Method and system for detecting characteristics of particles or particle environments, and method of training a machine learning model, computer progam or computer readable medium

Abstract

Methods and systems are disclosed for detecting characteristics of particles or particle environments. In one arrangement, trajectory data is received that represents detected diffusion trajectories of particles in one or more respective particle environments. Each detected diffusion trajectory comprises a sequence of trajectory segments. Characteristics of particles and / or of the one or more particle environments of the particles are identified by analysing the trajectory segments.

Description

[0001] METHOD AND SYSTEM FOR DETECTING CHARACTERISTICS OF PARTICLES OR PARTICLE ENVIRONMENTS, AND METHOD OF TRAINING A MACHINE LEARNING MODEL, COMPUTER PROGAM OR COMPUTER READABLE MEDIUM

[0002] The present disclosure relates to detecting characteristics of particles or particle environments by analysing diffusion behaviour. The disclosure is applicable to diagnosing or quantifying pathogens such as a viruses or bacteria, as well as to identify changes in particle compositions and / or transitions of particles between different environments, such as between different parts of a biological cell.

[0003] Viral outbreaks have affected the world's population for millennia. Despite hundreds of years of medical advancements, the world is still vulnerable to pandemics. Since the 20th century, there have been more than 50 epidemics, with six worldwide pandemics. In 2019-2021, COVID- 19 swept across the world.

[0004] While COVID-19 appears to have a lower fatality rate than SARS, it shows more severe symptoms in the vulnerable groups. In addition, a particular test is necessary to confirm the infection, as it shares similar symptoms with flu. Moreover, COVID-19 has a long incubation period during which patients can infect others even if they have not developed any signs of illness themselves. These properties have contributed to COVID- 19 spreading more rapidly compared to other viruses such as flu. Especially in many under developed regions, testing availability and speed has not been able to keep up with the expansion of the pandemic, causing a delay in test results and leaving many patients untested. There is an urgent need to provide improved testing methods that are rapid, sensitive and inexpensive, not only for tackling COVID-19 but also pandemics that may occur in the future.

[0005] Many viruses have an outer lipid envelope. Both coronaviruses and influenza viruses are enveloped by a lipid bilayer. Influenza is characterised by eight segmented, single-stranded RNA genome, where genome is a constitute of total genetic information in an individual organism. Depending on the genome, flu is further categorised into type A, B, or C in decreasing probability of causing an epidemic. Influenza A virus is responsible for the majority of the flu, especially those that spread globally and cause pandemics. Flu A is further divided into subtypes by its surface proteins: the hemagglutinin (HA) and the neuraminidase (NA). It has been found there are 18 possible subtypes for HA and 11 for NA, giving total 198 combinations, but only 3 HA (HI, H2, and H3) and two NA (N1 and N2) subtypes have caused human epidemics. In the case of coronavirus, it constitutes a non-segmented, single-stranded RNA genome. For the surface protein, coronavirus is covered by Spike protein with no further subtype.

[0006] There are various commercially applicable diagnostic tests for viruses with a range of sensitivity, specificity and waiting time.

[0007] Reverse Transcriptase PCR (RT-PCR) is the gold standard for current viral diagnosis, which directly detects the presence of the virus’s RNA genome in the clinical sample. The assay involves identifications and quantifications of the presence of infectious agents in a sample through the process of detection, amplification, and output measurement.

[0008] RT-PCT requires multiple temperature changes to drive cycles of reaction, involving sophisticated and expensive thermal equipment. Another way of doing this is to use an isothermal method, for instance Reverse Transcription Loop-Mediated Isothermal Amplification-Based Assay (RT- LAMP).

[0009] Rapid antigen test (RAT) detects virus surface proteins, usually coupled with lateral flow assay to display coloured results that can be read by eye. The process is much faster than RT-PCR, but it tends to suffer low sensitivity.

[0010] Enzyme-Linked Immunosorbent Assay (ELISA) is a microwell, plate -based assay technique designed to detect and quantify substances such as proteins, hormones, and antibodies (IgG / lgM). However, there might be a lag in time of antibody production at the initial stage of infection, which could lead to wrong negative results.

[0011] It is an object of the present disclosure to provide improved methods and systems for identifying pathogens such as viruses. More generally, it is also an object of the disclosure to provide methods that expand possibilities for monitoring particles and their environments.

[0012] According to an aspect of the invention, there is provided a computer-implemented method of detecting characteristics of particles or particle environments, comprising: receiving trajectory data representing detected diffusion trajectories of particles in one or more respective particle environments, each detected diffusion trajectory comprising a sequence of trajectory segments; and identifying characteristics of particles and / or of the one or more particle environments of the particles by analysing the trajectory segments.

[0013] Thus, a method is provided that uses multiple detected trajectory segments and can extract therefrom a wide range of information about particles and / or their environments. Using individual trajectory segments rather than an overall diffusion distance or diffusion coefficient allows more detailed information to be obtained about particles and / or their environments in comparison to alternative approaches, harnessing highly subtle aspects of the interplay between diffusion characteristics and the nature of the particles and their environments.

[0014] In an embodiment, the identifying of characteristics comprises identifying a particle type of one or more of the particles. The particles may comprise pathogens, optionally viruses or bacteria, and the identifying of a particle type comprises identifying a type of pathogen. The inventors have demonstrated that the method provides particularly high performance in the context of distinguishing between different types of pathogen. Furthermore, distinguishing between pathogens based on monitoring of diffusion trajectories can be implemented quickly and with equipment that is not excessively complex or expensive.

[0015] In an embodiment, one or more detectable labels is / are bound to each particle, the particles comprise enveloped virus particles and each detectable label is configured to bind to an envelope (e.g., lipid membrane) of the enveloped virus particles. This approach can be implemented particularly efficiently. Detectable labels can be formulated that are applicable to a wide range of enveloped viruses without requiring individual adaptation and the labelling process can be implemented almost instantaneously, providing distinct speed advantages over competing techniques for identifying viruses that may require significant incubation times etc. In one particularly convenient implementation, negatively charged nucleic acid segments are attached as labels to negatively charged lipid membranes of enveloped virus particles via positively charged calcium ions groups.

[0016] In an embodiment, the identifying of characteristics comprises identifying a particle environment type of one or more of the particle environments. The identifying of a particle environment type may comprise distinguishing between a plurality of predetermined zones within a biological cell. The inventors have demonstrated that the method is sufficiently sensitive to differences in diffusion characteristics that the existence of different environments can be detected even if the nature of the particle does not change. This allows sensitive measurements to be made of where target particles are in cells, allowing information to be extracted about concentrations of particles in different portions of cells and how the concentrations may change over time.

[0017] In an embodiment, the predetermined zones include zones within a bacterium, such as a periplasmic zone and a cytoplasmic zone. The inventors have demonstrated particularly high performance of the method in detecting whether proteins are in the periplasmic zone or the cytoplasmic zone of a bacterium, thereby providing a valuable tool for investigating biological mechanisms involving passage of proteins between these zones.

[0018] In an embodiment, the identifying of characteristics comprises detecting a change in a characteristic of a particle environment of a particle during a detected diffusion trajectory of the particle by analysing the trajectory segments of that detected diffusion trajectory. The inventors have demonstrated that the method is able to detect changes in a particle’s environment as a function of time. The inventors have demonstrated particularly high performance where the change in the characteristic of the particle environment comprises a transition between different zones of a biological cell, particularly transitions between the periplasmic zone and the cytoplasmic zone of a bacterium.

[0019] In an embodiment, the identifying of characteristics is performed by inputting the trajectory data to a trained machine learning model. The inventors have found that training a machine learning model to recognise links between sequences of trajectory segments and characteristics of particles and / or particle environments provides particularly high performance.

[0020] In an embodiment, the trained machine learning model is trained using a k-nearest neighbour machine learning algorithm. The inventors have demonstrated that k-nearest neighbour machine learning algorithms provide an advantageous balance between training efficiency and performance. Other classification algorithm can, however, be used.

[0021] In an embodiment, the method comprises pre-processing the trajectory data by a dimensionality reduction algorithm prior to inputting the trajectory data to the trained machine learning model, the dimensionality reduction algorithm optionally comprising principal component analysis. Using a dimensionality reduction algorithm in this context has been found to improve training efficiency and robustness.

[0022] In an embodiment, the method comprising capturing video data of the particles in the particle environment or particle environments and processing the video data to provide the trajectory data. The processing of the video data may comprise tracking positions of individual particles over a sequence of frames of video data to determine corresponding detected diffusion trajectories. The detected diffusion trajectories may be selectively included in the trajectory data, such that only a subset of detected particles contribute to the trajectory data. Selectively include trajectory data improves performance by allowing trajectory data of lower relevance and / or which may be ambiguous and / or erroneous to be omitted.

[0023] In an embodiment, the selection is based on a determined diffusion speed of the detected particles. The inventors have recognized that diffusion characteristics of target particles such as expected diffusion speed may be known in advance or calculated and that this can be exploited to exclude trajectory data that is less likely to correspond to a target particle of interest, thereby improving performance.

[0024] In an embodiment, the selection comprises excluding a detected diffusion trajectory for a particle on the basis of a determination that: a change in location of the particle between temporally adjacent frames, or frames separated by a predetermined number of frames, is higher than a predetermined threshold maximum; and / or a change in location of the particle between temporally adjacent frames, or frames separated by a predetermined number of frames, is lower than a predetermined threshold minimum. Thus, particles that are diffusing too quickly and / or particles that are diffusing too slowly can be omitted from the trajectory data.

[0025] In an embodiment, the selection comprises excluding a detected diffusion trajectory for a particle on the basis of a determination that the particle is located less than a predetermined minimum distance away from a detected diffusion trajectory of a different particle in one or more frames of the video data, optionally wherein the predetermined minimum distance is approximately equal to the predetermined threshold maximum. Excluding diffusion trajectories that are too close reduces or avoids the risk of overlapping diffusion trajectories, which can cause errors or ambiguity. According to an alternative aspect of the invention, there is provided a method of training a machine learning model, comprising: receiving labelled training data comprising data representing a plurality of detected diffusion trajectories of particles, each detected diffusion trajectory comprising a sequence of trajectory segments, and data representing a label associated with each detected diffusion trajectory, the label indicating a characteristic of a particle and / or a particle environment corresponding the diffusion trajectory; and training the machine learning model using supervised learning to identify characteristics of particles and / or particle environments of the particles based on the trajectory segments.

[0026] Embodiments of the disclosure will be further described by way of example only with reference to the accompanying drawings.

[0027] Figure 1 is a flow chart depicting elements of a method of detecting characteristics of particles or particle environments.

[0028] Figure 2 depicts an example system 2 for performing the method of Figure 1.

[0029] Figure 3 is a flow chart depicting elements of a method of training a machine learning model for use in a method of detecting characteristics of particles or particle environments.

[0030] Figure 4A depicts imaging of a sample using a wide field microscope in HILO mode.

[0031] Figure 4B is a magnified view of one of the virus particles 14 shown in Figure 4A.

[0032] Figure 4C schematically depicts single particle tracking allowing dynamics of virus particles to be observed.

[0033] Figure 5 schematically depicts the layout of an algorithm including track segmentation. One long track (diffusion trajectory) is split into smaller ones (containing fewer trajectory segments than the long track).

[0034] Figure 6 depicts KNN classification.

[0035] Figure 7 depicts example detected diffusion trajectories for: (A) SARS-CoV-2; (B) Udorn; (C)WSN; (D) IBV; (E) X31; (F) PR8. L = 1.153 pm.

[0036] Figure 8 depicts confusion matrices: viruses vs negative medium samples (with PCA on). The diagonal percentages indicate the classification performance for this class, e.g. for the true class of ALLA, 94.4% of it was predicted correctly and 5.6% incorrectly. The overall validation accuracy is obtained by averaging the diagonal numbers. (A) DMEM vs Udorn, the overall validation accuracy is 94.3%. (B) ALLA vs X31 vs IBV the overall validation accuracy is 94.2%. (C) MEM vs WSN vs SC2 vs WSN, the overall validation accuracy is 78.5%.

[0037] Figure 9 depicts confusion matrices: all influenza A viruses. 11,000 tracks were used for each viruses. (A) PCA off, overall accuracy 96.3%. (B) PCA on, overall accuracy 89.8%.

[0038] Figure 10 depicts confusion matrices: all influenza A viruses and IBV. PCA on. (A) 55,000 tracks were used for each virus, overall accuracy 87.3%. (B) 27,500 tracks were used for each virus, overall accuracy 84.0%.

[0039] Figure 11 depicts a colloidal electrolyte. V: virus particle. Fluorescently-labelled ssDNA and calcium ion are shown. The situation in the solution is complex. E.g., two viruses at certain distance close to each other may experience attractive force, and under the effect of Ca ions in between aggregation will take place.

[0040] Figure 12 depicts a confusion matrix depicting performance of a KNN model trained using a TorA-halotag (uninduced) as a periplasmic control and a TorAKK-Halotag (induced) as a cytoplasmic control. 70% of the data was used to train and validate the model.

[0041] Figure 13-16 are graphs showing relative classification scores for detected tracks (diffusion trajectories). In each figure, the upper figure shows variation of the relative classification score (vertical axis) as a function of sub-tracks (horizontal axis). The lower figure shows variation in an average step size (vertical axis) as a function of sub-tracks (horizontal axis). Figure 13 is an example of purely cytoplasmic track scoring nearly 1 for every sub-track as expected. Figure 14 is an example of a purely periplasmic track scoring nearly -1 for every sub-track as expected. Figures 15 and 16 depict examples of transition events detected in the datasets.

[0042] The present disclosure includes methods that are computer-implemented. Each step of such methods may be performed by a computer in the most general sense of the term, meaning any device capable of performing the data processing steps of the method, including dedicated digital circuits. The computer may comprise various combinations of known computer elements, including for example CPUs, RAM, SSDs, motherboards, network connections, firmware, software, and / or other elements known in the art that allow the computer to perform the required computing operations. The required computing operations may be defined by one or more computer programs. The one or more computer programs may be provided in the form of media or data carriers, optionally non-transitory media, storing computer readable instructions. When the computer readable instructions are read by the computer, the computer performs the required method steps. The computer may consist of a self-contained unit, such as a general-purpose desktop computer, laptop, tablet, mobile telephone, or other smart device. Alternatively, the computer may consist of a distributed computing system having plural different computers connected to each other via a network such as the internet or an intranet.

[0043] Methods are described below that use machine learning. The concept of machine learning encompasses a class of algorithms that can learn for themselves from data provided to them. Machine learning can be divided into two broad categories: supervised and unsupervised learning. In supervised learning, the training set fed into the algorithm is labelled. The goal for the supervised learning is then to find a function that can accurately map the input data to the desired output. In contrast, in unsupervised learning the training set is unlabelled, so the algorithm needs to leam and teach itself.

[0044] Example known supervised algorithms include k-Nearest Neighbour (KNN), Linear Regression, Decision Trees and Random Forest. Supervised learning may be used to solve classification and regression problems. Classification involves assigning labels to discrete outputs, while regression gives continuous output values (e.g. to predict quantities such as house price, temperature, size etc.). Some methods of the present disclosure aim to distinguish between different particles (e.g., pathogens), such as between different viruses, and / or between different particle environments, and may thus use a classification predictive model.

[0045] Figure 1 is a flow chart depicting elements of a method of detecting characteristics of particles or particle environments according to the disclosure. Figure 2 depicts an example system 2 for performing the method. The system 2 comprises a sample receiving device 4. The sample receiving device 4 is configured to receive a sample 6 to be tested. The sample received device 4 may comprise any suitable combination of components for allowing a sample 6 to be input to the device 4 and stored in a way that allows measurements to be performed on the sample 4. The sample 6 comprises the particles and / or particle environments of interest that are to be tested. Typically, the sample 6 will be a liquid, for example an aqueous solution.

[0046] In an embodiment, the system 2 comprises an image capturing arrangement 8. The image capturing arrangement 8 is configured to capture electromagnetic radiation, such as visible, IR or UV radiation, emitted from the sample 2. The system 2 further comprises a data processing arrangement 10. The data processing arrangement 10 comprises any suitable combination of data processing hardware, firmware, software, etc. necessary to perform the required data processing functionality. The data processing arrangement 10 may comprise various combinations of known computer elements, including for example CPUs, RAM, SSDs, motherboards, network connections, firmware, software, and / or other elements known in the art that allow the computer to perform the required computing operations. The required computing operations may be defined by one or more computer programs. The one or more computer programs may be provided in the form of media or data carriers, optionally non- transitory media, storing computer readable instructions. A computer readable medium may thus be provided that comprises instructions which, when the program is executed by a computer of the data processing arrangement 10, cause the computer to carry out methods of the present disclosure. The computer readable medium is an example of, and may be referred to as, a computer program product. The data processing arrangement 10 is configured to perform the computer-implemented methods described below. The data processing system 10 may output information, such as detected characteristics of particles and / or particle environments, to a display 12, or as an output data stream to an internal or external network or to a storage device.

[0047] In step SI of the method, trajectory data is received, for example by the data processing arrangement 10 or a data processor within the data processing arrangement 10. The trajectory data represents detected diffusion trajectories of particles in one or more respective particle environments. The trajectory data may represent detected diffusion trajectories of multiple particles in the same particle environment or multiple particles in multiple particle environments. Each detected diffusion trajectory comprises a sequence of trajectory segments. Each trajectory segment comprises a sub-portion of a longer continuous track followed by the respective particle. The continuous track is thus formed from multiple trajectory segments. The trajectory data may be derived from electromagnetic radiation captured by the image capturing arrangement 8. Further details for a specific example are given below with reference to Figure 4C.

[0048] In an embodiment, the image capturing arrangement 8 is configured to obtain video data of the particles in the particle environment or particle environments. In the example described below, and in other embodiments, the video data is captured using a wide field microscope, optionally configured to operate in a highly inclined and laminated optical sheet, HILO, mode.

[0049] The video data may be processed to obtain the trajectory data. The processing of the video data may be performed by the imaging capturing arrangement 8 or by the data processing arrangement 10 or both. The processing of the video data may comprise tracking positions of individual particles over a sequence of frames of video data.

[0050] In an embodiment, a detectable label is bound to each particle. The detectable label may, for example, be configured to be detectable by capture of electromagnetic radiation emitted from the detectable label. The detectable label may comprise a fluorescent label.

[0051] In an embodiment, the particles comprise enveloped virus particles and the detectable label is configured to bind to an envelope of the enveloped virus particles.

[0052] In step S2 of the method, characteristics of particles and / or one or more particle environments of the particles are identified by analysing the trajectory segments.

[0053] In an embodiment, the identifying of characteristics comprises identifying a particle type of one or more of the particles. In an embodiment, the particles comprise pathogens and the identifying of a particle type comprises identifying a type of pathogen. The pathogens may be viruses or bacteria.

[0054] In some embodiments, the identifying of characteristics is performed by inputting the trajectory data to a trained machine learning model. Figure 3 is a flow chart depicting elements of a method of training the machine learning model.

[0055] Step Ml of the method comprises receiving labelled training data. The labelled training data comprises data representing a plurality of detected diffusion trajectories of particles. The diffusion trajectories may be detected using an image capturing arrangement such as that described above within reference to Figure 2. Each detected diffusion trajectory comprises a sequence of trajectory segments. The labelled trained data further comprises data representing a label associated with each detected diffusion trajectory. The label indicates a characteristic of a particle and / or a particle environment corresponding to the diffusion trajectory (e.g., the particle that is diffusing along the diffusion trajectory and the particle environment within which the diffusion trajectory takes place.

[0056] Step M2 of the method comprises training the machine learning model using supervised learning to identify characteristics of particles and / or particle environments of the particles based on the trajectory segments.

[0057] Various types of machine learning model can in principle be used. In the example described below, the machine learning model is trained using a k-nearest neighbour machine learning algorithm. In some embodiments, the trajectory data is pre-processed by a dimensionality reduction algorithm prior to being input to the trained machine learning model. Applying a dimensionality reduction algorithm increases the speed of the training process and improves robustness. The dimensionality reduction algorithm may, for example, comprise principal component analysis.

[0058] In some embodiments, as exemplified below, one or more of the particle environments comprises an electrolyte. In such cases, electrolyte conditions for the trajectory data input to the trained machine learning model and electrolyte conditions used during training of the machine learning model may be arranged to be substantially the same in respect of at least the following parameters: pH; temperature; chemical composition. By keeping aspects of the particle environment that affect diffusion the same for the training data as for the particle of interest, the trained machine learning model can more effectively recognise aspects of the trajectory data that are indicative of the nature of the particle. This may be particularly important in the context of application of the method for identifying pathogens such as target bacteria or viruses.

[0059] Identifying pathogens example

[0060] To demonstrate methods of the disclosure, two coronaviruses and four strains of flu A were tested:

[0061] COVID-19 (SARS-CoV-2);

[0062] IBV (Infectious bronchitis virus); an H1N1 strain, PR8 (H1N1 A / Puerto Rico / 8 / 1934); another H1N1 strain WSN (H1N1 A / WSN / 1933); an H3N2 strain, Udom (H3N2 A / Udom / 1972); and another H3N2 strain, X31 (H3N2 A / X-31).

[0063] These viruses were compared with their negative growth media: Allantoic Fluid (ALLA), Minimum Essential Medium Eagle (MEM), and Dulbecco’s Modified Eagle Medium (DMEM).

[0064] As depicted schematically in Figure 4 A, viruses were imaged using a wide field microscope in HILO (highly inclined and laminated optical sheet) mode. The use of a wide field microscope in the HILO mode illuminates more viruses in the solution than evanescent wave in Total Internal Reflection Fluorescence. The microscope is an example of an image capturing arrangement 8. A glass slide 17 was used as a sample receiving device 4. The glass slide 17 was mounted on an objective 16 via an oil layer 18. A schematic trajectory of incident laser light is labelled 20. The laser incident angle was slightly smaller than the total internal reflection angle, so the light could penetrate into the sample 6.

[0065] Figure 4B is a magnified view of one of the virus particles 14 shown in Figure 4A. Figure 4B schematically depicts how calcium ions are used to facilitate an interaction between a negatively charged phospholipid layer and the phosphate group on the nucleic acid. The viruses were labelled by mixing three ingredients: the virus sample, buffered calcium chloride CaCl2solution and single-stranded DNA (ssDNA). Enveloped viruses have a net negative charge because of their lipid membrane and so does the ssDNA because of its phosphate backbone. The positively charged calcium ions group the negatively charged viral envelope and ssDNA, thus achieving labelling as depicted in Figure 4B. Negatively charged nucleic acid segments 24 are thus attached as labels to the lipid membrane of the enveloped viruses. This labelling method is universal (working effectively for any enveloped virus) and instantaneous (no incubation time is needed).

[0066] Fluorescently-labelled viruses were detected by single-particle tracking analysis software. The software located the fluorescent molecules by searching for the respective intensity peaks that are significantly above the background, then fitting the signal with a Gaussian function to obtain more precise localizations. As depicted schematically in Figure 4C, localizations for an individual virus particle can be connected over multiple frames into a continuous track 26, allowing the observation of particle dynamics. The continuous track 26 comprises a plurality of trajectory segments 28A-E. The trajectory segments 28A- E are obtained by tracking positions of individual particles 14 over a sequence of frames of video data. For example, each of the trajectories 28A-E may represent a movement of a particle by diffusion over a predetermined number of frames, such as between a respective pair of adjacent frames, of the video data. For example, where a sequence of frames is numbered consecutively, a sequence of trajectories may correspond to a sequence of displacements of a particle representing differences in position of the particle: between frames 1 and 2; between frames 2 and 3; between frames 3 and 4, etc. Alternatively, the differences in position of the particle may be: between frames 1 and 3; between frames 3 and 5; between frames 5 and 7, etc. A range of other choices would also work. However, typically, each trajectory 28A-E occurs over the same predetermined time period, which may correspond to a time period between frames in the video data or between a predetermined number of frames in the video data. Multiple particles can exist on the same frame. As described below, an exclusion radius (which may for example be defined on the basis of how far particles can travel in the time between two frames) may be used to assist with telling particles apart in subsequent frames. If we have 30 frames and a particle exists in all of them we get 30 localisations. A subset of them, i.e. anything between 2-29 localisations, can be defined as a trajectory segment.

[0067] The inventors have found that the extraction of relevant trajectory data from the video data can be made more efficient and / or robust by imposing constraints on which detected particles 14 contribute to the trajectory data. Thus, detected diffusion trajectories may be selectively included in the trajectory data, such that only a subset of detected particles 14 contribute to the trajectory data.

[0068] In one arrangement, the selection is based on a determined diffusion speed of the detected particles 14. The selection may, for example, seek to exclude particles 14 that appear to be diffusing significantly more quickly or significantly less quickly than would be expected for a particle 14 of interest. In an embodiment, the selection comprises excluding a detected diffusion trajectory for a particle 14 on the basis of a determination that a change in location of the particle between temporally adjacent frames, or frames separated by a predetermined number of frames, is higher than a predetermined threshold maximum. The predetermined threshold maximum may be referred to as a maximum step distance. Alternatively or additionally, the selection comprises excluding a detected diffusion trajectory for a particle 14 on the basis of a determination that a change in location of the particle between temporally adjacent frames, or frames separated by a predetermined number of frames, is lower than a predetermined threshold minimum. The predetermined threshold minimum may be referred to as a minimum step distance.

[0069] Imposing such maximum and / or minimum step distances may increase a probability that retained trajectory segments correspond to particles of interest. For example, as mentioned below, in the case where a particle of interest is a virus particle, imposing a maximum step distance may exclude particles that would be expected to diffuse more quickly than virus particles, such as free dye (e.g., ssDNA). Particles 14 that appear to move by more than the maximum step distance may also be more likely to be erroneous detections, for example corresponding to a situation where different particles detected in different video frames are erroneously considered to be the same particle.

[0070] Alternatively or additionally, the selection comprises excluding a detected diffusion trajectory for a particle on the basis of a determination that the particle is located less than a predetermined minimum distance away from a detected diffusion trajectory of a different particle in one or more frames of the video data. This may be referred to as an exclusion radius. Imposing such an exclusion radius reduces the risk of particle tracks overlapping with each other and causing confusion about which particle belongs to which trajectory segment.

[0071] In the present specific example, samples were prepared by mixing 3jUL of virus into 15^L buffered CaCl2solution with concentration of 0.45M, together with InM fluorescently-labelled DNA giving a final volume of about 20jUL. The glass slide 17 was imaged with green (532nm) and red (635nm) laser light of intensity 0.78kW / cm2. Exposure time for each movie (video) was 33.3ms.

[0072] Before running the point tracking in the single-particle tracking analysis software, two values were predefined for the maximum step distance and the exclusion radius. These parameters were used to ensure that the tracked particles were virus particles rather than any other kinds of particles, such as unbound DNA. In this example, the maximum step distance was used to exclude the free dye (ssDNA). The free dye was smaller and diffused faster in solution than the larger virus particles of interest. The exclusion radius was used to exclude tracks that crossed over to prevent ambiguity in trajectory assignment, as mentioned above. Particles expected to diffuse more slowly, such as cells or viral aggregates, could be excluded by imposing a suitable minimum step distance.

[0073] Appropriate values for the maximum step distance and exclusion radius can be determined by calculating the diffusing step of viruses from the Stokes-Einstein relation: kBT Ax2

[0074] D = - — = -

[0075] 6m]r 4At where D is the diffusion coefficient, kBis the Boltzmann’s constant, T is the temperature, Y is the viscosity, r is the radius of the virus particle, At is the exposure time (e.g., time between temporally adjacent frames), and Ax2is the mean square displacement.

[0076] Ax2was then calculated using the expression:

[0077] Ax2= 4DAt

[0078] This expression was then used to evaluate the maximum step size L in 2D according to:

[0079] L = V2Ax

[0080] The exclusion radius rexcwas estimated by: resc= O.lV iv where N is the average number of particles per unit area. If the exclusion radius is smaller than the maximum distance, then both parameters will use the maximum distance value. In this present example implementation, the exclusion radius was set to be equal to the maximum step size.

[0081] A machine learning model requires a large amount of data for training. During the training, the model is tuning its parameters to be able to map an input better to an output. However, it is often the case that data are only taken within a limited period of time, giving a small training set.

[0082] To ensure a sufficient amount of data can be used for training, the training set can be artificially expanded in the preprocessing stage. In the present implementation example, a minimum track length of 10 steps (e.g., corresponding to 10 trajectory segments) was chosen to make sure the data was normalized. Bigger tracks 51 in an original dataset 61 were divided into sub-tracks 52 in respective sub-datasets 62, as shown schematically in Figure 5. The sub-datasets 62 were used to form a table 70 comprising a validation set and a training set for training a machine learning model 72. Each track i contained steps (trajectory segments), then each track i was split into smaller tracks, vt, of 20 localizations (10 x coordinates, 10 y coordinates) per track such that

[0083] Smaller step sizes (5 and 7) were also tried. Although they produced large amount of data, the accuracies dropped slightly. For larger step size (e.g., 20), the data set was smaller, which also resulted in a decrease in accuracy. The 10-step size provided a good balance between a sufficient amount of data and good model accuracy. The reason why consecutive steps were used was to retain the information for the step size, which was likely to be picked up by the machine learning algorithm as a classification feature. The sub-datasets were then combined and converted into an expanded table 70 where it could be fed into the model 72 later.

[0084] K-Nearest Neighbour (KNN) is a supervised machine learning algorithm, mostly used for classification. It involves a basic assumption that similar inputs have similar outputs, which means data points close together are likely to be grouped into the same type, depicted schematically in Figure 6. The algorithm can be formalized as the following: for a dataset D and a test point x, there is a set of the k nearest neighbours x denoted as Sx, with SXD such that: dist(x,x') > max dist(x,x") where |SX| = k and (x',y') are points inside Sx, while (x",y") are points outside of Sxbut within D . x denotes position, and y represents the label to the point. KNN classifier can be defined as the following function h( ): (x) = mode {y"'- (x" , y") E Sx}) where mode( ) is to select the label with the highest number of votes. To summarize, KNN assigns the test point x with the label appearing the most number of times amongst its k nearest neighbour. The approach is illustrated schematically in Figure 6. The central circle 80 represents a track to be classified. When k = 1, the track corresponding to circle 80 gets one vote from the nearest neighbour belonging to the group 81 and will be assigned the label corresponding to group 81. When k = 3, the track corresponding to circle 80 gets two votes from nearest neighbours belonging to group 82 and only one vote from a nearest neighbour in group 81 and will be assigned the label corresponding to group 82. For the present example implementation, Fine KNN was used, where k = 1.

[0085] Recall that the KNN classifier assumes that points close to each other share similar labels. However, at higher dimension, data points start getting further apart. This problem is fundamentally due to not enough data being available for the expanding dimension and the counter intuitive geometrical properties at high-dimensional space. Increasing the dimension increases the size of data space, so to maintain the same density level the amount of data also needs to expand. However, the increment of data size to compensate the dimension expansion is disproportional. Furthermore, although it appears to be obvious how the KNN classifier works in 2D and 3D space, it is difficult to imagine the higher- dimension classification process. In our case, 20-dimensional data was used (10 x coordinates, 10 y coordinates). These issues can be addressed using dimensionality reduction. Principle component analysis (PCA) is an example implementation of dimensionality reduction and is described below.

[0086] Dimensionality reduction works by identifying the dimensions that truly matter in the data set.

[0087] Assume the data input to the KNN classifier has low intrinsic dimensionality: either data lies in a low dimensional sub-manifold or low dimensional subspace. For example, high dimensional points can be fitted to a 3D plane, then further assessment can be carried out by only looking at the projections in this 3D plane. Uniformly distributed data are of no use for classification.

[0088] Principal Component Analysis (PCA) is a method that firstly identifies the hyperplane that lies closest to the data, and then projects the data onto it. It first identifies an axis that accounts for the largest variance in the training set by maximizing the sum of the projection length square a . This is equivalent to minimizing the square distance between the data point to the projection axis b , but a computer would find it easier to manipulate with projection. If the training set has a total dimension d, PCA will find up to d number of mutually orthogonal axes. The unit vector defining the ithaxis is called the ithprincipal component (PC). To determine which axes will be used to construct the hyperplane that the points will project on, a variance ratio needs to be calculated. Variance is calculated by:

[0089] The variance for each PC is calculated, then the proportion of the dataset’s variance on each PC is evaluated via the variance ratio:

[0090] To set the classifier to account for at least 95% of the variance, we sum up the ratios from the largest up to 95%. The PCs chosen define the hyperplane. Therefore, instead of arbitrarily selecting the number of dimensions to reduce down to, this approach ensures the projection could preserve as much variance as possible.

[0091] The present model is used in the context of a clinical essay. To understand the utility of a clinical test, the following terms are introduced:

[0092] True positive TP): the patient has the disease and the test is positive;

[0093] False positive FP): the patient does not have the disease but the test is positive;

[0094] True negative (TN): the patient does not have the disease and the test is negative;

[0095] False negative (FN): the patient has the disease but the test is negative.

[0096] To evaluate a clinical test, sensitivity and specificity are used. Sensitivity indicates the ability of the test to correctly identify those patients with the disease. It is the ratio of the number of patients who are successfully detected with the disease to the true number of patients with the disease:

[0097] TP Sensitivity =Tp + pN

[0098] The higher the sensitivity, the better the test can identify all patients with the disease.

[0099] The specificity of a clinical test refers to the ability of the test to correctly identify those patients without the disease. It is the ratio of true number of negative cases to the total number of negative cases:

[0100] TN Specificity = -

[0101] H 7 yTN + FP

[0102] The likelihood of whether a patient has or does not have the disease with the result being positive or negative is evaluated by Positive Predictive Value (PPV) and Negative

[0103] Predictive Value (NPV) respectively: TP

[0104] PPV = -

[0105] TP + FP

[0106] TN

[0107] NPV = -

[0108] TN + FN

[0109] The conditions when conducting the experiment were: temperature T = 33 °C, viscosity was approximated as the viscosity of water at 33°C Y = 7.488 x 10-4Pas, the radius of the labelled viruses used r = 40 ~ 60nm, and the exposure time was At = 33.3ms. Substituting these values into the relevant equations for D and L given earlier yields

[0110] L = rexc= 1.153p,m ~ 1.412p,m

[0111] These values were inputted to the single-particle tracking analysis software to obtain primary track results, as exemplified in Figure 7. The tracks can been seen to vary in length, size and distribution. It is hard to distinguish their types by eye but the machine learning algorithm is able to extract useful information. Both of the example parameter values 1.153 and 1.412p.m were used as the basis for extracting tracks using single-particle tracking analysis software and were tested with the classifier. The results in terms of validation accuracies were similar but slightly better for the L = 1.153 ,m case, which involved exportation of more tracks from the single-particle tracking analysis software. Therefore, all of the tests in the following sections were conducted with the data obtained by using L = 1.153p,m from both green and red channels.

[0112] Before training the model, the dataset needed to be resized to include an equal contribution from all the training samples. The resizing process was done by identifying a suitable minimum virus data size after augmentation, then took out the same amount of data randomly from all groups of the viruses involved in the training, storing the rest for later tests. In the present example, 3000 tracks were taken for each virus for any training trail involving SARS- CoV-2 (SC2) due to its limited data size, and 11000 for each virus for the rest of the training.

[0113] The model was firstly examined on its ability to distinguish viruses from their negative environmental results. For lab cultured viruses, they were grown in different media as their nutrition requirements to support virus growing were different (see table below).

[0114] Table: Table of viruses and their corresponding negative controls.

[0115] negative. Hence, the experiment was organised as the following: MEM vs WSN vs PR8 vs SC2; DMEM vs Udom; ALLA vs X31 vs IBV.

[0116] After establishing its ability to distinguish the negative media, the model was further trained on its ability to differentiate viruses. Start with a simple one like, SC2 vs PR8 as they were grown on the same medium, MEM, then gradually add up with more viruses into the training.

[0117] KNN models with PCA on provided satisfactory results on distinguishing viruses with their negatives. Validation accuracies were all above 90% except the one with SC2 as shown in the confusion matrices in Figure 8. Low validation accuracy for the one with SC2 was mainly due to its limited data size. The training set randomly picked 3,000 out of 3,527 from SC2 with 3,000 were picked out of 30,680 from MEM. The model learnt from less than 10% of the amount of data from the MEM, from which very few features could be captured, leading to low accuracy in MEM identification (69.6%). Compared to that, a majority (84.6%) of SC2 data was used, giving a high validation accuracy of 87.2%. Two further tests were done on comparing viruses with MEM excluding SC2, and the outcomes were largely improved: MEM vs WSN gave 91.9%, and MEM vs WSN vs PR8 gave 88.1% accuracy.

[0118] For the comparison between viruses, although a small sample with only 6000 tracks in total for PR8 vs SC2, the overall result was good, giving on overall 92.3% validation accuracy. Then, all fluA samples (PR8, Udorn, WSN, X31) were mixed up, training the model to separate subtypes v virus (Figure 9B). The confusion matrix shows that the model can identify X31 and PR8 better than the other strains. However, it is notable that the validation accuracy is much higher when the PCA is off (Figure 9A), which implies there might be some properties being lost during the dimensionality reduction. The same situation happened when comparing influenza A viruses with IBV or SC2. The accuracy was 95.5% and 85% respectively with PCA off, while only 87.3% and 80.8% respectively with PCA on.

[0119] The loss of information due to PCA utilization can ultimately be overcome by increasing the training sample size. To further validate this point, the model was trained with different numbers of tracks from the same type of viruses (Figure 10). The reason why only IBV was included here is because this combination gave the largest data size. A model trained with more data is generally more reliable than one with less. The results in Figure 10 illustrate an improvement in model performance with larger training set: the overall validation accuracy dropped by 3.3% when the data size was halved.

[0120] In the end, all viruses were trained together. Again, due to limited SC2 sample size, only 3,000 tracks were used, and all of the other viruses need to keep the same number as SC2’s tracks. The overall training set was relatively small (total 18,000 tracks) with lower overall accuracy, 77.2% with PCA on and 82.2% with PCA off.

[0121] From the results discussed above, it appears that the model performs better with PCA off when testing speed is not considered. However, for practical applications the testing speed is an important factor.

[0122] From the table below, it can be seen that, compared to the case without PCA, the models with PCA hold lower accuracy, whereas the prediction speeds are far superior.

[0123] Table: Table concludes the model performance on virus differentiation. Test data size: All Flu A: 44,000 tracks; FluA vs IBV: 55,000 tracks; All viruses: 18,000 tracks. effect on doing thousands of tests will be non-negligible. Furthermore, the shortcoming in the aspect of accuracy can be resolved with the expansion of data volume, and on the other hand, it benefits the machine learning algorithm as a model is more reliable with more information feeding to it. Also, it can be noticed from the table above that the prediction speeds with PCA vary a lot with the training data size: all flu A and IBV (total 55,000 tracks) gives 2,300 object / s, all viruses (total 18,000 tracks) gives 7,800 object / s. The variation can be understood by the way how KNN works. Fine KNN model classifies the test point by finding the closest point and mapping the label to the test point. Searching in high dimensional space and finding distances to all data points around poses greater challenges than in the lower dimension. It will be certain that a model could produce perfect results without PCA, but the prediction speeds will be significantly hampered.

[0124] Therefore, to have a faithful and applicable model, a large amount of training data is inevitable, where in this case high prediction speed will be important, and the information loss in PCA will be compensated by more data input.

[0125] Some data was randomly excluded from the training. These unseen data can be used to check how models function. The tables below shows the resulting test accuracies, showing a relatively satisfactory result which means the model is able to distinguish tracks taken from the same trail of an experiment. The waiting time for all of these tests was less than 1 second using the model with PCA on.

[0126] Tables: the tables immediately below conclude the test accuracies for each model

[0127] Depending on the data size and virus types, the sensitivities for these models range from 71.6% to 91.7%, and specificities range from 94.2% to 97.3% (See tables below).

[0128] Tables: Tables of statistical results. SN: Sensitivity; SP: Specificity; PPV: Positive predictive value; NPV: Negative predictive value

[0129] Strictly speaking, viruses with ssDNA labelling in calcium chloride solution are colloidal electrolytes, where normal diffusion is no longer suitable for describing their motions. Colloidal particles or colloids are small solid particles that are suspended in a fluid phase. In our case, labelled viruses were colloids. Colloidal electrolytes are solutions of salts in which one ion has been replaced by a colloid. For instance, NaR, a sodium salt of an organic acid, R, the anion, form the colloid, and its dissociation can be written as: nNaR nNa++ R ~

[0130] Electrostatic interactions affect the virus motion, making the dynamics hard to predict. The influence of electrical forces makes the Stokes-Einstein inappropriate. The discrepancies between the theory and actual situation can be shown by substituting the experimental diffusion coefficient back into the Stokes-Einstein Equation, which yields a radius that is about 2 to 7 times larger than typical virus dimensions. The reasons for this may be due to charge interactions disrupting diffusion or viral aggregation. Furthermore, the presence of the freely moving ssDNAs and their interaction with the charged particles around would create more chaos, depicted schematically in Figure 11. Here, virus particles are labelled “V”, fluorescently labelled ssDNA are labelled “90”, and calcium ions are labelled “92”. The situation in such a solution is complex; e.g., two viruses V at a certain distance close to each other may experience an attractive force, and under the effect of Ca ions 92 in between aggregation will take place. So far, there is no accurate model that can describe such diffusion of colloidal electrolytes. However, approximating the situation to be a normal diffusion is useful for obtaining preliminary results on the tracks for analysis.

[0131] In the example described above, the identifying of characteristics in the method comprised identifying a particle type. In other embodiments, exemplified below, the identifying of characteristics comprises identifying a particle environment type. In the particular example below, the identifying of a particle environment type comprises distinguishing between a plurality of predetermined zones within a biological cell. Different zones within a biological cell may have properties that affect how particles, for example proteins, diffuse within them. The method uses these differences in properties to obtain information about where particles are at a given time. The approach is particularly useful in the context of bacteria. In this context, the predetermined zones may comprise different zones within a bacterium. As exemplified below, the different zones may comprise a periplasmic zone and a cytoplasmic zone. The method can thus be used to detect where a particle such as a protein is located within a bacterium, for example whether the protein is in the periplasmic zone or in the cytoplasmic zone, and monitor transitions through the membrane separating the two zones.

[0132] Protein transport across membranes is an essential process in bacteria, as it allows cells to properly compartmentalize proteins and other molecules within their cytoplasm and organelles. Protein transport can occur through both the cytoplasmic membrane and the outer membrane (periplasm) in bacterial cells.

[0133] There are several mechanisms by which proteins can be transported across membranes in bacteria. One common mechanism is the use of specialized proteintransporting machinery, such as the Sec machinery (general secretion pathway) or the Tat machinery (twin-arginine translocation). These systems use energy from ATP hydrolysis to power the transport of proteins across the membrane.

[0134] Other proteins can be transported across membranes via passive diffusion, where they simply diffuse through the lipid bilayer of the membrane. The movement of proteins through the membrane can be facilitated by the presence of channels or pores in the membrane, or by the presence of transport proteins that bind to the protein and help it cross the membrane.

[0135] Overall, protein transport across membranes is a complex and regulated process that is essential for the proper function and organization of bacterial cells. Methods of the present disclosure provide a valuable new tool for obtaining information about protein transport by making it possible to monitor movement of proteins with higher precision and / or in a wider range of scenarios than was previously possible at reasonable cost.

[0136] To exemplify efficacy of the method, the inventors focused on the twin-arginine translocation (Tat) pathway, which is a protein transport system that is found in many types of bacteria. It is used to transport folded proteins across the cytoplasmic membrane, which separates the cytoplasm of the cell from the external environment. It is named after the twin arginine residues that are found on the proteins that are transported via this pathway. These residues act as a signal that directs the protein to the Tat pathway for transport across the cytoplasmic membrane.

[0137] The Tat pathway is a complex process that involves several proteins and requires energy in the form of ATP. The protein to be transported is first recognized by the TatC protein, which is located in the cytoplasmic membrane. TatC then recruits the TatA and TatB proteins, which form a complex that spans the cytoplasmic membrane. The folded protein is then transferred to the TatA protein, which serves as a "chaperone" and helps the protein cross the membrane. Once across the membrane, the protein is released and transported to its final destination within the cell. The pathway is an important mechanism for transporting folded proteins across the cytoplasmic membrane in bacteria, and it plays a critical role in many cellular processes.

[0138] Diffusion in the periplasm is more complex than diffusion in the cytoplasma because the periplasm has additional complexity. The periplasma, contains the stationary peptidoglycan cell wall, which is porous and non-covalently attached to hundreds of thousands of OmpA molecules, a type of outer membrane protein found in many Gramnegative bacteria. These molecules would be expected to have very minimal lateral diffusion in the outer membrane. The periplasm also contains a large amount (depending on the osmolarity of the growth medium, up to 5 % of the total cell weight) of osmolarity regulated glycans, which are likely the cause of a more viscous environment in comparison to the cytoplasm.

[0139] The inventors have demonstrated that methods of the present disclosure can extract information about the complexity of periplasmic diffusion and thereby contribute to understanding how molecules behave in this compartment. This characterisation is important for classification of molecules as they move across membranes. The ability to track single-molecules and classify their compartmentalisation will enable measurements of the kinetics of protein transport across membranes.

[0140] The strains used in this example work are listed in Table SI below, the plasmids in Table S2 below, and the primers in Table S3 below. All constructs were verified by sequencing.

[0141] A plasmid to express HaloTag with a N-terminal TorA signal sequence and C- terminal SsrA tag (pQE-80 ssTorA-Halo-SsrA, periplasmic control) was constructed as follows. sstorA-gfpmut2 from plasmid pTGS (DeLisa et al,. 2002) was amplified using primers Pl and P2 and cloned between the BseRI and BamHI sites of pQE-80 (Qiagen) to create pQETG. An Agel restriction site was inserted between sstorA and gfpmut2 in pEQTG using Q5 site directed mutagenesis with primers P3 and P4 producing plasmid pQETAgeiG. The halotag gene was amplified from pHTC HaloTag® CMV-neo Vector (Promega) using primers P5 and P6 and used to replace the g / / v7u / / 2-containing Agel-BamHI fragment of pQETAgeiG, resulting in plasmid pQETH. The SsrA tag was inserted at the 3’ of the halotag gene in pQETH using Q5 site directed mutagenesis with primers P7 and P8.

[0142] To construct plasmid pQE-80 HaloTag (cytoplasmic control), the TorA signal sequence coding region and Agel restriction site were removed from pQETH using Q5 site directed mutagenesis with primers P9 and PIO. To create plasmid pQE-80 HaloTag-SsrA, the SsrA tag was inserted at the 3’ of the halotag gene of pQE-80 HaloTag using Q5 site directed mutagenesis with primers P7 and P8.

[0143] Strains expressing HaloTag protein constructs were grown from a single colony for 16 hrs at 37 °C and 180 rpm shaking in M9 minimal medium comprising M9 salts (Davis et al., 1986), 0.1 % (v / v) tryptone, 0.2 % (w / v) glucose, 2 mM MgSO4 and 0.1 mM CaC12, 0.1 mg / mL ampicillin. 100 pl was then sub-cultured into 5 ml of prewarmed M9 minimal medium and grown to mid-exponential phase (OD600 = 0.5). Janelia Fluor 646 HaloTag ligand was added to 1 mL of culture to a final concentration of 2 nM. Unreacted Halotag ligand was removed from the cells by resuspension and re-centrifugation with 1 ml of imaging buffer three times. The constructs containing a SsrA tag were further incubated for 30 mins at 37 °C with no shaking and then washed twice in imaging buffer. The washed cells were concentrated in 10-30 pl imaging buffer and then spotted onto 1% low- fluorescence agarose (Biorad) pads containing M9 salts, 0.2 % (w / v) glucose. The glass coverslips (#1.5 thickness) (Menzel-Glaser) used for imaging were pretreated by heating to 500 °C in a furnace to remove any fluorescent background particles.

[0144] Fluorescence images were acquired at 25 °C using a Nanoimager (Oxford Nanoimaging) equipped with a 640 nm 1W DPSS laser. Optical magnification was provided by a 100xoil-immersion objective (Olympus, numerical aperture (NA) 1.4) and images were acquired using an ORCA-Flash4.0 V3 CMOS camera (Hamamatsu), with a pixel size of 117 nm. The focal plane was manually positioned to the central cross-section of the cells (for highly inclined thin illumination (HiLo) imaging). Fluorescent foci that were acquired using a Nanoimager and had the super resolution fluorescence localisation performed using the Nano imager software (version 1.7.3). The track pre-processing and classification was identical to the methods used for the classification of viruses discussed above.

[0145] Table SI -E. coll strains used

[0146] Strains Genotype Source

[0147] MC4100 F", MacU169, araD139, Casadaban et al 1979 rpsL150, relAl, ptsF, rbsR, flbB5301

[0148] Table S2 - plasmids used in this study

[0149] Plasmid name Description Source pQE-80 Expression vector wtih Laci-repressible T5 promoter, Qiagen lacP, AmprpTGS sstorA-gfpmut2. (DeLisa et al,. 2002) pQETG pQE-80, expression of sstorA-gfpmut2. This work pQETH pQE-80, expression of sstorA-halotag. pQE-80 ssTorA- pQE-80, expression of sstora-halotag-ssra This work

[0150] Halo-SsrA pQE-80 Halo pQE-80, expression of halotag This work

[0151] Table S3 - DNA primers used in this study

[0152] Prime Primer sequence r name

[0153] “pi 5 ’-TAAAAGAGGAGAAATTAACTATGGACAATAACGATCTCTTTCAGGCATC-3 ’

[0154] P2 5’-TATTAGGATCCTTATTTGTATAGTTCATCCATGCCATGTGTAATCCC-3’

[0155] P3 5’-

[0156] GACTGACGCTACCGGTAGTAAAGGAGAAGAACTTTTCACTGGAGTTGTCCCAATT C-3’

[0157] P4 5’-GCCGCTTGCGCCGCAGTC-3’

[0158] P5 5 ’-TTACTACCGGTGAAATCGGTACTGGCTTTCCATTCG-3 ’

[0159] P6 5 ’-TGTAAGGATCCTTAACCGGAAATCTCCAGAGTAGACAG-3 ’

[0160] P7 5 ’-TACGCTTTAGCAGCTTAAGGATCCGCATGCGAGCTC-3 ’

[0161] P8 5’-GTTTTCGTCGTTTGCTGCACCGGAAATCTCCAGAGTAGAC-3’

[0162] Fine kNN models were trained using a TorA -halotag (uninduced) as a periplasmic control and a TorAKK-Halotag (induced) as a cytoplasmic control. 70% of the data was used to train and validate the model. The confusion matrix based on the validation set is shown in Figure 12. The overall accuracy of the trained model was 97.3%. The model was then used to classify sub-tracks that belong to the same track and plot the classification score per sub-track.

[0163] Figures 13-16 are graphs showing relative classification scores for detected tracks (diffusion trajectories). In each figure, the upper figure shows variation of the relative classification score (vertical axis) as a function of sub-tracks (horizontal axis). The lower figure shows variation in an average step size (vertical axis) as a function of sub-tracks (horizontal axis). Figure 13 is an example of purely cytoplasmic track scoring nearly 1 for every sub-track as expected. Figure 14 is an example of a purely periplasmic track scoring nearly -1 for every sub-track as expected. Figures 15 and 16 depict examples of transition events detected in the datasets. The signal transitions from purely cytoplasmic to purely periplasmic.

[0164] A classification score of 1 corresponds to purely 100% confidence (see Figure 13) from the model that the sub-track is cytoplasmic whereas a -1 classification score corresponds to purely periplasmic (see Figure 14). The idea is that, since the model was trained on purely cytoplasmic and periplasmic tracks, when the transition is happening the classification score should start dropping from 1 towards 0 which corresponds to a random decision and then towards -1 when the protein fully transitions to the periplasm (see Figures 15 and 16). This approach can be used to identify the exact point of transition, which provides information useful for understanding the mechanisms of protein transport across the cytoplasmic membrane. This functionality is an example of the identifying of characteristics of step S2 comprising detecting a change in a characteristic of a particle environment of a particle (in this case a protein) during a detected diffusion trajectory of the particle by analysing the trajectory segments (sub-tracks) of that detected diffusion trajectory. The change in the characteristic of the particle environment comprises a transition between different zones of a biological cell, in this case the different zones comprising a periplasmic zone and a cytoplasmic zone of a bacterium based on the Tat protein export pathway. However, the method is not limited to this. Transitions between different zones of other types of cell could be detected, as well as transitions using other mechanisms, such as the general secretion pathway (Sec).

[0165] The method can thus be used to study the dynamics of the Tat system, providing information for example about how certain pathogens, such as Staphylococcus aureus or Streptococcus pneumoniae, use this system to cause disease. The Tat system is also of interest in biotechnology because it can be used to transport proteins across the cytoplasmic membrane, which could be useful for the production of recombinant proteins. By understanding the dynamics of the Tat system, researchers could also gain insights into the mechanisms of protein transport across the cytoplasmic membrane and could use these insights to develop new therapies for diseases caused by disrupted protein transport.

[0166] The above examples demonstrate use of the method for distinguishing between different types of particle based on diffusion trajectories, exemplified particularly in the context of distinguishing between different viruses, and for detecting changes in characteristics of particle environments, exemplified particularly in the context of detecting transitions of proteins between different zones in biological cells, such as across membranes in bacteria. The method is not, however, limited to these applications. Based on the same principles, the identifying of characteristics of step S2 of the method may comprise detecting a change in a characteristic of a particle during a detected diffusion trajectory of the particle by analysing the trajectory segments of that detected diffusion trajectory. The inventors have demonstrated that the method is sensitive to small changes in the nature of the particle (which is necessary to distinguish between different viruses). Changes to a particle can thus also be detected. For example, the method can be used to detect a change in the folding of a protein, which can be driven by variations in the composition of the environment around the protein, such as variations in pH or salt concentration. This application thus also provides information about variations in the particle environment. The method can also be used to detect a change in a composition of a particle that affects its diffusion characteristics. For example, the method can detect a change in composition that arises at least partly due to attachment of material to the particle. The attachment of material to the particle may, for example comprise aggregation of viruses into a composite particle. Thus, the particle being monitored may comprise a single virus at first and the method may detect when viruses aggregate with the single virus to form a composite particle comprising plural viruses. Alternatively or additionally, the method may be used to detect a change in composition of a particle that arises at least partly due to detachment of material from the particle, such as a transition from a composite particle comprise multiple viruses to a particle consisting of a single virus or fewer viruses than were present in the original composite particle.

[0167] Methods of the present disclosure thus provide a tool that could be used to obtain new insights into fundamental biological processes and thereby contribute to developing new therapies and biotechnology products.

Claims

CLAIMS1. A computer-implemented method of detecting characteristics of particles or particle environments, comprising: receiving trajectory data representing detected diffusion trajectories of particles in one or more respective particle environments, each detected diffusion trajectory comprising a sequence of trajectory segments; and identifying characteristics of particles and / or of the one or more particle environments of the particles by analysing the trajectory segments.

2. The method of claim 1, wherein one or more of the particle environments comprises an electrolyte.

3. The method of claim 1 or 2, wherein the identifying of characteristics comprises identifying a particle type of one or more of the particles.

4. The method of claim 3, wherein the particles comprise pathogens, optionally viruses or bacteria, and the identifying of a particle type comprises identifying a type of pathogen.

5. The method of any preceding claim, wherein the identifying of characteristics comprises identifying a particle environment type of one or more of the particle environments.

6. The method of claim 5, wherein the identifying of a particle environment type comprises distinguishing between a plurality of predetermined zones within a biological cell.

7. The method of claim 6, wherein the particles comprise proteins.

8. The method of claim 6 or 7, wherein the predetermined zones include zones withina bacterium.

9. The method of claim 8, wherein the zones within the bacterium comprise a periplasmic zone and a cytoplasmic zone.

10. The method of any preceding claim, wherein the identifying of characteristics comprises detecting a change in a characteristic of a particle during a detected diffusion trajectory of the particle by analysing the trajectory segments of that detected diffusion trajectory.

11. The method of claim 10, wherein the change in a characteristic of the particle comprises a change in the folding of a protein.

12. The method of claim 10 or 11, wherein the change in a characteristic of the particle comprises a change in the composition of the particle.

13. The method of claim 12, wherein the change in the composition arises at least partly due to attachment of material to the particle, optionally wherein the particle comprises a virus and the attachment of material to the particle comprises aggregation of viruses into a composite particle.

14. The method of claim 12 or 13, wherein the change in the composition arises at least partly due to detachment of material from the particle.

15. The method of any preceding claim, wherein the identifying of characteristics comprises detecting a change in a characteristic of a particle environment of a particle during a detected diffusion trajectory of the particle by analysing the trajectory segments of that detected diffusion trajectory.

16. The method of claim 15, wherein the change in the characteristic of the particle environment comprises a transition between different zones of a biological cell, wherein,optionally: the particle comprises a protein; the biological cell comprises a bacterium; and / or the different zones comprise a periplasmic zone and a cytoplasmic zone of a bacterium.

17. The method of claim 16, wherein the transition between different zones arises via the twin-arginine translocation, Tat, protein export pathway or the general secretion pathway, Sec.

18. The method of any preceding claim, wherein the identifying of characteristics is performed by inputting the trajectory data to a trained machine learning model.

19. The method of claim 18, wherein: one or more of the particle environments comprises an electrolyte; and electrolyte conditions for the trajectory data and electrolyte conditions during training of the machine learning model are substantially the same in respect of at least the following parameters: pH; temperature; chemical composition.

20. The method of claim 18 or 19, wherein the trained machine learning model is trained using a k-nearest neighbour machine learning algorithm.

21. The method of any of claims 18-20, comprising pre-processing the trajectory data by a dimensionality reduction algorithm prior to inputting the trajectory data to the trained machine learning model, the dimensionality reduction algorithm optionally comprising principal component analysis.

22. The method of any preceding claim, comprising processing video data to provide the trajectory data.

23. The method of any preceding claim, comprising capturing video data of theparticles in the particle environment or particle environments and processing the video data to provide the trajectory data.

24. The method of claim 22 or 23, wherein the processing of the video data comprises tracking positions of individual particles over a sequence of frames of video data to determine corresponding detected diffusion trajectories.

25. The method of claim 24, wherein detected diffusion trajectories are selectively included in the trajectory data, such that only a subset of detected particles contribute to the trajectory data.

26. The method of claim 25, wherein the selection is based on a determined diffusion speed of the detected particles.

27. The method of claim 26, wherein the selection comprises excluding a detected diffusion trajectory for a particle on the basis of a determination that: a change in location of the particle between temporally adjacent frames, or frames separated by a predetermined number of frames, is higher than a predetermined threshold maximum; and / or a change in location of the particle between temporally adjacent frames, or frames separated by a predetermined number of frames, is lower than a predetermined threshold minimum.

28. The method of any of claims 25-27, wherein the selection comprises excluding a detected diffusion trajectory for a particle on the basis of a determination that the particle is located less than a predetermined minimum distance away from a detected diffusion trajectory of a different particle in one or more frames of the video data, optionally wherein the predetermined minimum distance is approximately equal to the predetermined threshold maximum.

29. The method of any of claims 25-28, wherein the video data is captured using a widefield microscope, optionally configured to operate in a highly inclined and laminated optical sheet, HILO, mode.

30. The method of any preceding claim, wherein one or more detectable labels is / are bound to each particle.

31. The method of claim 30, wherein each detectable label is configured to be detectable by capture of electromagnetic radiation emitted from the detectable label, the detectable label optionally comprising a fluorescent label.

32. The method of claim 30 or 31 , wherein the particles comprise enveloped virus particles and the detectable label is configured to bind to an envelope of the enveloped virus particles.

33. A method of training a machine learning model, comprising: receiving labelled training data comprising data representing a plurality of detected diffusion trajectories of particles, each detected diffusion trajectory comprising a sequence of trajectory segments, and data representing a label associated with each detected diffusion trajectory, the label indicating a characteristic of a particle and / or a particle environment corresponding the diffusion trajectory; and training the machine learning model using supervised learning to identify characteristics of particles and / or particle environments of the particles based on the trajectory segments.

34. A system for detecting characteristics of particles or particle environments, the system comprising a data processing arrangement configured to perform the method of any of claims 1-32.

35. The system of claim 34, further comprising: a sample receiving device configured to receive a sample to be tested; and an image capturing arrangement configured to capture electromagnetic radiationemitted from the sample, the system being configured to derive the trajectory data from the captured electromagnetic radiation.

36. A computer program or computer readable medium comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any of claims 1-32.

Images