Image classification of dose-response plots

Abstract

A computer-implemented method of classifying images, the images comprising dose-response plots obtained from a dose-response experiment. The method comprises receiving, at a curve shape classifier model, an input comprising image data, the image data comprising a plurality of pixels, wherein the image data represents an image of a dose-response plot indicative of a relationship between a concentration of a compound and its activity. The curve shape classifier model comprises a neural network model configured for classifying images of dose-response plots into a plurality of dose-response plot classes related to curve shapes. The method further comprises generating, using the neural network model, a classification output for the image represented by the received image data, the generating comprising processing the image data according to parameters associated with one or more layers of the neural network model using the one or more layers.

Description

Technical Field

[0001] This specification relates to a computer-implemented method for classifying images, including dose-response maps obtained from multiple dose-response experiments. Background Technology

[0002] In the early stages of a drug discovery project, potential drug targets are identified. Drug targets are molecules in the body that are intrinsically linked to specific disease processes. Depending on the disease to be treated, targets can be proteins (such as receptor proteins, enzymes, ion channels, or transport proteins) or nucleic acids (such as DNA). Researchers hypothesize that altering the activity of the target with a drug will produce the desired therapeutic effect.

[0003] Once a target has been identified, a testing system is identified or developed that generates a detectable signal to assess the effect of a compound on the drug target. This testing 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, the compound is tested at multiple concentrations, and a dose-response (DR) plot is generated. Analysis of the DR plot allows researchers to determine whether the compound is active and at what concentration.

[0004] High-throughput screening (HTS) is frequently used when a large number of potential compounds need to be tested. This utilizes robots, data processing / control software, liquid handling devices, and sensitive detectors, enabling researchers to rapidly perform thousands or even millions of screening tests. When all compounds are tested at multiple concentrations to generate DR data, this is called quantitative HTS (qHTS).

[0005] However, the large amount of data generated in the dose-response (DR) step of HTS activities requires careful analysis by researchers to detect artifacts and correct erroneous data points before validation experiments. This step requires expert review of each DR experiment, which can be time-consuming and prone to human error or inconsistencies. Summary of the Invention

[0006] This specification describes a computer-implemented method for classifying images, including dose-response maps obtained from dose-response experiments. The method includes receiving input including image data comprising a plurality of pixels at a curve shape classifier model, wherein the image data represents an image of a dose-response map indicating 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 dose-response map image into multiple dose-response map categories associated with curve shapes. Specifically, each dose-response map category may be associated with a corresponding curve shape that defines the dose-response relationship for that category.

[0007] The method includes generating a classification output of an image represented by received image data using the neural network model, the generation including processing the image data using one or more layers of the neural network model based on parameters associated with the one or more layers. The neural network model may include a convolutional neural network model.

[0008] Classifying dose-response map images (rather than the raw dose-response map data) reduces the impact of any lack of homogeneity between different inputs to the classifier model or between the input used for prediction and the training data. Therefore, the claimed method provides a flexible classifier capable of classifying dose-response map images with, for example, different numbers of data points (e.g., 8, 10, or 12 concentrations), missing points, and / or different numbers of copies.

[0009] The classification output may include a probability vector, where each probability represents the likelihood that the input, represented by the received image data, belongs to the corresponding category. Alternatively or additionally, the classification output may indicate the category with the highest probability. In some examples, the classification output may indicate two (or more) most probable categories.

[0010] In an example implementation, the dose-response map may initially be classified into a first or second discrete category, wherein the curve shape classifier model is used to process the image data representing the dose-response map only if the dose-response map is classified into the first discrete category. A discrete classifier may be used to classify the dose-response map into either the first or second discrete category. The discrete classifier may include a binary classifier. The second discrete category may be a high discrete category, and the first category may be a category for other dose-response maps that do not belong to the second category (e.g., lower discrete maps). The discrete classifier may classify the dose-response map into the first or second discrete category (e.g., higher or lower discrete category) based on the difference between activity measurements at the same concentration.

[0011] Dose-response maps can be classified by a discrete classifier using interquartile ranges and / or other interquartile values ​​between copies at all concentrations. The advantage of this approach is that there is no limit to the number of data points to consider, meaning the classifier is effective even if some data points are missing, or if the number of concentrations differs between the inputs to the discrete classifier (or between the inputs used for prediction and the training data). This benefit is synergistic with using images as input to a curve-shaped classifier and contributes to a flexible classification pipeline that reduces the influence of any heterogeneity in the input data.

[0012] Discrete classifiers can include trained machine learning models. In one example implementation, the discrete classifier includes a multilayer perceptron neural network model.

[0013] The multiple dose-response categories may include one or more (or all) of the following:

[0014] For the "Top" category, which exhibits high activity across the entire concentration range;

[0015] The "NoBottom" category is defined for S-shaped curves where the upper asymptote is visible but the lower asymptote is not.

[0016] The "Sigmoid" category is for well-formed S-shaped curves that include both the lower asymptote and the upper asymptote.

[0017] For the "Active No Top" category, which is an S-shaped curve in which the lower asymptote is visible but the upper asymptote is not, and which reaches the 50% activity threshold, the dose-response plot is visible after the inflection point;

[0018] The "No Top" category is for weakly active compounds within the concentration range of the dose-response diagram.

[0019] The "Non-Active" category is defined for inactive compounds within the concentration range of the dose-response diagram.

[0020] In response to EC 50 High-slope category of S-shaped curves with high slope;

[0021] In response to EC 50 Low-slope category of S-shaped curves with low slope;

[0022] The "partial" category is defined for S-shaped curves where the difference between A(c) at the upper asymptote and A(c) at the lower asymptote is less than 70%.

[0023] The activity exhibits alternating increases and decreases with concentration, categorized as a "wave" pattern; and

[0024] The “lastup” category includes those that show no activity other than the highest concentration or two highest concentrations.

[0025] The multiple dose-response categories may include:

[0026] Bell curve categories, and / or

[0027] Toxicity category based on the S-shaped curve indicating reduced inhibitory effect at higher concentrations.

[0028] The bell curve category and the toxicity category can be fused together into a single category for inference, but kept separate during training, allowing the model to independently learn slightly different visual patterns associated with each of these categories. Experts are not always able to easily distinguish between these two categories. However, in the implementation described in this specification, these categories can be advantageously distinguished by a curve shape classifier model because it operates on low-level features in the form of pixel data.

[0029] In another example implementation, this specification describes a computer-implemented method for generating a curve shape classifier model for classifying dose-response maps obtained from dose-response experiments. The method includes: receiving a plurality of training images at a neural network model, wherein each training image is an image of a dose-response map indicating 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 the neural network model, according to parameters associated with the one or more layers, and updating the parameters based on an objective function, the objective function including a comparison between the output generated for each training image and corresponding label data associated with the training image, the label data indicating that the training image belongs to one or more dose-response map categories related to curve shape. The neural network model may include a convolutional neural network model.

[0030] In another example implementation, this specification describes a data processing apparatus comprising one or more processors configured to perform any of the computer-implemented methods described herein.

[0031] In another example implementation, this specification describes a computer-readable storage medium including instructions that, when executed by one or more processors, cause the one or more processors to perform any computer-implemented method described herein.

[0032] As used herein, the term "dose-response plot" refers to a graph indicating the dose-response relationship between a compound concentration and its activity. A dose-response plot may, for example, comprise a set of data points plotted relative to Cartesian axes (e.g., a horizontal (X) axis and a vertical (Y) axis). In various embodiments, dose-response plots may be obtained from high-throughput screening processes. However, dose-response plots may alternatively be obtained from other sources, such as other screening processes (e.g., low-throughput screening). In the art, "dose-response plot" may alternatively be referred to as a "dose-response curve," but it is understood that the term "dose-response curve" does not necessarily imply the existence of a continuous curve.

[0033] Term EC 50 This refers to the half-maximum effective concentration (IC) of a compound, which is the concentration at which a 50% maximum response occurs. The term IC is also relevant. 50 This refers to the half-maximal inhibitory concentration (WMC), which is the concentration at which 50% inhibition occurs. It will be understood that dose-response mapping can measure activities other than inhibition; for example, it can measure the ability of a compound to act as an agonist of a biomolecule or organism. The claimed method can be used to classify dose-response mapping for measuring any kind of activity. Therefore, the term EC... 50 The term IC will be used throughout this description, but is to be understood as including implementation schemes that measure inhibition, and may be used with respect to the term IC. 50 .

[0034] As used herein, the term A(c) can be understood to refer to the activity (A) observed at a specific concentration (c). Similarly, the term I(c) can be understood to refer to the inhibitory effect (I) observed at a specific concentration (c). The term A(c) will be used throughout this description, but can be understood to include embodiments that measure the inhibitory effect, and the term I(c) may be used therein.

[0035] Understandably, the term "activity" can be understood as referring to biological or biophysical activity. For example, a dose-response graph can indicate 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 additionally, a dose-response graph can indicate the relationship between the concentration of a compound and its ability to bind to a biomolecule or organism. A biomolecule or organism can be or may include nucleophilic structures. A biomolecule or organism can be selected from amino acids, peptides, affinity molecules, proteins, glycoproteins, lipopolysaccharides, antibodies or fragments thereof, nucleic acids, organic polymers, viruses, bacteria, parasites, cells, and cell-associated structures. Proteins can be or include receptors, enzymes, ion channels, and / or transport proteins.

[0036] It is understood that the computer-implemented methods described herein can be used to aid in the identification of compounds with desired activity. These compounds can serve as potential drug candidates. Alternatively, medicinal chemists can prepare modified compounds based on or comprising the identified active compounds. The invention extends to any drug comprising or based on compounds already tested in dose-response experiments, and classifying the results of said experiments using the computer-implemented methods described herein. The invention also extends to any drug comprising or based on compounds already tested in dose-response experiments, and classifying the results of said experiments using the computer-implemented methods described herein for therapeutic purposes. The drug may be a compound already tested in dose-response experiments.

[0037] The computer-implemented methods described herein can be used to identify small molecules and / or biomolecules, such as, but not limited to, molecular biomarkers. The invention is also extended to small molecules or biomolecules already tested in dose-response experiments and the results of those experiments can be classified using the computer-implemented methods described herein. Attached Figure Description

[0038] To facilitate a clearer understanding of the invention, embodiments thereof are now described by way of example only with reference to the accompanying drawings, in which:

[0039] Figure 1 illustrates a set of dose-response map categories;

[0040] Figure 2 illustrates the classification pipeline based on the example implementation;

[0041] Figure 3 illustrates the architecture of the neural network shape classifier according to the example implementation;

[0042] Figure 4 illustrates the architecture of a neural network discrete classifier based on the example implementation; and

[0043] Figure 5 is a schematic diagram of a system / device used to perform the methods described herein. Detailed Implementation

[0044] Overview

[0045] Various example implementations involve systems for classifying DR plots based on a neural network acting on a normalized image of the dose-response (DR) plot. These systems can annotate thousands of curves across multiple categories within minutes to aid high-throughput screening (HTS) researchers in their analysis. Categories may be associated with active or inactive compounds, or with features of interest such as noise between replicates, weaker effects at high doses that may be related to toxicity, aggregation, or dissolution issues, or suspiciously weak or strong slopes at inflection points on the DR plot of the active substance.

[0046] introduce

[0047] High-throughput screening (HTS) is one of the primary strategies used in the pharmaceutical industry to find hit rates. Recently, screening techniques have become more sophisticated, leading to approaches such as quantitative HTS and strategies that use more reverse screening or selective assays to quantify hit rates. These techniques, in turn, increase the number of dose-response (DR) results generated. Other large dose-response datasets are obtained by querying protein libraries using selection techniques such as phage display, yeast display, or fluorescence-activated cell sorting (FACS). The quality of dose-response data depends on screening conditions, protocols, and overall assay robustness, and also on the performance of the compounds. Basic automated analysis of dose-response relationships relies on fitting algorithms, which can be unreliable in suboptimal settings due to outliers caused by interference effects or other technical artifacts. In practice, human review and action on dose-response data are required to make decisions about subsequent actions for the corresponding compounds in the project. Therefore, the visual inspection step is time-consuming, especially when hit rates are high, and the outcome of this step depends on the quality of the curves, the expert's experience, and the time available for analysis. This approach can delay projects when processing large volumes of results and may lead to a lack of consistency and robustness in the analysis over time.

[0048] When the active compound has EC within the measured concentration range 50 At concentrations that result in a 50% maximal response, whether for inhibition or other activity measurements, the ideal shape of the dose-response curve is either a full S-shape (with both its lower and upper asymptotes clearly defined) or a portion thereof. An ideal dose-response curve can be flat, either because the compound exhibits maximal activity across the entire concentration range or because it has no activity at all. However, DR experiments can be perturbed for a variety of reasons, causing the curve to deviate from these ideal shapes. Some are solely related to the nature of the compound (e.g., colloidal aggregation or toxicity issues at high concentrations), while others depend on the cell line or experimental protocol (e.g., interference from staining and fluorescent compounds in luminescence assays).

[0049] The standard post-processing workflow for DR plots involves using the Hill equation to fit the suppression percentage A(c) into an S-shape and extracting several parameters, such as EC. 50 and its confidence interval, EC 50The slope at the top and bottom asymptotes, A(c). The Hill equation itself is known to those skilled in the art and will not be described herein. 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, 20, 296-300.

[0050] Individually extracted parameters are insufficient to fully characterize the activity of a compound. Therefore, visual inspection of the DR plot is often performed, which may involve masking outliers to generate a better fit, confirming or adjusting the fitted parameters (top and bottom asymptotes, EC...). 50 The process involves identifying invalid experiments (slope), retesting them, annotating valid curves that show defects or specificities, and labeling each curve with a final decision label: "Active" (A), "Inactive" (NA), or "Invalid" (NV). This organizational step by experts faces challenges: it is time-consuming, may depend on the expert, and even an expert may face consistency issues when annotating critical cases at different times.

[0051] To alleviate these difficulties, this specification describes a computer-implemented method for automatically classifying DR images based on their visual characteristics. Specifically, as described in more detail below, the classification can be based on a visual pattern of dose-response relationship, 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, which is trained and configured to classify received DR image images into multiple DR image categories.

[0052] In the example implementation described below, the system can provide 14 different categories defined by experts and their classification probabilities. This solution allows experts to group similar DR (Digital Recognition) maps into categories with interpretable labels based on visual similarity for batch processing and easily identify poorly predicting curves for in-depth review. The system improves the speed of the inspection process and the robustness and consistency of the final decision.

[0053] Data Strategy

[0054] Figure 1 shows examples of DR diagrams across 14 categories, along with the associated category labels. Each category has an associated label, and the two terms (“category” and “label”) will be used synonymously.

[0055] These 14 categories comprise 13 subcategories (sets A, B, and C) that can be associated with well-defined curve shapes.

[0056] Among them, six categories form the first group (set A), which have different activity levels and no defects (e.g., no defects or extreme parameters):

[0057] - "Top" indicates an extremely potent compound that exhibits complete signal suppression across the entire concentration range.

[0058] - "No bottom" corresponds to potent compounds with an S-shaped inhibition curve, where the upper asymptote is visible, but the lower asymptote is not.

[0059] - "S-shaped" refers to the benign S-shaped curve of an active compound, including the lower asymptote and the upper asymptote.

[0060] - "Active without a peak" represents a DR plot that includes low asymptotes but no high asymptotes and reaches the 50% inhibition threshold, where the portion of the DR plot after the inflection point is visible.

[0061] - "No top" refers to weakly reactive compounds within the measured concentration range.

[0062] - An "inactive" DR plot is for compounds that are inactive in the assay.

[0063] The three categories in the second group (set B) correspond to S-shaped DR plots with some extreme parameters:

[0064] - "High slope" represents a full S-shaped DR plot, in EC 50 It has a high slope (typically nHill>4),

[0065] - "Low slope" conversely describes the situation in EC 50 A full S-shaped DR plot with a low slope (typically nHill < 0.5).

[0066] These patterns can be observed when a synergistic effect occurs in the system.

[0067] - The "partial" label refers to full S-shaped DR plots where the relative maximum inhibition (the difference between A(c) at the upper asymptote and A(c) at the lower asymptote) is less than 70%, which is often observed in cytometric assays.

[0068] The third group (set C) includes DR diagrams, which illustrate the defects commonly observed in practice:

[0069] - In a "bell-shaped" DR plot, A(c) decreases at its highest concentration, resulting in a bell-shaped curve. This pattern may be caused by signal interference issues in the fluorescence assay format or by compound aggregation problems.

[0070] The "toxicity" ("reverse effect at high concentrations") tag also applies to full S-shaped inhibition curves, where the inhibitory effect decreases sharply at the highest concentration due to the compound's toxicity, a phenomenon frequently observed in cellular assays. The "toxicity" and "bell-shaped" tags can be combined in post-processing, as discussed below.

[0071] - In the waveform DR plot, A(c) shows alternating increases and decreases. This could be due to various issues, such as interference in the signal readings or problems in the compound dilution series.

[0072] - The "Last Rise" label refers to a DR plot where no inhibition was observed in the titration curve except for the highest compound concentration or the two highest compound concentrations, resulting in an invalid curve.

[0073] These categories can be associated with three higher-order ensembles: “clearly active,” “clearly inactive,” and “requires review,” the latter set targeting all curves with specificity or defects. The “top,” “bottomless,” “S-shaped,” and “active without a top” categories are associated with the “clearly active” ensemble. The “inactive” category is associated with the “clearly inactive” ensemble. The “bottomless,” “high-slope,” “low-slope,” “partial,” “bell-shaped,” “toxic,” “waveform,” “last rise,” and “discrete” (see below) categories are associated with the “requires review” ensemble.

[0074] Human experts can visually examine DR plots and assign them to one of 13 categories (sets A, B, and C) based on the shape of the curve defining the trend of the dose-response relationship. For example, example DR plot 110 can be assigned to the "S-shaped" category because the curve defining the trend of the dose-response relationship is S-shaped. DR plot 120 can be assigned to the "waveform" category because the curve has a waveform shape, with alternating increases and decreases in activity. Thus, it can be intuitively understood from Figure 1 that each of the 13 categories (sets A, B, and C) can be associated with a corresponding curve shape that defines the dose-response relationship for that category. Therefore, these categories may be referred to herein as "shape categories".

[0075] Note that some categories may be associated with the same or similar curve shapes. For example, “top” and “inactive” have the same (horizontal) curves and may differ only in the magnitude of the reaction at each concentration. “Last rise” and “no top” also have some similarities, but the difference may be that the curve of “last rise” rises more steeply than that of “no top”.

[0076] As described in more detail below, an image classifier model can learn to classify images of DR maps based on the curve shape by training the classifier on images of DR maps that have been previously classified by experts.

[0077] Figure 1 also shows the “discrete” category (set D), where noise is observed between replicates. A replicate is two or more measurements taken at the same concentration value. Unlike the other categories, the “discrete” category is not related to a specific shape, but can still be processed separately as part of a classification pipeline, as will be described now.

[0078] Classification Assembly Line

[0079] Figure 2 illustrates an example classification pipeline 200, which includes a discrete classifier 210 and a shape classifier 220. The discrete classifier 210 acts as an initial filter, and only those DR maps that pass through this filter are submitted to the shape classifier 220. More specifically, the discrete classifier can output the probability p (a value between 0 and 1) that the dose-response map belongs to the "discrete" category. If the probability is greater than 0.5, the dose-response map is classified as the "discrete" category. Otherwise, a normalized image of the dose-response map is generated and fed into the shape classifier, which classifies the image according to the curve shape. As shown, images of dose-response maps classified with a probability below 0.9 can be assigned to a specific "low probability" category.

[0080] The shape classifier can be configured to classify DR 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 merged together (i.e., considered as a single category), because it has been found that experts are not always able to distinguish between these two categories.

[0081] The shape classifier 220 may include a convolutional neural network (CNN) that takes an image of the DR image 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, 521, 436-444.

[0082] The classification output generated by the CNN can include the probability of each of 12 shape categories, each probability being the likelihood that the DR image belongs to the corresponding category. Alternatively or additionally, the classification output can include an indication of the category with the highest probability, unless that probability is less than 0.9, in which case the DR image is classified as "low probability".

[0083] Discrete classifier 210 is a binary classifier because it classifies into two possible categories. It can include a multilayer perceptron (MLP) classifier, which takes statistical features extracted from the raw data of the DR map as input and outputs the probability that the DR map belongs to a discrete category. For example, to process the DR map using a discrete classifier, the positive difference of A(c) between replicates at each concentration can be extracted, and the quartiles q1, q2, and q3, as well as the interquartile range, can be calculated for this distribution. These four descriptors can be normalized between 0 and 1 using MinMax scaling, and these four normalized values ​​can be used as input to the classifier.

[0084] Shape classifier preprocessing

[0085] The shape classifier can receive images either during the training or prediction phase. During training, training images are used to train the model; during prediction, "new" DR image images not seen during training are classified. In either phase, input images can be generated from the raw DR image data by generating grayscale images that depict the set of data points of the DR image relative to the Cartesian (e.g., XY) axes. The generated images can be normalized because each image can be generated with the same size (e.g., 150×150 pixels), and the X and Y axes are located at the same position in each image. Furthermore, the generated images can be normalized because the Y-axis (activity) at the same position in each image is labeled with the same scale or value (e.g., values ​​between -50 and 150). On the other hand, the scale or value can be omitted from the X-axis (density), and different images do not need to involve the same range of density values ​​and can have different numbers of data points. To depict the data on a normalized "frame" defined by the images, the raw data can be normalized by removing / filtering data points falling outside that "frame." Various examples of DR images are shown in Figure 1.

[0086] Converting DR maps into images for CNN processing reduces the impact of any lack of homogeneity between different DR map samples (e.g., between DR map samples used in inference) compared to the training set. It provides a flexible classifier that can handle DR maps with, for example, different numbers of data points (e.g., 8, 10, or 12 concentrations), missing points, and / or different numbers of copies (which results in different numbers of Y values ​​for some concentration values).

[0087] Training data

[0088] DR maps used for training purposes can be obtained from existing data sources or generated by algorithms. DR maps can be manually labeled by experts into one of the 14 categories shown in Figure 1. The resulting labels can be represented as vectors and stored as "ground truth" data.

[0089] Generating DR maps algorithmically can involve using the Hill equation and parameters within a specific parameter space for each category label to generate synthetic dose-response data. The parameter space for the category labels may include, for example, the location of inflection points, the slope at those inflection points, and the locations of the top and bottom asymptotes. Noise and / or defects may be introduced.

[0090] Specifically, DR plots belonging to the "significantly active" and "significantly inactive" ensembles can be constructed based on the Hill equation. "Bell-shaped" and "waveform" curves can be generated by combining two or three Hill functions, respectively. "Toxicity" and "Last Rise" can be generated by adding appropriate noise to A(c) at the highest concentration, using curves from the "significantly active" and "significantly inactive" categories, respectively. Ten concentrations can be used, with two replicates generated for each concentration using logarithmic noise. To obtain a realistic noise distribution between replicates, statistics can be determined based on the differences between replicates using the results of (multiple) experimental HTS DR activities. Based on the fit of the noise distribution to the logarithmic law, associated parameters can be extracted, and the resulting parameterized functions can be used to model the noise between replicates. Normal noise can be added at the concentrations, and uniform noise can be added to A(c) across the entire concentration range.

[0091] Training images for training the shape classifier can be obtained from existing or algorithmically generated DR maps using the preprocessing steps described above. Thus, each training image may comprise a 150×150 pixel grayscale image depicting the set of data points of the DR map relative to the Cartesian axis. Each training image may be stored along with a "ground truth" label indicating the classification decision made by an expert (or more than one expert) regarding the DR map represented by the image.

[0092] Shape classifier architecture and training

[0093] An example architecture for a shape classifier is shown in Figure 3. In summary, it includes:

[0094] - A block containing six 2D convolutional layers using 3,3 filters (32, 32, 64, 64, 128, 128) with ReLU activation and L2 regularization;

[0095] - Four 2D max-pooling layers of size (2, 2);

[0096] - Five batch normalization layers;

[0097] - A flat layer;

[0098] - A dense layer consisting of 256 neurons with ReLU activation and L2 regularization;

[0099] - A batch normalization layer;

[0100] - A dropout layer with a coefficient of 0.5;

[0101] - A final dense layer with 13 output categories on Softmax activation.

[0102] Those skilled in the art will understand that many variations and modifications to the architecture shown in Figure 3 are possible.

[0103] Note that the "?" in Figure 3 refers to the number of training examples processed by the neural network at one time, and indicates that any suitable number can be used.

[0104] A shape classifier can be trained using a training set comprising several thousand (e.g., 5000) DR map images for each category. Training images can be generated from DR maps produced by an algorithm, manually labeled with "ground truth" classification tags as described above. Training images can be generated for each of the 13 shape categories discussed above.

[0105] During training, training images are received at the input layer of the neural network model (see Figure 3). Each training image is processed through subsequent layers in the neural network model according to the parameters of the model to generate a probability of belonging to each of the 13 shape categories shown in sets A, B, and C of Figure 1. The parameters of the neural network (i.e., neuron weights and biases) can be updated by optimizing an objective function. The objective function includes a loss that depends on a comparison between the output generated for each training image and the corresponding label data associated with that training image. The label data may include vectors representing “ground truth” labels, which are applied after a visual inspection of the training images by an expert (or a group of experts).

[0106] The loss can measure the mean squared error between the output for each training image and the ground truth label data. The objective function may additionally include a regularization term; for example, the objective function may 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 gradient-based methods such as the Adam optimizer, stochastic gradient descent, mini-batch gradient descent, or batch gradient descent.

[0107] In one example, the Adam optimizer is used with a learning rate of 1.0E-4 and a learning rate decay of 3.0E-7. Training is measured by validation loss and runs for a maximum of 300 epochs, stopping early at the total validation loss (patience = 5, delta = 0.001). The input data is split 80 / 20 between the training and validation sets.

[0108] The training process produces a trained shape classifier model. Once trained, this model can be used to classify “new” DR imagery. For a tested DR imagery, the trained model can generate a classification output that includes probabilities for each of 12 shape categories. These 12 shape categories are the same as the 13 shape categories in sets A, B, and C of Figure 1, except that “bell” and “toxic” (which are kept separate during training) are combined for inference (i.e., treated as a single category). Alternatively or additionally, the classification output may include an indication of the category with the highest probability, or an indication of two (or more) most probable categories.

[0109] Using convolutional neural networks for shape classifiers allows the classifier to focus on the overall shape of the curve rather than the details. However, alternative implementations can use other neural network architectures, such as fully connected neural networks.

[0110] Discrete classifier architecture and training

[0111] An example architecture of a discrete classifier is shown in Figure 4. The classifier includes a multilayer perceptron (MLP) classifier. It consists of two hidden layers, each with five neurons and ReLU activation.

[0112] Those skilled in the art will understand that many changes and modifications to this architecture are possible.

[0113] Note that the "?" in Figure 4 refers to the number of training examples processed by the neural network at one time, and indicates that any suitable number can be used.

[0114] The classifier can be trained using a training set comprising several thousand (e.g., 5000) DR maps per category. For each DR map in the training set, the positive difference between two replicates at each concentration can be extracted, and the q1, q2, q3 quartiles and interquartile ranges can be calculated from this distribution. This can be extended to any number of doses by calculating the positive difference between all replicates at each concentration and by using the distribution of all these differences when calculating the q1, q2, and q3 quartiles and interquartile ranges.

[0115] These four descriptors can be normalized between 0 and 1 using MinMax scaling, and these four normalized values ​​can be used as input to an MLP classifier.

[0116] The advantage of this approach is that there is no limit to the number of data points to consider, meaning that the classifier is effective even if some data points are missing, or if the concentrations differ between the inputs to the discrete classifier (or between the inputs used for prediction and the training data). This benefit is synergistic with using images as input to the shape classifier and contributes to a flexible classification pipeline that reduces the impact of any heterogeneity in the input data.

[0117] The input is processed through layers of the classifier based on the classifier's parameters to generate a classified output. The classifier's parameters (i.e., neuron weights and biases) can be updated by optimizing an objective function. The objective function includes a loss that depends on a comparison between the output generated for each training DR map and the corresponding label data associated with that training DR map. The label data may include binary values ​​representing "ground truth" labels, which are applied after an expert performs a visual inspection of the training images (i.e., whether the DR map is classified as a "discrete" category by the expert).

[0118] The loss can measure the mean squared error between the output for each training image and the ground truth label data. The objective function may additionally 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 gradient-based methods (e.g., stochastic gradient descent). Training can be performed using an 80 / 20 split between the training and validation sets.

[0119] It has been found that using MLP architectures for discrete classifiers is advantageous; however, other machine learning models, such as random forest models, can also be used.

[0120] Figure 5 illustrates a schematic example of a system / device for performing the methods described herein. The system / device shown is an example of a computing device. Those skilled in the art will understand that other types of computing devices / systems (such as distributed computing systems) can be used alternatively to implement the methods described herein.

[0121] Device (or system) 500 includes one or more processors 502. The one or more processors control the operation of other components of the system / device 500. The one or more processors 502 may, for example, include general-purpose processors. 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.

[0122] The system / device includes working or volatile memory 504. The 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, such as static RAM (SRAM), dynamic RAM (DRAM), or the volatile memory can include flash memory, such as an SD card.

[0123] The system / device includes non-volatile memory 506. Non-volatile memory 506 stores a set of operation instructions 508 for controlling the operation of processor 502 in the form of computer-readable instructions. Non-volatile memory 506 can be any type of memory, such as read-only memory (ROM), flash memory, or magnetically driven memory.

[0124] The one or more processors 502 are configured to execute operation instructions 508 causing the system / device to perform any of the methods described herein. Operation instructions 508 may include code relating to hardware components of the system / device 500 (i.e., drivers) and code relating to the basic operations of the system / device 500. Generally, the one or more processors 502 execute one or more instructions of operation instructions 508, which are permanently or semi-permanently stored in non-volatile memory 506, thereby using volatile memory 504 to temporarily store data generated during the execution of the operation instructions 508.

[0125] The methods described herein can be implemented in digital electronic circuit systems, integrated circuit systems, specially designed ASICs (Application-Specific Integrated Circuits), computer hardware, firmware, software, and / or combinations thereof. These may include computer program products (such as software stored on, for example, a disk, optical disk, memory, or programmable logic device), the computer program products including computer-readable instructions that, when executed by a computer (such as those described with respect to FIG. 5), cause the computer to perform one or more of the methods described herein.

[0126] The terms “drug” or “pharmaceutical” are used synonymously herein and describe pharmaceutical preparations containing one or more active pharmaceutical ingredients or their pharmaceutically acceptable salts or solvates, and optionally a pharmaceutically acceptable carrier. In the broadest sense, an active pharmaceutical ingredient (“API”) is a chemical structure that has a biological effect on humans or animals. In pharmacology, a drug or pharmaceutical preparation is used to treat, cure, prevent, or diagnose a disease, or to otherwise enhance physical or mental health. Drugs or pharmaceutical preparations may be used for a limited duration or periodically for chronic disorders.

[0127] As described below, a drug or pharmaceutical preparation may include at least one API or combination thereof in various types of formulations for the treatment of one or more diseases. Examples of APIs may include small molecules (having 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; and 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 may be incorporated into molecular delivery systems such as vectors, plasmids, or liposomes. Mixtures of one or more drugs are also contemplated.

[0128] Drugs or pharmaceutical preparations may be contained in primary packaging or "drug containers" suitable for use with drug delivery devices. Drug containers may be, for example, cartridges, syringes, reservoirs, or other robust or flexible vessels configured to provide suitable chambers for storing (e.g., short-term or long-term storage) one or more drugs. For example, in some cases, the chambers may be designed to store the drug for at least one day (e.g., from 1 day to at least 30 days). In some cases, the chambers may be designed to store the drug for about one month to about two years. Storage may occur at room temperature (e.g., about 20°C) or at refrigerated temperatures (e.g., from about -4°C to about 4°C). In some cases, drug containers may be or may include dual-chamber cartridges configured to separately store two or more components of a pharmaceutical preparation to be administered (e.g., an API and a diluent, or two different drugs), one component in each chamber. In such cases, the two chambers of a dual-chamber cartridge may be configured to allow mixing between the two or more components before and / or during administration to a human or animal. For example, the two chambers can be configured such that they are in fluid communication with each other (e.g., via a conduit between the two chambers), and allow the user to mix the two components as needed before dispensing. Alternatively or additionally, the two chambers can be configured to allow mixing during dispensing of the components into a human or animal body.

[0129] Drugs or agents contained in drug delivery devices as described herein can be used to treat and / or prevent many different types of medical disorders. Examples of disorders include, for example, diabetes or diabetes-related complications (such as diabetic retinopathy), thromboembolic disorders (such as deep vein or pulmonary thromboembolism). Other examples of disorders are acute coronary syndrome (ACS), angina pectoris, myocardial infarction, cancer, macular degeneration, inflammation, hay fever, atherosclerosis, and / or rheumatoid arthritis. Examples of APIs and drugs are those described in the following manuals: such as Rote Liste 2014 (e.g., but not limited to, main group 12 (antidiabetic drugs) or 86 (oncology drugs)) and Merck Index, 15th edition.

[0130] Examples of APIs used to treat and / or prevent type 1 or type 2 diabetes or complications associated with type 1 or type 2 diabetes include insulin (e.g., human insulin, or human insulin analogs or derivatives); glucagon-like peptide-1 (GLP-1), GLP-1 analogs or GLP-1 receptor agonists, or analogs or derivatives thereof; dipeptidyl peptidase-4 (DPP4) inhibitors, or pharmaceutically acceptable salts or solvates thereof; or any mixture thereof. As used herein, the terms “analyte” and “derivative” refer to a polypeptide having a molecular structure that is formally derived from the structure of a naturally occurring peptide (e.g., the structure of human insulin) by deletion and / or exchange of at least one amino acid residue present in a naturally occurring peptide and / or by addition of at least one amino acid residue. The added and / or exchanged amino acid residues may be encoding amino acid residues or other naturally occurring residues or purely synthetic amino acid residues. Insulin analogs are also referred to as “insulin receptor ligands”. Specifically, the term "derivative" refers to a polypeptide having a molecular structure that is formally derived from the structure of a naturally occurring peptide (e.g., the structure of human insulin), wherein one or more organic substituents (e.g., fatty acids) are bound to one or more amino acids. Optionally, one or more amino acids present in a naturally occurring peptide may have been deleted and / or substituted with other amino acids (including non-coding amino acids), or amino acids (including non-coding amino acids) may have been added to a naturally occurring peptide.

[0131] Examples of insulin analogs are Gly(A21), Arg(B31), Arg(B32) human insulin (glargine insulin); Lys(B3), Glu(B29) human insulin (glutamate insulin); Lys(B28), Pro(B29) human insulin (lispro insulin); Asp(B28) human insulin (aspart insulin); human insulin wherein the proline at position B28 is replaced by Asp, Lys, Leu, Val, or Ala and wherein Lys at position B29 can be replaced by Pro; Ala(B26) human insulin; Des(B28-B30) human insulin; Des(B27) human insulin and Des(B30) human insulin.

[0132] Examples of insulin derivatives are, for example, B29-N-myristoyl-des(B30) human insulin, Lys(B29)(N-tetradecanoyl)-des(B30) human insulin (detemir insulin, B29-N-palmitoyl-des(B30) human insulin; B29-N-myristoyl human insulin; B29-N-palmitoyl human insulin; B28-N-myristoyl LysB28ProB29 human insulin; B28-N-palmitoyl-LysB28ProB29 human insulin; B30-N-myristoyl-ThrB29LysB30 human insulin; B30-N-palmitoyl-ThrB29LysB30 human insulin; B29-N-(N-palmitoyl-γ-glutamyl)-des(B30) human insulin, B29-N-ω-carboxypentadecanoyl-γ-L-glutamyl-des(B30) human insulin (degludec insulin) ); B29-N-(N-lithochyl-γ-glutamyl)-des(B30) human insulin; B29-N-(ω-carboxyheptadecanoyl)-des(B30) human insulin and B29-N-(ω-carboxyheptadecanoyl) human insulin.

[0133] Examples of GLP-1, GLP-1 analogs, and GLP-1 receptor agonists include, for example, lixilatin. Exenatide (exocetide-4, Liraglutide, a 39-amino acid peptide produced by the salivary glands of the Gila monster. Semaglutide, Taspoglutide, Albiglutide Dulaglutide rExendin-4, CJC-1134-PC, PB-1023, TTP-054, Langelatide / HM-11260C (Efpeglenatide), HM-15211, CM-3, GLP-1Eligen, ORMD-0901, NN-9423, NN-9709, NN-9924, NN-9926, NN-9927, Nodexen, Viador-GLP-1, CVX-096, ZYO G-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, tezapatide (LY3298176), pamadotide (SAR425899), exenatide-XTEN, and glucagon-Xten.

[0134] An example of an oligonucleotide is, for instance, mipomersen sodium. It is an antisense cholesterol-lowering agent used to treat familial hypercholesterolemia, or RG012 used to treat Allport syndrome.

[0135] Examples of DPP4 inhibitors include linagliptin, vedagliptin, sitagliptin, denagliptin, saxagliptin, and berberine.

[0136] Examples of hormones include pituitary hormones or hypothalamic hormones or regulatory active peptides and their antagonists, such as gonadotropins (follicle-stimulating hormone, luteinizing hormone, human chorionic gonadotropin, fertility-stimulating hormone), growth hormone (growth hormone), desmopressin, terlipressin, gosorelin, triptorelin, leuprorelin, buserorelin, nafarelin, and goserelin.

[0137] Examples of polysaccharides include glucosaminoglycans, hyaluronic acid, heparin, low molecular weight heparin or ultra-low molecular weight heparin or their derivatives, or sulfated polysaccharides (e.g., polysulfated forms of the above polysaccharides), and / or their pharmaceutically acceptable salts. An example of a pharmaceutically acceptable salt of polysulfated low molecular weight heparin is enoxaparin sodium. An example of a hyaluronic acid derivative is Hylan GF 20. It is a type of sodium hyaluronate.

[0138] As used herein, the term "antibody" refers to an immunoglobulin molecule or its antigen-binding portion. Examples of antigen-binding portions of immunoglobulin molecules include F(ab) and F(ab')2 fragments that retain the ability to bind antigens. Antibodies can be polyclonal antibodies, monoclonal antibodies, recombinant antibodies, chimeric antibodies, deimmunized or humanized antibodies, fully human antibodies, non-human (e.g., murine) antibodies, or single-chain antibodies. In some embodiments, antibodies have effector function and can fix complement. In some embodiments, antibodies have reduced or no ability to bind Fc receptors. For example, antibodies can be isotypes or subtypes, antibody fragments, or mutants that do not support binding to Fc receptors, for example, they have a mutagenic or missing Fc receptor-binding region. The term antibody also includes antigen-binding molecules based on tetravalent bispecific tandem immunoglobulins (TBTI) and / or dual variable region antibody-like binding proteins with cross-binding region orientation (CODV).

[0139] The term "fragment" or "antibody fragment" refers to a polypeptide (e.g., antibody heavy chain and / or light chain polypeptide) derived from an antibody polypeptide molecule that does not contain the full-length antibody polypeptide but still contains at least a portion of the full-length antibody polypeptide capable of binding an antigen. Antibody fragments may contain cleaved portions of the full-length antibody polypeptide, although the term is not limited to such cleaved fragments. Antibody fragments that can be used in this invention include, for example, Fab fragments, F(ab')2 fragments, scFv (single-chain Fv) fragments, linear antibodies, monospecific or multispecific antibody fragments (such as bispecific, trispecific, tetraspecific, and multispecific antibodies (e.g., double-chain, triple-chain, and quadruple-chain antibodies)), monovalent or multivalent antibody fragments (such as bivalent, trivalent, quadruvalent, and multivalent antibodies), microantibodies, chelated recombinant antibodies, tri- or bispecific antibodies, intracellular antibodies, nanobodies, small modular immunopharmaceuticals (SMIPs), binding domain immunoglobulin fusion proteins, camel-derived antibodies, and antibodies containing VHH. Further examples of antigen-binding antibody fragments are known in the art.

[0140] The term "complementarity-determining region" or "CDR" refers to a short polypeptide sequence within the variable region of both heavy and light chain polypeptides, primarily responsible for mediating specific antigen recognition. The term "frame region" refers to an amino acid sequence within the variable region of both heavy and light chain polypeptides that is not a CDR sequence, and is primarily responsible for maintaining the correct positioning of the CDR sequence to allow antigen binding. Although frame regions themselves typically do not directly participate in antigen binding, as is known in the art, certain residues within the frame regions of some antibodies can directly participate in antigen binding or can affect the ability of one or more amino acids in the CDR to interact with the antigen.

[0141] Examples of antibodies are anti-PCSK-9 mAbs (e.g., alizumab), anti-IL-6 mAbs (e.g., sarerumab), and anti-IL-4 mAbs (e.g., dupilumab).

[0142] Pharmaceutically acceptable salts of any API described herein are also contemplated for use in drugs or pharmaceutical preparations in drug delivery devices. Pharmaceutically acceptable salts are, for example, acid addition salts and basic salts.

[0143] Those skilled in the art will understand that various modifications (additions and / or removals) can be made to the APIs, formulations, instruments, methods, systems, and embodiments described herein without departing from the full scope and spirit of the invention, which covers such variations and any and all equivalents.

[0144] Example drug delivery devices may involve needle-based injection systems as described in Table 1 of Section 5.2 of ISO 11608-1:2014(E). As described in ISO 11608-1:2014(E), needle-based injection systems can be broadly categorized into multi-dose container systems and single-dose (with partial or complete discharge) container systems. The container may be a replaceable container or an integrated, non-replaceable container.

[0145] As further described in ISO 11608-1:2014(E), a multi-dose container system can relate to a needle-based injection device with replaceable containers. In such a system, each container holds multiple doses, the size of which can be fixed or variable (preset by the user). Another multi-dose container system can relate to a needle-based injection device with an integrated, non-replaceable container. In such a system, each container holds multiple doses, the size of which can be fixed or variable (preset by the user).

[0146] As further described in ISO 11608-1:2014(E), a single-dose container system can relate to a needle-based injection device with a replaceable container. In one example of such a system, each container contains a single dose, thereby discharging the entire deliverable volume (complete discharge). In another example, each container contains a single dose, thereby discharging a portion of the deliverable volume (partial discharge). As also described in ISO 11608-1:2014(E), a single-dose container system can relate to a needle-based injection device with an integrated, non-replaceable container. In one example of such a system, each container contains a single dose, thereby discharging the entire deliverable volume (complete discharge). In another example, each container contains a single dose, thereby discharging a portion of the deliverable volume (partial discharge).

[0147] Many modifications and variations of the embodiments described herein will be apparent to those skilled in the art, and these modifications and variations fall within the definition of the following claims.

Claims

1. A computer-implemented method for classifying images, said images including dose-response maps obtained from dose-response experiments, the method comprising: The curve shape classifier model receives input including image data, which comprises multiple pixels, wherein the image data represents an image of a dose-response plot indicating the relationship between the concentration and activity of a compound, wherein the curve shape classifier model includes a neural network model configured to classify the dose-response plot image into multiple shape categories, wherein each shape category is associated with a corresponding curve shape, the curve shape defining the dose-response relationship of that category, and The neural network model is used to generate a classification output of an image represented by the received image data, the generation including processing the image data using one or more layers of the neural network model according to parameters associated with the one or more layers.

2. The computer-implemented method according to claim 1, wherein, The curve shape classifier neural network model includes a convolutional neural network model.

3. The computer-implemented method according to claim 1 or claim 2, comprising classifying dose-response maps into a first or second discrete category based on differences in activity measurements at the same concentration, wherein, The curve shape classifier model is used to process the image data representing the dose-response map only if the dose-response map is classified in the first discrete category.

4. The computer-implemented method according to claim 3, wherein, The dose-response map is classified into a first or second discrete category based on the interquartile values ​​of the difference in activity measurements as a function of concentration.

5. The computer-implemented method of claim 3, comprising classifying the dose-response map into a first or second discrete category using a discrete classifier comprising a multilayer perceptron neural network model.

6. The computer-implemented method according to claim 1 or 2, wherein, The multiple shape categories include one or more of the following: For categories that exhibit high activity across the entire concentration range; This refers to the category of S-shaped curves where the upper asymptote is visible but the lower asymptote is not; For categories of well-formed S-shaped curves that include both lower and upper asymptote portions; For the category of S-shaped curves where the lower asymptote is visible but the upper asymptote is not, and which reach the 50% activity threshold, the dose-response plot is visible after the inflection point; For the category of weakly active compounds within the concentration range of the dose-response diagram; The category of inactive compounds within the concentration range of the dose-response diagram; In response to EC 50 High-slope category of S-shaped curves with high slope; In response to EC 50 Low-slope category of S-shaped curves with low slope; For the category of S-shaped curves where the difference between A(c) at the upper asymptote and A(c) at the lower asymptote is less than 70%; The activity varies with concentration, showing alternating increases and decreases. Among them, no activity was shown except for the highest concentration or the two highest concentrations. EC 50 It is the concentration of the compound that causes a 50% maximum reaction. Where A(c) is the activity (A) observed at a specific concentration (c).

7. The computer-implemented method of claim 1 or 2, further comprising performing preprocessing, the preprocessing including receiving raw data representing a set of data points of the dose-response map and generating an image of the dose-response map based on the raw data.

8. The computer-implemented method according to claim 7, comprising: Receive raw data representing a set of data points for each of multiple dose-response graphs; For each dose-response map, a corresponding image is generated, wherein the image comprises multiple pixels and depicts at least some of the corresponding data point sets relative to the Cartesian axis, wherein each image is generated with the same pixel height and pixel width, and wherein the Cartesian axis is positioned at the same location in each image. The image data for each corresponding image is received at the curve shape classifier model.

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

10. A computer-implemented method for generating a curve shape classifier model for classifying dose-response maps obtained from a dose-response experiment, the method comprising: Multiple training images are received at the neural network model, where each training image is a dose-response map indicating the relationship between the concentration of a compound and its activity. Generate output for each training image, wherein generating the output for the training image includes processing the training image through one or more layers of the neural network model, processing the training image according to parameters associated with the one or more layers, and The parameters are updated based on an objective function that includes a comparison between an output generated for each training image and corresponding label data associated with the training image, the label data indicating that the training image belongs to one or more shape categories related to a curve shape, wherein each shape category is associated with a corresponding curve shape that defines the dose-response relationship of the category.

11. The computer-implemented method according to claim 10, wherein, Each training image depicts a corresponding set of data points relative to the Cartesian axis, wherein each training image has the same pixel height and pixel width, and wherein the Cartesian axis is located at the same position in each image.

12. The computer-implemented method according to claim 10 or claim 11, wherein, The shape categories include: Bell curve categories, and Toxicity category.

13. A data processing apparatus comprising one or more processors configured to perform the method according to any one of the preceding claims.

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 according to any one of claims 1 to 12.

Images