Systems and methods for designing vaccines

Abstract

A system for designing a vaccine is disclosed that includes one or more processors and computer memory storing executable computer instructions that, when executed by the one or more processors, cause the one or more processors to perform one or more operations. The one or more operations include applying a plurality of driver models configured to generate output data representing one or more molecular sequences to a first time series dataset. The one or more operations include, for each of the plurality of driver models, training the driver model. The one or more operations include selecting a set of trained driver models from the plurality of driver models based on one or more trained translational responses. The one or more operations include selecting a subset of trained driver models from the set of trained driver models based on second translational response data.

Description

[0001] Cross-references to related applications

[0002] This application claims priority to U.S. Provisional Patent Application Serial No. 62 / 924,096, filed October 21, 2019, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This published text generally relates to systems and methods for generating vaccines. Background Technology

[0004] The mammalian immune system uses two common mechanisms to protect the body from environmental pathogens. When it encounters molecules originating from a pathogen, an immune response is activated to ensure protection against that pathogen.

[0005] The primary immune system mechanism is a nonspecific (or innate) inflammatory response. The innate immune system recognizes specific molecules present on pathogens rather than on the body itself.

[0006] The second immune system mechanism is the specific or acquired (or adaptive) immune response. For each injury or infection, the innate response is essentially the same. In contrast, the acquired response occurs specifically in response to molecules in or derived from the pathogen. The immune system recognizes and responds to structural differences between self- and non-self (e.g., pathogen- or pathogen-derived) proteins. Proteins that the immune system recognizes as non-self are called antigens. Pathogens typically express a large number of highly complex antigens. The acquired immune system utilizes two mechanisms: the first is the production of immunoglobulins (antibodies) in response to the many different molecules (called antigens) present in the pathogen. The second is the recruitment of receptors to bind processed forms of antigens, which are presented on the cell surface so that other cells can identify the infected cell.

[0007] In summary, acquired immunity is mediated by specialized immune cells called B and T lymphocytes (or simply B and T cells). Acquired immunity has a specific memory of antigenic structure. Repeated exposure to the same antigen increases the response, which can increase the level of induced protection against that specific pathogen. B cells generate and mediate their function through the action of antibodies. B cell-dependent immune responses are called "humoral immunity" because antibodies are present in body fluids. T cell-dependent immune responses are called "cell-mediated immunity" because effector activity is directly mediated by the local action of effector T cells. The local action of effector T cells is amplified through synergistic interactions between T cells and secondary effector cells, such as activated macrophages. The result is that pathogens are killed and disease is prevented.

[0008] Similar to pathogens, vaccines work by activating an innate immune response at the vaccination site and by activating antigen-specific T and B cells, which can generate long-term memory cells in secondary lymphoid tissues. The precise interaction between the vaccine and cells at the vaccination site, as well as with T and B cells, is crucial for the ultimate success of the vaccine.

[0009] Determining whether a candidate antigen is a functional and effective vaccine typically requires a rigorous testing and evaluation protocol. Traditionally, candidate antigens are tested preclinically using a method that evaluates them through in vitro assays, ex vivo assays, and the use of various animal models (e.g., mouse models, ferret models, etc.).

[0010] One exemplary type of assay that can be used to measure a biological response is the hemagglutination inhibition assay (HAI). HAI utilizes the hemagglutination process in which sialic acid receptors on the surface of red blood cells (RBCs) bind to hemagglutinin glycoproteins present on the surface of influenza viruses (and several other viruses), creating a network or lattice structure of interconnected RBCs and virus particles called hemagglutination, which occurs on the virus particles in a concentration-dependent manner. This is a physical measurement, serving as an alternative to the ability of viruses to bind to pathogens targeting cells with similar sialic acid receptors. The introduction of antiviral antibodies occurs in a human or animal immune response against another virus (which may be genetically similar to or different from the virus used in the assay for binding RBCs). These antibodies interfere with virus-RBC interactions and alter the concentration of the virus sufficiently to change the concentration at which hemagglutination is observed in the assay. One objective of HAI can be to characterize the concentration of antibodies in antiserum or other antibody-containing samples, relating to their ability to induce hemagglutination in the assay. The highest dilution of the antibody that prevents hemagglutination is called the HAI titer (i.e., the measured response).

[0011] Another approach to measuring biological responses is to measure a potentially larger group of antibodies elicited by a human or animal immune response, which may not necessarily affect hemagglutination in HAI assays. For this purpose, a common method utilizes enzyme-linked immunosorbent assay (ELISA), in which viral antigens (e.g., hemagglutinin) are immobilized on a solid surface, and antibodies from antiserum are then allowed to bind to the antigens. Readout measures the catalytic action of the exogenous enzyme on a substrate, which conjugates with antibodies from the antiserum or other antibodies that bind to the antiserum itself. The catalytic action on the substrate produces an easily detectable product. Many variations of this in vitro assay exist. One such variation is called antibody forensics (AF); this is a multiplex bead array technique that allows for the simultaneous measurement of a single serum sample against many antigens. Compared to HAI titers, these measurements characterize concentration and total antibody recognition, with HAI titers considered more specifically related to interference by hemagglutinin molecules on sialic acid binding. Therefore, in some cases, the antibodies of antiserum may have a proportionally higher or lower HAI titer compared to the corresponding HAI titer of a hemagglutinin molecule of one virus (relative to the hemagglutinin molecule of another virus); in other words, the two measurements of AF and HAI are usually not linearly related.

[0012] Currently, routine candidate antigen testing can only be performed under the condition of evoking the intended "protective" immune response. That is, if an animal or assay fails to demonstrate an appropriate response to a candidate antigen, the candidate antigen is typically "downgraded" (i.e., abandoned as a production candidate). For example, a sequential selection protocol is commonly used to test influenza antigens, where the antigen is first evaluated by an in vitro assay to ensure it is readily available for large-scale production. Conditioned that the antigen meets these requirements, it is then evaluated by immunizing, for example, mice, to measure its ability to elicit a protective immune response in the mice. This response is generally considered to be a protective response against the antigen itself and various other viral strains and / or viral components, with the expectation of protection against these strains and / or viral components. Subsequently, ferrets can be evaluated in a similar manner, provided that previous measurements in mice or other animals have demonstrated what can be considered a protective response. Following evaluation in humans, in vitro platforms such as replicas of the human immune system or non-human primates can be evaluated; again, this is conditional on the success of the preceding steps. Summary of the Invention

[0013] In one aspect, a system for designing vaccines is provided. The system includes one or more processors. The system includes a computer memory storing executable computer instructions, wherein when executed by the one or more processors, the executable computer instructions cause the one or more processors to perform one or more operations. The one or more operations include applying a plurality of driver models to a first time-series dataset, the plurality of driver models being configured to generate output data representing one or more molecular sequences, the first time-series dataset indicating: one or more molecular sequences, and for each of the one or more molecular sequences, one or more times of the prevalence of a pathogenic strain of that molecular sequence as a natural antigen. The one or more operations include, for each of the plurality of driver models, training the driver model in the following manner: i) receiving output data from the driver model, the output data representing one or more predicted molecular sequences based on a received first time-series dataset; ii) applying a translation model to the output data representing the one or more predicted molecular sequences, the translation model being configured to predict biological responses to molecular sequences with respect to a plurality of translation axes based on the one or more predicted molecular sequences of the output data, to generate first translation response data representing one or more first translation responses corresponding to a specific translation axis among the plurality of translation axes; iii) adjusting one or more parameters of the driver model based on the first translation response data; and iv) repeating steps i-iii multiple times iteratively to generate trained translation response data representing one or more trained translation responses corresponding to the specific translation axis. The one or more operations include selecting a set of trained driver models from the plurality of driver models based on the one or more trained translation responses. The one or more operations include: for each trained driver model in the group, applying the trained driver model to a second time series dataset to generate trained output data representing one or more predicted molecular sequences for a specific season; applying the translation model to the final output data to generate second translation response data, the second translation response data representing one or more second translation responses about each of the plurality of translation axes; and selecting a subset of trained driver models in the group based on the second translation response data.

[0014] At least one of the plurality of driver models may include a recurrent neural network. At least one of the plurality of driver models includes a long short-term memory recurrent neural network.

[0015] The output data representing one or more predicted molecular sequences based on a received first time-series dataset may include output data representing antigens for each of multiple disease-causing seasons. The output data representing antigens for each of multiple disease-causing seasons may include antigens determined by predicting molecular sequences that will generate a maximized aggregate biological response across all pathogenic strains circulating in a particular season. The output data representing antigens for each of multiple disease-causing seasons may include antigens determined by predicting molecular sequences that will generate an effective immune response against the maximum number of viruses circulating in a particular season.

[0016] The plurality of translation axes may include at least one of the following: ferret antibody evidence (AF) axis, ferret hemagglutination inhibition assay (HAI) axis, mouse AF axis, mouse HAI axis, human replica AF axis, human AF axis, or human HAI axis. The number of iterations may be based on a predetermined number of iterations. The number of iterations may be based on a predetermined error value. The one or more first translation responses may include at least one of the following: predicted ferret HAI titer, predicted ferret AF titer, predicted mouse AF titer, predicted mouse HAI titer, predicted human replica AF titer, predicted human AF titer, or predicted human HAI titer.

[0017] Selecting the group of trained driver models from the plurality of driver models may include assigning each of the plurality of driver models to a class of driver models, wherein each class is associated with a specific translation axis among a plurality of translation axes used to train the driver model. Selecting the group of trained driver models from the plurality of driver models may include, for each of the plurality of driver models, comparing one or more trained translation responses of that driver model with one or more trained translation responses of at least one other driver model assigned to the same class as that driver model.

[0018] The operation may further include: for each trained driver model in a subset of the trained driver models, validating the trained driver model by comparing second translational response data corresponding to the trained driver model with observed experimental response data; and generating a vaccine in response to validating the trained driver model, the vaccine comprising one or more molecular sequences represented by training output data corresponding to the trained driver model.

[0019] In one aspect, a system is provided. The system includes a computer-readable storage memory comprising computer-executable instructions. The system includes at least one processor configured to execute executable logic including at least one machine learning model trained to predict one or more molecular sequences, wherein while the at least one processor is executing the computer-executable instructions, the at least one processor is configured to perform one or more operations. The one or more operations include receiving time-series data indicating: one or more molecular sequences, and for each of the one or more molecular sequences, one or more times of prevalence of a pathogenic strain of that molecular sequence as a natural antigen. The one or more operations include processing the time-series data through one or more data structures stored in one or more portions of the executable logic included in the machine learning model to predict one or more molecular sequences based on the time-series data.

[0020] Predicting one or more molecular sequences based on the time-series data may include predicting one or more immunological properties that the predicted molecular sequences will confer for use at future times. Predicting the one or more molecular sequences based on the time-series data may include predicting one or more molecular sequences that will generate a maximum aggregate biological response across all pathogenic strains in the time-series data. Predicting the one or more molecular sequences based on the time-series data may include predicting one or more molecular sequences that will generate a biological response that will effectively cover the maximum number of pathogenic strains in the time-series data. The predicted one or more molecular sequences can be used to design vaccines against pathogenic strains prevalent during periods following one or more times of prevalence in the time-series data.

[0021] Machine learning models can include recurrent neural networks.

[0022] These and other aspects, features, and implementations may be expressed as methods, apparatus, systems, components, program products, methods of performing transactions, means or steps for performing functions, and other means, and will become clear from the following description, including the claims.

[0023] The implementation of this disclosure can provide one or more of the following advantages: Compared to conventional techniques, vaccines can be designed for future disease seasons to provide greater protection in terms of the amount of biological response to at least one pathogenic strain in that future disease season. Compared to conventional techniques, vaccines can be designed for future disease seasons to provide greater protection in terms of the breadth of effective coverage of multiple pathogenic strains in that future disease season (i.e., eliciting an effective immune response against multiple pathogenic strains in the future disease season). Unlike conventional techniques, rarely observed strains (which may provide “greater protection” because they cross-react with more strains compared to frequently observed strains) can be evaluated, and their vaccination efficacy can be predicted.

[0024] These and other aspects, features, and implementation methods can be expressed as methods, devices, systems, components, program products, means, or steps for performing functions, as well as other means.

[0025] These and other aspects, features, and implementations will become clear from the following description (including the claims). Attached Figure Description

[0026] Figure 1 shows an example of a system used for designing vaccines.

[0027] Figures 2A and 2B show flowcharts of methods for designing systems for vaccine design.

[0028] Figure 3 shows a flowchart of the method used to design vaccines.

[0029] Figure 4 shows a flowchart of a method for training one or more driver models to design vaccines.

[0030] Figure 5 shows a graph depicting the improvement along the translation axis compared to conventional techniques used for vaccine design.

[0031] Figure 6 shows an example of a system for predicting biological responses using machine learning techniques.

[0032] Figure 7 shows a flowchart illustrating an example of a method for predicting biological responses using machine learning techniques.

[0033] Figure 8 shows an example of data used to train a machine learning model to predict biological responses.

[0034] Figure 9 shows a flowchart of an example used to train a machine learning model to predict biological responses. Detailed Implementation

[0035] Conventional methods for selecting candidate vaccines (CVs) and / or antigens expressing recombinant proteins for said candidate vaccines often rely on several assumptions. As an illustrative example, in the case of influenza, conventional methods for selecting CVs may assume that: (1) for any given disease season, there exists a “dominant strain”; (2) uninfected ferrets are an accurate model of influenza drift (i.e., cross-reactivity in ferrets demonstrates whether a CV as an antigen would provide protection against other circulating influenza strains); and (3) increased ferret cross-reactivity can be a reliable predictor of increased efficacy of human vaccines. Based on these assumptions, conventional methods for selecting CVs may involve: (1) selecting CVs that protect against the dominant strain; (2) establishing protective associations using, for example, ferret HAIs; and (3) assessing the cross-reactivity of clinical isolates in ferrets. Furthermore, conventional methods for selecting CVs typically include selecting CVs that were prevalent in the year preceding the year the vaccine was recommended, and evaluating (usually using ferrets) the selected CVs against other frequently observed pathogenic strains.

[0036] While these assumptions might have been helpful for effective CVV selection when 1 to 10 pathogenic isolates were observed in a year 50 years or more ago, they are unlikely to be helpful for effective CVV selection in the current disease season, in which thousands of pathogenic isolates may be observed and reported. This is because it may be difficult to extend ferret evaluation to thousands of pathogenic isolates. The result could be, for example, that current selection of seasonal influenza vaccines typically achieves less than 50% vaccine efficacy (i.e., the percentage reduction in severe illness in the vaccinated group compared to the unvaccinated group among individuals seeking cases).

[0037] The systems and methods described in this specification can be used to mitigate one or more of the aforementioned drawbacks of conventional CV selection techniques. According to the systems and methods described in this disclosure, one or more molecular sequences (e.g., antigen sequences) predicted to perform well on at least one translation axis are selected using a subset of initial multiple machine learning models (which may be referred to as driver models in this specification). A translation axis may refer to a measure of the biological response of a human or non-human model to, for example, an antigen (e.g., the HAI titer obtained from mice exposed to a specific antigen or the HAI titer obtained from collected human serum). A subset of driver models can be selected for use in a reasonable manner by first assigning each of the initial multiple driver models to a class of translation axes, wherein each class of translation axes corresponds to one of the multiple translation axes (e.g., at least one of ferret AF, ferret HAI, mouse AF, mouse HAI, human replica AF, human AF, or human HAI).

[0038] In some implementations, based on time-series data (representing multiple molecular sequences and, for each molecular sequence, the prevalence time of the pathogenic strain as a natural antigen), each driver model is trained to predict molecular sequences that will generate an extreme (e.g., maximized) biological response (e.g., maximized mouse HAI titer) across all pathogenic strains prevalent in a particular disease season, or will generate a response that will effectively cover the maximum number of pathogenic strains prevalent in a particular disease season. In some implementations, for each driver model, a translation model configured to predict biological responses with respect to multiple translation axis pairs of molecular sequences is used to provide feedback in the form of translation response data, which represents one or more translation responses corresponding to a translation axis class assigned to the driver model.

[0039] This process is performed in multiple iterations, where, for each iteration, the driver model updates one or more parameters (often referred to as weights and biases) based on feedback from the translation model. After multiple iterations, a set of trained driver models is selected. For each translation axis, the selected set of trained driver models may include trained driver models that predict molecular sequences that result in the expected (typically: highest) aggregate (e.g., average) biological response (e.g., immune response) as predicted by the translation model for that translation axis. For each trained driver model in the set, the antigen predicted by that trained driver model can then be applied to the translation model, which predicts the response to that antigen with respect to each translation axis.

[0040] Then, a subset of trained driver models from the group is selected. Selecting a subset of trained driver models may include, for each translation axis, selecting a trained driver model from the group that predicts the antigen that elicits the highest aggregate biological response (as predicted by the translation model for that translation axis) across all pathogenic strains during a particular disease season. Each trained driver model in the subset of trained driver models is validated using data observed from human or non-human experiments. If a trained driver model is validated, it can be used to design vaccines based on antigens predicted by the validated trained driver models.

[0041] In the accompanying drawings, for ease of description, specific arrangements or sequences of schematic elements are shown, such as those representing devices, modules, instruction blocks, and data elements. However, those skilled in the art should understand that the specific order or arrangement of schematic elements in the drawings does not imply a requirement for a particular order or sequence of processing or process separation. Furthermore, the inclusion of schematic elements in the drawings does not imply that such elements are necessary in all implementations, or that features represented by such elements may not be included in or combined with other elements in some implementations.

[0042] Additionally, in the accompanying drawings, when connecting elements such as solid or dashed lines or arrows are used to illustrate connections, relationships, or associations between two or more other schematic elements, the absence of any such connecting element does not imply the absence of a connection, relationship, or association. In other words, some connections, relationships, or associations between elements are not shown in the drawings to avoid obscuring the text of this disclosure. Furthermore, for ease of illustration, a single connecting element is used to represent multiple connections, relationships, or associations between elements. For example, when a connecting element represents communication of signals, data, or instructions, those skilled in the art will understand that such an element represents one or more signal paths (e.g., a bus), as needed, to influence communication.

[0043] Reference will now be made in detail to the implementations, embodiments of which are illustrated in the accompanying drawings. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the various described implementations. However, it will be apparent to those skilled in the art that the various described implementations can be practiced without these specific details. In other instances, well-known methods, processes, components, circuits, and networks have not been described in detail to avoid unnecessarily obscuring various aspects of the implementations.

[0044] Several features are described below, each of which can be used independently of each other or in any combination with other features. However, any single feature may not solve any of the problems discussed above, or may only solve one of the problems discussed above. Some of the problems discussed above may not be fully solved by any of the features described herein. Although headings may be provided, data related to a particular heading but not found in the section with that heading may also be found elsewhere in this specification.

[0045] Figure 1 illustrates an example of a system 100 for designing vaccines. System 100 includes a computer processor 110. The computer processor 110 includes a computer-readable storage device 111 and computer-readable instructions 112. System 100 also includes a machine learning system 150. The machine learning system 150 includes a machine learning model 120. The machine learning system 150 may be separate from or integrated with the computer processor 110.

[0046] Computer-readable storage device 111 (or computer-readable medium) may include any type of data storage technology suitable for the local technical environment, including but not limited to semiconductor-based memory devices, magnetic memory devices and systems, optical memory devices and systems, fixed memory, removable memory, disk memory, flash memory, dynamic random access memory (DRAM), static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), etc. In one implementation, computer-readable storage device 111 includes a code segment having executable instructions.

[0047] In some implementations, computer processor 110 includes a general-purpose processor. In some implementations, computer processor 110 includes a central processing unit (CPU). In some implementations, computer processor 110 includes at least one application-specific integrated circuit (ASIC). Computer processor 110 may also include a general-purpose programmable microprocessor, an application-specific programmable microprocessor, a digital signal processor (DSP), a programmable logic array (PLA), a field-programmable gate array (FPGA), special-purpose electronic circuits, or combinations thereof. Computer processor 110 is configured to execute program code means, such as computer-executable instructions 112. In some implementations, computer processor 110 is configured to execute a machine learning model 120.

[0048] Computer processor 110 is configured to receive time-series dataset 161. Time-series dataset 161 may include data representing one or more molecular sequences, and for each of the one or more molecular sequences, one or more times of the prevalence of a pathogenic strain for which the molecular sequence is a natural antigen. As an illustrative example, for A / SINGAPORE / INFIMH160019 / 2016, A / MISSOURI / 37 / 2017, A / KENYA / 105 / 2017, A / MIYAZAKI / 89 / 2017, A / ETHIOPIA / 1877 / 201, A / OSORNO / 60580 / 2017, A / BRISBANE / 1059 / 2017, and A / VICTORIA / 11 / 2017, time-series dataset 161 may indicate the molecular sequences and the time of prevalence (e.g., specific month, specific season of onset, etc.). Although only eight pathogenic strains are described, the time-series dataset 161 can include molecular sequence information and epidemic times corresponding to billions of pathogenic strains. The time-series dataset 161 can be obtained in one or more ways, such as wired or wireless communication with a database (including cloud-based environments), fiber optic communication, Universal Serial Bus (USB), CD-ROM, etc.

[0049] Machine learning system 150 uses machine learning techniques to train machine learning model 120. When the machine learning model is applied to input data, the machine learning model generates an indication of whether the input data item has one or more associated properties, such as the probability that the input data item has a specific Boolean property, an estimate of a scalar property, or an estimate of a vector (i.e., an ordered combination of multiple scalars).

[0050] As part of training the machine learning model 120, the machine learning system 150 can form a training set of input data by identifying a positive training set of input data items that have been determined to have the features in question, and in some implementations, form a negative training set of input data items that lack the features in question.

[0051] Machine learning system 150 extracts feature values ​​from the input data of the training set. These features are variables that are considered potentially related to whether an input data item possesses one or more relevant properties. The ordered list of features of the input data is referred to here as the feature vector of the input data. In some implementations, machine learning system 150 applies dimensionality reduction (e.g., via linear discriminant analysis (LDA), principal component analysis (PCA), deep features learned from a neural network, etc.) to reduce the amount of data in the feature vector of the input data to a smaller, more representative dataset.

[0052] In some implementations, the machine learning system 150 uses supervised machine learning to train the machine learning model 120, where feature vectors from both the positive and negative training sets are used as input. Different machine learning techniques are used in some implementations, such as linear support vector machines (linear SVMs), augmentations for other algorithms (e.g., AdaBoost), neural networks, logistic regression, Naive Bayes, memory-based learning, random forests, bagged trees, decision trees, augmented trees, or augmented stumps. When applied to feature vectors extracted from input data items, the machine learning model 120 outputs an indication of whether the input data item possesses the property in question, such as a Boolean yes / no estimate, a scalar value representing a probability, a vector representing the scalar values ​​of multiple properties, or a nonparametric distribution representing the scalar values ​​of multiple properties that are distinct and not a fixed prior number, which can be explicitly or implicitly represented in a Hilbert or similar infinite-dimensional space.

[0053] In some implementations, the validation set consists of additional input data besides those in the training set, which have been determined to have or lack the feature in question. The machine learning system 150 applies the trained machine learning model 120 to the validation set data to quantify the accuracy of the machine learning model 120. Common measures used in accuracy measurement include: precision = TP / (TP+FP) and recall = TP / (TP+FN), where precision is the number of correct predictions (TP or true positives) made by the machine learning model 120 out of its total predictions (TP+FP or false positives), and recall is the number of correct predictions (TP) made by the machine learning model 120 out of the total number of input data items that actually have the feature in question (TP+FN or false negatives). The F-score (F-score = 2*PR / (P+R)) unifies precision and recall into a single metric. In some implementations, the machine learning system 150 iteratively retrains the machine learning model 120 until a stopping condition occurs, such as an accuracy measurement indicating that the model 120 is sufficiently accurate, or multiple training rounds have already taken place.

[0054] In some implementations, machine learning model 120 includes a neural network. In some implementations, the neural network includes a recurrent neural network (RNN). RNNs typically describe a class of artificial neural networks where connections between nodes form a directed graph along a time series, allowing it to exhibit temporal dynamics. Unlike feedforward neural networks, RNNs can use their internal states (memories) to process sequences of input. In some implementations, RNNs include a long short-term memory (LSTM) architecture. LSTM refers to an RNN architecture with feedback connections, capable of processing not only single data points (such as images) but also entire sequences of data (such as speech or video). Machine learning model 120 may include other types of neural networks, such as convolutional neural networks, radial basis function neural networks, physical neural networks (e.g., optical neural networks), and so on. Exemplary methods for designing and training machine learning model 120 will be discussed in more detail later with reference to Figures 2A through 4.

[0055] Machine learning model 120 is configured to predict one or more molecular sequences based on a received time series dataset 161, and what immunological properties the predicted one or more molecular sequences will confer for use at future times. As an illustrative example, assume the received time series dataset 161 includes data representing multiple pathogenic strains, where each pathogenic strain was found to be prevalent at one or more times between January 1, 2014, and December 31, 2018. Machine learning model 120 can predict one or more molecular sequences (e.g., antigens) based on the pathogenic strains found to be prevalent at one or more times between January 1, 2014, and December 31, 2018, which will generate a maximized aggregate biological response (e.g., a maximized average human HAI titer) across all viruses prevalent between January 1, 2019, and May 31, 2019. Alternatively or concurrently, machine learning model 120 may predict one or more molecular sequences based on pathogenic strains found to be circulating between January 1, 2014, and December 31, 2018. These molecular sequences would generate a biological response that would effectively cover (e.g., effectively immunize against) the maximum number of viruses circulating between January 1, 2019, and May 31, 2019. The predicted molecular sequences could be used to design vaccines against viruses circulating during future time periods (such as January 1, 2019, to May 31, 2019, as in the previous example).

[0056] Figures 2A and 2B illustrate flowcharts of an architecture 200 designed for a system used to design vaccines. Architecture 200 includes multiple driver models 210, a translation model 220, and a feedback and selection module 230. First, the multiple driver models 210 are initiated. Each of the multiple driver models 210 is configured to generate data representing one or more molecular sequences (e.g., antigens) and predictions about what immune properties each molecular sequence will confer for use, as previously discussed with reference to the machine learning model 120 of Figure 1. In the illustrated implementation, the multiple driver models 210 include a first driver model 210a, a second driver model 210b, a third driver model 210c, a fourth driver model 210d, a fifth driver model 210e, a sixth driver model 210f, a seventh driver model 210g, an eighth driver model 210h, a ninth driver model 210i, and a tenth driver model 210j. Although ten driver models are shown, multiple driver models 210 may include more or fewer driver models (e.g., 5 driver models, 30 driver models, 100 driver models, etc.). One or more of the driver models may be, for example, RNNs as described above with reference to Figure 1.

[0057] Translation model 220 is configured to predict biological responses to molecular sequences across multiple translation axes. In the implementation shown, translation model 220 includes a ferret HAI translation axis 220a, a ferret AF translation axis 220b, a mouse HAI axis 220c, a mouse AF translation axis 220d, and a human clone AF translation axis 220e. Although specific translation axes are shown, the implementation is not limited to these specific translation axes. For example, the translation model may additionally or alternatively include a human HAI translation axis, a human AF translation axis, a human clone HAI axis, or combinations thereof. Some implementations of translation model 220 will be discussed in more detail later with reference to Figures 6 through 9.

[0058] Referring to Figure 2A, each of the multiple actuator models 210 is assigned a specific translation axis to the translation model 220. In the implementation shown, the first actuator model 210a and the third actuator model 210c are assigned to the ferret HAI translation axis 220a, the second actuator model 210b and the sixth actuator model 210f are assigned to the ferret AF translation axis 220b, the fourth actuator model 210d and the eighth actuator model 210h are assigned to the mouse HAI translation axis 220c, the fifth actuator model 210e and the ninth actuator model 210i are assigned to the mouse AF translation axis 220d, and the seventh actuator model 210g and the tenth actuator model 210j are assigned to the human replica AF translation axis 220e.

[0059] Each of the multiple driver models 210 receives a first time-series dataset 201. The first time-series dataset 201 may include multiple molecular sequences and the epidemic time of a pathogenic strain containing at least one of the multiple molecular sequences as a natural antigen. As an illustrative example, the first time-series dataset 201 may include molecular sequences and epidemic times of all observed pathogenic strains circulating between January 1, 2014 and December 31, 2018 (which may be referred to as the "pathogenic period"). Based on the received first time-series dataset 201, each of the multiple driver models 220 is capable of generating output data representing one or more molecular sequences. For example, the output data may represent molecular sequences (such as antigens) for each pathogenic season of the pathogenic period. For each pathogenic season, the molecular sequences may be determined based on temporal strain data from one or more pathogenic seasons prior to that pathogenic season by predicting molecular sequences that will generate a maximum aggregate biological response across all viruses circulating in that pathogenic season, and / or will generate a response that effectively covers (e.g., effectively vaccinated against) the maximum number of viruses circulating in that pathogenic season.

[0060] Translation model 220 is capable of receiving output data from each of the plurality of driver models 210 and generating first translation response data for each of the plurality of driver models 210, the first translation response data representing one or more translation responses corresponding to a specific translation axis assigned to that driver model. In the example shown, translation model 220 may receive output data representing one or more predicted molecular sequences from a first driver model 210a and, based on ferret HAI translation axis 220a, predict ferret HAI titers against each of the one or more molecular sequences across all pathogenic strains prevalent in each pathogenic season (that is, predicting the immune response of ferrets exposed to the pathogenic strain after immunization with the predicted molecular sequence for each pathogenic strain in a particular pathogenic season).

[0061] Feedback and selection module 230 receives first translational response data corresponding to each of the plurality of driver models 210, comparing the predicted response for each disease season with a threshold response. For example, for each driver model, feedback and selection module 230 may aggregate (e.g., average) the predicted biological responses for all viruses across each disease season, compare the aggregated responses with a threshold aggregated response, and generate an error value based on the comparison. Alternatively or additionally, for each driver model, feedback and selection module 230 may compare the number of viruses that would be effectively vaccinated against in each disease season with a threshold number, and generate an error value based on the comparison. Feedback and selection module 230 may then cause each driver model to adjust one or more parameters (such as their weights and biases) based on the error value for each disease season. This process is repeated iteratively multiple times. The number of iterations may be a set number of iterations or determined based on a threshold error value (i.e., the process will continue until the threshold error value is exceeded). Therefore, at a high level: (1) each driver model can predict one or more molecular sequences for immunization against pathogens of a specific disease season for a specific disease season based on pathogens of a previous disease season; (2) the performance of each driver model can be evaluated for each disease season; and (3) the parameters of each driver model can be adjusted based on its performance in each disease season.

[0062] After multiple iterations, the performance of each of the driver models (now referred to as the trained driver model) is compared with other driver models assigned to the same translation axis as that driver model, and the driver model exhibiting the best performance is selected to generate a selected set of trained driver models 240. For example, after multiple iterations, the total predicted ferret HAI titer of the molecular sequence predicted by the first driver model 210a can be compared with the total predicted ferret HAI titer of the molecular sequence predicted by the third driver model, and the feedback and selection module 230 can select the driver model corresponding to the highest total predicted ferret HAI titer (or the highest number of pathogenic strains effectively combated by vaccination) across all or some of the pathogenic periods. In the implementation shown, the selected set of driver models 240 includes the first driver model 210a, the second driver model 210b, the fifth driver model 210e, the seventh driver model 210g, and the tenth driver model 210j.

[0063] Referring to Figure 2B, each of the selected set of driver models 240 receives a second time-series dataset 202 and generates trained output data representing one or more molecular sequences for a specific disease-causing season based on the second time-series dataset 202. Similar to the first time-series dataset 201, the second time-series dataset 202 may include data representing molecular sequences and epidemic times of all observed pathogenic strains circulating within a given disease-causing time period. The disease-causing time period of the second time-series dataset 202 may be the same as or different from the disease-causing time period of the first time-series dataset 201. Each of the driver models in the selected set of driver models 240 is capable of predicting one or more molecular sequences (e.g., antigens) for one or more disease-causing seasons. In some implementations, the predicted one or more molecular sequences are one of the disease-causing seasons for a time period (e.g., the most recent disease-causing season). As an illustrative example, assume that the received second time-series dataset 202 includes data representing multiple pathogenic strains, where each pathogenic strain was found to be circulating at one or more times between January 1, 2014 and April 31, 2018. Based on pathogenic strains found to be prevalent during the previous disease season between January 1, 2014 and September 30, 2017, each of the selected set of driver models 240 can predict one or more molecular sequences (e.g., antigens) that will generate a maximized aggregate biological response across all viruses prevalent between October 1, 2017 and April 31, 2019. Alternatively or additionally, based on pathogenic strains found to be prevalent during the previous disease season between January 1, 2014 and September 30, 2017, each of the selected set of driver models 240 can predict one or more molecular sequences that will generate a biological response that will effectively cover (e.g., effectively immunize against) the maximum number of viruses prevalent between October 1, 2017 and April 31, 2018.

[0064] Translation model 220 receives trained output data from each of the selected set of driver models 240 and generates second translation response data for each driver model based on the trained output data. For each driver model, the second translation response data represents one or more translation responses across all translation axes of translation model 220 based on one or more predicted molecular sequences of that driver model. As an illustrative example, translation model 220 may receive trained output data from a first driver model 210a representing one or more molecular sequences. Translation model 220 may predict ferret HAI titers, ferret AF titers, mouse HAI titers, mouse AF titers, and human replica AF titers for one or more molecular sequences predicted by the first driver model 210a across all strains. Feedback and selection module 230 receives the second translation response data for each driver model of the selected set of driver models 240. Feedback and selection module 230 is capable of comparing the performance of each driver model for each translation axis and selecting the driver model with the highest performance on each axis or combination of axes to generate a selected subset 250 of driver models. Using the previous illustrative example, regarding the ferret HAI axis 220a, the feedback and selection module 230 can compare the total HAI titer for one or more molecular sequences predicted by each of the driver models from a selected set of driver models 240 across all pathogenic strains circulating between January 1, 2019 and May 31, 2019. The feedback and selection module 230 can then select the driver model found to have the highest total HAI titer across all pathogenic strains. In the implementation shown, the selected subset 250 of driver models includes a second driver model 210b and a tenth driver model 210j. One or more of the selected subset 250 of driver models can be included in the machine learning model 120 previously discussed with reference to Figure 1.

[0065] Each of the driver models in the selected subset 250 of driver models can then be validated based on observations from real-world experiments. For example, second translation response data corresponding to the second driver model 210b can be compared with biological responses observed in human HAI experiments (or ferret HAI experiments, mouse HAI experiments, etc.), where human subjects are vaccinated with one or more molecular sequences predicted by the second driver model 210b and exposed to one or more pathogenic strains circulating between October 1, 2017 and April 31, 2018. The feedback and selection module 230 can compare the predicted response with the observed response to generate an error value, and the feedback and selection module 230 can determine, based on the error value, whether one or more translation axes corresponding to the second driver model 210b (e.g., ferret HAI translation axis 220a if the second driver model 210b is selected based on its performance on the ferret HAI translation axis 220a) is a good or bad predictor of human responses. If the error value meets the error threshold, one or more molecular sequences predicted by the second driver model 210b can be used to design vaccines for at least the disease seasons of October 1, 2017 and April 31, 2018, or for disease seasons even after those. For example, if the second driver model 210b is validated using a real-world ferret HAI experiment, the determined error value can be used to adjust the parameters of the translation model 220, the second driver model 210b, or both.

[0066] Figure 3 illustrates a flowchart of method 300 for designing a vaccine. For illustrative purposes, method 300 will be described as being performed by the architecture 200 previously described with reference to Figures 2A to 2B. The method includes: applying multiple driver models to a first time series dataset (box 310), training each driver model using the first time series dataset (box 320), selecting a set of trained driver models (box 330), applying the selected set of trained driver models to a second time series dataset (box 340), and selecting a subset of the trained driver models (box 350).

[0067] At box 310, each of the plurality of driver models 210 receives a first time series dataset 201. Based on the received first time series dataset 201, each of the plurality of driver models 220 can generate output data representing one or more molecular sequences.

[0068] At box 320, for each of the driver models 210, the driver model is trained using the translation axis of the translation model 220 assigned to that driver model. Figure 4 shows a flowchart of a method 400 for training one or more driver models to design a vaccine. Referring to Figure 4, method 400 includes receiving output data from each of the driver models of the plurality of driver models 210 (box 410), applying the translation model 220 to the output data to generate first translation response data for each of the driver models of the plurality of driver models 210 according to the translation axis assigned to that driver model (box 420), adjusting one or more parameters of that driver model based on the first translation response data corresponding to that driver model for each of the plurality of driver models 210 (box 430), and repeating boxes 410-430 multiple times (box 440).

[0069] At box 330, for each translation axis of translation model 220, a selected set of driver models 240 is generated based on the performance of the driver models assigned to that translation axis. For example, after multiple iterations, the total predicted ferret HAI titer for a molecular sequence predicted by the first driver model 210a can be compared with the total predicted ferret HAI titer for a molecular sequence predicted by the third driver model 210c, and the feedback and selection module 230 can select the driver model corresponding to the highest total predicted ferret HAI titer (or the highest number of pathogenic strains effectively vaccinated).

[0070] At box 340, each of the selected set of driver models 240 receives a second time series dataset 202 and generates trained output data representing one or more molecular sequences for a specific disease-causing season based on the second time series dataset 202.

[0071] At box 350, translation model 220 receives trained output data from each of the selected set of driver models 240 and generates second translation response data for each driver model based on the trained output data. For each driver model, the second translation response data represents one or more translation responses across all translation axes of translation model 220 based on one or more predicted molecular sequences of that driver model. As an illustrative example, translation model 220 may receive trained output data from a first driver model 210a representing one or more molecular sequences. Translation model 220 may predict ferret HAI titers, ferret AF titers, mouse HAI titers, mouse AF titers, and human replica AF titers for one or more molecular sequences predicted by the first driver model 210a. Feedback and selection module 230 receives the second translation response data for each driver model of the selected set of driver models 240. Feedback and selection module 230 is capable of comparing the performance of each driver model for each translation axis and selecting the driver model with the highest performance on each axis to generate a selected subset 250 of driver models.

[0072] Figure 5 shows a graph depicting the improvement over conventional techniques used for vaccine design along translation axes. In an exemplary experiment, five (5) different vaccine candidates (MO / 17, OS / 17, MI / 17, ET / 17, and KE / 17, homologous to strains A / MISSOURI / 37 / 2017, A / OSORNO / 60580 / 2017, A / MIYAZAKI / 89 / 2017, A / ETHIOPIA / 1877 / 2017, and A / KENYA / 105 / 2017, respectively) were selected using the aforementioned process and evaluated against five (5) different translation axes (displayed across the x-axis) relative to a conventionally selected CV, A / SINGAPORE / INFIMH160019 / 2016. Each of the five different CVs selected using the system and methods described in this specification is shown as a labeled label for each of the translation axes, with slight jitter within each translation axis for visual clarity. For each translation axis, the y-axis indicates the fraction of clinical isolates (referred to in the sequence as “seasonal representative strains”) reported in the Global Initiative on Sharing All Influenza Data (GISAID) global database as of April 15, 2018, that were predicted to be better protected against a specific antigen than conventionally selected CVs (A / SINGAPORE / INFIMH160019 / 2016), which were the standard of care (SOC) for H3N2 as of March 2018. For example, the leftmost column (ferret HAI) shows that the translation model predicted that A / MISSOURI / 37 / 2017 would produce antibodies in ferrets with consistently higher HAI titers against all of these seasonal representative strains compared to conventionally selected CVs. As another example, in the far right column (human serum antibody evidence (AF)), A / ETHIOPIA / 1877 / 2017 and A / OSORNO / 60580 / 2017 were predicted to be non-inferior to CVs selected in the conventional manner. These results, taken together, suggest that these five candidates will exhibit dissimilar and different non-inferiority patterns of induced immune responses, as assessed by different shift axes.

[0073] Exemplary translation model:

[0074] Figure 6 illustrates an example of a system 600 that uses machine learning techniques to predict biological responses according to one or more embodiments of this disclosure. As previously discussed, system 600 can be used as a translation model. System 600 includes a computer processor 610. Computer processor 610 includes computer-readable storage 611 and computer-readable instructions 612. System 600 also includes a machine learning system 650. Machine learning system 650 includes a machine learning model 620. Machine learning system 650 can be decoupled from or integrated with computer processor 610.

[0075] Computer-readable storage device 611 (or computer-readable medium) may include any type of data storage technology suitable for the local technical environment, including but not limited to semiconductor-based memory devices, magnetic memory devices and systems, optical memory devices and systems, fixed memory, removable memory, disk memory, flash memory, dynamic random access memory (DRAM), static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), etc. In some implementations, computer-readable storage device 611 includes a code segment with executable instructions.

[0076] In some implementations, the computer processor 610 includes a general-purpose processor. In some implementations, the computer processor 610 includes a central processing unit (CPU). In some implementations, the computer processor 610 includes at least one application-specific integrated circuit (ASIC). The computer processor 610 may also include a general-purpose programmable microprocessor, an application-specific programmable microprocessor, a digital signal processor (DSP), a programmable logic array (PLA), a field-programmable gate array (FPGA), special-purpose electronic circuits, or combinations thereof. The computer processor 610 is configured to execute program code means, such as computer-executable instructions 612. In some implementations, the computer processor 610 is configured to execute a machine learning model 620.

[0077] Computer processor 610 is configured to acquire first molecular sequence data 661 of a first molecular sequence and second molecular sequence data 662 of a second molecular sequence. First molecular sequence data 661 may include amino acid sequence data of a candidate antigen (e.g., an inoculated strain). The candidate antigen may correspond to, for example, the H3N1 virus. Second molecular sequence data 662 may include amino acid sequence data of a known viral strain for which protection is sought. For example, the second molecular sequence may be a known viral strain that emerged in 2001. In some implementations, as will be explained in further detail later with reference to FIG9, computer processor 610 is also configured to receive non-human biological response data associated with the first and second molecular sequences. Non-human biological response data may include, for example, a biological response readout (e.g., antibody titer) measuring the biological response of a non-human model (e.g., mouse, ferret, human immune system replica, etc.) to the second molecular sequence after inoculation with the first molecular sequence. As discussed in further detail later with reference to FIG9, in some implementations, computer processor 610 is capable of encoding first molecular sequence data 661 and second molecular sequence data 662 as amino acid mismatches. The aforementioned data can be obtained in one or more ways, such as wired or wireless communication with databases (including cloud-based environments), fiber optic communication, Universal Serial Bus (USB), CD-ROM, etc.

[0078] The machine learning system 650 uses machine learning techniques to train a machine learning model 620. When the machine learning model is applied to input data, the machine learning model generates an indication of whether an input data item has one or more associated properties, such as the probability that the input data item has a specific Boolean property, or an estimate of a scalar property.

[0079] As part of training the machine learning model 620, the machine learning system 650 can form a training set of input data by identifying a positive training set of input data items that have been determined to have the features in question, and in some implementations, form a negative training set of input data items that lack the features in question.

[0080] The machine learning system 650 extracts feature values ​​from the input data of the training set. These features are variables that are considered potentially related to whether an input data item possesses one or more relevant properties. The ordered list of features of the input data is referred to here as the feature vector of the input data. In some implementations, the machine learning system 650 applies dimensionality reduction (e.g., via linear discriminant analysis (LDA), principal component analysis (PCA), deep features learned from a neural network, etc.) to reduce the amount of data in the feature vector of the input data to a smaller, more representative dataset.

[0081] In some implementations, the machine learning system 650 uses supervised machine learning to train the machine learning model 620, where feature vectors from both the positive and negative training sets are used as input. Different machine learning techniques are used in some implementations, such as linear support vector machines (linear SVMs), augmentations for other algorithms (e.g., AdaBoost), neural networks, logistic regression, Naive Bayes, memory-based learning, random forests, bagged trees, decision trees, augmented trees, or augmented stumps. When applied to feature vectors extracted from input data items, the machine learning model 620 outputs an indication of whether the input data item possesses the property in question, such as a Boolean yes / no estimate, a scalar value representing a probability, a vector representing the scalar values ​​of multiple properties, or a nonparametric distribution representing the scalar values ​​of multiple properties that are distinct and not a fixed prior number, which can be explicitly or implicitly represented in a Hilbert or similar infinite-dimensional space.

[0082] In some implementations, the validation set consists of additional input data besides those in the training set, which have been determined to have or lack the feature in question. The machine learning system 650 applies the trained machine learning model 620 to the validation set data to quantify the accuracy of the machine learning model 620. Common measures used in accuracy measurement include: precision = TP / (TP+FP) and recall = TP / (TP+FN), where precision is the number of correct predictions (TP or true positives) made by the machine learning model 620 out of its total predictions (TP+FP or false positives), and recall is the number of correct predictions (TP) made by the machine learning model 620 out of the total number of input data items that actually have the feature in question (TP+FN or false negatives). The F-score (F-score = 2*PR / (P+R)) unifies precision and recall into a single metric. In some implementations, the machine learning system 650 iteratively retrains the machine learning model 620 until a stopping condition occurs, such as an accuracy measurement indicating that the model 620 is sufficiently accurate, or multiple training rounds have already taken place.

[0083] In some implementations, the machine learning model 620 includes a neural network. In some implementations, the neural network includes a convolutional neural network. The machine learning model 620 may include other types of neural networks, such as recurrent neural networks, radial basis function neural networks, physical neural networks (e.g., optical neural networks), and so on. Specific methods for training machine learning models according to one or more implementations of this disclosure will be discussed in more detail later with reference to Figures 8 and 9.

[0084] Machine learning model 620 is configured to predict a biological response 663 to a second molecular sequence based on the received data. For example, suppose first molecular sequence data 661 represents the amino acid sequence of a candidate antigen to be used for vaccination, and second molecular sequence data 662 represents the amino acid sequence of a viral strain known to have been circulating in 2012. If the human immune system is inoculated with the first molecular sequence (i.e., the candidate antigen), machine learning model 620 can predict the biological response (e.g., antibody titer) that the human immune system will produce upon encountering the second molecular sequence (e.g., the known viral strain).

[0085] Figure 7 illustrates a flowchart depicting an example of a method 700 for predicting a biological response using machine learning techniques according to one or more implementations of this disclosure. For illustrative purposes, method 700 is described as being performed by system 600, which uses machine learning techniques previously discussed with reference to Figure 6 to predict a biological response. Method 700 includes receiving first sequence data of a first molecular sequence (box 710), receiving second sequence data of a second molecular sequence (box 720), and predicting a biological response to the second molecular sequence (box 730).

[0086] At box 710, computer processor 710 receives first molecular sequence data 161 of the first molecular sequence. As previously indicated, the first molecular sequence data 161 may include amino acid sequence data of a candidate antigen (e.g., an inoculum). For example, the candidate antigen may correspond to the H3N1 virus.

[0087] At box 720, computer processor 720 receives second molecular sequence data 662 for the second molecular sequence. The second molecular sequence data 662 may include amino acid sequence data for a known viral strain for which protection is sought. For example, the second molecular sequence may be a known viral strain that emerged in 2001.

[0088] In some implementations, method 700 further includes encoding the first molecular sequence data 661 and the second molecular sequence data 662 as amino acid mismatches. For example, similar regions of the first and second molecular sequences can be compared, and each mismatched amino acid pair in a region can be encoded with a "1" value, while each matching amino acid pair in a region can be encoded with a "0" value. Thus, the dissimilarity between the first and second molecular sequences, as defined by mismatched amino acids at positions within similar regions between the molecular sequences, can be provided to machine learning model 620.

[0089] In some implementations, method 700 further includes receiving non-human biological response data associated with the first molecular sequence and the second molecular sequence. The non-human biological response data may include, for example, a biological response readout (e.g., antibody titer) measuring the biological response of a non-human model (e.g., mouse, ferret, human immune system clone, etc.) to the second molecular sequence after inoculation with the first molecular sequence.

[0090] At box 730, machine learning model 620 predicts a biological response to the second molecular sequence based on the received data. For example, if a human immune system has been inoculated with the first molecular sequence (i.e., the candidate antigen), machine learning model 620 can predict the biological response (e.g., antibody titer) that the human immune system will produce upon encountering the second molecular sequence (i.e., the known viral strain). In some implementations, machine learning model 620 is configured to predict a non-human biological response to the second molecular sequence. For example, if an animal's immune system has been inoculated with the first molecular sequence, the machine learning model can predict the antibody titer that the animal's immune system (e.g., a mouse, a ferret, etc.) will produce upon encountering the second molecular sequence.

[0091] Methods for training machine learning models to predict biological responses:

[0092] Methods for training a machine learning model 620 to predict biological responses will now be described. Figure 8 illustrates examples of data used to train a machine learning model for predicting biological responses, according to one or more implementations of this disclosure. As shown, data from thousands (or millions, billions, etc.) of experiments can be used to construct data from, for example, ferrets, mice, and in vitro replicas of the human immune system (e.g., A comprehensive repository of biological response readouts and viral sequence data for a model is provided. In the illustrated embodiment, the data includes antigen sequence data, viral sequence data, and biological response readouts measured by hemagglutination inhibition assay (HAI) and antibody evidence collection (AF). The viral sequence data includes a set of known viral strains (referred to as the "readout" group). Experiments may be divided into batches referred to as "cycles" (e.g., cycle 1 and cycle 2). In each cycle, the model system is excited with a selected molecular sequence (e.g., H3 protein, vaccine formulation, etc.), and the ability of the model system to generate an immune response against a set of "readout" viral strains (referred to as the "readout group") is measured. The viral readout group can be selected to represent a broad sampling of influenza strains circulating during a defined period (e.g., 1950 to 2016).

[0093] To correlate model experiments with human outcomes, human serum can be measured against the "readout" group. In the example shown, for each pair of antigen / readout strains tested in the model system, there is not always a corresponding pair in the human serum measurement. This is because human samples may be collected from vaccinated individuals, and the vaccination period does not cover all years used in each cycle. Therefore, the machine learning model can be restricted to antigens and readouts tested only in human serum, and the vector of human readout titers can be chosen as the target vector of the machine learning model. Human AF readouts can come from human serum collected on day 21 post-vaccination, a time typically sufficient for the subject to undergo seroconversion after vaccination.

[0094] Using the data obtained from the experiments described above, models can be trained to predict biological responses. In some implementations, linear models can be used.

[0095] Figure 9 illustrates a flowchart of an example of training a machine learning model to predict biological responses according to one or more implementations of this disclosure. As shown, a data matrix 900 is first prepared, where each row corresponds to a pair of viral antigens, such as an antigen strain and an H3 region of a “readout” strain. The columns (or features) of the matrix include specific columns for AF readout titers 902 for ferret models and AF readout titers 903 for mouse models. In some implementations, missing titer data is estimated using the average of the columns. However, many standard methods can be used to estimate missing titer data. Sequence column 901 represents the amino acid sequence difference (SeqDiff) between the antigen strain and the “readout” strain in a selected region, which in the example shown includes the H3 regions of the antigen strain and the “readout” strain. The SeqDiff is prepared by checking at each position of the H3 amino acid sequence alignment whether the amino acids are the same or different between the antigen strain and the “readout” strain. If the amino acids between the two strains are different, a “1” can be encoded. If the amino acids between the two strains are the same, a “0” can be encoded. Encoding two sequences as amino acid mismatches essentially creates a protein Hamming distance measure, which typically reflects the number of different positions corresponding to the amino acids. In some implementations, columns that are consistently "0" throughout the training set are discarded. Linear regression is used to associate columns 901, 902, and 903 of each row with the corresponding human titer 904.

[0096] Columns 902 and 903, including the titer readouts, can undergo z-score transformation, for example, before fitting a linear regression model. The z-score can refer to the linearly transformed data value with a mean of zero and a standard deviation of one, and can indicate how many standard deviations above or below the mean the observation is. Because the encoding of a SeqDiff representation can be sparse, in some cases, principal component analysis (PCA) can be used to reduce the dimension of the SeqDiff vector to five components. PCA is a statistical process that uses orthogonal transformations to convert observations of a set of potentially correlated variables into values ​​of a set of linearly uncorrelated variables called principal components. PCA can be used to emphasize variation, highlight strong patterns in a dataset, and reduce a large number of variables to a smaller set without losing a significant amount of information from a larger set. Linear models can be trained on various combinations of data to better understand the relative ability of mouse titers, ferret titers, and sequence data to predict human responses.

[0097] While machine learning models can be constructed as linear models to predict biological responses, as previously mentioned, nonlinear relationships may exist between data features and human biological responses. Therefore, using data from the aforementioned experiments, models employing deep neural networks or other nonlinear models can be constructed that 1) leverage nonlinear relationships in the data to make relatively accurate predictions compared to the aforementioned linear models, and 2) simultaneously predict both animal and human titers. Predicting all titers together leverages the understanding that strong signals of the immune response can be directly encoded in the protein sequences of both the antigenic strain and the “readout” strain. By training the model to predict titers for both humans and animals solely from sequences, the machine learning model can be forced to search for sequence-function relationships driving cross-species immunogenicity. In statistical terms, this can be called “borrowing strength” and allows the model to better utilize the large amount of available data for one type of model (e.g., the ferret model) to generate more robust predictions of human responses. This strategy can be adapted to more viral antigens and the construction of data matrices with over 13,000 example rows. Similar to the linear model, the SeqDiff representation of the H3 region for each virus strain and readout pair can be used as input data.

[0098] Although in some implementations, the target vector is the human titer for linear models, nonlinear neural network models can represent a multi-objective regression problem using, for example, seven output columns (ferret HAI and AF titers, mouse HAI and AF titers, MIMIC AF, human HAI, and human AF). Since the detection limit for HAI experiments is typically 40 (or, when expressed as a dilution, 1:40), any measurement below this value can be set to 40. Similarly, if the AF measurement is below this value, it can be set to 10000. HAI can be expressed as log2(titer / 10), and AF as log2(titer). Human and human replica data can have additional complexity if measurements are taken at inoculation (day 0) and after seroconversion (day 21). Therefore, human and human replica titers can be expressed as a log2-fold change from day 0 to day 21. In cases where titer values ​​are missing from the target vector, these values ​​can be set to zero, and the loss function in the neural network can be masked for these locations. This ensures that the prediction of missing values ​​does not affect the model's fitness during training.

[0099] In some implementations, a neural network with two 128-node dense layers with ReLU activations and a 7-node dense output layer can be used. A portion of the data (e.g., 15%) can be randomly excluded as a test set, and the neural network can be trained for multiple epochs (e.g., 400, 500, 1000, etc.). In some implementations, the following parameters are used: learning rate = 0.001; weight decay = 0.0001; batch size = 128.

[0100] In some implementations, the L2 loss function is used for human replicas, human AF, and human HAI target vectors. Typically, the L2 loss function minimizes the squared difference between the estimated target value and the existing target value. In some implementations, the Huber loss function can be used for ferret and mouse data. The Huber loss function is typically used in robust regression, and in at least some cases, it is less sensitive to outliers in the data compared to the L2 loss function. To further bias the model, an explicit weighting scheme can be used to impose additional penalties on misclassified human samples. For example, at each epoch of training, the following weights can be multiplied by each target loss: ferret HAI = 0.8; ferret AF = 1; mouse HAI = 1; mouse AF = 1; human HAI = 2; human AF = 2; MIMIC = 1.5.

[0101] Although for illustrative purposes, the preceding descriptions may sometimes describe pathogenic strains within the context of influenza strains, the term pathogenic should be interpreted broadly to encompass any infectious agent. For example, a pathogenic strain can refer to a viral strain, bacterial strain, protozoan strain, prion strain, viroid strain, or true bacterial strain. A pathogenic strain can correspond to respiratory syncytial virus and other paramyxoviruses. A pathogenic strain can correspond to pertussis, diphtheria, or tetanus, etc.

[0102] While the foregoing description may sometimes refer to the causal season within the context of influenza season, the term causal season should be interpreted broadly to encompass any discrete time interval. For example, a causal season can refer to a specific month, a specific week, a specific series of weeks, a specific series of months, a specific series of days, and so on. Furthermore, consecutive causal seasons can be consistent or variable. For instance, two consecutive causal seasons can both be one month long, or one causal season can be one month long while the second causal season can be four days long.

[0103] While the foregoing descriptions have outlined certain translational axes / bioresponses, such as ferret HAI titers and mouse AF titers, implementation is not limited to these. For example, a bioresponse / translation axis can correspond to antibody characterization (such as affinity and / or affinity for specific antigens and / or sets of antigen fragments (e.g., protein arrays, phage display libraries, etc.)), functional profiling (such as identifying anti-drug antibodies, immune complement interactions (e.g., phagocytosis, inflammation, membrane attack), antibody-dependent cytotoxicity (ADCC), or similar Fc-mediated effector functions), profiling of formed immune complexes (e.g., receptor binding profiles), immunoprecipitation assays, or combinations thereof. A bioresponse / translation axis can also correspond to antibodies competing with other antibodies or antisera for binding targets. The biological response / translation axis can correspond to antiserum characterization (which can correspond to those of the aforementioned antibody characterization), as well as functional assays (such as microneutralization assays, hemagglutination inhibition, and neuraminidase inhibition), binding assays (such as hemagglutination assays), enzyme reaction assays (such as enzyme-linked lectin assays (ELLA)), ligand binding assays (such as binding of sialic acid derivatives and their mimics), and fluorescence readout assays (such as cleavage of 20-(4-methylumbelliferyl)-aDN-acetylneuraminic acid (MUNANA)).

[0104] The biological response / translation axis can correspond to in vivo assessment using monoclonal or polyclonal antibodies, achieved through passive transfer and / or exogenous expression or transfer via one or more of the following mechanisms: retroviral infection-mediated transfection or endogenous expression, or host genome modification (such as via CRISPR), fluid transfer between two individuals, or a combination of these. The biological response / translation axis can correspond to in vivo assessment of immunity induced by immunization to evaluate antigenicity. The biological response / translation axis can correspond to characterization such as binding / affinity measurements of linear peptide antigens on major histocompatibility complex (MHC) classes I and II, and can also assess productive T cell epitope display regarding T cell recognition. The biological response / translation axis can correspond to characterization, such as affinity for antigen fragment sets (e.g., protein arrays, phage display libraries, etc.), to identify recognized epitopes. The biological response / translation axis can correspond to ex vivo and / or in vitro functional profiling, such as determining T cell responses and / or mediated responses. The bioresponse / translation axis can correspond to in vivo and / or in situ measurements of the proliferation (e.g., abundance in tissue compartments) of adaptive response-related T cells (e.g., αβ or γδ T cells) in response to natural infection and / or provocation and / or immunity. The bioresponse / translation axis can also correspond to in vitro and / or ex vivo measurements of recognition specificity of adaptive response-related T cells (e.g., αβ or γδ T cells) in response to natural infection and / or provocation and / or immunity, measured through competition with other epitopes.

[0105] A biological response / translation axis can correspond to in situ, in vitro, and / or in vivo assessments of morphological or physiological changes in tissue formation, tissue repair, or tissue invasion caused by a pathogen or surrogate to be antagonized (such as a pseudovirus or bacteria). A biological response / translation axis can correspond to in situ, in vitro differences in protein, gene expression, and / or non-coding RNA levels relative to other antigens and / or physiological states, characterized, for example, by biomarkers such as age, sex, frailty, nominal serological status, ethnicity, haplotype, and geographic location. A biological response / translation axis can correspond to in situ assessments of protective, transmission, or other overall physiological responses to infections that occur naturally or are transmitted through humans or model organisms (such as, but not limited to, mice, rats, rabbits, ferrets, guinea pigs, pigs, cattle, chickens, sheep, porpoises, bats, dogs, cats, zebrafish and other bony fish, and non-human primates such as monkeys and apes).

[0106] Regarding the response to intentional infection (i.e., challenge) with homotypic and / or heterotypic infectious agents included in controlled human challenge studies, the bioresponse / translation axis may correspond to the in situ, in vitro, and / or in vivo assessment of proteins or metabolites present in blood or tissues, where proteins may be cytokines, hormones, or signaling molecules, and where metabolites may be vitamins, cofactors, or other metabolic byproducts. The bioresponse / translation axis may correspond to the in situ, in vitro, and / or in vivo assessment of the microbiome that may be affected by or influence the immune response. The bioresponse / translation axis may correspond to the in vitro functional profile, in vivo phenotypic profile, and / or functional T cell response profile (receptor expression, cytokine production, cytotoxic potential) in response to a single antigen or in combination with an antigen and innate immune cells (such as natural killer (NK) cells, dendritic cells (DCs), neutrophils, macrophages, monocytes, etc.). The bioresponse / translation axis may correspond to epigenetic analyses performed using samples collected or generated using the techniques or methods described above.

[0107] While the foregoing descriptions outlined some methods and data used to train machine learning models to predict biological responses, other methods and data can also be used. For example, neural network models may include more or fewer layers than the previously described models, with each layer potentially having more or fewer nodes.

[0108] In the foregoing description, numerous specific details have been described with reference to the implementation, which may vary depending on the implementation. Therefore, the specification and drawings are intended to be illustrative, not restrictive. The sole and exclusive indication of the scope of this disclosure, and the scope of this disclosure as desired by the applicant, is the literal and equivalent scope of the claims (including any subsequent amendments) given in the specific form set forth in the claims of this application. Any definitions expressly set forth herein for terms included in the claims will determine the meaning of those terms as used in the claims. Furthermore, when we use the term “further includes” in the foregoing description or in the subsequent claims, what follows this phrase may be an additional step or entity, or a sub-step / sub-entity of the foregoing step or entity.

Claims

1. A method for designing a vaccine, comprising: Multiple driver models are applied to a first time series dataset, the multiple driver models being configured to generate output data representing one or more molecular sequences, the first time series dataset indicating: one or more molecular sequences, and for each of the one or more molecular sequences, including one or more times of the prevalence of a pathogenic strain in which the molecular sequence is a natural antigen; For each of the plurality of driver models, the driver model is trained in the following manner: i) Receive output data from the driver model, the output data representing one or more predicted molecular sequences based on the received first time series dataset; ii) Applying a translation model to output data representing one or more predicted molecular sequences, the translation model being configured to predict biological responses to molecular sequences with respect to multiple translation axes based on the one or more predicted molecular sequences in the output data, to generate first translation response data representing one or more first translation responses corresponding to a specific translation axis among the multiple translation axes; iii) Adjust one or more parameters of the driver model based on the first translation response data; as well as iv) Repeat steps i-iii multiple times to generate trained translation response data representing one or more trained translation responses corresponding to the specific translation axis; Based on the one or more trained translation responses, select a set of trained driver models from the plurality of driver models; For each trained driver model in the group: The trained driver model is applied to a second time series dataset to generate trained output data representing one or more predicted molecular sequences for a specific season; The translation model is applied to the trained output data to generate second translation response data, the second translation response data representing one or more second translation responses about each of the plurality of translation axes; as well as Based on the second translation response data, a subset of trained driver models from the trained driver models of the group is selected.

2. The method of claim 1, wherein at least one of the plurality of driver models comprises a recurrent neural network.

3. The method of claim 1, wherein at least one of the plurality of driver models comprises a long short-term memory recurrent neural network.

4. The method of claim 1, wherein the output data representing one or more predicted molecular sequences based on the received first time series dataset includes output data representing antigens for each of a plurality of pathogenic seasons.

5. The method of claim 4, wherein the output data representing the antigens of each of a plurality of pathogenic seasons comprises antigens determined by predicting molecular sequences that will generate a maximum aggregate biological response across all pathogenic strains prevalent in a particular season.

6. The method of claim 4, wherein the output data representing the antigens of each of a plurality of disease-causing seasons comprises antigens determined by predicting molecular sequences that will generate an effective immune response against the maximum number of viruses circulating in a particular season.

7. The method of claim 1, wherein the plurality of translation axes comprises at least one of the following: ferret antibody evidence (AF) axis, ferret hemagglutination inhibition assay (HAI) axis, mouse AF axis, mouse HAI axis, human clone AF axis, human AF axis, or human HAI axis.

8. The method of claim 1, wherein the number of iterations is based on a predetermined number of iterations.

9. The method of claim 1, wherein the number of iterations is based on a predetermined error value.

10. The method of claim 1, wherein the one or more first translation responses comprise at least one of: a predicted ferret HAI titer, a predicted ferret AF titer, a predicted mouse AF titer, a predicted mouse HAI titer, a predicted human replica AF titer, a predicted human AF titer, or a predicted human HAI titer.

11. The method of claim 1, wherein selecting the group of trained driver models from the plurality of driver models comprises: Each of the plurality of driver models is assigned to a class of driver models, wherein each class is associated with a specific translation axis among a plurality of translation axes used to train the driver model; as well as For each of the plurality of driver models, one or more trained translation responses of that driver model are compared with one or more trained translation responses of at least one other driver model of the same class as that driver model.

12. The method of claim 1, further comprising, for each trained driver model in the subset of the trained driver models: The trained driver model is validated by comparing the second translational response data corresponding to the trained driver model with the observed experimental response data; and In response to validating the trained driver model, a vaccine is generated, the vaccine comprising one or more molecular sequences represented by training output data corresponding to the trained driver model.

13. A system for designing a vaccine, comprising: One or more processors; as well as A computer memory storing executable computer instructions, wherein when the executable computer instructions are executed by the one or more processors, the one or more processors cause the one or more processors to perform operations including: Multiple driver models are applied to a first time series dataset, the multiple driver models being configured to generate output data representing one or more molecular sequences, the first time series dataset indicating: one or more molecular sequences, and for each of the one or more molecular sequences, including one or more times of the prevalence of a pathogenic strain in which the molecular sequence is a natural antigen; For each of the plurality of driver models, the driver model is trained in the following manner: i) Receive output data from the driver model, the output data representing one or more predicted molecular sequences based on the received first time series dataset; ii) Applying a translation model to output data representing one or more predicted molecular sequences, the translation model being configured to predict biological responses to molecular sequences with respect to multiple translation axes based on the one or more predicted molecular sequences in the output data, to generate first translation response data representing one or more first translation responses corresponding to a specific translation axis among the multiple translation axes; iii) Adjust one or more parameters of the driver model based on the first translation response data; as well as iv) Repeat steps i-iii multiple times to generate trained translation response data representing one or more trained translation responses corresponding to the specific translation axis; Based on the one or more trained translation responses, select a set of trained driver models from the plurality of driver models; For each trained driver model in the group: The trained driver model is applied to a second time series dataset to generate trained output data representing one or more predicted molecular sequences for a specific season; The translation model is applied to the trained output data to generate second translation response data, the second translation response data representing one or more second translation responses about each of the plurality of translation axes; as well as Based on the second translation response data, a subset of trained driver models from the trained driver models of the group is selected.

14. The system of claim 13, wherein at least one of the plurality of driver models comprises a recurrent neural network.

15. The system of claim 13, wherein at least one of the plurality of driver models comprises a long short-term memory recurrent neural network.

16. The system of claim 13, wherein the output data representing one or more predicted molecular sequences based on a received first time series dataset includes output data representing antigens for each of a plurality of pathogenic seasons.

17. The system of claim 16, wherein the output data representing the antigens of each of a plurality of pathogenic seasons comprises antigens determined by predicting molecular sequences that generate a maximum aggregate biological response across all pathogenic strains prevalent in a particular season.

18. The system of claim 16, wherein the output data representing antigens for each of a plurality of disease-causing seasons comprises antigens determined by predicting molecular sequences that will generate an effective immune response against the maximum number of viruses circulating in a particular season.

19. The system of claim 13, wherein the plurality of translation axes comprises at least one of: ferret antibody evidence (AF) axis, ferret hemagglutination inhibition assay (HAI) axis, mouse AF axis, mouse HAI axis, human clone AF axis, human AF axis, or human HAI axis.

20. The system of claim 13, wherein the number of iterations is based on a predetermined number of iterations.

21. The system of claim 13, wherein the number of iterations is based on a predetermined error value.

22. The system of claim 13, wherein the one or more first translation responses comprise at least one of: a predicted ferret HAI titer, a predicted ferret AF titer, a predicted mouse AF titer, a predicted mouse HAI titer, a predicted human replica AF titer, a predicted human AF titer, or a predicted human HAI titer.

23. The system of claim 13, wherein selecting the group of trained driver models from the plurality of driver models comprises: Each of the plurality of driver models is assigned to a class of driver models, wherein each class is associated with a specific translation axis among a plurality of translation axes used to train the driver model; as well as For each of the plurality of driver models, one or more trained translation responses of that driver model are compared with one or more trained translation responses of at least one other driver model of the same class as that driver model.

24. The system of claim 13, wherein the operation further comprises, for each trained driver model in the subset of the trained driver models: The trained driver model is validated by comparing the second translational response data corresponding to the trained driver model with the observed experimental response data; and In response to validating the trained driver model, a vaccine is generated, the vaccine comprising one or more molecular sequences represented by training output data corresponding to the trained driver model.

25. A system comprising: A computer-readable storage medium, the computer-readable storage medium including computer-executable instructions; as well as At least one processor, the at least one processor being configured to perform operations including: Receive time series data indicating: a plurality of molecular sequences, and for each of the plurality of molecular sequences, one or more times during which the molecular sequence was prevalent as a natural antigen of a pathogenic strain in a previous pathogenic period; Machine learning training techniques are used to train a machine learning model by performing the following operations in each of multiple training iterations: The time series data is processed using the machine learning model to generate one or more candidate molecular sequences; For each candidate molecular sequence, a biological response score is generated, which defines the predicted biological response of an entity inoculated with the candidate molecular sequence to the pathogen. as well as Adjust one or more parameters of the machine learning model to optimize the loss function that depends on the biological response score; and By processing the time-series data using a trained machine learning model, one or more molecular sequences are predicted that will generate a maximum aggregate biological response to pathogenic strains prevalent in future time periods following the previous pathogenic period.

26. The system of claim 25, wherein predicting the one or more molecular sequences comprises predicting one or more molecular sequences that generate a maximum aggregate biological response across all pathogenic strains of the time series data.

27. The system of claim 25, wherein predicting the one or more molecular sequences comprises predicting one or more molecular sequences that will generate a biological response that will effectively cover the maximum number of pathogenic strains in the time series data.

28. The system of claim 25, wherein one or more predicted molecular sequences are capable of being used to design a vaccine against a pathogenic strain that is prevalent in a future time period following the previously pathogenic time period.

29. The system of claim 25, wherein the machine learning model comprises a recurrent neural network.

Images