Classification of images of dose-response graphs

Abstract

A computer-implemented method for classifying an image comprising a dose-response graph obtained from a dose-response experiment. The method includes receiving, in a curve shape classifier model, an input comprising image data comprising a plurality of pixels, where the image data represents an image of a dose-response graph illustrating the relationship between the concentration of a compound and its activity. The curve shape classifier model includes a neural network model configured to classify the image of the dose-response graph into a plurality of dose-response graph categories related to the curve shape. The method further includes using the neural network model to generate a classification output for the image represented by the received image data, where the generating includes processing the image data using one or more layers of the neural network model according to parameters associated with the one or more layers.

Description

[Technical Field]

[0001] The present specification relates to a computer-implemented method for classifying images containing dose-response graphs obtained from multiple dose-response experiments. [Background technology]

[0002] Early in a drug discovery project, potential drug targets are identified. Drug targets are molecules in the body that are intrinsically linked to a particular disease process. Depending on the disease to be treated, the target can be a protein (e.g., a receptor protein, an enzyme, an ion channel, or a transport protein) or a nucleic acid (e.g., DNA). Researchers hypothesize that by using a drug to modify the activity of the target, the desired therapeutic effect will be achieved.

[0003] Once a target has been identified, a test system that produces a detectable signal is identified or developed to evaluate the effect of a compound on the drug target. This test system is called an assay. Once an assay has been identified or developed, researchers can use it to identify compounds with the desired activity. Typically, compounds are tested at multiple concentrations, and a dose-response (DR) graph can be generated. Analysis of the DR graph allows researchers to determine whether a compound is active, and at what concentrations.

[0004] When it is desirable to test a large number of potential compounds, high-throughput screening (HTS) is often used. It uses robotics, data processing / control software, liquid handling devices, and sensitive detectors, allowing researchers to rapidly conduct thousands or even millions of screening tests. When all compounds are tested at multiple concentrations to generate DR data, this is referred to as quantitative HTS (qHTS). Summary of the Invention [Problem to be solved by the invention]

[0005] However, the large amount of data generated during the dose-response (DR) phase of an HTS campaign requires careful analysis by researchers to detect artifacts and correct erroneous data points before validating the experiment. This process, which requires expert review of each DR experiment, is time-consuming and prone to human error or inconsistency. [Means for solving the problem]

[0006] This specification describes a computer-implemented method for classifying images containing dose-response graphs obtained from dose-response experiments. The method includes receiving, in a curve shape classifier model, input including image data comprising a plurality of pixels, where the image data represents an image of a dose-response graph showing the relationship between the concentration of a compound and its activity. The curve shape classifier model includes a neural network model configured to classify the image of the dose-response graph into a plurality of dose-response graph categories related to curve shapes. In particular, each dose-response graph category can be associated with a respective curve shape that defines the dose-response relationship for that category.

[0007] The method includes generating a classification output for an image represented by received image data using a neural network model, wherein the generating includes processing the image data using one or more layers of the neural network model according to parameters associated with the one or more layers. The neural network model can include a convolutional neural network model.

[0008] Classifying images of dose-response graphs (rather than raw dose-response graph data) reduces the impact of any lack of homogeneity between different inputs to the classifier model or between the inputs used for prediction and the training data. Thus, the claimed approach provides a more flexible classifier that can classify dose-response graph images with, for example, different numbers of data points (e.g., 8, 10, or 12 concentrations), missing data points, and / or different numbers of replicates.

[0009] The classification output can include a vector of probabilities, each probability representing the likelihood that the input represented by the received image data belongs to a respective one of the categories. Alternatively, or in addition, the classification output can indicate the category with the highest probability. In some examples, the classification output can indicate two (or more) most probable categories.

[0010] In an example embodiment, a dose-response graph can be first classified into a first or second variance category, and image data representing the image of the dose-response graph is processed using a curve shape classifier model only if the dose-response graph is classified into the first variance category. A variance classifier can be used to classify the dose-response graph into the first or second variance category. The variance classifier can include a binary classifier. The second variance category can be a high variance category, and the first category can be a category for other dose-response graphs that are not in the second category (e.g., low variance graphs). The variance classifier can classify the dose-response graph into the first or second variance category (e.g., high variance category or low variance category) based on the difference between measurements of activity at the same concentration.

[0011] Dose-response graphs can be classified by a variance classifier using the interquartile range and / or other quartiles between replicates across all concentrations. The advantage of this approach is that there are no constraints on the number of data points to consider, meaning that the classifier is valid even if some data points are missing or the number of concentrations varies between inputs to the variance classifier (or between the inputs used for prediction and the training data). This benefit synergizes with using images as input to the curve shape classifier, facilitating a flexible classification pipeline that reduces the impact of any heterogeneity in the input data.

[0012] The distributed classifier may include a trained machine learning model, hi one example embodiment, the distributed classifier includes a multi-layer perceptron neural network model.

[0013] Multiple dose-response categories are: A "Top" category with high activity across the entire concentration range; A "No Bottom" category for sigmoid curves where the upper asymptote is visible but the lower asymptote is not; and the "sigmoid" category of well-behaved sigmoid curves, including lower and upper asymptote portions; an "Active No Top" category of sigmoid curves, where the lower asymptote is visible but the upper asymptote is not, and the sigmoid curve reaches the 50% activity threshold, but the portion after the inflection point of the dose-response graph is visible; A "No Top" category for weakly active compounds in the concentration range of the dose-response graph; and an "inactive" category for inactive compounds in the concentration range of the dose-response graph; EC 50 and the high gradient category of sigmoid curves with large slopes at; EC 50 and the low-slope category of sigmoid curves with small slopes at; a "partial" category of sigmoid curves where the difference between A(c) at the upper and lower asymptote is less than 70%; with the "wave" category, where there is an alternating rise and fall of activity with respect to concentration; a "last up" category in which no activity is shown except at the highest or two highest concentrations; may include one or more (or all) of:

[0014] Multiple dose-response categories are: Bell curve category, and / or Toxicity category of a sigmoid curve showing a decrease in inhibition rate at higher concentrations may include:

[0015] The bell curve category and the toxicity category can be fused into a single category for inference, but can be kept separate during training so that the model can independently learn the slightly different visual patterns associated with each of these categories. These two categories are not necessarily easily distinguishable by experts. However, in the embodiments described herein, the curve shape classifier model can advantageously distinguish between these categories because it operates on low-level features in the form of pixel data.

[0016] In another example embodiment, this specification describes a computer-implemented method for generating a curve shape classifier model for classifying dose-response graphs obtained from dose-response experiments. The method includes receiving a plurality of training images in a neural network model, each training image being an image of a dose-response graph showing the relationship between the concentration of a compound and its activity; generating an output for each training image, where generating the output for the training image includes processing the training image through one or more layers of the neural network model according to parameters associated with the one or more layers; and updating the parameters based on an objective function including comparing the generated output for each training image with corresponding label data associated with the training image, where the label data indicates that the training image belongs to one or more dose-response graph categories related to the curve shape. The neural network model can include a convolutional neural network model.

[0017] In another example embodiment, this specification describes a data processing apparatus including one or more processors configured to perform any of the computer-implemented methods described herein.

[0018] In another example embodiment, this specification describes a computer-readable storage medium containing instructions that, when executed by one or more processors, cause the one or more processors to perform any of the computer-implemented methods described herein.

[0019] The term "dose-response graph" as used herein refers to a graph showing the dose-response relationship between the concentration of a compound and its activity. A dose-response graph can include, for example, a set of data points plotted against Cartesian axes, e.g., a horizontal axis (X) and a vertical axis (Y). In various embodiments, the dose-response graph can be obtained from a high-throughput screening process. However, the dose-response graph can alternatively be obtained from other sources, e.g., from other screening processes, such as low-throughput screening. In the art, a "dose-response graph" may alternatively be referred to as a "dose-response curve," although it is understood that the term "dose-response curve" does not necessarily imply the existence of a continuous curve.

[0020] EC 50 The term IC refers to the half maximal effective concentration of a compound, i.e., the concentration that leads to 50% of the maximal response. 50 The term "half maximal inhibitory concentration" refers to the concentration that leads to 50% inhibition. Dose-response graphs can measure activities other than inhibition; for example, a dose-response graph can measure the ability of a compound to act as an agonist of a biomolecule or organism. The claimed methods can be used to classify dose-response graphs that measure any type of activity. Thus, throughout this specification, EC 50 Although the term is used, inhibition is measured and IC 50 The term "embodiment" can be used and can be understood to include embodiments.

[0021] As used herein, the term A(c) can be understood to refer to the observed activity (A) at a particular concentration (c). Similarly, the term I(c) can be understood to refer to the observed inhibition (I) at a particular concentration (c). Although the term A(c) is used throughout this specification, it can be understood to include embodiments in which inhibition is measured and the term I(c) can be used.

[0022] It can be appreciated that the term "activity" can be understood to refer to organic or organophysical activity. For example, a dose-response graph can show the relationship between the concentration of a compound and its ability to act as an agonist, antagonist, or allosteric modulator of a biomolecule or organism. Alternatively, or in addition, a dose-response graph can show the relationship between the concentration of a compound and its ability to bind to a biomolecule or organism. The biomolecule or organism can be or contain a nucleophilic structure. The biomolecule or organism can be selected from the group consisting of an amino acid, a peptide, an affimer, a protein, a glycoprotein, a lipopolysaccharide, an antibody or fragment thereof, a nucleic acid, an organic polymer, a virus, a bacterium, a parasite, a cell, and a cell-associated structure. The protein can be or contain a receptor, an enzyme, an ion channel, and / or a transporter.

[0023] It can be appreciated that the computer-implemented methods described herein can be used to aid in the identification of compounds with desired activity. These compounds can be advanced as potential drug candidates. Alternatively, medicinal chemists can produce modified compounds based on or containing the identified active compounds. The present invention extends to any drug that includes or is based on a compound that has been tested in a dose-response experiment, the results of which have been classified using the computer-implemented methods described herein. The present invention also extends to any drug that includes or is based on a compound that has been tested in a dose-response experiment, the results of which have been classified using the computer-implemented methods described herein for use in therapy. The drug can be a compound that has been tested in a dose-response experiment.

[0024] The computer-implemented methods described herein can be used to identify small molecules and / or biomolecules, including but not limited to molecular biomarkers. The present invention also extends to small molecules or biomolecules that have been tested in dose-response experiments, the results of which have been classified using the computer-implemented methods described herein.

[0025] In order that the present invention may be more readily understood, embodiments thereof will now be described, by way of example only, with reference to the accompanying drawings, in which: [Brief explanation of the drawings]

[0026] [Figure 1-1] FIG. 1 illustrates an example set of dose-response graph categories. [Figure 1-2] Continued from Figure 1-1. [Figure 1-3] Continued from Figure 1-2. [Figure 2] FIG. 1 illustrates a classification pipeline according to an example embodiment. [Figure 3-1] FIG. 1 illustrates the architecture of a neural network shape classifier according to an example embodiment. [Figure 3-2] Continued from Figure 3-1. [Figure 4] FIG. 1 illustrates the architecture of a neural network distributed classifier according to an example embodiment. [Figure 5] FIG. 1 is a schematic diagram of a system / apparatus for carrying out the methods described herein. DETAILED DESCRIPTION OF THE INVENTION

[0027] overview Various example embodiments relate to a system for classifying dose-response (DR) graphs based on a neural network operating on normalized images of the DR graphs. This system allows thousands of curves to be annotated into multiple categories in minutes, aiding high-throughput screening (HTS) researchers in their analysis. Categories can be associated with active or inactive compounds, or with features of interest, such as the presence of noise between replicates, decreased efficacy at high doses that may be related to toxicity, aggregation, or solubility issues, or a suspiciously weak or strong slope at the inflection point of the DR graph for an active compound.

[0028] preface High-throughput screening (HTS) is one of the major strategies used for hit compound discovery in the pharmaceutical industry. Recently, screening techniques have become more sophisticated, leading to strategies such as quantitative HTS and the use of more counterscreens or selectivity assays to qualify hit compounds. Furthermore, these techniques have also increased the quantity of dose-response (DR) results generated. Other large-scale dose-response datasets are obtained after interrogation of protein libraries with selection techniques such as phage display, yeast display, or fluorescence-activated cell sorting (FACS). The quality of dose-response data depends on the robustness of the screening conditions, protocols, and overall assay, as well as on the behavior of the compound. Basic automated analysis of dose-response relationships relies on fitting algorithms that may be unreliable at suboptimal settings due to the presence of outliers resulting from interference effects or other technical artifacts. In practice, dose-response data must be manually examined and addressed to guide decisions regarding the pursuit of corresponding compounds in projects. Therefore, the visual inspection process is time-consuming, especially when the hit rate is high, and the outcome of this process depends on the quality of the curves, the experience of the experts, and the time available for the analysis. When dealing with large numbers of results, this approach can slow down the project and, over time, can lead to a lack of consistency and robustness in the analysis.

[0029] Active compounds have an EC 50The ideal shape of a DR curve (the concentration that leads to 50% of the maximal response, whether for inhibition or other activity measurements) is either a perfect sigmoidal function with clearly defined lower and upper asymptote, or some subset thereof. An ideal dose-response graph would be flat, as the compound would exhibit either maximal activity or no activity at all across the entire concentration range. However, there are many reasons why perturbations in DR experiments can result in curves that do not belong to these ideal shapes. Some are related solely to the properties of the compound (e.g., colloidal aggregation or toxicity issues at high concentrations), while others are due to the cell line or experimental protocol (e.g., colored and fluorescent compounds interfering with luminescence assays).

[0030] The standard post-processing workflow for DR graphs involves fitting the percent inhibition A(c) to a sigmoid function using the Hill equation and calculating the EC 50 and its confidence interval, EC 50 The Hill equation consists of extracting several parameters, such as the slope at σ, A(c) at the upper and lower asymptote, etc. The Hill equation is known per se to those skilled in the art and will not be described here. See "Handbook of Drug Screening"; Seethala, R.; Zhang, L. (eds.); Drugs and the Pharmaceutical Sciences; CRC Press: 2009, and Shockley, KR, "Quantitative high-throughput screening data analysis: challenges and recent advances", Drug Discovery Today 2015, Vol. 20, pp. 296-300.

[0031] The extracted parameters alone are not sufficient to fully characterize the activity of a compound. Therefore, a visual inspection of the DR graph is usually performed, which involves masking outliers to obtain a better fit and analyzing the fitted parameters (upper and lower asymptote, EC 50This may involve checking or adjusting the curves (e.g., slopes), finding and retesting invalid experiments, annotating valid curves that indicate defects or idiosyncrasies, and tagging each curve with a final decision label: "active" (A), "inactive" (NA), or "invalid" (NV). This expert curation process presents challenges: it is time-consuming, can be expert-dependent, and even a single expert may face consistency issues when annotating borderline cases at different times.

[0032] To alleviate these problems, this specification describes a computer-implemented method for automatically classifying DR graphs according to their visual characteristics. In particular, as described in more detail below, the classification can be based on visual patterns of dose-response relationships, which can be defined by the shape of the dose-response curve. The system includes a neural network image classifier in the form of a convolutional neural network model that is configured, through training, to classify received DR graph images into multiple DR graph categories.

[0033] In the example embodiment described below, the system can provide 14 different expert-defined categories along with classification probabilities. This solution allows experts to group similar DR graphs into categories with interpretable labels according to their visual similarities for performing batch operations on them, and easily identify curves with poor prediction accuracy for thorough review. This system speeds up the inspection process and improves the robustness and consistency of the final decision.

[0034] Data Strategy Figure 1 shows an example of a DR graph divided into 14 categories, along with the associated category labels. Each category has an associated label, and the two terms ("category" and "label") are used interchangeably.

[0035] The 14 categories include 13 categories (sets A, B and C) that can be associated with distinct curve shapes.

[0036] Of these, six categories form a first group (Set A) with different levels of activity and no defects (e.g., no defects or extreme parameters): - "Top" represents ultra-active compounds showing complete signal inhibition along the entire concentration range; - "no lower" corresponds to a highly active compound with a sigmoidal inhibition curve, with a visible upper asymptote but no visible lower asymptote; - "Sigmoid" refers to a sigmoid curve that shows good behavior of active compounds, including upper and lower asymptote; - "Topless activity" represents a DR graph that includes the lower asymptote but not the upper asymptote, reaching the 50% inhibition threshold, and the portion of the DR graph after the inflection point is visible; - "No top" is for weakly active compounds within the assay concentration range, - The "inactive" DR graph is for compounds that are inactive in the assay.

[0037] The second group of three categories (set B) corresponds to sigmoid DR graphs with some extreme parameters: - "High gradient" is EC 50 shows a perfect sigmoidal DR graph with a large slope at (typically nHill>4), - "Low gradient" means, on the other hand, 50 Describes a perfect sigmoidal DR graph with a small slope at (typically, nHill<0.5). These patterns can be observed when cooperative effects occur within a system: - The "partial" label is for a fully sigmoidal DR graph where the relative maximum inhibition (the difference between the A(c) of the upper and lower asymptote) is less than 70%, as is frequently observed in cellular assays.

[0038] The third group (Set C) contains DR graphs showing defects commonly observed in practice: - In a "bell" DR graph, A(c) decreases at the highest concentrations, forming a bell-shaped curve. This pattern may be due to signal interference issues in the fluorescent assay format or compound aggregation issues. The "toxicity" ("adverse effect at high concentrations") label also refers to a perfectly sigmoidal inhibition curve, where the rate of inhibition drops off sharply at the highest concentration due to compound toxicity, as is frequently observed in cellular assays. The "toxicity" and "bell" labels can be merged in post-processing, as described below. - In a "wave" DR graph, there are alternating rises and falls in A(c), which can be due to a variety of issues, such as interference with the signal readout or problems with the compound dilution series. - The "last rise" label is for DR graphs where no inhibition is observed except at the highest or two highest compound concentrations in the titration curve, resulting in a non-effective curve.

[0039] These categories can be associated with three higher-order ensembles: "definitely active," "definitely inactive," and "needs further investigation," this latter set being for all curves with idiosyncrasies or defects. The "upper," "no lower," "sigmoid," and "active without upper" categories are associated with the "definitely active" ensemble. The "inactive" category is associated with the "definitely inactive" ensemble. The "no upper," "high slope," "low slope," "partial," "bell," "toxic," "wave," "last rise," and "dispersion" (see below) categories are associated with the "needs further investigation" ensemble.

[0040] A human expert can visually inspect a DR graph and assign it to one of 13 categories (sets A, B, and C) based on the shape of the curve that defines the trend of the dose-response relationship. For example, DR graph example 110 can be assigned to the "sigmoid" category because the curve that defines the trend of the dose-response relationship is sigmoid. DR graph 120 can be assigned to the "wave" category because the curve has a wave shape, with activity alternately rising and falling. Thus, as can be visually seen from FIG. 1 , each of the 13 categories (sets A, B, and C) can be associated with a respective curve shape that defines the dose-response relationship for that category. Accordingly, these categories can be referred to herein as "shape categories."

[0041] Note that several categories may be associated with the same or similar curve shapes. For example, "top" and "inactive" may be the same (horizontal) curve, differing only in the magnitude of the response at each concentration. "End rising" and "no top" may also be somewhat similar, but differ in that the curve for "end rising" rises more steeply compared to "no top."

[0042] As will be described in more detail below, an image classifier model can be learned to classify images of DR graphs based on curve shape by training the classifier on images of DR graphs that have been previously classified by experts.

[0043] Figure 1 also shows a "variance" category (set D), where noise is observed across replicates. Replicates are two or more measurements made at the same concentration value. Unlike the other categories, the "variance" category is not associated with a specific shape, but can nevertheless be processed separately as part of a classification pipeline such as the one described here.

[0044] Classification Pipeline FIG. 2 illustrates an example classification pipeline 200, including a variance classifier 210 and a shape classifier 220. The variance classifier 210 acts as an initial filter; only DR graphs that pass this filter are submitted to the shape classifier 220. More specifically, the variance classifier can output a probability p (value between 0 and 1) that the dose-response graph is in the "variance" category. If the probability is greater than 0.5, the dose-response graph is classified into the "variance" category. Otherwise, a normalized image of the dose-response graph is generated and input into a shape classifier, which classifies the image by curve shape. Images of dose-response graphs classified with a probability lower than 0.9 can be assigned to a specific "low probability" category, as shown.

[0045] A shape classifier can be configured to classify DR graph images into 12 shape categories. These 12 shape categories are the same as the 13 shape categories in Sets A, B, and C in Figure 1, except that "bell" and "toxic" are fused together (i.e., treated as one combined category) because they have been found not to be always distinguishable by experts.

[0046] The shape classifier 220 may include a convolutional neural network (CNN) that takes an image of the DR graph as input and generates a classification output. Convolutional neural networks are known to those skilled in the art and will not be described in detail here. See LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, Vol. 521, pp. 436-444.

[0047] The classification output generated by the CNN can include a probability for each of the 12 shape categories, where each probability is the likelihood that the DR graph belongs to the respective category. Alternatively, or in addition, if the probability is less than 0.9 and the DR graph is not classified as "low probability," the classification output may include an indication of the category with the highest probability.

[0048] The variance classifier 210 is a binary classifier in that it classifies into two possible categories. It can include a multilayer perceptron (MLP) classifier that takes as input statistical features extracted from the raw data for the DR graph and outputs the probability that the DR graph belongs to the variance category. For example, to process a DR graph using the variance classifier, the positive difference in A(c) between replicates at each concentration can be extracted, and the quartiles of q1, q2, and q3 for this distribution can be calculated along with the interquartile range. These four descriptors can be normalized between 0 and 1 using MinMax scaling, and the four normalized values ​​can be used as input to the classifier.

[0049] Preprocessing the shape classifier The shape classifier can receive images during the training phase, where training images are used to train the model, or during the prediction phase, where "new" DR graph images not seen during training are classified. In either phase, input images can be generated from the raw DR graph data by generating grayscale images that depict a set of DR graph data points relative to Cartesian (e.g., XY) axes. The generated images can be normalized in that each image can be generated to have the same size (e.g., 150 x 150 pixels) and the X and Y axes can be at the same location in each image. Additionally, the generated images can be normalized in that the Y axis (activity) is labeled with the same scale or values ​​(e.g., values ​​from -50 to 150) at the same location in each image. Meanwhile, the scale or values ​​can be omitted from the X axis (concentration), and different images need not be associated with the same range of concentration values ​​and can have different numbers of data points. To depict the data within a normalized "window" defined by the image, the raw data can be normalized by removing / filtering data points that fall outside that "window." Figure 1 shows various examples of DR graph images.

[0050] Converting the DR graph to an image for processing by the CNN reduces the effect of any lack of homogeneity between different DR graph samples, e.g., between the DR graph samples used for inference compared to the training set. This provides a flexible classifier that can handle DR graphs with, for example, different numbers of data points (e.g., 8, 10, or 12 concentrations), missing data points, and / or different numbers of replicates (resulting in different numbers of Y-values ​​for some concentration values).

[0051] training data The DR graph for training purposes can be obtained from an existing data source or generated algorithmically. The DR graph can be manually labeled by an expert into one of the 14 categories shown in Figure 1. The resulting labels can be represented as vectors and stored as "ground truth" data.

[0052] Algorithmically generating a DR graph can include generating synthetic dose-response data using the Hill equation using parameters within a specific parameter space for each category label. The parameter space for the category label can include, for example, the location of the inflection point, the slope at the inflection point, and the location of the upper and lower asymptote. Noise and / or imperfections can be added.

[0053] In particular, DR graphs for categories belonging to the "definitely active" and "definitely inactive" ensembles can be constructed based on the Hill equation. "Bell" and "wave" curves can be generated by combining two or three Hill functions, respectively. "Toxic" and "End-Rise" curves can be generated by adding appropriate noise to A(c) at the highest concentration using curves derived from the "definitely active" and "definitely inactive" categories, respectively. Ten concentrations can be used, with two replicates per concentration generated using logarithmic noise. To obtain realistic noise distributions between replicates, the results of an experimental HTS DR campaign can be used to determine statistics based on the differences between replicates. Based on fitting the noise distribution to a logarithmic law, relevant parameters can be extracted, and the resulting parameterized function can be used to model noise between replicates. Normal noise can be added to the concentration, and uniform noise can be added to A(c) across the entire concentration range.

[0054] Training images for training a shape classifier can be obtained from existing or algorithmically generated DR graphs using the preprocessing steps described above. Thus, each training image can include a 150x150 pixel grayscale pixel image depicting a set of data points of the DR graph relative to Cartesian axes. Each training image can be stored along with a "ground truth" label for the training image that indicates the classification decision of an expert (or two or more experts) for the DR graph represented by the image.

[0055] Shape Classifier Architecture and Training One example of a shape classifier architecture is shown in Figure 3. In summary, this architecture: - One block containing six two-dimensional (2D) convolutional layers with Relu activation and l2 regularization using a 3,3 filter (32,32,64,64,128,128); - 4 2D maxpooling layers of size (2,2); - 5 batch normalization layers; - 1 flattening layer; - 1 dense layer of 256 neurons with Relu activation and l2 regularization; - 1 batch normalization layer; - 1 dropout layer with coefficient 0.5; - 1 final dense layer with 13 output categories in softmax activation Includes.

[0056] Those skilled in the art will recognize that many variations and modifications to the architecture shown in FIG. 3 are possible.

[0057] Note that the "?" in FIG. 3 refers to the number of training examples that the neural network processes at one time, and indicates that any suitable number can be used.

[0058] The shape classifier can be trained using a training set containing thousands (e.g., 5,000) of DR graph images per category. The training images can be generated from algorithmically generated DR graphs that are manually labeled with "ground truth" classification labels as described above. Training images can be generated for each of the 13 shape categories described above.

[0059] During training, training images are received at the input layer of the neural network model (see Figure 3). Each training image is processed by subsequent layers of the neural network model according to the neural network model's parameters to generate a probability of being in each of the 13 shape categories shown in sets A, B, and C in Figure 1. The neural network's parameters (i.e., neuron weights and biases) can be updated by optimizing an objective function. The objective function includes a loss determined by a comparison between the generated output for each training image and the corresponding label data associated with the training image. The label data can include vectors representing "ground truth" labels applied following visual inspection of the training images by an expert (or group of experts).

[0060] The loss can measure the mean squared error between the output for each training image and the "ground truth" label data. The objective function can further include a regularization term; for example, the objective function can be a linear combination of the loss and the regularization term. Other weighted losses can be included as part of the objective function. The objective function can be optimized using a gradient-based method such as the Adam optimizer, stochastic gradient descent, mini-batch gradient descent, or batch gradient descent.

[0061] In one example, the Adam optimizer was used with a learning rate of 1.0E-4 and a learning rate decay of 3.0E-7. Training was run for up to 300 epochs, measured by validation loss, and stopped early at a total validation loss (Patience=5, delta 0.001). The input data was split 80 / 20 between the training and validation sets.

[0062] This training process produces a trained shape classification model. The trained model can be used to classify "new" DR graph images. For test DR graph images, the trained model can generate a classification output that includes the probability of each of 12 shape categories, which are the same as the 13 shape categories in Sets A, B, and C of FIG. 1, except that "bell" and "toxic" (which are kept separate in training) are fused together (i.e., treated as one combined category) for inference. Alternatively, or in addition, the classification output may include an indication of the category with the highest probability, or may include indications of the two (or more) most probable categories.

[0063] The use of a convolutional neural network for the shape classifier is beneficial in that it allows the classifier to focus on the overall curve shape rather than on the details, however, in alternative implementations, other neural network architectures can be used, such as a fully connected neural network.

[0064] Distributed Classifier Architecture and Training One example architecture for a distributed classifier is shown in Figure 4. This classifier includes a multi-layer perceptron (MLP) classifier. The classifier includes two hidden layers, each with five neurons and a Relu activity.

[0065] Those skilled in the art will recognize that many variations and modifications to this architecture are possible.

[0066] Note that the "?" in FIG. 4 refers to the number of training examples that the neural network processes at one time, and indicates that any suitable number can be used.

[0067] The classifier can be trained using a training set containing thousands (e.g., 5,000) of DR graphs per category. For each DR graph in the training set, the positive difference between two replicates at each concentration can be extracted, and from this distribution, the quartiles of q1, q2, and q3, along with the interquartile range, can be calculated. This can be extended to any number of doses by calculating the positive difference between all replicates at each concentration and using the distribution of all of these differences in calculating the quartiles and interquartile ranges of q1, q2, and q3.

[0068] The four descriptors can be normalized between 0 and 1 using MinMax scaling, and the four normalized values ​​can be used as inputs for the MLP classifier.

[0069] An advantage of this approach is that there is no constraint on the number of data points to consider, which means that the classifier is valid even if some data points are missing or the number of cardinalities varies between the inputs to the distributed classifier (or between the inputs used for prediction and the training data). This benefit synergizes with using images as input to the shape classifier, facilitating a flexible classification pipeline that reduces the impact of any heterogeneity in the input data.

[0070] Inputs are processed through layers of classifiers according to the classifier parameters to generate classified outputs. The classifier parameters (i.e., neuron weights and biases) can be updated by optimizing an objective function. The objective function includes a loss that depends on comparing the generated output for each training DR graph with the corresponding label data associated with the training DR graph. The label data may include a binary value (i.e., whether the DR graph was classified by the expert into the "distributed" category) that represents a "ground truth" label applied after visual inspection of the training images by an expert.

[0071] The loss can measure the mean squared error between the output for each training image and the "ground truth" label data. The objective function can further include a regularization term, e.g., the objective function can be a linear combination of the loss and the regularization term. Other weighted losses can be included as part of the objective function. The objective function can be optimized using a gradient-based method, e.g., stochastic gradient descent. Training can be performed using an 80 / 20 split between the training set and the validation set.

[0072] Although it has been found advantageous to use MLP architectures for distributed classifiers, other machine learning models can also be used, for example random forest models.

[0073] 5 shows a schematic example of a system / apparatus for performing the methods described herein. The illustrated system / apparatus is an example of a computing device. Those skilled in the art will appreciate that other types of computing devices / systems, such as distributed computing systems, may alternatively be used to implement the methods described herein.

[0074] The device (or system) 500 includes one or more processors 502. The one or more processors control the operation of the other components of the system / device 500. The one or more processors 502 may include, for example, a general-purpose processor. The one or more processors 502 may be single-core or multi-core devices. The one or more processors 502 may include a central processing unit (CPU) or a graphics processing unit (GPU). Alternatively, the one or more processors 502 may include dedicated processing hardware, such as a RISC processor, or programmable hardware with embedded firmware. Multiple processors may be included.

[0075] The system / device includes working or volatile memory 504. One or more processors can access the volatile memory 504 to process data and can control the storage of data in the memory. The volatile memory 504 can include any type of RAM, for example, static RAM (SRAM), dynamic RAM (DRAM), or can include flash memory, such as an SD card.

[0076] The system / apparatus includes a non-volatile memory 506. The non-volatile memory 506 stores a set of operating instructions 508 in the form of computer-readable instructions for controlling the operation of the processor 502. The non-volatile memory 506 can be any type of memory, such as read-only memory (ROM), flash memory, or magnetic drive memory.

[0077] The one or more processors 502 are configured to execute operational instructions 508 that cause the system / device to perform any of the methods described herein. The operational instructions 508 may include code (i.e., drivers) associated with hardware components of the system / device 500 and code associated with basic operations of the system / device 500. Generally, the one or more processors 502 execute one or more of the operational instructions 508 that are permanently or semi-permanently stored in non-volatile memory 506, using volatile memory 504 to temporarily store data generated during execution of the operational instructions 508.

[0078] Implementations of the methods described herein can be realized in digital electronic circuitry, integrated circuits, specially designed ASICs (application-specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These can include a computer program product (e.g., software stored on a magnetic or optical disk, memory, programmable logic device, etc.) containing computer-readable instructions that, when executed by a computer such as that described in connection with FIG. 5, cause the computer to perform one or more of the methods described herein.

[0079] The terms "drug" or "medicament" are used interchangeably herein to describe a pharmaceutical formulation containing one or more active pharmaceutical ingredients or pharmaceutically acceptable salts or solvates thereof, and optionally a pharmaceutically acceptable carrier. An active pharmaceutical ingredient ("API"), in its broadest sense, is a chemical structure that has a biological effect on humans or animals. In pharmacology, drugs or medications are used to treat, cure, prevent, or diagnose disease or otherwise improve physical or mental well-being. Drugs or medications can be used for a limited duration or periodically for chronic disorders.

[0080] As described below, drugs or pharmaceutical agents may contain at least one API or a combination thereof in various types of formulations for the treatment of one or more diseases. Examples of APIs include small molecules with a molecular weight of 500 Da or less, polypeptides, peptides, and proteins (e.g., hormones, growth factors, antibodies, antibody fragments, and enzymes), carbohydrates and polysaccharides, as well as nucleic acids, double-stranded or single-stranded DNA (including naked and cDNA), RNA, antisense nucleic acids, such as antisense DNA and RNA, small interfering RNA (siRNA), ribozymes, genes, and oligonucleotides. Nucleic acids can be incorporated into molecular delivery systems such as vectors, plasmids, or liposomes. Mixtures of one or more drugs are also contemplated.

[0081] The drug or agent can be contained in a primary package or "drug container" adapted for use in a drug delivery device. The drug container can be, for example, a cartridge, syringe, reservoir, or other rigid or flexible vessel configured to provide a chamber suitable for storage (e.g., short-term or long-term storage) of one or more drugs. For example, in some cases, the chamber can be designed to store the drug for at least one day (e.g., from one day to at least 30 days). In some cases, the chamber can be designed to store the drug for about one month to about two years. Storage can occur at room temperature (e.g., about 20°C) or refrigerated temperatures (e.g., from about -4°C to about 4°C). In some cases, the drug container can be or include a dual-chamber cartridge configured to separately store two or more components of a pharmaceutical formulation to be administered (e.g., an API and a diluent, or two different drugs), one in each chamber. In such cases, the two chambers of the dual-chamber cartridge can be configured to allow mixing between two or more components prior to and / or during administration to the human or animal body. For example, the two chambers can be configured to be in fluid communication with each other (e.g., via a conduit between the two chambers) and to allow mixing of the two components by a user, if desired, prior to administration. Alternatively or additionally, the two chambers can be configured to allow mixing upon administration of the components to the human or animal body.

[0082] The drugs or agents contained in the drug delivery devices described herein can be used to treat and / or prevent many different types of medical disorders. Examples of disorders include, for example, diabetes or complications associated with diabetes, such as diabetic retinopathy, and thromboembolic disorders, such as deep vein thromboembolism or pulmonary embolism. Further examples of disorders include acute coronary syndrome (ACS), angina, myocardial infarction, cancer, macular degeneration, inflammation, hay fever, atherosclerosis, and / or rheumatoid arthritis. Examples of APIs and drugs are those listed in handbooks such as Rote Liste 2014 (e.g., but not limited to, Main Group 12 (antidiabetic agents) or 86 (oncology agents)) and the Merck Index, 15th edition.

[0083] Examples of APIs for the treatment and / or prevention of type 1 or type 2 diabetes or complications associated with type 1 or type 2 diabetes include insulin, e.g., human insulin, or a human insulin analog or derivative; glucagon-like peptide (GLP-1), a GLP-1 analog or GLP-1 receptor agonist, or an analog or derivative thereof; a dipeptidyl peptidase-4 (DPP4) inhibitor; or a pharmaceutically acceptable salt or solvate thereof, or any mixture thereof. As used herein, the terms "analog" and "derivative" refer to a polypeptide having a molecular structure that is formally derivable from the structure of a naturally occurring peptide, e.g., the structure of human insulin, by deletion and / or replacement of at least one amino acid residue present in the naturally occurring peptide and / or by addition of at least one amino acid residue. The added and / or replaced amino acid residue can be either a codable amino acid residue, another naturally occurring residue, or a purely synthetic amino acid residue. Insulin analogs are also referred to as "insulin receptor ligands." In particular, the term "derivative" refers to a polypeptide having a molecular structure formally derivable from the structure of a naturally occurring peptide, for example, the molecular structure of human insulin in which one or more organic substituents (e.g., fatty acids) are attached to one or more of the amino acids. Optionally, one or more amino acids present in the naturally occurring peptide are deleted and / or replaced by other amino acids, including non-codable amino acids, or amino acids, including non-codable ones, are added to the naturally occurring peptide.

[0084] Examples of insulin analogues are Gly(A21), Arg(B31), Arg(B32) human insulin (insulin glargine); Lys(B3), Glu(B29) human insulin (insulin glulisine); Lys(B28), Pro(B29) human insulin (insulin lispro); Asp(B28) human insulin (insulin aspart); human insulin in which the proline at position B28 is replaced by Asp, Lys, Leu, Val or Ala and the Lys at position B29 may be replaced by Pro; Ala(B26) human insulin; Des(B28-B30) human insulin; Des(B27) human insulin and Des(B30) human insulin.

[0085] Examples of insulin derivatives are, for example, B29-N-myristoyl-des(B30) human insulin, Lys(B29)(N-tetradecanoyl)-des(B30) human insulin (insulin detemir, Levemir®); B29-N-palmitoyl-des(B30) human insulin; B29-N-myristoyl human insulin; B29-N-palmitoyl human insulin; B28-N-myristoylLysB28ProB29 human insulin; B28-N-palmitoyl-LysB28ProB29 human insulin; B30-N-myristoyl-ThrB29LysB30 human insulin. B30-N-palmitoyl-ThrB29LysB30 human insulin; B29-N-(N-palmitoyl-gamma-glutamyl)-des(B30) human insulin, B29-N-omega-carboxypentadecanoyl-gamma-L-glutamyl-des(B30) human insulin (insulin degludec, Tresiba®); B29-N-(N-lithocholyl-gamma-glutamyl)-des(B30) human insulin; B29-N-(ω-carboxyheptadecanoyl)-des(B30) human insulin and B29-N-(ω-carboxyheptadecanoyl) human insulin.

[0086] Examples of GLP-1, GLP-1 analogs and GLP-1 receptor agonists are, for example, lixisenatide (Lyxumia®), exenatide (exendin-4, Byetta®, Bydureon®, a 39 amino acid peptide produced by the salivary glands of the flathead monster), liraglutide (Victoza®), semaglutide, taspoglutide, albiglutide (Syncria®), dulaglutide (Trulicity®), rexendin-4, CJC-1134-PC, PB-1023, TTP-054, langrenatide / HM-11260C (efpegrenatide). , HM-15211, CM-3, GLP-1 Erigen, ORMD-0901, NN-9423, NN-9709, NN-9924, NN-9926, NN-9927, Nodexene, Viador-GLP-1, CVX-096, ZYOG-1, ZYD-1, GSK-2374697, DA-3091, MAR-701, MAR709, ZP-2929, ZP-3022, ZP -DI-70, TT-401 (Pegapamodtide), BHM-034, MOD-6030, CAM-2036, DA-15864, ARI-2651, ARI-2255, Tirzepatide (LY3298176), Bamadutide (SAR425899), Exenatide-XTEN, and Glucagon-Xten.

[0087] Examples of oligonucleotides are, for example, the cholesterol-lowering antisense therapeutic mipomersen sodium (Kynamro®) for the treatment of familial hypercholesterolemia, or RG012 for the treatment of Alport syndrome.

[0088] Examples of DPP4 inhibitors are linagliptin, vidagliptin, sitagliptin, denagliptin, saxagliptin, berberine.

[0089] Examples of hormones include pituitary or hypothalamic hormones or regulatory active peptides and their antagonists, such as gonadotropins (follitropin, lutropin, chorion gonadotropin, menotropin), somatropine (somatropin), desmopressin, terlipressin, gonadorelin, triptorelin, leuprorelin, buserelin, nafarelin, and goserelin.

[0090] Examples of polysaccharides include glycosaminoglycans, hyaluronic acid, heparin, low-molecular-weight heparin or ultra-low-molecular-weight heparin or derivatives thereof, or sulfated polysaccharides, such as the polysulfated forms of the aforementioned polysaccharides, and / or pharmaceutically acceptable salts thereof. An example of a pharmaceutically acceptable salt of polysulfated low-molecular-weight heparin is enoxaparin sodium. Examples of hyaluronic acid derivatives include Hylan G-F20 (Synvisc®) and sodium hyaluronate.

[0091] As used herein, the term "antibody" refers to an immunoglobulin molecule or an antigen-binding portion thereof. Examples of antigen-binding portions of immunoglobulin molecules include F(ab) and F(ab')2 fragments that retain antigen-binding ability. An antibody can be a polyclonal antibody, a monoclonal antibody, a recombinant antibody, a chimeric antibody, a deimmunized or humanized antibody, a fully human antibody, a non-human (e.g., murine) antibody, or a single-chain antibody. In some embodiments, an antibody has effector function and is capable of fixing complement. In some embodiments, an antibody has reduced or no binding ability to Fc receptors. For example, an antibody can be an isotype or subtype, antibody fragment, or mutant that does not support Fc receptor binding, e.g., has a mutation or deletion of the Fc receptor binding region. The term antibody also includes antigen-binding molecules based on tetravalent bispecific tandem immunoglobulins (TBTIs) and / or dual variable region antibody-like binding proteins (CODVs) with a crossover binding region orientation.

[0092] The term "fragment" or "antibody fragment" refers to a polypeptide (e.g., an antibody heavy and / or light chain polypeptide) derived from an antibody polypeptide molecule that does not include the full-length antibody polypeptide but comprises at least a portion of the full-length antibody polypeptide that is still capable of binding to antigen. Antibody fragments can include truncated portions of a full-length antibody polypeptide, but the term is not limited to such truncated fragments. Antibody fragments useful in the present invention include, for example, Fab fragments, F(ab')2 fragments, scFv (single-chain Fv) fragments, linear antibodies, monospecific or multispecific antibody fragments, e.g., bispecific, trispecific, tetraspecific, and multispecific antibodies (e.g., diabodies, triabodies, tetrabodies), monovalent or multivalent antibody fragments, e.g., bivalent, trivalent, tetravalent, and multivalent antibodies, minibodies, chelating recombinant antibodies, tribodies or bibodies, intrabodies, nanobodies, small modular immunopharmaceuticals (SMIPs), binding domain immunoglobulin fusion proteins, camelized antibodies, and VHH-containing antibodies. Additional examples of antigen-binding antibody fragments are known in the art.

[0093] The term "complementarity determining region" or "CDR" refers to short polypeptide sequences within the variable regions of both heavy and light chain polypeptides that are primarily responsible for mediating specific antigen recognition. The term "framework region" refers to amino acid sequences within the variable regions of both heavy and light chain polypeptides that are not CDR sequences and that are primarily responsible for maintaining the proper orientation of the CDR sequences to enable antigen binding. Although the framework regions themselves typically do not directly participate in antigen binding, as is known in the art, certain residues within the framework regions of a particular antibody may be directly involved in antigen binding or may affect the ability of one or more amino acids within the CDRs to interact with the antigen.

[0094] Examples of antibodies are anti-PCSK-9 mAb (e.g., alirocumab), anti-IL-6 mAb (e.g., sarilumab), and anti-IL-4 mAb (e.g., dupilumab).

[0095] Pharmaceutically acceptable salts of any of the APIs described herein are contemplated for use as drugs or medicaments in drug delivery devices. Pharmaceutically acceptable salts include, for example, acid addition salts and base salts.

[0096] Those skilled in the art will understand that modifications (addition and / or removal) of various components of the APIs, formulations, devices, methods, systems, and embodiments described herein may be made without departing from the full scope and spirit of the invention, which encompasses such modifications and any and all equivalents thereof.

[0097] Example drug delivery devices can include needle-based injection systems such as those described in Table 1 of Chapter 5.2 of ISO 11608-1:2014(E). As described in ISO 11608-1:2014(E), needle-based injection systems can be broadly distinguished into multi-dose container systems and single-dose container systems (with partial or full ejection). The container can be an exchangeable container or an integrated, non-exchangeable container.

[0098] As further described in ISO 11608-1:2014(E), a multi-dose container system can include a needle-based injection device with an exchangeable container. In such a system, each container holds multiple doses, and the dose size can be fixed or variable (pre-set by the user). Another multi-dose container system can include a needle-based injection device with an integrated, non-exchangeable container. In such a system, each container holds multiple doses, and the dose size can be fixed or variable (pre-set by the user).

[0099] As further described in ISO 11608-1:2014(E), a single-dose container system can include a needle-based injection device with replaceable containers. In one example for such a system, each container holds a single dose, thereby discharging the entire deliverable volume (full discharge). In a further example, each container holds a single dose, thereby discharging a portion of the deliverable volume (partial discharge). Also as described in ISO 11608-1:2014(E), a single-dose container system can include a needle-based injection device with an integrated, non-replaceable container. In one example for such a system, each container holds a single dose, thereby discharging the entire deliverable volume (full discharge). In a further example, each container holds a single dose, thereby discharging a portion of the deliverable volume (partial discharge).

[0100] Many modifications and variations to the embodiments described herein that fall within the scope of the following claims will become apparent to those skilled in the art.

Claims

1. 1. A computer-implemented method for classifying images containing dose-response graphs obtained from dose-response experiments, comprising: receiving, in a curve shape classifier model, an input including image data including a plurality of pixels, the image data representing an image of a dose-response graph showing the relationship between the concentration of a compound and its activity, the curve shape classifier model including a neural network model configured to classify the image of the dose-response graph into a plurality of dose-response graph categories related to curve shape; and generating a classification output for the image represented by the received image data using the neural network model; The computer-implemented method, wherein the generating includes processing the image data using one or more layers of a neural network model according to parameters associated with the one or more layers.

2. The computer-implemented method of claim 1 , wherein the curve shape classifier neural network model comprises a convolutional neural network model.

3. 3. The computer-implemented method of claim 1, further comprising classifying the dose-response graph into a first or second dispersion category based on the difference between measurements of activity at the same concentration, wherein image data representing an image of the dose-response graph is processed using the curve shape classifier model only if the dose-response graph is classified into the first dispersion category.

4. 4. The computer-implemented method of claim 3, wherein the dose-response graph is classified into first or second dispersion categories based on quartiles for the difference in measured activity as a function of concentration.

5. 5. The computer-implemented method of claim 3 or 4, comprising classifying the dose-response graph into a first or second dispersion category using a dispersion classifier comprising a multi-layer perceptron neural network model.

6. Multiple dose-response categories are: A category of high activity across the entire concentration range; A category of sigmoid curves in which the upper asymptote portion is visible but the lower asymptote portion is not; A category of well-behaved sigmoid curves, including lower and upper asymptote portions; a category of sigmoid curves in which the lower asymptote is visible but the upper asymptote is not, and which reach a 50% activity threshold, with the portion after the inflection point of the dose-response graph being visible; Categories of weakly active compounds in the concentration range of the dose-response graph; and categories of inactive compounds in the concentration range of the dose-response graph; a high slope category of sigmoid curves with large slopes at EC50; a low slope category of sigmoid curves with small slopes at EC50; The category of sigmoid curves where the difference between A(c) at the upper and lower asymptote is less than 70%; a category in which there is an alternating increase and decrease in activity with respect to concentration; A category in which no activity is shown except at the highest or two highest concentrations. The computer-implemented method of any one of claims 1 to 5, comprising one or more of:

7. 7. The computer-implemented method of claim 1, further comprising performing pre-processing, the pre-processing comprising receiving raw data representing a set of data points of a dose-response graph, and generating an image of the dose-response graph based on the raw data.

8. receiving raw data representing a set of data points for each of a plurality of dose-response graphs; generating a respective image for each dose-response graph, the image including a plurality of pixels and showing at least some of the respective sets of data points against Cartesian axes, each image generated with the same pixel height and pixel width, and the Cartesian axes located in the same location on each image; receiving image data for each respective image in a curve shape classifier model; The computer-implemented method of claim 7, comprising:

9. The computer-implemented method of claim 8 , wherein each image has a vertical axis of the same scale.

10. 1. A computer-implemented method for generating a curve shape classifier model for classifying dose-response graphs obtained from dose-response experiments, comprising: receiving a plurality of training images in the neural network model, each training image being an image of a dose-response graph showing the relationship between the concentration of a compound and its activity; generating an output for each training image, wherein generating the output for the training image includes processing the training image through one or more layers of a neural network model according to parameters associated with the one or more layers; updating the parameters based on an objective function that includes comparing the generated output for each training image with corresponding label data associated with the training image, the label data indicating that the training image belongs to one or more dose-response graph categories associated with curve shape; The computer-implemented method includes:

11. 11. The computer-implemented method of claim 10, wherein each training image depicts a respective set of data points relative to Cartesian axes, each training image having the same pixel height and pixel width, and the Cartesian axes located in the same location in each image.

12. The dose-response categories are: Bell curve categories and Toxicity category and 12. The computer-implemented method of claim 10 or 11, comprising:

13. A data processing apparatus comprising one or more processors configured to carry out a method according to any one of claims 1 to 12.

14. A computer readable storage medium comprising instructions that, when executed by one or more processors, cause the one or more processors to perform the method of any one of claims 1 to 12.

Images