thick gold bottom electrode was patterned by UV irradiation and formed by electron beam evaporation.

[0097] Step 6: Evaporation of the nano-gap insulating layer. A 1 nm Al2O3 film was obtained using ALD deposition technology.

[0098] Step 7: Nanogap insulating layer patterning. Using photoresist as a mask, ICP etching is used to retain the Al2O3 film above the bottom electrode.

[0099] Step 8 Top electrode fabrication. UV irradiation was used for top electrode patterning, and electron beam vapor deposition and lift-off techniques were used to obtain a 50 nm Au film.

[0100] Step 9: Backside UV Lithography

[0101] Step 9: Backside UV lithography technology: In order to pre-etch the backside Si, photoresist is needed to produce an etching window pattern, so the backside jacket method is used to obtain the etching window pattern.

[0102] Step 10 Etching window fabrication: ICP etching of SiO2 is used to expose the etching material Si.

[0103] Step 11: Photoresist removal: Use acetone to remove the etch window photoresist.

[0104] Step 12: Anti-corrosion protective layer vapor deposition. Since KOH solution can also be used as an etchant for the sacrificial layer Al in the self-curling structure, it is necessary to evaporate the anti-etching protective layer on the substrate surface. PECVD is used to evaporate a 130nm SiO2 protective layer.

[0105] Step 13: Etching of the bottom Si substrate. The substrate is placed in a KOH solution at room temperature for etching of the substrate Si. Since the thickness of the micro-cantilever needs to be guaranteed, the etching time of the KOH solution needs to be strictly controlled to ensure that the cantilever beam Si substrate has sufficient thickness.

[0106] Steps 14-15 UV irradiation and etching of top etched window. To transfer the etched window pattern to the top of the substrate, 130nm of SiO2 protective layer is etched using UV lithography together with ICP.

[0107] Step 16: Photoresist removal

[0108] Step 16: Photoresist removal: Use acetone to remove the etch window photoresist.

[0109] Step 17: Self-curling release window pattern. Ultraviolet irradiation is used to obtain a self-curling release window pattern.

[0110] Step 18: Release from the curled structure: NMD solution is used to etch the sacrificial layer Al to gradually release the curled structure.

[0111] Step 19: Removing the photoresist: The sample with the self-curling structure is placed in an acetone solution to remove the photoresist.

[0112] Step 20: Si etching. KOH is used to etch the silicon substrate. Only the Si at the connection points of the upper and lower release windows can be etched away.

[0113] Step 21: Removal of SiO2 protective layer. The sample is placed in a BOE etching solution to completely etch the exposed SiO2, and then placed in an aqueous solution to obtain an AFM probe with a self-curling structure.

[0114] Based on strain engineering, a pair of opposing forces (F1 and F2) that form a three-dimensional structure [PV2000] are used to form a pre-strained two-dimensional nanometer-thick heteroepitaxial bilayer due to curling ( Fig.14A ) can create three-dimensional structures. The inventors have successfully demonstrated the controllable fabrication and characterization of various rolled-up three-dimensional spiral nanostructures with nanoscale features, such as SiGe / Si tubes ( Fig. 14B ), (diameter between 10nm and 10μm), Si / Cr ring ( Fig. 14C )[ZL2008](See reference [SO2005] for InGaAs / metal structures), SiGe / Si coils[ZL2005]( Fig.14D )、InGaAs / GaAs coils[BD2006]( Fig.14E ), small pitch InGaAs / GaAs coils ( Fig.14F )、Si / Cr claw( Figure 14G )、Si / Cr spiral [ZL2006a]( Fig.14H ), and small pitch SiGe / Si / Cr coils [ZL2006b]( Fig.14I The inventors have developed special processes such as coiling / uncoiling and winding / unwinding for manipulating various rolled-up three-dimensional helical nanostructures such as SiGe / Si and InGaAs / GaAs tubes [BD2006], coils, rings and spirals, as well as minimal robotic swimmers from prestressed bilayers such as SiGe / Si and InGaAs / GaAs for movement in the low Reynolds number region, and many other nanosystems [DL2006][DL2009].

[0115] Currently, the GRF project (11219419) is supporting the research of "Memory Ambient Oxygen Nanosensor for Medical Microrobotics" (PI: Lixin Dong, Period: 01Oct 19-30Sep 22), one of the goals of which is the stress engineering of microrobotics with nanoelectrode pairs. The semi-curled structure has the following features: Fig.14J and 14K The two stable states shown, horizontal and curled, allow this structure to be used as a cantilever. This approach can omit FIB-based electrode writing. The idea of ​​converting the gap between the cantilever tips into the thickness of the insulator layer has significant commercial value. In theory, it is possible to achieve sub-nanometer gaps based on ALD or using 2D materials.

[0116] Machine learning approach to iterative DNA sequencing. As the end effector of the nanorobot moves along the DNA chain, the electrical signal is measured. Since the highest occupied molecular orbital (HOMO) energies of the four DNA bases are as follows: G (-5.7eV) > A (-5.9eV) > C (-6.1eV) > T (-6.6eV) [TO2012], the differences between these HOMO energy values ​​form the basis for DNA sequencing using nanorobots. Specifically, the two electrodes on the end effector can measure the HOMO energy of the base. In theory, if the electrodes are small enough and the spacing between the electrodes is equal to the size of the base, each base can be measured independently by moving the end effector. However, based on the manufacturing technology of the prior art, the minimum distance between the electrodes will be larger than the base. For example, as Fig.15 Each electrode is 50 nm and has a 5 nm space in between. Each base is about 0.5 nm in size, so the space between two electrodes spans about 10 bases and each electrode will cover about 100 bases. Therefore, the actual measurement scenario will be like Fig.16 shown.

[0117] The essence of HOMO energy is electric potential energy, so a single DNA chain can be regarded as a battery chain with different voltages. These batteries covered by electrodes can be regarded as short circuits, such as Fig.17 As shown. Therefore, the potential difference between the two electrodes will be the sum of the 10 bases between the two electrodes. Based on this understanding, the physical size of the electrodes can be ignored and the contact between the electrodes and the DNA chain can be regarded as a point contact without affecting the final measurement results, such as Fig.18 shown.

[0118] Because one measurement is the sum of 10 bases, those bases in the middle of the DNA chain can exert influence on the 10 measurements, while the bases at the beginning and end of the DNA chain have much less influence. To balance this unfairness, nine known bases are connected to the beginning and end of the DNA chain. Since the HOMO energy of the T base is the easiest to distinguish, the known base is chosen to be T. This concatenation operation is like Fig.19 shown.

[0119] Assuming that the total number of bases after T is connected in series at the beginning and end of the DNA chain is n, n-9 measurements are required in one measurement cycle. Noise-free measurements should meet Fig. 20 The equations shown here, where x represents the HOMO energy value of each base and b represents the noise-free measured value. The first 18 equations represent the known true values ​​of the connected T bases.

[0120] In practical measurement situations, noise is unavoidable. For the case of DNA sequence measurements, there are two types of noise. The first type of noise is the contact noise between the electrode and the DNA strand or string, which can be approximated by adding Gaussian noise to the measurement. The second type of noise comes from the motion error of the nanorobot end effector, which is modeled using beta noise, as Fig.21 As shown. The electrode may not move exactly to the junction of two adjacent bases, but there will be some deviation δ. Assume that the deviation does not exceed the length of one base. The probability distribution of different deviation values ​​is approximately the β distribution

[0121]

[0122] Due to the elasticity of the DNA strands, the deviations of the two electrodes may be different. Two independent random variables δ1 and δ2 are assigned to each electrode. This deviation directly affects the measurement matrix. The actual measurement matrix will no longer be Fig. 20 one of them, but becomes Fig. 22 However, the β noise value is unknown, so when solving the DNA sequence, the measurement matrix used is still Fig. 20 However, the measured value is affected by both Gaussian noise and β noise, such as Fig.23 shown.

[0123] Fig.23 The equation in is a hyperstatic equation and can be solved using a neural network. The neural network structure used is as follows Fig.24 As shown. The input of the neural network is the position sequence of each base, such as 1, 2, 3..., and the output of the network is the HOMO energy value of the base at the corresponding position. If the equation is abbreviated as Ax = d, the loss function of training the neural network is defined as By minimizing the loss value, the optimal DNA sequence can be found.

[0124] Due to the presence of noise, one measurement cycle is not enough to obtain a convincing DNA sequence. Multiple measurements are necessary. Fig.23 The measurement equation in is easily scaled up. New measurements can be added directly to its end as new rows. Once a new measurement cycle is completed, new results about the DNA sequence can be calculated. To improve efficiency, the training results after the previous measurement are used as the initial value for the next calculation. In this way, after each new measurement cycle, a new DNA candidate sequence will be generated. As the number of measurements increases, the difference between the generated DNA sequences will become smaller and smaller until it no longer changes. The difference between two adjacent results can be used to calculate the confidence interval of the generated DNA sequence. This iterative DNA sequencing framework is shown in Fig.25 shown.

[0125] A simulation study was conducted to test the DNA sequencing framework under Gaussian contact noise and beta motion error. The first 30 bases of the green fluorescent protein DNA sequence were used as the true values, which are “TTACCAATGCTTAATCAGTGAGGCACCTAT”. First, only the measurements were contaminated with Gaussian contact noise, and the Gaussian noise used followed an N(0,0.04) distribution. The average DNA sequencing accuracy with respect to the number of measurements is shown in Fig.26 Second, only the β motion error of the nanorobot end effector is considered, and the β error is equal to

[0126]

[0127] The average DNA sequencing accuracy with respect to the number of measurements is Fig. 27 Finally, Gaussian contact noise and β motion error are considered, which are the same as those used in the first and second simulations. The simulation results are shown in Fig.28 As shown. In all cases, the accuracy increases with the number of measurements, which proves the effectiveness of the DNA sequencing framework of the present invention. More importantly, it shows the significance of using multiple measurements to improve sequencing accuracy.

[0128] The above simulation results are based on ideal contact noise and motion error models. In fact, contact noise may not completely conform to Gaussian distribution, and motion error may not completely conform to β distribution. Gaussian model and β model can only be used as a good reference. The true distribution of noise must be calibrated. To this end, the present invention uses a machine learning-based calibration method as follows: several known DNA sequences are measured to separate noise values. Through this method, some data about noise can be obtained, although very limited. Then, a new calibration framework based on the newly developed Generative Adversarial Three Model (GAT model) [SW2020] can be used to calibrate the noise distribution. The GAT model can promote machine learning using an approximate qualitative model to reduce the demand for training data.

[0129] The basic idea of ​​the GAT model is as follows Fig.29 As shown. Obviously there are two types of models, namely machine learning models and analytical models. The specific machine learning model can be any kind of supervised learning model or unsupervised learning model, not just limited to neural networks. These machine learning models can be trained independently until they converge with labeled or unlabeled data, or even without data. The analytical model can rely on any kind of knowledge about the machine learning output, such as physical knowledge or mathematical knowledge. Therefore, the analytical model can be used to correct the flawed output of the machine learning model. Next, the machine learning model will be reinitialized based on the corrected output and trained again until it converges from this new initialization. The whole process forms a cycle and converges when the machine learning output is consistent with the analytical model to an acceptable degree.

[0130] Machine learning models are often used to represent some complex functional relationships, such as the noise distribution in the current DNA sequencing problem. It is expected that the distribution will be obtained by training the data. However, training data may be difficult to obtain, so there may be only a small amount of data, or biased data. There are many reasons why the machine learning results are not perfect or do not meet some necessary constraints. However, if some additional knowledge about the desired function is known, this knowledge can be used to adjust the temporal machine learning output. Therefore, the machine learning model is reinitialized based on the adjusted output and further training is performed again. In the GAT model of the present invention, the functional relationship is generated by the machine learning model, so it acts as a generator. The machine learning model and the analysis model optimize this function in turn in an adversarial manner. The two models are competitive because the analysis model wants the function to be more consistent with the additional knowledge, while the machine learning model wants the function to be more consistent with the loss function in terms of smaller loss values. But this competition is a positive-sum competition, not a zero-sum competition like most generative adversarial networks (GAN) [CA2018], because this functional relationship is expected to conform to both the loss function and the additional knowledge.

[0131] The most critical connection point in the GAT model is how to reinitialize the machine learning model when adjusting the output. Fig.24 A feasible method has been developed using the neural network shown. This fully connected network with only one hidden layer is able to reproduce any function, so it can be used to generate the desired noise distribution. The reproduction process is as follows Fig.30 As shown. For any function (represented by the blue curve), a series of points can be sampled from it, (0,y0), (x1,y1(, (x2,y2), and (x3,y3). The original function can be approximated by a piecewise linear function consisting of line segments connecting every two adjacent points. The extensions of these line segments are represented by l1, l2, and l3. For Fig.30 In the specific case on the right side, the piecewise linear function can be expressed as l1+ReLU(l2-l1)-ReLU(l2-l3), which naturally conforms to the functional relationship represented by the neural network.

[0132] In general, if the sampling data points are (x0, y0), (x1, y1), …, (x n ,y n ), a formula algorithm has been developed to initialize each parameter in the neural network so that the neural network can reproduce the piecewise linear function connecting these sampled points. This algorithm is described in pseudocode as follows, where sgn(·) is the symbolic function.

[0133] Algorithm 1:

[0134]

[0135] b1 [1] ←-w1 [1] ·x0

[0136]

[0137] for i from 2 to n

[0138]

[0139] b i [1] ←-w i [1] ·x i-1

[0140]

[0141] end

[0142] b [2] ←y0

[0143] Algorithm 1 reinitializes the neural network by using the output adjusted by the analytical model so that the network can improve the adjusted output through further training. As the machine learning model and the analytical model alternate, the GAT model output also evolves. Several indices about the output of the GAT model can be defined to represent the output properties at different stages. These indices can be combined to form different states. The operation of the GAT model can then be viewed from the perspective of another state transition, such as the discrete event dynamic system (DEDS) model. Therefore, the convergence and stability of the GAT model can be analyzed.

[0144] The specific application of GAT model in noise calibration is as follows: Fig.31 As shown. In the GAT model, a neural network is used as a generator to generate a probability distribution of noise. The input of the network is different noise values, and the output is its corresponding probability density value. The network first trains experimental noise data from random initialization until convergence. The generated distribution is then used to quantize the unknown constants and their fusion weights in the Gaussian model and the β model. The neural network is then reinitialized using Algorithm 1 using the quantized noise model. The network is then trained again with experimental data. When the loss value is less than the threshold ε, the entire GAT model converges. In this way, the distribution that the real measurement noise conforms to can be calibrated. Based on the noise level, the number of measurements required for certain confidence levels (e.g., 99%) in sequencing can be estimated.

[0145] In summary, the proposed nanorobotic system will be able to repeatedly measure DNA strands and calibrate the system before sequencing. These key capabilities provide a basis for applying the proposed GAT model-based machine learning approach to analyze sequencing data and estimate the confidence of the obtained sequences. Therefore, it will significantly improve the accuracy and reliability of sequencing results.

[0146] The experimental testing and validation of the nanorobot system of the present invention is divided into two phases. The first phase focuses on evaluating the technical capabilities of the system, such as sequencing accuracy, sequencing speed, and software robustness. The second phase focuses on comparative studies to compare the performance of the system with sequencing methods and systems available on the market (e.g., BotSeqS, NanoSeq, and ION).

[0147] Use the following sample preparation protocol to create individual DNA strands for sequencing:

[0148] Isolate target single cells by mechanical dissociation and / or enzymatic dissociation

[0149] Extraction of genomic DNA

[0150] Removes any remaining RNA, such as lncRNA and miRNA, and proteins, such as histones, to generate straight single-stranded DNA and prevent base pair recombination

[0151] Use a known strand to join the two ends so they can be labeled with a recognizable barcode

[0152] Straighten and fix single strands on a flat substrate and keep all bases in the same direction.

[0153] In the first step, known DNA sequences are used to calibrate and validate the system. After that, unknown sequences are sequenced. At the same time, they are also sequenced using other techniques to validate the results.

[0154] In addition, the following studies were conducted in Phase II studies:

[0155] Human EPSCs can first be subjected to random genome mutations by gamma-ray or UV irradiation. Alternatively, they can be subjected to precise genome editing by CRISPR-Cas9-mediated knockout or knockin. Single-cell DNA libraries can be prepared and run through PCR, nanopores, and nanorobots in parallel to demonstrate the efficiency and precision of the nanorobotic system.

[0156] Cancer cells and primary tumors can be expanded in vitro and subjected to nanorobot-based genomic DNA sequencing to demonstrate superior efficiency and accuracy over other sequencing methods. DNA sequencing can be used to infer cell lineage and trace back to cancer stem cell populations.

[0157] The ION system is based on biological nanopores. When single-stranded DNA in a liquid passes through the nanopore, the ion flow between the inside and outside of the chamber changes due to the blockage of the nanopore by the DNA. The ion flow can be measured and provide information on the sequence of the DNA chain.

[0158] The main technical difference between the present invention and the ION system is that the present invention fixes the DNA chain to be measured on a substrate and moves the measuring electrode instead of moving the DNA chain so that it passes through a fixed nanopore with an electrode. This technical difference overcomes the main difficulties and shortcomings faced by the ION system, such as the inability to control the speed of DNA movement, measurement in a liquid environment, and loss of bases. Obviously, the mobile nanorobot end effector is much easier to control than the mobile DNA chain. More importantly, the DNA chain can only pass through the nanopore once in the ION system, and the measurement cannot be repeated. However, the nanorobot end effector of the present invention can move back and forth repeatedly on the fixed DNA chain. This allows the application of advanced data analysis, such as machine learning to significantly improve measurement accuracy.

[0159] In addition, the ION system has a huge problem when sequencing DNA chains with the same base. Because the DNA blockage in the nanopore is the same, the same base will not cause any change in ion flow when passing through the nanopore. Therefore, the short-read sequencing accuracy is low and the single nucleotide variation (SNV) is not ideal [NK2019][SA2020]. The main reason is that ION can only use the change in ion flow to detect the base, but does not know the position of the base or whether the base has passed through the nanopore. In contrast, the nanorobot of the present invention can not only repeat the measurement, but also register the position of the base while measuring. This advantage easily solves the problems faced by ION.

[0160] Furthermore, the microfluidic cell used for ION can only be used once. In addition to increasing the cost of sequencing, more importantly it makes it impossible to calibrate the device before using it. This can significantly increase the error of sequencing. However, the present invention has no consumables. The nanorobot end effector can be used multiple times and can be carefully calibrated in advance. Therefore, the cost of sequencing is greatly reduced and the sequencing error is significantly reduced.

[0161] References

[0162] The references cited in this application are all incorporated herein by reference, and their contents are as follows:

[0163] Fabrication and characterization of three-dimensional InGaAs / GaAs nanosprings ([BD2006] DJ Bell, LX Dong, BJ Nelson, M. Golling, L. Zhang, and D. Grutzmacher, "Fabrication and characterization of three-dimensional InGaAs / GaAs nanosprings," Nano Letters, vol. 6, no. 4, pp. 725-729, April 2006.)

[0164] Atomic force microscope ([BG1986] Binnig, G., C. F. Quate, and C. Gerber, "Atomic force microscope," Physical Review Letters, Vol. 56, p. 930, (1986).)

[0165] Generative adversarial networks: An overview ([CA2018] A.Creswell, T.White, V.Dumoulin, K.Arulkumaran, B.Sengupta, and AABharath, "Generative adversarial networks: An overview," IEEE Signal Processing Magazine, vol.35, no.1, pp.53-65, 2018.)

[0166] [CD2010b]XD Cui,WH Zhang,D. Erni,and LX Dong, "Optical properties of a nanomatch-like plasmonic structure,"Journal of the Optical Society of America a-Optics Image Science and Vision,vol.27,no.8,pp.1783-1790,Aug 2010.)

[0167] Nanopore research at Oxford Nanopore ([CG2016] CG Brown, and J. Clarke, "Nanopore development at Oxford nanopore," Nature biotechnology, vol. 34, no. 8, pp. 810-811, 2016.)

[0168] [CH2010] Chen, H., N. Xi, KW Lai, CK Fung, and R. Yang, "Development of infrared detectors using single carbon-nanotube-based field-effect transistors," IEEE transactions on nanotechnology, Vol. 9, pp. 582-589, (2010).

[0169] [CX2010a] X.Cui, LXDong, W.Zhang, W.Wu, Y.Tang, and D.Erni, "Numerical investigations of a multi-walledcarbon nanotube-based multi-segmented optical antenna," Applied Physics B-Lasers and Optics, vol.101, no.3, pp.601-609, Nov 2010.

[0170] [DB2020]https: / / www.pharmiweb.com / press-release / 2020-11-23 / dna-sequencing-next-generation-sequencing-market-2020-forecast-with-top-companies-growth-factors.

[0171] Electron beam-induced deposition of carbon nanotube emitters ([DL2002] LX Dong, F. Arai, and T. Fukuda, "Electron-beam-induced deposition with carbon nanotube emitters," Applied Physics Letters, vol. 81, no. 10, pp. 1919-1921, Sept. 2002.)

[0172] [DL2004] LX Dong, F. Arai, and T. Fukuda, "Destructive constructions of nanostructures with carbon nanotubes through nanorobotic manipulation," IEEE / ASME Transactions on Mechatronics, vol. 9, no. 2, pp. 350-357, June 2004.

[0173] Nanotube Linear Servo Motor ([DL2006] LX Dong, BJ Nelson, T. Fukuda, and F. Arai, "Towards Nanotube Linear Servomotors," IEEE Transactions on Automation Science and Engineering, vol. 3, no. 3, pp. 228-235, July 2006.)

[0174] Carbon nanotubes for nanorobotics ([DL2007a] LX Dong, A. Subramanian, and B. J. Nelson, "Carbon nanotubes for nanorobotics," Nano Today, vol. 2, no. 6, pp. 12-21, Dec 2007.)

[0175] Nanorobotics for creating NEMS from 3D helical nanostructures ([DL2007b] Dong, L., L. Zhang, DJ Bell, D. Grützmacher, and BJ Nelson, "Nanorobotics for creating NEMS from 3D helical nanostructures," in Journal of Physics: Conference Series, p. 257, (2007).)

[0176] [DL2009] LX Dong, L. Zhang, BE Kratochvil, KY Shou, and BJ Nelson, "Dual-Chirality Helical Nanobelts: Linear-to-Rotary Motion Converters for Three-Dimensional Microscopy," Journal of Microelectromechanical Systems, vol. 18, no. 5, pp. 1047-1053, Oct. 2009.

[0177] [FA2021] F. Abascal, LM Harvey, E. Mitchell, A. R. Lawson, SV Lensing, P. Ellis, A. J. Russell, R. A. Alcantara, A. Baez-Ortega, and Y. Wang, "Somatic mutation landscapes at single-molecule resolution," Nature, pp. 1-6, 2021.)

[0178] [FC2009] Fung, CKM, N.Xi, B.Shanker, and KWCLai, "Nanoresonant signal boosters for carbonnanotube based infrared detectors," Nanotechnology, Vol. 20, p. 185201, (2009).

[0179] [FT2013] F. Traversi, C. Raillon, S. Benameur, K. Liu, S. Khlybov, M. Tosun, D. Krasnozhon, A. Kis, and A. Radenovic, "Detecting the translocation of DNA through a nanopore using graphene nanoribbons," Nature nanotechnology, vol. 8, no. 12, pp. 939-945, 2013.

[0180] Single-cell genome sequencing: current state of the science ([GC2019] Gawad C. et.al Single-cell genome sequencing: current state of the science. Nat Rev Genet. 2016 Mar; 17(3): 175-88.)

[0181] [GH2014] H.Gao, Y.W.Hu, Y.Xuan, J.Li, Y.L.Ang, RVMartinez, CYLi, J.Luo, MHQi, and G.J.Cheng, "Large-scalenanoshaping of ultrasmooth 3D crystalline metallic structures," Science, vol.346, no.6215, pp.1352-1356, Dec 2014.

[0182] [GM2017]https: / / www.statista.com / statistics / 940953 / oxford-nanopore-sales-by-geographical-market-united-kingdom-uk /

[0183] Graphene as a subnanometre trans-electrode membrane ([GS2010] S. Garaj, W. Hubbard, A. Reina, J. Kong, D. Branton, and JA Golovchenko, "Graphene as a subnanometre trans-electrodemembrane," Nature, vol. 467, no. 7312, pp. 190-193, Sep 2010.)

[0184] [GV2020]https: / / www.grandviewresearch.com / press-release / global-single-cell-genome-sequencing-market.

[0185] Establishment of porcine and human expanded potential stem cells ([GX2019] Gao X. et. Al. Establishment of porcine and human expanded potential stem cells. Nat Cell Biol. 2019 Jun; 21(6): 687-699.)

[0186] [HC2018] C.Hou, Y.Wang, L.Yang, B.Li, Z.Cao, Q.Zhang, Y.Wang, Z.Yang, and L.Dong,"Position sensitivity of opticalnano-antenna arrays on optoelectronic devices,"Nano Energy, vol.53, pp.734-744, 2018.)

[0187] [HP2010] Harris, P. V., D. Welner, K. McFarland, E. Re, J.-C. Navarro Poulsen, K. Brown, R. Salbo, H. Ding, E. Vlasenko, and S. Merino, "Stimulation of lignocellulosic biomass hydrolysis by proteins of glycoside hydrolase family 61: structure and function of a large, enigmatic family," Biochemistry, Vol. 49, pp. 3305-3316, (2010).

[0188] Graphene nanodevices for DNA sequencing ([HS2016] SJ Heerema and C. Dekker, "Graphene nanodevices for DNA sequencing," Nature Nanotechnology, vol. 11, no. 2, pp. 127-136, Feb 2016.)

[0189] [HY2016] YW Hu, Y. Xuan, XL Wang, BW Deng, M. Saei, SY Jin, J. Irudayaraj, and GJ Cheng, "Superplastic Formation of Metal Nanostructure Arrays with Ultrafine Gaps," Advanced Materials, vol. 28, no. 41, pp. 9152-9162, Nov 2016.

[0190] [IS2013] Ido, S., K. Kimura, N. Oyabu, K. Kobayashi, M. Tsukada, K. Matsushige, and H. Yamada, "Beyond the helixpitch: direct visualization of native DNA in aqueous solution," ACS nano, Vol. 7, pp. 1817-1822, (2013).

[0191] Ion-beam sculpting at nanometre length scales ([JL2001] J. Li, D. Stein, C. McMullan, D. Branton, M. A. Ziz, and J. A. Golovchenko, “Ion-beam sculpting at nanometre length scales,” Nature, vol. 412, no. 6843, pp. 166-169, 2001.)

[0192] [JS2017] J. Shendure, S. Balasubramanian, G.M. Church, W. Gilbert, J. Rogers, J.A. Schloss, and R.H. Waterston, "DNA sequencing at 40: past, present and future," Nature, vol. 550, no. 7676, pp. 345-353, 2017.)

[0193] [KR2004] R. Kometani, T. Morita, K. Watanabe, T. Hoshino, K. Kondo, K. Kanda, Y. Haruyama, T. Kaito, J. I. Fujita, M. Ishida, Y. Ochiai, and S. Matsui, "Nanomanipulator and actuator fabrication on glass capillary by focused-ion-beam-chemical vapordeposition," Journal of Vacuum Science & Technology B, vol. 22, no. 1, pp. 257-263, Jan-Feb 2004.

[0194] [LA2015] LAE Leal, and O. Lopez-Acevedo, "On the interaction between gold and silver metal atoms and DNA / RNA nucleobases–a comprehensive computational study of ground state properties," Nanotechnology Reviews, vol. 4, no. 2, pp. 173-191, 2015.

[0195] [LC2017]C Li, Y Cheng, S Bi, Y Cai, N Xi, learning object recognition based on compressive sampling, 2017 IEEE International Conference on Robotics and Biomimetics (ROBIO), Macau, China, 5-8 Dec. 2017.

[0196] Content-Based Compressive Sensing ([LC2018]C Li,Y Cheng,Z Sun,P He,S Bi,N Xi,Content-Based Compressive Sensing,2018IEEE 8th Annual International Conference onCYBER Technology in Automation,Control,and Intelligent Systems(CYBER),Tianjin,China,19-23July 2018.)

[0197] Inference and effects of barcode multiplets in droplet-based single-cell assays ([LC2020] Lareau, CA, Ma, S., Duarte, FM & Buenrostro, JD Inference and effects of barcode multiplets in droplet-based single-cell assays. Nat. Commun. 11, 866 (2020).)

[0198] [LG2014] Gongxin Li, Peng Li, Yuechao Wang, Wenxue Wang, Ning Xi, Lianqing Liu, “Efficient Imaging and Real-Time Display of Scanning Ion Conductance Microscopy Based on Block Compressive Sensing”, International Journal of Optomechatronics 01 / 2014; 8(3).

[0199] [LG2015] Gongxin Li; Wenxue Wang; Yuechao Wang; Shuai Yuan; Wenguang Yang; Ning Xi; Lianqing Liu, "Nano-Manipulation Based on Real-Time Compressive Tracking" IEEE Transactions on Nanotechnology, Volume: 14, Issue: 5, September 2015

[0200] [LK2009a] Lai, KWC, N.Xi, CKM Fung, H.Chen, and T.-J.Tarn, "Engineering the band gap of carbon nanotubefor infrared sensors," Applied Physics Letters, Vol. 95, p. 221107, (2009).)

[0201] Automated nanomanufacturing system to assemble carbon nanotube based devices ([LK2009b] Lai, KWC, N.Xi, CKM Fung, J.Zhang, H.Chen, Y.Luo, and UC Wejinya, "Automated nanomanufacturing system to assemble carbon nanotube based devices," The International Journal of Robotics Research, Vol. 28, pp. 523-536, (2009).)

[0202] [LL2002] Lynd, LR, PJ Weimer, WH Van Zyl, and I SP Retorius, "Microbial cellulose utilization: fundamentals and biotechnology," Microbiol. Mol. Biol. Rev., Vol. 66, pp. 506-577, (2002).

[0203] [LY2011] Liu, Y.-S., JOBaker, Y. Zeng, M.E. Himmel, T. Haas, and S.-Y. Ding, "Cellobiohydrolase hydrolyzes crystalline cellulose on hydrophobic faces," Journal of Biological Chemistry, Vol. 286, pp. 11195-11201, (2011).)

[0204] [ML2016] ML Hoang, I. Kinde, C. Tomasetti, KW McMahon, TA Rosenquist, APG Rollman, KW Kinzler, B. Vogelstein, and N. Papadopoulos, "Genome-wide quantification of rare somatic mutations in normal human tissues using massively parallel sequencing," Proceedings of the National Academy of Sciences, vol. 113, no. 35, pp. 9846-9851, 2016.)

[0205] Preparation of three-dimensional nanostructures by focused ion beam chemical vapor deposition ([MS2000] S.Matsui, T.Kaito, J.Fujita, M.Komuro, K.Kanda, and Y.Haruyama, "Three-dimensional nanostructure fabrication by focused-ion-beam chemical vapor deposition," Journal of Vacuum Science & Technology B, vol. 18, no. 6, pp. 3181-3184, Nov-Dec 2000.)

[0206] [MT2019] MT Noakes, H. Brinkerhoff, AH Laszlo, IM Derrington, KW Langford, JW Mount, JL Bowman, KS Baker, KM Doering, and BITickman, "Increasing the accuracy of nanopore DNA sequencing using a time-varying cross membrane voltage," Nature biotechnology, vol. 37, no. 6, pp. 651-656, 2019.)

[0207] [NK2019] Kono N, Arakawa K. Nanopore sequencing: Review of potential applications in functional genomics. Dev Growth Differ. 2019 Jun; 61(5): 316-326.

[0208] [NN2018]Nikooienejad, N., A. Alipour, M. Maroufi, and SR Moheimani,"Video-Rate Non-Raster AFM Imaging With Cycloid Trajectory,"IEEE Transactions on Control Systems Technology, 2018).

[0209] [NT2015]https: / / www.nature.com / articles / nrg.2015.16.

[0210] [PP2018] P. Pungetmongkol, "Speculation of Nano-gap Sensor for DNA sequencing technology: A Review on Synthetic Nanopores," Engineering Journal, vol. 22, no. 6, pp. 229-250, 2018.

[0211] [PS2020]https: / / www.prnewswire.com / news-releases / next-generation-sequencing-data-analysis-market-size-worth-1-72-billion-by-2028-grand-view-research-inc-301229618.html

[0212] Free-standing and overgrown InGaAs / GaAs nanotubes, nanohelices and their arrays ([PV2000] VYPrinz, VASeleznev, AKGutakovsky, AVChehovskiy, VVPreobrazhenskii, MAPutyato, and TAGavrilova, "Free-standing and overgrown InGaAs / GaAs nanotubes, nanohelices and their arrays," Physica E, vol. 6, no. 1-4, pp. 828-831, Feb 2000.)

[0213] [RR2019] RR Wick, L M Judd, and K E Holt, "Performance of neural network base calling tools for Oxford Nanopore sequencing," Genome biology, vol. 20, no. 1, pp. 1-10, 2019.

[0214] [SA2020] Amarasinghe SL, Su S, Dong X, Zappia L, Ritchie ME, Gouil Q. Opportunities and challenges in long-read sequencing data analysis. Genome Biol. 2020 Feb 7; 21(1): 30.

[0215] [SB2014] B. Song, J. Zhao, N. Xi, H. Chen, KWC Lai, R. Yang, L. Chen, “Compressive Feedback-Based Motion Control for Nanomanipulation—Theory and Applications”, IEEE Transactions on Robotics 01 / 2014; 30(1):103-114.

[0216] [SJ2020]https: / / satijalab.org / costpercell / .

[0217] [SO2005] O.Schumacher, S.Mendach, H.Welsch, A.Schramm, C.Heyn, and W.Hansen, "Lithographically defined metal-semiconductor-hybrid nanoscrolls," Applied Physics Letters, vol.86, no.14, Apr2005.

[0218] Computational identification of cell doublets in single-cell transcriptomic data ([SW2019]Wolock, SL, Lopez, R. & Klein, A. M.Scrublet: computational identification of cell doublets in single-cell transcriptomic data. Cell Syst. 8, 281–291 (2019).)

[0219] [SW2020] S.Wang and N.Xi, "Calibration ofHaptic Sensors Using Transfer Learning," IEEE Sensors Journal, vol.21, no.2, pp.2003-2012, 2020.

[0220] Classification of low quality cells from single-cell RNA-seq data ([TL2016] Ilicic, T. et al. Classification of low quality cells from single-cell RNA-seq data. Genome Biol. 17, 29 (2016).)

[0221] Complementary base pairs for electrically pinpointing complementary nucleobases facilitate electron tunneling ([TO2006] T. Ohshiro, and Y. Umezawa, "Complementary base-pair-facilitated electron tunneling for electrically pinpointing complementary nucleobases," Proceedings of the National Academy of Sciences, vol. 103, no. 1, pp. 10-14, 2006.)

[0222] [TO2012] T.Ohshiro, K.Matsubara, M.Tsutsui, M.Furuhashi, M.Taniguchi, and T.Kawai,"Single-molecule electrical random resequencing of DNA and RNA,"Scientific reports, vol.2, no.1, pp.1-7, 2012.

[0223] [WA2018] A. Wasfi, F. Awwad, and AIAyesh, "Graphene-based nanopore approaches for DNA sequencing: A literature review," Biosensors & Bioelectronics, vol. 119, pp. 191-203, Nov 2018.

[0224] [WQ2013] Wu, Q., Y. Meng, K. Concha, S. Wang, Y. Li, L. Ma, and S. Fu, "Influence of temperature and humidity on nano-mechanical properties of cellulose nanocrystal films made from switchgrass and cotton," Industrial Crops and Products, Vol. 48, pp. 28-35, (2013).

[0225] [WR2016] Wagner, R., RJ Moon, and A. Raman, "Mechanical properties of cellulose nanomaterials studied by contact resonance atomic force microscopy," Cellulose, Vol. 23, pp. 1031-1041, (2016).

[0226] [WY2019] Wen, Y., H. Lu, Y. Shen, and H. Xie, "Nanorobotic Manipulation System for 360° Characterization Atomic Force Microscopy," IEEE Transactions on Industrial Electronics, 2019.)

[0227] Deconstruction of lateral DNA conductance from first principles ([XG2006] X.-G. Zhang, P.S. R. JC Wells, and M. Fuentes-Cabrera, "First-principles transversal DNA conductancedeconstructed," Biophysical journal, vol.91, no.1, pp.L04-L06, 2006.)

[0228] [XN2007]Xi, N., J. Zhang, H. Szu, and G. Li, "Nanorobot assembly of carbon nanotubes for mid-IR sensor," in Independent Component Analyses, Wavelets, Unsupervised Nano-Biomimetic Sensors, and Neural Networks V, p. 65760K, (2007).

[0229] [XN2010]Xi, N., KWCLai, H.Chen, and CKMFung, "Carbon nanotube-based noncryogenic cooledmultispectrum focal plane array," in Infrared Technology and ApplicationsXXXVI, p.766032, (2010).)

[0230] Augmented Reality for Nano Manipulation ([XN2011]Xi, N., B. Song, R. Yang, and K. Lai, "Augmented Reality for Nano Manipulation," in Handbook of Augmented Reality, ed: Springer, 2011, pp. 435-447.)

[0231] [XN2013]Xi, N; Song, B; Yang, R; Lai, KWC; Chen, H; Qu, C; Chen, L, “Video Rate Atomic Force Microscopy: Use of compressive scanning for nanoscale video imaging”, IEEE Nanotechnology Magazine, Vol. 7, Issue 1, Page(s): 4–8, March 2013.

[0232] Establishment of mouse expanded potential stem cells ([YJ2017] Yang J. et. al. Establishment of mouse expanded potential stem cells. Nature. 2017 Oct 19; 550(7676): 393-397.)

[0233] [YR2015] Yang, R., B. Song, Z. Sun, KWC Lai, CKM Fung, KC Patterson, K. Seiffert-Sinha, AA Sinha, and N. Xi, "Cellular level robotic surgery: Nanodissection of intermediate filaments in live keratinocytes," Nanomedicine: Nanotechnology, Biology and Medicine, Vol. 11, pp. 137-145, (2015).

[0234] [ZJ2008] Zhang, J., N. Xi, H. Chen, KWC Lai, G. Li, and UC Wejinya, "Design, manufacturing, and testing of single-carbon-nanotube-based infrared sensors," IEEE transactions on nanotechnology, Vol. 8, pp. 245-251, (2008).

[0235] Controllable fabrication of SiGe / Si and SiGe / Si / Cr helical nanobelts ([ZL2005] L. Zhang, E. Deckhardt, A. Weber, C. Schonenberger, and D. Grutzmacher, "Controllable fabrication of SiGe / Si and SiGe / Si / Cr helical nanobelts," Nanotechnology, vol. 16, no. 6, pp. 655-663, Jun 2005.)

[0236] Preparation and characterization of independent Si / Cr micro-nanospirals ([ZL2006a] L. Zhang, LX Dong, DJ Bell, BJ Nelson, C. and D.Grützmacher, "Fabrication and Characterization of Freestanding Si / Cr Micro-and Nanospirals," MicroelectronicEngineering, vol.83, no.4-9, pp.1237-1240, April-Sept.2006.)

[0237] [ZL2006b] L. Zhang, E. Ruh, D. Grutzmacher, LX Dong, DJ Bell, BJ Nelson, and C. Schonenberger, "Anomalouscoiling of SiGe / Si and SiGe / Si / Cr helical nanobelts," Nano Letters, vol. 6, no. 7, pp. 1311-1317, July 2006.

[0238] Ring closure of rolled-up Si / Cr nanoribbons ([ZL2008] L. Zhang, LX Dong, and BJ Nelson, "Ringclosure of rolled-up Si / Cr nanoribbons," Applied Physics Letters, vol. 92, no. 14, Apr 2008.)

[0239] When the present invention is explained as being related to certain embodiments, it is to be understood that various modifications thereof will become apparent to those skilled in the art who read the specification. Therefore, it should be understood that the present disclosure is intended to cover modifications falling within the scope of protection of the appended claims.

Claims

1. A nanoscale robotic system for single-cell DNA sequencing of a DNA strand positioned on a glass slide, comprising: an atomic force microscope (AFM) having an end effector in the form of a probe with a tip, which engages in a scanning process when the slide is positioned on a sample stage of the AFM, causing its probe tip to scan the base pairs of the DNA strand; a pair of spaced apart electrodes at the tip that contact opposite ends of the DNA strand; an electric current measuring system connected to the electrodes, the measuring system measuring the electric current between the bases of the DNA strand; and An artificial intelligence based data analysis system for determining a DNA sequence based on current from an electric current measurement system. 2 . The nanoscale robotic system according to claim 1 , wherein the current measurement system is a high-precision tunnel current measurement system.

3. The nanoscale robotic system of claim 1, wherein the artificial intelligence-based data analysis system utilizes a generative adversarial triad (GAT) machine learning approach.

4. The nanoscale robotic system of claim 1, wherein the atomic force microscope comprises: A first control circuit moves the tip in the XY plane of the slide so as to move along the DNA strip, A laser that directs a laser beam to reflect off the flat top of the probe, a light detector configured to receive the light beam after it is reflected from the top of the probe and to generate a deflection signal based on a change in the reflected light, the deflection signal being related to a deflection of the probe in the Z direction toward or away from the slide, and A second control circuit moves the probe in the Z direction based on the deflection signal to create a feedback signal that keeps the probe in a consistent position over the DNA strand.

5. The nanoscale robotic system of claim 4, wherein the light detector is a position sensitive photodiode (PSPD).

6. The nano-scale robotic system according to claim 1, wherein the atomic force microscope enables the probe tip to scan the base pairs of the DNA chain through semantic compression motion control based on a non-vector space control method, comprising: Atomic force microscopy output from semantically compressed samples of real intensity local scan images of DNA; Compressing a source of a desired intensity local scan image; a comparator for comparing the compressed desired intensity local scan image with the compressed actual intensity local scan image and generating a difference signal; as well as A non-vector space controller receives the difference signal and causes the compressed actual intensity local scan image to converge to the compressed desired intensity local scan image in the set space by reducing the difference signal, wherein the output of the controller causes the atomic force microscope to change the compressed actual intensity local scan image as part of a feedback loop.

7. The nano-scale robotic system of claim 6, wherein the AFM scanning process uses a compressive scanning mode.

8. The nano-scale robotic system of claim 6, wherein prior knowledge of image content is used in semantic compression during local scanning of the image.

9. The nano-scale robotic system of claim 8, wherein prior knowledge of the shape of the DNA strand is used for semantic compression.

10. The nano-scale robotic system of claim 6, wherein the semantic compressed motion control utilizes a sensor matrix learned from a set of DNA strand images, wherein the sensor matrix training process is performed by a universal deep learning strategy.

11. The nanoscale robotic system of claim 10, wherein the semantic content of the atomic force microscopy images of DNA strands is learned using a non-negative matrix factorization (NMF) representation.

12. A nano-scale robotic single-cell DNA sequencing method, the method positioning a DNA chain on an atomic force microscope slide, moving an end effector in the form of a probe with a tip along the DNA chain, wherein the spacing between the tip electrodes is greater than the spacing between the bases of the DNA chain, so that the potential difference between the electrodes is the sum of the molecular orbital (HOMO) energies of all bases between the two electrodes, comprising the following steps: Prepare DNA strands for sequencing; Add the energy values ​​of 9 T bases to the beginning and end of the DNA chain; n-9 measurements are performed as a measurement cycle, where n is the total number of bases after the T bases at the beginning and end of the DNA chain are connected in series. Gaussian noise was added to the measurements to account for the contact noise between the electrode and DNA; Beta noise is added to the measurements to account for errors in the end-effector motion; A neural network is used to solve the DNA sequence. The neural network input is the position sequence of each base, and the output is the molecular orbital energy value of the base at the corresponding position. The loss function for training the neural network is defined as The optimal DNA sequence can be determined by minimizing the loss value; and Repeating the measurement cycle, where the result after the previous measurement is used as the initial value for the next calculation; and The measurement cycle is repeated until there is no change in the resulting DNA sequence.

13. The method according to claim 12, wherein the Gaussian noise and beta noise are calibrated by the following steps: Several known DNA sequences were measured to determine the noise value of the separation; We create a new framework based on the generative adversarial triad to calibrate the noise distribution, where: Generative adversarial triads include machine learning models and analytical models and utilize approximate qualitative models to promote machine learning and thus reduce the need for training data; Independently train machine learning models until convergence; Base analytical models on any kind of knowledge about machine learning outputs; as well as The machine learning model initialization state is reinitialized based on the corrected output, and the machine learning model is trained again until it converges from the reinitialized value, whereby the entire process forms a cycle and converges when the machine learning output is consistent with the analytical model to an acceptable degree.

14. The method according to claim 12, wherein the step of preparing a DNA strand comprises the following steps: Isolate target single cells by mechanical separation and / or enzymatic dissociation; Extract genomic DNA; Remove any residual RNA, as well as proteins; Generates linear single-stranded DNA and prevents base pair recombination; connecting the two ends of the chain with a known chain to label the chain with a recognizable barcode; as well as Straighten and fix single strands on a flat substrate and keep all bases in the same direction.

15. The method according to claim 14, wherein the removed residual RNA includes lncRNA and miRNA; and / or The proteins removed include histones.

Images