Systems and methods for machine learning predictive models to predict molecular targets of small molecules

Abstract

A computer-based method may involve collecting a set of known drug compounds from a database and creating an augmented dataset containing labeled and unlabeled chemical data, as well as biological data. The computer-based method may involve training a neural network using the augmented training set to generate predictive scores of the biological activity of candidate drug compounds against biological targets, wherein the contribution of a second set of unlabeled samples during training may be scaled by parameters. The computer-based method may output one or more improved neural network models capable of generating candidate drug compounds with predicted activity. The computer-based method may generate candidate drug compounds by inputting the chemical data associated with the candidate drug compounds into the trained neural network model.

Description

Detailed description of the invention

[0001] [Technical field] Incorporation by reference to any priority application This application claims the benefit of U.S. Provisional Patent Application No. 63 / 517,572, filed on 3 August 2023, which is incorporated herein by reference. [Background technology]

[0002] This disclosure relates to predicting the chemobiological activity of small molecules. More specifically, this disclosure relates to predicting the biological activity of small molecule drugs against DNA, RNA, and protein targets when existing biological data for a particular group of biological targets is scarce.

[0003] Description of related technologies Currently, efforts are underway to computationally predict the chemical and biological activity of drug molecules. Some of these efforts include quantitative structure-activity relationship (QSAR) models, deep learning models, molecular docking models, and generative models, all of which fall under the field of machine learning. All of these models use large compound libraries to screen potential drug candidates.

[0004] QSAR models analyze the relationship between the chemical structure of molecules and their biological activity. Specific machine learning algorithms such as linear regression, random forests, and support vector machines are trained on known chemical structure and corresponding biological activity data to predict the activity of new molecules. Deep learning models aim to extract complex features from molecular structures and learn complex patterns in large datasets. Techniques such as graph convolutional networks (GCNs) and recurrent neural networks (RNNs) have been applied to predict biological activity based on molecular graphs or sequential data. Molecular docking models are used to predict binding affinity between small molecules and target proteins. Machine learning approaches can improve the accuracy of docking by incorporating additional features, such as target-based descriptors, into the docking process.

[0005] The goal of generative models, such as generative adversarial networks (GANs) and variational autoencoders (VAEs), is to generate novel drug-like molecules. GANs attempt to understand the influence of molecular structure on biological activity and generate novel compounds as candidates by training on large chemical databases. While these models are useful for exploring chemical spaces and discovering novel drug candidates, they may still be limited in that they only generate drug candidates that are similar to or nearly identical to the original training set. Another limitation is that these models often generate compounds that cannot be synthesized.

[0006] Disease models of interest may have only small datasets, which presents significant challenges for machine learning, including but not limited to insufficient representation, overfitting, significant variance, and limited feature space coverage. Small datasets typically lack the diversity and quantity of examples needed to effectively capture the complexity of the problem, resulting in ineffective generalization and unreliable predictions. Models trained on small datasets can overfit, remembering limited data and performing poorly on unseen examples. High variance and limited feature space coverage hinder the model's ability to recognize true patterns, leading to extreme sensitivity to small parameter changes. [Overview of the prefecture]

[0007] One aspect of this disclosure relates to methods and systems for predicting the chemobiotic activity of small molecules. For example, some embodiments relate to targeting miRNAs using small molecules, which are involved in human diseases ranging from cancer to COVID-19. This disclosure provides a generalized deep learning framework for predicting the chemobiotic activity of small molecules against miRNA targets based on the chemical structure of the small molecules and the sequence information of the miRNA targets. A novel objective function is used that enables the neural network to learn the chemical space from a large amount of chemical structures, thereby overcoming the limitation of having little information about the chemobiotic activity of small molecules against miRNAs.

[0008] One aspect of this disclosure relates to scaling the size of a training set when a dataset of bioactive chemicals is only available to a limited extent. Embodiments of this disclosure include an improved neural network trained with the expanded dataset, which has been experimentally validated. This disclosure provides a loss function that scales the contribution of any unlabeled dataset to the training of the neural network. Another aspect of this disclosure relates to a framework for predicting bioactive small molecules that are chemically dissimilar to those available for training. One aspect relates to generating novel compounds, or existing but unused compounds, for activity against biological targets in a model organism.

[0009] For the purpose of summarizing the merits achieved with respect to this disclosure and all related research, certain objectives and merits of this disclosure have been described herein. Needless to say, it should be understood that not all of these objectives or merits can necessarily be achieved according to any particular embodiment of this disclosure. Therefore, for example, a person skilled in the art will recognize that this disclosure can be embodied or implemented in a manner that achieves or optimizes one or a group of merits taught herein, without necessarily achieving other objectives or merits that can be taught or suggested herein.

[0010] In some examples, the techniques described herein relate to computer-based methods for automatically training an artificial intelligence engine that generates candidate drug compounds, the computer-based methods comprising: collecting a set of known drug compounds from a database; creating a first training set containing chemical and biological data associated with the set of known drug compounds, wherein the chemical data contains structural information of the drug compounds and the biological data contains information on the biological activity of the drug compounds against their biological targets; creating a second training set containing chemical data associated with drug compounds having unknown biological activity against their biological targets; and combining the first and second training sets to form an extended training set. This includes training a neural network that generates predictive scores of the biological activity of candidate drug compounds against biological targets using an extended training set, wherein the contribution of a second set of unlabeled samples in training the neural network is scaled by a parameter α such that α is greater than 0 and less than 1; outputting one or more improved neural network models capable of generating candidate drug compounds having predicted activity; generating candidate drug compounds by inputting chemical data associated with the candidate drug compounds into the trained neural network models; and evaluating the predicted activity of the candidate drug compounds using the trained neural network models.

[0011] In some examples, the techniques described herein relate to computer-implemented methods for increasing the size of training set data for an artificial intelligence engine, the computer-implemented methods comprising: providing a first training set comprising a plurality of first training samples, each training sample comprising an associated input data label and a corresponding output label; and providing a second training set comprising a plurality of second partially labeled samples, each partially labeled sample comprising an associated input data label without a corresponding data output label, the second training set being larger than the first training set. The method includes: providing and optionally applying data augmentation techniques to an initial training set that generates an expanded training sample, wherein the data augmentation techniques introduce variation into the input data while retaining the corresponding output labels; combining a first training set and a second training set to form an expanded training set; and training a neural network using the expanded training set to generate predictive scores of the bioactivity of candidate drug compounds, wherein in training the neural network, the contribution of a second set of unlabeled samples is scaled by a parameter α, where α is greater than 0.

[0012] In some examples, the techniques described herein relate to computer-based methods for training a neural network engine to generate candidate drug compounds that affect one or more biological targets, wherein the computer-based methods include: collecting a set of known drug compounds from a database; creating a first training set containing chemical and biological data associated with the set of known drug compounds, wherein the chemical data contains structural information of the drug compounds and the biological data contains information on the bioactivity of the drug compounds against one or more biological targets, with each biological target having associated sequence information; and creating a second training set containing chemical data associated with drug compounds having unknown bioactivity against one or more biological targets. This includes: combining a first training set and a second training set to form an extended training set; calculating sequence similarity scores between biological targets based on sequence information of biological targets; training a neural network using the extended training set to generate predictive scores of the biological activity of candidate drug compounds against biological targets, wherein the contribution of each biological target to other biological targets is weighted by the sequence similarity score, and in training the neural network, the contribution of a second set of unlabeled samples is reduced compared to the first set of labeled data; and outputting one or more improved neural network models capable of generating candidate drug compounds with predicted activity.

[0013] All of these embodiments are intended to be included within the scope of the Disclosure disclosed herein. These and other embodiments of the Disclosure will be readily apparent to those skilled in the art from the modes for carrying out the inventions of the following preferred embodiments, and the Disclosure is not limited to any specific preferred embodiment(s) disclosed.

[0014] The features and advantages of the methods and compositions described herein will become more readily apparent by referring to the following description and the appended claims together with the accompanying drawings. These drawings illustrate only some embodiments of the present disclosure and should not be construed as limiting its scope. In the drawings, unless otherwise indicated in the context, similar reference numbers or symbols typically identify similar components. In some cases, the drawings may not be drawn to scale. [Brief explanation of the drawing]

[0015] [Figure 1A] This document illustrates an exemplary machine learning system for analyzing chemical and biological data. [Figure 1B] This document illustrates exemplary machine learning methods for analyzing chemical and biological data. [Figure 2A] The first and second training sets are shown. [Figure 2B] This describes a method by which the first training set and the second training set can be combined to form an extended training set. [Figure 3] This shows an extended training set that can be used to train a neural network model. [Figure 4] Figures A and B show the performance of neural networks trained using the methods and systems described herein, compared to other machine learning models. [Figure 5] Figures A and B show the performance of neural networks trained using the methods and systems described herein, compared to other machine learning models. [Figure 6] This plot shows the average predicted rank of small molecules with known biological activity. [Figure 7] Figures A and B demonstrate the predictive performance in a single-drop procedure when reading small molecules with activity against 50 randomly selected miRNA targets in Homo sapiens from a pool of 4,000 small molecules. [Figure 8A] A heatmap showing the most common mechanisms of action (MoA) and indications of small molecules predicted to target miRNAs in Homo sapiens by a neural network trained using the methods and system (sChemNET) described herein. [Figure 8B] A heatmap showing the most common mechanisms of action (MoA) and indications of small molecules predicted to target miRNAs in Homo sapiens by a neural network trained using the methods and system (sChemNET) described herein. [Figure 8C] A heatmap showing the most common mechanisms of action (MoA) and indications of small molecules predicted to target miRNAs in Homo sapiens by a neural network trained using the methods and system (sChemNET) described herein. [Figure 9A] An example of a trained neural network for predicting small molecules that bind to target miRNAs in other species is shown. [Figure 9B] An example of a trained neural network for predicting small molecules that bind to target miRNAs in other species is shown. [Figure 10] A shows the performance of a machine learning model in predicting the bioactivity of small molecules in Mus musculus (mouse), and B shows the performance of a machine learning model in predicting the bioactivity of small molecules in Rattus norvegicus (rat).

BEST MODE FOR CARRYING OUT THE INVENTION

[0016] In the embodiments for carrying out the invention described herein, refer to the accompanying drawings, which constitute part thereof. In the drawings, unless otherwise indicated in the context, similar symbols typically identify similar components. Thus, in some embodiments, part numbers may be used for similar components in multiple drawings, or different part numbers may be used in different drawings. The exemplary embodiments described in the embodiments for carrying out the invention, drawings, and claims are not intended to be limiting. Other embodiments may be used, and other modifications may be made, without departing from the spirit or scope of the subject matter presented herein. It will be readily apparent that the embodiments of this disclosure, as generally described herein and shown in the drawings, can be arranged, substituted, combined, and designed in a wide variety of different configurations. All of these are expressly assumed and constitute part of this disclosure.

[0017] MicroRNAs (miRNAs) are a type of non-coding RNA that plays a role in post-transcriptional gene regulation by regulating the levels of transcripts. MiRNA dysregulation has been associated with various metabolic and cardiovascular disorders, cancer, hepatitis, and infectious diseases such as COVID-19. MicroRNAs are highly stable and can circulate in the blood of individuals with disease, providing biomarkers and potential therapeutic targets for these diseases.

[0018] Oligonucleotide inhibitors with structures complementary to characterized miRNAs associated with specific diseases are being developed to inhibit these miRNA targets. For example, oligonucleotides complementary to miR-122 are being developed to treat hepatitis C virus, and antisense oligonucleotides against miR-2392 are being studied to treat COVID-19. Similarly, synthetic miRNAs are being developed for cancer treatment to control certain oncogenes and inhibit tumor growth. For example, mimics of miR-34 have been designed for the treatment of various human cancers. However, using oligonucleotides to target miRNAs or mimic miRNA activity has proven challenging due to the high toxicity of oligonucleotides and the challenges associated with their delivery.

[0019] Alternatively, miRNAs can be targeted by small molecules. However, information on the biological activity of small molecules that target miRNAs is limited. Currently, the biological activity of 131 small molecules has been evaluated against 126 human miRNAs. For most of these 126 miRNAs, fewer than 10 small molecules with biological activity have been identified. If this is taken correctly, the human genome contains more than 2,600 mature miRNAs, and there are many more small molecules that exist or can be synthesized to target each of these human miRNAs. The lack of information on small molecules and their effects on miRNAs presents a technical challenge in predicting the biological activity of new, unevaluated small molecules that have one or more miRNA targets. Existing models for predicting miRNA targets of small molecules have limited applicability and have only been able to predict the biological activity of 131 known small molecules against 126 known miRNA targets. These 126 miRNA targets had already been identified experimentally.

[0020] This application, in part, relates to a system, method, and computer-readable medium for predicting the bioactivity of known and unknown small molecules having miRNA targets. This specification describes a novel deep learning framework that, based on the chemical characteristics of the small molecule, generates, with improved accuracy and efficiency, the probability that a given small molecule can modulate a given miRNA target. In contrast to existing methods, the systems and methods described herein can be applied to identify novel bioactive chemicals from any chemical library. For example, novel small molecules may be identified based on information regarding the 2D chemical structure of a compound. The systems and methods described herein can be applied across species to identify small molecules for targeting diseases in human patients.

[0021] This disclosure provides a model capable of integrating chemical structure information for small molecules with unknown effects on miRNAs. Surprisingly, even when trained on small, known datasets, the disclosed model achieves good predictive performance for the biological activity of a broad range of small molecules. This disclosure further provides the application of interspecies data integration and integration of miRNA sequence information to complement smaller chemical datasets currently available for model organisms.

[0022] A system for predicting molecular targets of small molecules in datasets of small-scale biologically active chemicals. This disclosure relates to a novel machine learning system designed to predict the biological activity of a broad range of small molecules. Figure 1A shows an exemplary machine learning system 100 for analyzing chemical data 10 and biological data 20. Chemical data 10 may include labeled chemical data 12. Chemical data 10 may include unlabeled chemical data 14. Chemical data 10 may include both labeled chemical data 12 and unlabeled chemical data 14. In an example, labeled chemical data 12 contains structural information of small molecules and is labeled with paired biological data 20 for the small molecules. For example, a first small molecule may be clinically observed to have biological activity to a first biological target, no biological activity to a second biological target, and unknown biological activity to a third biological target. Unlabeled chemical data 14 may contain structural information of small molecules. In an example, chemical data 14 may lack paired biological data for the small molecules. The machine learning system 100 can apply a machine learning model 150 that utilizes advanced machine learning algorithms to analyze chemical data 10 and biological data 20.

[0023] The machine learning system 100 can apply a machine learning model 150 to train on chemical data 10 and biological data 20. The trained machine learning model 151 can be applied to a novel candidate small molecule 70. In an example, the trained machine learning model 151 can predict the biological activity of the small molecule 70. The molecule shown in Figure 1A is an example of a candidate small molecule, and the specific compound is 1-(4-(4-amino-6-(methoxymethyl)-7-(morpholinomethyl)pyrrolo[2,1-f][1,2,4]triazine-5-yl)-2-fluorophenyl)-3-(2-fluoro-5-(trifluoromethyl)phenyl)urea. This non-limiting example was determined to have biological activity against a miRNA target, while other examples included in the unlabeled dataset were not predicted to have biological activity.

[0024] In the example, the machine learning system 100 may predict novel candidate small molecules 70 by applying a machine learning model 150 to train on labeled chemical data 12 and unlabeled chemical data 14. In the example, the novel candidate small molecules 70 may be included in the unlabeled chemical data 14. In the example, a trained model 151 (not shown) may provide updated predictions for previously unlabeled candidate small molecules 170. However, it should be noted that machine learning models are typically not trained on the exact data they are used to predict in order to maintain a clear separation between training data and test data. In the example, the machine learning system 100 may also generate novel candidate small molecules 70 by applying a generative machine learning model 152 (shown as one embodiment of 150) to train on chemical data 10 and biological data 20. The generative machine learning model 152 may learn the underlying structure of molecules and generate novel compounds 170 having a desired bioactivity profile. The generative machine learning model 152 may include the same architecture as shown in Figure 1A, but may also include different architectures, with or without a neural network architecture.

[0025] As an example, this disclosure provides a deep learning predictive model that incorporates information on small molecules with known and previously unknown biological activity against miRNAs into a neural network model to predict small molecules that target miRNAs (or their downstream targets) based solely on the chemical structure of the small molecules. For the unlabeled chemical data 14, approximately 2,400 “unlabeled” small molecules were used, along with a smaller number of “labeled” small molecules 12. The labeled chemical data 12 consisted of a group of small molecules known to directly or indirectly affect the expression levels of miRNAs. The machine learning model 50 in this example included a two-layer neural network to which the chemical structure information of the labeled and unlabeled small molecules was fed and distributed into a group of hidden layers in the nodes. The output layer of the network represented each miRNA, and the model output a predictive score for each miRNA based on the chemical characteristics of the given small molecule. Thus, the trained machine learning model generates a mapping of scores between the chemical characteristics of small molecules and the group of miRNAs.

[0026] In some embodiments, the machine learning system may include a small molecule-miRNA interaction predictor that comprises a multitask, two-layer feedforward neural network. In an example, the machine learning system may include a set of input chemical features. In an example, the machine learning system may include a set of hidden units fully coupled to the input features (with a dropout parameter p, batch normalization, and a ReLU activation function). In an example, the machine learning system may include a set of hidden units fully coupled to the input features, which may then be coupled to a set of output units.

[0027] The output units may represent individual miRNAs. The output units may be fully bound to hidden units (with a dropout parameter p and a sigmoid activation function). The model can be trained with input features and output units by minimizing an appropriate loss function using the corresponding optimizer.

[0028] In accordance with this disclosure and the examples provided, the following hyperparameters, namely the number of hidden units (n), the unlabeled regularization parameter (α), the number of epochs (e), the learning rate (lr), and the dropout (p), can be optimized simultaneously or independently. In some embodiments, a Bayesian optimization method may be used for hyperparameter search using the following boundary: e∈[100,300], lr∈[0.0001,0.1], p∈[0.1,0.5], α∈[0.001,0.3]. The number of hidden units can be tested for discrete sets n∈{8,16,32}.

[0029] When predicting the biological activity of small molecules, one or more loss functions may be employed to train the model and improve the accuracy of predictions. For regression tasks, mean squared error (MSE) is commonly used. This measures the mean of the squares of the differences between the predicted activity value and the true activity value. Another option is mean absolute error (MAE), which calculates the mean of the absolute values ​​of the differences. For binary classification tasks (active or inactive), binary cross-entropy is often employed to assess the dissimilarity between the predicted probability and the true label. In the case of multi-class classification, multi-class cross-entropy is frequently used to encourage the model to assign a high probability to the correct class. In addition, if the objective is to rank molecules based on their activity levels, ranking losses such as pairwise ranking loss or listwise loss can be employed. As mentioned above, the choice of a specific loss function depends on the nature of the problem and the desired behavior of the model. However, a preferred loss function may include elements from the aforementioned loss functions and integrate chemical structure information of small molecules that were previously unknown to directly affect miRNAs or the transcriptional programs regulated by miRNAs. Specific embodiments will be described in more detail below. Briefly, the loss function provides an approach that can be applied to infer novel biologically active chemicals from various chemical libraries using information about the 2D chemical structure of compounds.

[0030] In some embodiments, machine learning models may be trained and tested using small chemical datasets containing labeled information about the biological activity of each small molecule against miRNA. Small Molecule to miRNA (SM2miR) databases can be used to obtain curated information about small molecules known to affect the expression levels of specific miRNAs or their corresponding mRNA targets. For each target miRNA, the number of biologically active small molecules in the dataset may vary, for example, between 5 and 35, and their distribution may follow a long-tail pattern. In the example, other datasets may be used that may contain more or fewer entries with different distributions. A wide range of 2-D structural information can be obtained for unlabeled drug compounds from a drug repositioning hub containing a large population of structurally and therapeutically diverse small molecules that have reached clinical drug development, including most FDA-approved drugs. 6,302 unique unlabeled small molecules can be used together with 131 small molecules from the SM2miR database to construct a collection of 6,433 unique small molecules. Each small molecule can then be represented by a 167-dimensional chemical feature vector based on its MACCS chemical fingerprint. For example, sequence similarity between miRNAs can be obtained by calculating the Needlemanbunsch score using their mature sequences from the miRBase database.

[0031] The use of a large amount of chemical information by sChemNET allows for the simulation of a realistic scenario in which small molecules with biological activity against miRNAs can be recovered from a large pool of chemicals. For this purpose, for each known biologically active small molecule-miRNA association, a test set can be constructed containing a large number of small molecules, of which only a few, or even just one, may be experimentally confirmed to be biologically active, while the rest may be randomly selected small molecules whose effect on the target miRNA is not yet known. For example, for each known biologically active small molecule-miRNA association, a test set containing 4,000 small molecules may be constructed, of which only one may be experimentally confirmed to be biologically active, and 3,999 may be randomly selected small molecules whose effect on the target miRNA is not yet known. In this example, high performance of the model can be observed when evaluated based on the proportion of known biologically active small molecules that can be read out from the top predicted small molecules.

[0032] The specific examples described herein offer one of many possible approaches to solving machine learning problems, while alternative methods can be employed at various steps in the process. In accordance with this disclosure, a diverse range of algorithms, techniques, and frameworks may be used to improve the performance of machine learning models, for example, by predicting novel compounds, integrating sequence information, scaling the size of training sets, and efficiently handling unlabeled datasets. The machine learning models described herein include specific steps tailored to the structure and dimensionality of the features the machine learning model receives as input data or generates as output predictions / classifications. For example, the size, shape, and type of input and output data for machine learning models may vary depending on the specific problem being addressed. Therefore, depending on the nature of the problem, alternative machine learning techniques may replace certain steps.

[0033] The size of the input data can refer to the number of observations available for the model to learn. In one example, biological data may include sequence information that can be reduced and represented by a sequence similarity coefficient. In another example, the sequence information may include polynucleotide sequences for a biological target, which allows for the extraction of potential patterns from the association between a drug and its sequence.

[0034] In the examples, sequence similarity coefficients and polynucleotide sequences have different structures and dimensionalities and may influence the architecture and design of machine learning models in accordance with this disclosure. The type of input data can vary widely, from numerical to categorical data (e.g., labels, text), or even to more complex forms such as sequences or graphs. In classification problems, the output may be a categorical label indicating the class to which each input sample belongs. For regression tasks, the output is typically a numerical value representing a predicted or estimated value. Other types of tasks may have different outputs, such as generating sequences, ranking small molecules for their potential impact on a biological target, or suggesting chemicals to be tested in "design of experiment" format.

[0035] Types of machine learning models The methods and systems provided herein may employ machine learning models 150 from one or more categories. For example, a machine learning model 50 could be a neural network system that introduces a multi-layer architecture including interconnected nodes, where these nodes mimic processing and learning mechanisms. The system may incorporate deep learning techniques, thereby enabling the network to automatically learn hierarchical representations of data features. By iteratively processing and transforming input data through multiple layers, the neural network system may capture meaningful patterns. The neural network system may incorporate advanced techniques such as regularization, dropout, and batch normalization to mitigate overfitting, improve generalization ability, and enhance the system's robustness when handling unseen data. Embodiments of the disclosure include deep learning models, which are subsets of neural networks comprising multiple layers of interconnected artificial neurons. Furthermore, the disclosure provides the incorporation of loss functions detailed herein, as well as activation functions such as rectified linear functions (ReLU), softmax, and sigmoid, to introduce nonlinearity and enhance the system's ability to accurately model complex data distributions.

[0036] Machine learning models can leverage both supervised and unsupervised learning algorithms to adaptively adjust their weights to fit the training data. Generally, supervised learning models are trained on labeled datasets, where input data is paired with corresponding target labels. These models learn to map input features to desired outputs, thereby enabling them to make predictions when applied to new, unseen data. Examples include linear regression, logistic regression, decision trees, random forests, support vector machines (SVMs), and neural networks.

[0037] Unsupervised learning models are trained on unlabeled data. This means that these models lack predefined target labels. These models aim to discover hidden patterns, structures, or relationships within the data. Clustering algorithms such as k-means clustering, hierarchical clustering, and Gaussian mixture models can be used to group similar data points together, representing either chemical or biological features. In other examples, dimensionality reduction techniques such as principal component analysis (PCA) and t-SNE can extract essential features and / or reduce the complexity of the data.

[0038] In some embodiments, semi-supervised learning models may use a combination of labeled and unlabeled data to train a machine learning model. These models may leverage limited labeled data to guide the learning process, while leveraging unlabeled data to improve generalization and accuracy. The machine learning models according to this disclosure may also leverage knowledge learned from one task or domain to improve performance in different but related tasks or domains. Such models may transfer pre-trained knowledge from a first dataset to a second dataset, thereby enabling better generalization on limited data.

[0039] One example of this disclosure concerns scaling the size of a training set to complement an initial small training set, and handling unlabeled data. Therefore, the machine learning models presented in this disclosure may be open-ended and can be combined with other machine learning techniques to make final predictions. By leveraging the diversity and collective insights of multiple models, ensemble methods improve accuracy, reduce overfitting, and enhance generalization.

[0040] A two-layer fully connected neural network model, also known as a shallow neural network, may include an input layer, a hidden layer, and an output layer. While a two-layer fully connected neural network model is considered shallow (i.e., having fewer layers compared to deeper architectures), it can still capture fundamental patterns and associations in data. Each layer may consist of interconnected artificial neurons or nodes, where the connections between neurons have associated weights. The input layer is responsible for receiving input data or features. Each neuron in the input layer represents one feature, and the number of neurons in this layer corresponds to the dimensionality of the input data. The values ​​of the input neurons can be directly determined by the input features.

[0041] From a model architecture perspective, a hidden layer can be placed between the input and output layers and can be responsible for processing and transforming the input data. Each neuron in the hidden layer can be configured to receive input from all neurons in the previous layer (in this case, the input layer) and apply an activation function to generate an output value. The activation function can introduce nonlinearity into the model, thereby enabling the model to learn complex patterns in the data. The number of neurons in the hidden layer can be determined by the complexity of the problem and the desired capacity of the model.

[0042] The output layer can be configured to receive input from neurons in the hidden layer and generate the final output of the neural network. The number of neurons in the output layer depends on the nature of the problem. For example, in a binary classification task, the output layer typically has one neuron, representing the probability or prediction of one class. In a multi-class classification task, the number of neurons may correspond to the number of classes, and the output may be represented as a probability (for example, using an activation function such as softmax).

[0043] Each neuron in the hidden and output layers can be connected to all neurons in the previous layer (fully connected). These connections have associated weights, which represent the strength or importance of the connection. During training, the model may be configured to iteratively adjust these weights using optimization algorithms such as gradient descent to minimize the difference between the predicted output and the true label. The weight values ​​can be learned by backpropagation, in which the error can be propagated backward from the output layer to the hidden layer, thereby updating the weights to improve the model's performance.

[0044] The models described herein can be modified and extended in various ways. For example, a model may be modified by incorporating regularization techniques, selecting different activation functions, and / or using different optimization algorithms to improve its performance and generalization ability.

[0045] Types of chemical data Chemical data 10 may include labeled chemical data 12 and unlabeled chemical data 14. Chemical data 10 may include, for example, molecular structures represented through atomic bond graphs or Simplified Molecular Input Line Notation (SMILES) strings, as well as descriptors that quantify the properties of a compound, including molecular weight, partition coefficients, and topological indices. Chemical fingerprints may encode binary patterns to represent the presence or absence of specific substructures, as well as physicochemical properties such as solubility and boiling point. In addition, some data, such as toxicity data and bonding coefficients, may be included as chemical data 10.

[0046] The size, shape, and type of input data for a machine learning model can vary depending on the specific problem being addressed. The shape of the input data represents the structure and dimensionality of the features and is considered in determining the architecture and design of a given model. Generally, a machine learning model takes input data, often called features, and produces output predictions or classifications. The size of the input data refers to the number of observations or samples available for the model to learn from. For example, if the task is to classify small molecules, the size of the input data corresponds to the total number of molecules in the dataset and the number of features provided for each molecule. The type of input data can range from numerical to categorical data (e.g., labels and text).

[0047] Chemical data can be represented using various descriptors, such as MACCs (Molecular Access System) keys. MACCs are a set of binary fingerprints that encode the presence or absence of specific chemical substructures or features within a molecule. The number of binary fingerprints can vary, but in some embodiments, a MACCs key may contain 166 fingerprints or 127 fingerprints. Each MACC key corresponds to a specific chemical pattern, such as the presence of a particular functional group, a ring system, or a specific atomic arrangement. The MACCs descriptor system provides a compact representation of chemical structures, thereby enabling efficient analysis of molecules. Each MACC key can be considered as a bit within a binary fingerprint, where "1" indicates the presence of the corresponding chemical feature and "0" indicates its absence.

[0048] In the example, the described process and system may be configured to provide predictions for the bioactivity of small molecules that may be chemically dissimilar to the molecules in the training set. Several methods exist for calculating chemical similarity. For example, the Tanimoto coefficient is an index of similarity between two sets. In the context of chemical informatics, the Tanimoto coefficient is commonly used to quantify the similarity between two chemical compounds or fingerprints. The Tanimoto coefficient can be calculated by dividing the intersection of the two sets by the union of the sets. Mathematically, it can be expressed as the following formula: Tanimoto(A,B) = (A∩B) / (|A|+|B|-A∩B)

[0049] A and B represent two sets, and |A| and |B| indicate the size (concentration) of set A and set B, respectively. When applied to chemical compounds, sets A and B typically represent the presence or absence of specific chemical features, which are often encoded as binary fingerprints or descriptors. The Tanimoto coefficient quantifies the degree of overlap between chemical features present in two compounds. The resulting Tanimoto coefficient values ​​range from 0 to 1, where 0 indicates no similarity (no common features) and 1 indicates perfect similarity (identical features). Therefore, a higher Tanimoto coefficient suggests a greater degree of similarity between chemical compounds. MACCs can also be used for similarity searches. MACC fingerprints allow for the identification of molecules with specific chemical features of interest, and by comparing MACC fingerprints, researchers can identify molecules that share similar chemical substructures.

[0050] The following chemical dataset is a non-limiting example of preferred chemical data 10. In some embodiments, the labeled chemical data 12 may include the SM2miR database 20 as of April 27, 2015. The database in this example contains manually curated associations between small molecules and miRNAs. Each small molecule in the labeled chemical data 12 may be mapped to a corresponding PubChem identifier (CID). In addition, the corresponding miRNA may be mapped to a miRbase identifier. A total of 4,244 small molecule-miRNA associations were extracted across 18 different species through the SM2miR database.

[0051] In one example, each organism under investigation may only retain miRNAs with a minimum of five small molecule associations. For Homo sapiens, 1,102 associations involving 131 small molecules and 126 miRNA targets may be available. For mice, 272 associations involving 44 small molecules and 43 miRNAs may be obtained. Brown rats contributed 78 associations, including 32 small molecules and 13 miRNAs.

[0052] In this example, a drug repositioning hub may be used to expand a chemical library of small molecules that do not have known activity against miRNAs. This hub contains a diverse collection of small molecules that have undergone clinical trials for various indications. Specifically, "non-duplicate" small molecules may be selected, thereby ensuring that they have different chemical structures than those available in the SM2miR database. As a result, the final collection of small molecules for Homo sapiens, mice, and brown rats reached 6,302, 6,281, and 6,294, respectively.

[0053] Types of biological data The labeled chemical data 12 contains structural information of small molecules and may be labeled with paired biological data 20 for each small molecule. The size and type of biological data 20 used as output data depend on the specific task at hand. In a classification problem, the output may be a category label indicating the class to which each input sample belongs, for example, whether it is biologically active or not. For a regression task, the output may be a numerical value representing a prediction. Other types of tasks may have different output requirements, such as generating sequences, ranking items, or detecting anomalies.

[0054] Specific biological data 20 may include, but are not limited to, gene expression data, protein sequences, and / or 3D structures. This type of information can contribute to understanding protein function, interactions, and drug binding. In some embodiments, some biological information may be used as labels for chemical data 10 and not strictly as output data. For example, the three-dimensionality of protein sequences and biological structures contributes to understanding protein function, interactions, and drug binding. Therefore, the protein sequence of a biological target may be incorporated into the input layer of the model by labeling it, for example, that a chemical may be active against a protein, and by including the protein sequence.

[0055] In some embodiments, biological data, such as miRNA sequence similarity, may be incorporated as weights in the loss function. Therefore, in some embodiments, chemical and biological data may be integrated without being explicitly included in the input or output layers. Mature miRNA sequences for biological targets may be obtained from the miRBase database using miRNA identifiers. miRNA sequence similarity may be calculated using mature miRNA sequences with global alignment in BioPython v1.76. Sequence similarity scores may then be normalized between 0 and 1.

[0056] training In the example, a method for training machine learning model 50 is described. For example, machine learning model 50 can be constructed by preparing representative and diverse datasets of the types described above. These datasets include input features and corresponding target labels, and furthermore, to ensure reliable model training, the dataset may be divided into training subsets, validation subsets, and test subsets, thereby enabling proper evaluation of the performance of the trained model on unseen data.

[0057] Prior to training, feature engineering techniques may be employed to optimize the input features. These techniques include normalization, scaling, encoding of categorical variables, handling of missing values, and dimensionality reduction. By performing feature engineering, the input features can be suitably transformed and prepared for optimal model performance, thereby enabling the model to learn from the most relevant information in the data. In some preferred embodiments, relevant information includes chemical data in MACCs key format 112 and biological data in binary categorical labels of active / inactive.

[0058] This disclosure provides for selecting an appropriate model architecture that is suited to a specific problem and data characteristics. As mentioned above, various types of models can be selected, such as neural networks, decision trees, support vector machines, or ensemble methods, each having its own strengths and suitability for different tasks. The model architecture determines the arrangement of the model's layers, nodes, and connections and affects the model's ability to learn relationships in the data. In some embodiments, the model architecture may include a set of input chemical features, a set of hidden units fully coupled to the input features (having a dropout parameter p, batch normalization, and a ReLU activation function), followed by a set of output units representing each of the miRNAs and fully coupled to the hidden units (having a dropout parameter p and a sigmoid activation function).

[0059] In some embodiments, the system may initiate the training process by initializing model parameters such as weights and biases. Proper initialization sets a starting point for the model to learn and improve its parameters based on the provided input data. This initialization may be performed using techniques such as random initialization or pre-training with relevant datasets. In some embodiments, unlabeled chemical data 14 may be initialized with an initial value of zero. In some embodiments, zero may be acceptable as an initial value for float types. In some embodiments, the value of zero may be fixed depending on the availability of known negative outcomes, and may be required to correspond to the absence of biological activity.

[0060] During the training phase, training algorithms such as gradient descent, stochastic gradient descent, or evolutionary algorithms may be employed. The algorithm iteratively updates the model parameters to minimize a loss function that quantifies the discrepancy between the model's predictions and the actual target labels. Optimization techniques such as learning rate scheduling, adaptive learning rates, or momentum may be employed to improve convergence and avoid falling into local optima.

[0061] When predicting the biological activity of small molecules, one or more loss functions may be employed to evaluate the accuracy of the prediction. For regression tasks, mean squared error (MSE), which measures the mean of the squared differences between the predicted activity value and the true activity value, is commonly used. Another option is mean absolute error (MAE), which calculates the mean of the absolute values ​​of the differences. For binary classification tasks (active or inactive), binary cross-entropy is often employed to evaluate the dissimilarity between the predicted probability and the true label. In the case of multi-class classification, multi-class cross-entropy is frequently used to encourage the model to assign a high probability to the correct class. In addition, if the objective is to rank molecules based on their activity levels, ranking losses such as pairwise ranking loss or listwise loss can be employed. In some embodiments, the loss function may be minimized using the ADAM optimizer with default parameters (beta1=0.9, beta2=0.999, epsilon=1e-7) in Tensorflow / Keras v2.8.0.

[0062] As further described herein, this disclosure provides specific loss functions that can integrate chemical and biological information. For example, the loss function may include a first summation term that applies fitting constraints to labeled chemical information. This first term may enable a machine learning model 50 to learn a high predictive score for known associations between small molecules and miRNAs.

[0063] One example of a loss function could include available labeled information for all other miRNAs, thereby allowing the relative learning contribution of each unknown small molecule to a biological target to be weighted based on the sequence similarity of the miRNA target to other targets.

[0064] Another term in the loss function may include hyperparameters to control the relative importance of unlabeled small molecules, which may be assigned lower prediction scores to each targeted miRNA during training / initialization.

[0065] Hyperparameter tuning can be a delicate step in the training process. Hyperparameters such as the learning rate, regularization parameters, or network architecture configuration can be fine-tuned to optimize the model's performance. Techniques such as grid search, random search, or Bayesian optimization can be employed to systematically explore the hyperparameter space and identify the optimal configuration for the model.

[0066] Evaluation and validation also play a role in the training process. Model performance can be periodically evaluated on separate validation datasets. Metrics such as loss, accuracy, or other relevant performance indicators can be monitored to assess the model's progress. Based on this evaluation, adjustments may be made to the model architecture or hyperparameters to improve the model's performance and generalization ability.

[0067] To prevent overfitting and determine an ideal stopping point, stopping criteria can be defined. These criteria may be based on the convergence of the loss function or an early stopping strategy that leverages the model's performance on a validation dataset. By adopting appropriate stopping criteria, the training process can be effectively managed, ensuring that the model is trained to the point where its predictive power is maximized without overfitting the training data.

[0068] Prediction of small molecule activity One embodiment may focus on determining the activity of small molecules against known biological targets. The majority of FDA-approved drugs are known to exert their therapeutic effects by binding to specific proteins within cells. Of all human proteins, only an estimated 700 have been successfully used in drug discovery to date. Therefore, this disclosure provides an in silico approach that can predict miRNA targets targeted by small molecules based solely on the chemical structure of the small molecule.

[0069] The model referred to herein as sChemNET is a non-limiting example prepared in accordance with this disclosure. sChemNET can be trained using labeled and unlabeled small molecules, and the trained model can be used to rank 4,000 small molecules in a test set by their predicted activity scores. sChemNET can achieve good performance with n=16, α=0.286, dropout=0.174, and learning rate=0.0346. The architecture of sChemNET may include one or more hyperparameters, namely, the number of hidden units (n), an unlabeled regularization parameter (α), the number of epochs, the learning rate, and the dropout. For hyperparameter exploration, a Bayesian optimization approach based on LOOCV of small molecules known to target miR-224-5p was employed. miR-224-5p is a miRNA that was randomly selected and subsequently excluded from evaluation analyses.

[0070] sChemNET's performance was evaluated based on the proportion of small molecules with known biological activity that could be read out from the predicted top 100, 300, 500, or 1000 small molecules. This was compared with and without integrating sequence similarity information. uvThe performance of sChemNET in (1) can be compared to other machine learning methods trained with the same input feature information as sChemNET, namely XGBoost, logistic regression (LR), random forest (RF), and feedforward neural networks (FNN), as well as two other approaches that rank each of the 4,000 small molecules in the test set based on (i) the highest Tanimoto chemical similarity (chemical similarity) with the bioactive small molecules in the training set, or (ii) a random score (random) assigned to each small molecule when sampling from a uniform distribution between 0 and 1. sChemNET outperformed the comparison methods in different numbers of predictions read: 1–9% for the top 100 small molecules read from the test set, 7–21% (for the top 300), 5–33% (for the top 500), and 8–29% for the top 1000. sChemNET achieves good prediction performance even without using sequence similarity information in its loss function, but with a slight decrease in prediction performance of approximately 1.81–3.62% across different top-K thresholds.

[0071] In some embodiments, the model provides a mechanism for interpreting and explaining the reasoning behind its predictions, thereby enabling users to understand the factors influencing the results and increasing transparency. In other embodiments, the model may only predict whether a small molecule can affect miRNA, but may not provide information about the molecular mechanism of action of the small molecule.

[0072] A method for predicting molecular targets of small molecules in datasets of small-scale biologically active chemicals. Figure 1B shows an example of a computer-based process for predicting the probability that a small molecule will bind to a target miRNA. In some embodiments, process 101 may include more or fewer steps. For example, one or more steps of process 101 may be performed in a different order or simultaneously with one or more other steps of process 101. In some embodiments, different machine learning methods may include more or fewer steps. For example, a classification model may include different steps or experimental designs than a regression model to generate predicted biological activity of a novel small molecule.

[0073] Process 101 may begin with step 110, in which a processor of one or more computing systems can read a group of small molecules from a database. Small molecules in the database may be known to have biological activity against at least one miRNA target. Small molecules known to have biological activity against at least one miRNA may be referred to herein as “labeled” small molecules. In some embodiments, the database may store chemical properties associated with each of the small molecules known to have biological activity against at least one miRNA target. These chemical properties may include the chemical fingerprint of each known small molecule. For example, this chemical fingerprint may be a 2-D chemical fingerprint such as a MACCS fingerprint, a PubChem fingerprint, or a custom fingerprint scheme.

[0074] In some embodiments, one or more processors may receive user input that includes a group of small molecules. In some embodiments, one or more processors may receive a chemical formula or structural formula for the small molecule(s). One or more processors may determine the chemical fingerprint of the small molecule(s) based on this chemical formula and / or structural formula.

[0075] In some embodiments, one or more processors may read the known biological activity of at least one small molecule against at least one target miRNA. A small molecule with biological activity may inhibit or promote the expression of miRNA, inhibit or promote the activity of the target miRNA, and so on. In some embodiments, the biological activity of a small molecule against miRNA may be assigned a biological activity score. The biological activity score may be selected from a range of values, e.g., 0 to 1, where 1 represents maximum biological activity and 0 represents no biological activity. The biological activity score may be selected from a category of values, e.g., 0 or 1, where 1 represents biological activity and 0 represents no biological activity. For example, the biological activity of a first small molecule known to affect a first miRNA target may be stored in a database, and one or more processors may read the biological activity of the first small molecule. In some embodiments, one or more processors receive the biological activity of a small molecule(s) via user input.

[0076] Process 101 may proceed to step 120, in which one or more processors may create a first training set. The first training set may be referred to herein as the training set of “labeled small molecules”. One or more processors may create the first training set using the information and data collected in step 110. One or more processors may compile chemical and biological data associated with a group of small molecules that have known biological activity against a target miRNA(s). Chemical data may include chemical properties associated with the small molecules, such as chemical fingerprints and molecular similarity scores. Biological data may include information on the biological activity of the small molecules against biological targets, such as miRNA targets.

[0077] The first training set may include at least one feature vector associated with at least one small molecule having known biological activity against at least one target miRNA. The feature vector may include chemical and / or biological properties associated with the small molecule(s). The feature vector may include the chemical fingerprint of at least one small molecule. The chemical fingerprint can be converted into an "N"-D chemical feature vector. The chemical feature vector may include multiple binary values ​​indicating the presence or absence of a chemical structure or motif. The length of the chemical feature vector may be based on the chemical fingerprint adopted. For example, in embodiments where a MACCS fingerprint is used, the chemical feature vector may be at least 166-dimensional.

[0078] In some embodiments, the feature vector may include additional properties associated with the small molecule, such as a miRNA sequence similarity score. In some embodiments, the miRNA sequence similarity score may be determined by calculating a Needlemanbunsch score. The Needlemanbunsch score for the small molecule and a specific miRNA target may be calculated based on the mature sequence of the target miRNA. For example, the chemical feature vector may be a 167-dimensional chemical feature vector. The 167-dimensional chemical feature vector may include 166 MACCS keys to which the Needlemanbunsch score is appended. In some embodiments, the sequence similarity score may be stored in a database and retrieved in step 110. In some embodiments, one or more processors may retrieve the mature forms of miRNA targets from the miRbase database and calculate the sequence similarity score for the small molecule.

[0079] The training set may include bioactivity scores for small molecules (or more) known to be bioactive to at least one miRNA. These bioactivity scores can be stored separately from the feature vectors as "true" values ​​that can be used to train the model.

[0080] Biological activity can be stored in a binary vector the length of the total number of miRNAs (e.g., N=126 in Homo sapiens data). In some examples, the vector takes the form =[1,0,0,0,1,…..].

[0081] Process 101 may proceed to step 130, in which case one or more processors may create a second training set. The second training set may be referred to herein as the training set of “unlabeled small molecules”. The second training set may include chemical and / or biological data associated with at least one small molecule having unknown biological activity against at least one target miRNA. Similar to the first training set, one or more processors may retrieve chemical data associated with the small molecule(s) having unknown biological activity against the target miRNA(s). For example, a database may store chemical data associated with small molecules having unknown biological activity, which may include chemical formulas, structural formulas, chemical fingerprints such as MACCS fingerprints, etc. In some embodiments, data associated with small molecules(s) having unknown biological activity against the target miRNA(s) may be stored in a separate database from data associated with small molecules(s) known to have biological activity against the target miRNA(s).

[0082] As described above, one or more processors can use chemical data to create chemical feature vectors for each small molecule that has unknown biological activity against a target miRNA(s). For example, one or more processors can read the MACCS fingerprint for each small molecule and, using the MACCS fingerprint, add a calculated sequence similarity score, such as the Needlemanbunsch score, to create a 167-dimensional feature vector. In some embodiments, one or more processors can set the biological activity of a small molecule with unknown biological activity to a negligible value, such as zero.

[0083] In some embodiments, the processor may proceed to step 140, in which one or more processors may combine the first training set and the second training set to form an extended training set. The extended training set may include at least one feature vector from the first training set and multiple feature vectors from the second training set. In some embodiments, the extended training set may include feature vectors from the first training set that may be known to have biological activity against target miRNAs. For example, a feature vector associated with a first small molecule known to target a first miRNA may be included in the extended training set. In another example, a first feature vector associated with a first small molecule known to target a first miRNA and a second feature vector associated with a second small molecule known to target a second miRNA may be included in the extended training set.

[0084] In some cases, the distribution of known "bioactive" small molecules differs for each miRNA (e.g., 5–35 in Homo sapiens data), so the number of labeled small molecules differs for each miRNA. In some cases, this number does not change in the expanded training set. This is because one possible purpose of the expanded training set is to increase the number of small molecules without adding new information about the labels.

[0085] Feature vectors of small molecules with unknown bioactivity from a second training set can be added to feature vectors from the first training set to create an extended training set.

[0086] In some cases, data can be supplemented by row-level and / or column-level concatenation. In some cases, rows have sub-units and columns have features, but the selection of row and column labels is arbitrary. Row-level and column-level concatenation are operations used in data manipulation and matrix processing. Row-level concatenation involves combining two or more datasets or matrices by stacking them vertically, for example, placing one dataset or matrix below another. This operation increases the number of rows in the resulting dataset or matrix while maintaining the original number of columns. Column-level concatenation, on the other hand, involves merging datasets or matrices horizontally, for example, placing one dataset or matrix next to another. These types of operations increase the number of columns in the resulting dataset or matrix while maintaining the original number of rows.

[0087] In some embodiments, feature vectors from a second training set may be randomly selected. We randomly select a subset of small molecules along with their corresponding feature vectors.

[0088] For a given miRNA u, we randomly select unlabeled small molecules (those not yet known to affect miRNA u). These are over 2,400 small molecules retained as part of the training set, and for the remainder, we use them for prediction (or LOOCV methods). The use of random selection is to prevent the model from learning the details of the chemical feature->label mapping when label=0. Rather, we want the model to learn what constitutes a chemical feature->label when label=1 (having biological activity); that is, to learn meaningful chemical features from those that are biologically active, while also using unlabeled small molecules to better identify those that are biologically active in the chemical space.

[0089] In some embodiments, the number of feature vectors from a second training set may be proportional to the number of feature vectors from a first training set. For example, the number of feature vectors from a second training set may be at least 5 times, 10 times, 15 times, 20 times, 25 times, 30 times, 35 times, 40 times, 45 times, 50 times, or any other suitable multiple compared to the number of feature vectors from a first training set.

[0090] In some embodiments, the number of feature vectors from the second training set may represent a proportion of the total number of feature vectors in the second training set. For example, this could include about 5%, about 10%, about 15%, about 20%, about 25%, about 30%, about 33.33%, or 1 / 3, about 35%, about 40%, about 45%, about 50%, or any other preferred proportion of the feature vectors in the second training set. The remaining portion of the second training set not included in the extended training set can be used as a test set. In some embodiments, feature vectors associated with small molecules having unknown bioactivity may be analyzed using the methods described herein and excluded from the extended training set.

[0091] Process 101 may proceed to step 150, in which one or more processors may train a model that generates predicted bioactivity scores for candidate small molecules against at least one target miRNA using an augmented training set. In step 150, the method performed by the computer may include training a neural network that generates predicted bioactivity scores for candidate drug compounds against biological targets using an augmented training set, wherein the contribution of a second plurality of unlabeled samples to the training of the neural network is scaled by a parameter α, where α is greater than 0 and less than 1. The model may be a neural network consisting of multiple layers. The neural network may include an input layer, at least one hidden layer, and an output layer. In some embodiments, the neural network may include at least two hidden layers, at least three hidden layers, at least four hidden layers, or at least five hidden layers.

[0092] The input layer may include binary inputs corresponding to each value in the small molecule's feature vector. At least one hidden layer may include multiple nodes. In some embodiments, at least one hidden layer may include at least 5 nodes, at least 10 nodes, at least 15 nodes, at least 20 nodes, at least 25 nodes, at least 35 nodes, or at least 40 nodes. The number of nodes in at least one hidden layer may be optimized over time based on training and subsequent model performance. The output layer may include predicted bioactivity values ​​for each target miRNA. The predicted bioactivity values ​​may represent the percentage probability that the small molecule will affect the expression level of the target miRNA. In some embodiments, the predicted bioactivity values ​​may represent the percentage probability that the small molecule will bind to the target miRNA and thereby affect the expression level of the target miRNA or its mRNA target.

[0093] In some examples, and throughout this disclosure, the prediction score and output probability may be used interchangeably as two different ways of representing a model's prediction for a given input. The prediction score may be a real number representing the model's estimate for a specific outcome. This score may be continuous or discrete, depending on the type of problem being solved. For example, in a regression task, the prediction score may represent a numerical value predicting, such as the price of a house or the binding constant of a miRNA target. In a binary classification task, the prediction score may represent the model's confidence in assigning an instance to a particular class (e.g., 0 for one class and 1 for the other), or any value in between.

[0094] The output probability can be a value between 0 and 1, and in classification problems, it represents the likelihood or confidence that a given input belongs to a particular class. In binary classification, the output probability represents the probability of belonging to the positive class (Class 1), and by extension, the value obtained by subtracting that probability from 1 represents the probability of belonging to the negative class (Class 0). For example, an output probability of 0.8 for the positive class means that the model is 80% confident that the input belongs to that class.

[0095] One difference where applicable is that prediction scores are often used in regression tasks or binary classification tasks when the model is not probabilistic, while output probabilities are generally used in probabilistic models such as logistic regression, softmax regression, or deep learning models with softmax activation in the output layer.

[0096] In some embodiments, the model may provide a ranking of each small molecule based on its predicted bioactivity value.

[0097] Each layer may contain multiple nodes or neurons, each of which can be connected to at least one neuron in each neighboring layer. In some embodiments, the neural network may be fully connected, and each node may be connected to all nodes in each neighboring layer. For example, each node in an input layer may be connected to all nodes in at least one hidden layer, and each node in at least one hidden layer may be connected to each node in the output layer.

[0098] A neural network may include one or more hyperparameters. These hyperparameters may include the number of nodes, the number of epochs, the learning rate, the dropout rate, and / or the loss function hyperparameter of at least one loss function. In some embodiments, one or more hyperparameters may be optimized using a set of small molecules known to bind to a specific target miRNA. The target miRNAs used to optimize the hyperparameters may be selected based on target miRNAs and / or candidate drug compounds evaluated by the neural network after training. For example, if a set of candidate drug compounds is evaluated against a first target miRNA, a second target miRNA may be used to optimize the hyperparameter(s).

[0099] In some embodiments, a Bayesian optimization approach may be used to identify desirable or optimal values ​​for one or more hyperparameters. For example, a set of small molecules known to couple with miR-224-5p may be input to the neural network, and the values ​​of one or more hyperparameters may be identified by using one-miss cross-validation (LOOCV) and a Bayesian optimization approach. In this embodiment, the number of nodes in each hidden layer was identified as 16, the learning rate as 0.0346, and the dropout rate as 0.174. The values ​​of one or more loss function hyperparameters may also be identified as described above.

[0100] A neural network can be trained using a loss function. In some embodiments, the loss function may be a mean squared error function. The loss function may separately weight the predicted bioactivity of small molecules with known bioactivity and small molecules with unknown bioactivity. In some embodiments, the loss function may include a first weighted sum for labeled small molecules with known bioactivity and a second weighted sum for unlabeled small molecules with unknown bioactivity.

[0101] As described above, the loss function may include at least one loss function hyperparameter. The loss function hyperparameter may include one or more sequence similarity weighting parameters and / or unknown bioactivity weighting parameters. The sequence similarity weighting parameter may be expressed as the term "s". The sequence similarity weighting parameter may weight the difference between the predicted bioactivity and the true bioactivity of a small molecule known to have bioactivity for at least one target miRNA, using the sequence similarity with small molecules that have known bioactivity for at least one target miRNA. The unknown bioactivity weighting parameter may be expressed as alpha, or "α". The unknown bioactivity weighting parameter may weight the predicted bioactivity for small molecule compounds that have unknown bioactivity.

[0102] In some embodiments, the loss function may be defined by the following equation:

[0103]

number

[0104] "i" may represent the i-th small molecule, and "u" may represent the target miRNA.

[0105]

number

[0106]

Number

[0107]

Number

[0108] As described above, the term "s" in the above formula may represent a sequence similarity weighting parameter, and the term "α" in the above formula may represent an unknown bioactivity weighting parameter. "s" may be set based on the sequence similarity score for the i-th small molecule. In some embodiments, "s" may be equal to the sequence similarity score read out and / or calculated for the i-th small molecule in steps 110, 120, and / or 130, e.g., the Needlemanbunsch score. In some embodiments, "α" may be a value greater than or equal to 0 but less than or equal to 2. In some embodiments, "α" may be a value greater than or equal to 0 but less than 1.

[0109] In some embodiments, s and / or α can be optimized using the Bayesian optimization approach described above. Returning to the example above, a set of small molecules known to bind with miR-224-5p may be input to the neural network, and optimal values ​​for loss function hyperparameters such as α can be determined by using one-miss cross-validation (LOOCV) and the Bayesian optimization approach. In some embodiments, α may be greater than 0.2 but less than 0.3. In some embodiments, α may be greater than 0.25 but less than 0.3. In some embodiments, α may be approximately equal to 0.286.

[0110] Using optimized hyperparameters and loss functions, the neural network can be trained with the expanded training set created above. The neural network can be trained and validated using one or more of the aforementioned optimized hyperparameters, such as the number of hidden units, the number of epochs, the learning rate, and the dropout rate. One or more parameters may be modified during training to minimize the neural network's loss function. In some embodiments, one-miss cross-validation (LOOCV) may be employed to evaluate the performance of the neural network during training.

[0111] Process 101 may proceed to step 160, in which one or more processors may output one or more improved neural network models capable of predicting the bioactivity of a candidate drug compound against at least one target miRNA. One or more processors may train one or more neural networks to predict the bioactivity of small molecules against at least one target miRNA. In some embodiments, one or more processors may generate neural networks that predict the bioactivity of input small molecules against multiple target miRNAs. For example, one or more processors may generate a neural network model that predicts the bioactivity of at least one candidate drug compound against at least 125 known human miRNAs. In some embodiments, one or more processors may train multiple neural networks, each configured to predict the bioactivity of a candidate drug compound against one target miRNA. In some embodiments, one or more improved neural networks may be stored in computer-readable memory.

[0112] In step 170, one or more processors may generate candidate drug compounds by inputting chemical data associated with the candidate drug compounds into a trained neural network model. As the chemical data propagates through the neural network model, it undergoes mathematical operations and activation functions within each layer. These operations include weighted sums, nonlinear transformations, and activation functions that introduce nonlinearity and enable the model to capture complex relationships within the data. The model's output is generated, representing the predicted properties or features of the candidate drug compounds.

[0113] The output generated from a neural network model can be post-processed to interpret it and convert it into actionable information. This may include deciphering the model's predictions, applying thresholds, or converting the output into a suitable chemical or biological representation. Based on the predicted properties or features, the processor(s) generate or select candidate drug compounds that exhibit desirable characteristics such as high efficacy, low toxicity, or target specificity.

[0114] An iterative optimization process may be employed to refine the generated candidate drug compounds. This process may involve modifying the specific chemical characteristics or properties of the compound while maintaining its desirable features. Techniques such as genetic algorithms, evolutionary strategies, or gradient-based optimization can be used to iteratively explore the chemical space and generate improved candidate compounds.

[0115] In step 170, one or more processors may provide the generated candidate drug compounds as output, while these candidates may then be further evaluated through in silico methods such as molecular docking, molecular dynamics simulation, or ADME (absorption, distribution, metabolism, and excretion) prediction to assess their potential for further development. This evaluation helps determine the likelihood that the candidate compounds will exhibit the desired bioactivity, pharmacokinetic properties, and safety profile.

[0116] Process 101 may proceed to step 180, in which one or more processors may evaluate candidate drug compounds. One or more processors may evaluate candidate drug compounds by predicting the bioactivity of the candidate drug compounds using one or more improved neural network models. As described later, one or more improved neural network models may output the probability that a candidate drug compound has bioactivity against at least one target miRNA. In some embodiments, the neural network may rank multiple candidate drug compounds based on the predicted bioactivity of the candidate drug compound against at least one target miRNA. The ranking may include candidate drug compounds with the highest bioactivity, for example, the top 100, top 200, top 250, top 300, top 400, top 500, top 600, top 700, top 750, top 800, top 900, top 1,000, etc. In some embodiments, the ranking may include one or more small molecule / candidate drug compounds that are known to have bioactivity against the target miRNA. In step 160, the method may proceed by outputting one or more trained neural networks that may be capable of generating candidate drug compounds having predicted activity against biological targets. The disclosure offers several advantages useful for predicting novel small molecules, any of which may be used individually or in combination. For example, the disclosure may be used to create an augmented dataset and then to train a neural network, which may then be used to evaluate or propose novel drug compounds. The trained neural network may be used in a variety of ways, including evaluating novel drugs that can be fed into the model (during or after training) or proposing novel drugs (usually after training), as described herein.

[0117] Based on predictions from machine learning models, chemical compounds may be identified as having potential biological activity against specific targets or diseases. The selection process for choosing any of the chemical compounds may include factors such as predicted efficacy, selectivity, safety profile, and relevance to the target or disease of interest. Chemical compounds may undergo clinical trials after preclinical evaluations have been conducted to assess their safety and efficacy in vitro and in animal models. This step may involve conducting experiments in relevant biological systems to determine the pharmacokinetics (absorption, distribution, metabolism, and excretion) and pharmacodynamics (mechanism of action, target engagement, and biological effects) of the compound.

[0118] In some embodiments, some or all of the steps described above may be carried out by a computer as a way to increase the size of the training set data for an artificial intelligence engine. Alternatively, or in combination, some of the steps described above may be carried out by a computer as a way to train a neural network engine to generate candidate drug compounds that affect one or more biological targets.

[0119] Generally, exemplary methods for training a neural network engine to generate candidate drug compounds that affect one or more biological targets may include: collecting a set of known drug compounds from a database; creating a training set containing chemical and biological data associated with the set of known drug compounds, wherein the chemical data contains structural information of the drug compounds, and the biological data contains information on the bioactivity of the drug compounds against one or more biological targets, each of which has associated sequence information; calculating sequence similarity scores between biological targets based on the sequence information of the biological targets; training a neural network using the expanded training set to generate predictive scores of the bioactivity of candidate drug compounds against the biological targets, wherein the contribution of each biological target to other biological targets is weighted by the sequence similarity scores; and outputting one or more improved neural network models capable of generating candidate drug compounds with predicted activity. Examples of such models have shown that they outperform comparable models by 6.18–24.67% in the top 300 (7.5%) of the retrieved predictions, and by 2.74–20.50% in the top 1000.

[0120] Figure 2A shows a first training set 202 and a second training set 204. The first training set may be stored in a first database, such as the SM2miR database 206 shown in Figure 2A. As described above, the first training set may include at least one small molecule having known biological activity against at least one target miRNA. For example, as shown in Figure 2A, the first training set 202 includes at least four small molecules having known biological activity against target miRNA 208. The first small molecule 210 and the fourth small molecule 216 may increase the activity of target miRNA 208, while the second small molecule 212 and the third small molecule 214 may decrease the activity of target miRNA 208. In some embodiments, the first training set 202 may include chemical and / or biological information associated with small molecules known to have biological activity against at least one target miRNA 208. For example, the first training set may include feature vectors for each of the small molecules 210, 212, 214, and 216. The feature vectors may include chemical fingerprints such as MACCS fingerprints, and sequence similarity scores between each of the small molecules 210, 212, 214, and 216 for the target miRNA 208.

[0121] The second training set 204 may include multiple small molecules 218 having unknown biological activity against at least one target miRNA, such as target miRNA 208. The second training set may include feature vectors for each of the multiple small molecules 218. Similar to the first training set, the feature vectors may include chemical fingerprints of the small molecules, such as MACCS fingerprints.

[0122] Figure 2B illustrates how the first training set 202 and the second training set 204 can be combined to form an extended training set 220. As described above, the extended training set 220 may include at least one feature vector from the first training set 202 and multiple feature vectors from the second training set 204. For example, the extended training set 220 may include a feature vector associated with a first small molecule 210 having known biological activity against the target miRNA 208, and multiple feature vectors associated with multiple small molecules 218 having unknown biological activity against the target miRNA 208. For example, the extended training set may include 3,999 feature vectors associated with the feature vector associated with the first small molecule 210 and small molecules belonging to the multiple small molecules 218 having unknown biological activity.

[0123] Figure 3 shows an extended training set 302 that may be used to train the neural network model 304. As discussed herein, the extended training set 302 may include at least one feature vector from a first training set 306 containing information about at least one small molecule having known biological activity against a target miRNA, and multiple feature vectors from a second training set 308 containing information about multiple small molecules having unknown biological activity against a target miRNA.

[0124] The extended training set 302 may be used to train the neural network model 304. As described above, the neural network model 304 may be a neural network consisting of multiple layers. The neural network 304 may include an input layer 310, at least one hidden layer 312, and an output layer 314. Each layer may include multiple nodes or neurons, each of which can be connected to at least one neuron in each neighboring layer. In some embodiments, the neural network may be fully connected. The input layer 310 may include binary inputs corresponding to each value in the feature vector of a small molecule. The at least one hidden layer 312 may include multiple nodes. In some embodiments, the at least one hidden layer 312 may include at least 5 nodes, at least 10 nodes, at least 15 nodes, at least 20 nodes, at least 25 nodes, at least 35 nodes, at least 40 nodes, or any other preferred number of nodes. The output layer 314 may include predicted bioactivity values ​​for each target miRNA 316. In some embodiments, Model 304 may provide ranking of small molecules based on the predicted bioactivity value for each small molecule.

[0125] Figures 4A, 4B, 5A, and 5B show the performance of the neural network 402 trained using the method and system described herein, compared to other machine learning models. The trained neural network may be referred to as the “sChemNET” model 402 in this specification and in the drawings. Figures 4A, 4B, 5A, and 5B show the performance of the sChemNET model 402. SuvThe sChemNET model 402 is compared with other versions of the sChemNET model, such as the sChemNET model 404, where =1. In addition, Figures 4A, 4B, 5A, and 5B compare the sChemNET model 402 with existing models including the XGBoost model 406, the logistic regression model 408, the random forest model 410, the feedforward neural network ("FNN") model 412, the chemical similarity model 414, and the random model 416, which randomly selects small molecules. The chemical similarity model 414 predicts the bioactivity of small molecules using the highest Tanimoto chemical similarity with a group of active small molecules in the training set.

[0126] In Figures 4A, 4B, 5A, and 5B, the models are, from left to right, sChemNET model 402, sChemNET Suv-1 These are shown as Model 404, XGBoost Model 406, Logistic Regression Model 408, Random Forest Model 410, FNN Model 412, Chemical Similarity Model 414, and Random Model 416.

[0127] Figure 4A shows the predictive performance of sChemNET 402 in reading small molecules that alter the activity of 125 miRNA targets in Homo sapiens. sChemNET 402 outperformed the baseline method by 1–9% for the top 100 small molecules read from the test set, 7–21% for the top 300, 5–33% for the top 500, and 8–29% for the top 1000, across different numbers of small molecules read. Suv=1 404 also achieves good prediction performance without using array similarity information in the loss function, but it comes with a decrease in prediction performance of approximately 1.81 to 3.62% across different thresholds.

[0128] In the example, the following methods may be used to score small molecules in the test set: Chemical similarity baseline. Each small molecule in the test set may be scored based on its maximum chemical similarity to an active small molecule in the training set. Chemical similarity was calculated using 2D Tanimoto chemical similarity based on binary fingerprints. Random baseline. Each small molecule in the test set may be assigned a random score sampled from a uniform distribution between 0 and 1. Machine learning baseline. The machine learning baseline may be implemented using sklearn running on the same dataset as sChemNET. These models include logistic regression, random forest (best hyperparameter set, "n_estimators":2, "min_samples_split":10, "min_sample_leaf":3, "max_features":2, "max_depth":50, "bootstrap":True), and XGBoost (best hyperparameter set, "subsample":0.5, "n_estimators":1000, "min_samples_split":5, "min_samples_leaf":5, "max_depth":3, "learning_rate":0.02).

[0129] Figure 4B shows the predictive performance of sChemNET 402 for bioactive small molecules in the test set that were chemically dissimilar (Tanimoto similarity < 0.6) to the small molecules available in the training set. sChemNET 402 significantly outperformed the baseline method by 5–9% (top 100), 10–24% (top 300), 10–40% (top 500), and 12–34% (top 1000) in identifying the top 100 small molecules read from the test set. These findings suggest that sChemNET 402 is useful for discovering novel small molecules targeting miRNAs.

[0130] Figures 5A and 5B show the average predicted rank for small molecules with known biological activity when ranking the probability scores of 4,000 small molecules in the test set. In Figure 5A, the analysis was performed without considering chemically similar compounds between the training set and the test set. In contrast, in Figure 5B, the analysis was performed taking into account the chemical similarity of compounds between the training set and the test set. In Figure 5B, only instances in which the active small molecules in the test set were chemically dissimilar (Tanimoto chemical similarity less than 0.6) to the active small molecules present in the training set were analyzed.

[0131] Similar to Figures 5A and 5B, Figure 6 shows sChemNET 602, sChemNET Suv=1 The plot 604 shows the average predicted rank of small molecules with known bioactivity by other models, and plot 600 shows the average predicted rank by other models. Other existing models include the XGBoost model 606, the Random Forest model 608, the Logistic Regression model 610, and the Feedforward Neural Network ("FNN") model 612. The average rank of small molecules with known bioactivity for each model is shown based on the number of small molecules with known bioactivity in the training set. The average rank for each model is shown as a box plot.

[0132] From left to right on the x-axis, plot 600 shows the average rank of small molecules with known bioactivity when the training set includes small molecules with 4 known bioactivity, 5 known bioactivity, 6 known bioactivity, 7 known bioactivity, 8 known bioactivity, 9 known bioactivity, 10 known bioactivity, 11 known bioactivity, 12 known bioactivity, 13 known bioactivity, 14 known bioactivity, 16 known bioactivity, 17 known bioactivity, 20 known bioactivity, and 34 known bioactivity. For each group, the models shown from left to right are sChemNET model 602 and sChemNET. Suv=1 These are Model 604, XGBoost Model 606, Random Forest Model 608, Logistic Regression Model 610, and FNN Model 612.

[0133] Figures 7A and 7B show the predictive performance when reading small molecules with activity against 50 randomly selected miRNA targets in Homo sapiens from a pool of 4,000 small molecules using a single-drop procedure. Figures 7A and 7B show sChemNET model 702 and sChemNET. Suv=1 This shows the proportion of small molecules with known bioactivity found by Model 704, Chemical Similarity Model 706, and Random Model 708. As mentioned above, Chemical Similarity Model 706 predicts the bioactivity of small molecules using the highest Tanimoto chemical similarity with the group of active small molecules in the training set. Random Model 708 randomly assigns a predicted bioactivity of 0 or 1 to small molecules in the test set.

[0134] Figures 7A and 7B show the proportion of small molecules with known bioactivity found by each model among the top 100 small molecules ranked by each model, the top 300 small molecules ranked by each model, the top 500 small molecules ranked by each model, and the top 1,000 small molecules ranked by each model. These figures show the proportion of small molecules with known activity read out as a function of the number of unlabeled small molecules used for training the model. Similar to the figures above, Figure 7A shows an analysis without controlling for chemistry similarity in training and testing, and Figure 7B shows rankings for only chemically dissimilar small molecules (2D Tanimoto chemistry similarity < 0.6). These figures demonstrate the practical applicability of the approach of this disclosure to predict novel small molecules with bioactivity that are chemically dissimilar to small molecules available for training the model.

[0135] Figures 8A, 8B, and 8C are heatmaps showing the most common mechanisms of action (MoA) and indications for small molecules predicted to target miRNAs in Homo sapiens by a neural network trained using the method and system described herein (sChemNET). In Figure 8A, miRNAs are clustered using hierarchical clustering based on their expression levels in human tissues using a tissue atlas dataset, where darker colors represent higher expression. Figure 8B shows the most common indications for the predicted small molecules. Figure 8C shows the most common mechanisms of action (MoA) for the predicted small molecules. The colors represent:

[0136]

number

[0137]

number

[0138] The p-values ​​adjusted by the Benjamin-Hochberg method can be calculated using Fisher's exact test to maintain overall significance below 0.05. White areas indicate non-significant correlations. These figures demonstrate the effectiveness of sChemNET in computationally predicting small molecules with biological activity against miRNAs, and furthermore, in mapping specific miRNAs to the pharmacological and clinical spaces of drugs, i.e., the spaces of drug mechanism of action (MoA) and indications, respectively. Thus, insights into additional biological targets and possible mechanisms, and subsequent in vivo experiments, may be provided.

[0139] Figures 9A and 9B show examples of trained neural networks for predicting small molecules that bind to target miRNAs in other species, such as mice and rats. Little is known about miRNA targets and the small molecules that bind to them in humans, but even less is known about miRNA targets in other species, such as mice and rats. Figure 9A shows that in mice, only 44 small molecules with bioactivity and 43 target miRNAs have been identified. Similarly, in rats, only 32 small molecules with little bioactivity and 13 target miRNAs have been identified.

[0140] The systems and methods described herein may be used to train models for predicting the bioactivity of small molecules against target miRNAs in other species, such as mice and rats. Similar to the process described above, a first training set 910 of small molecules with known bioactivity may be collected or read. The first training set 910 may include at least one feature vector associated with small molecules with known bioactivity against target miRNAs. In some embodiments, the first training set may include at least one feature vector associated with small molecules with known bioactivity against a first target miRNA in a first species, and at least one feature vector associated with small molecules with known bioactivity against a second target miRNA in a second species.

[0141] For example, as shown in Figure 9B, the first training set 910 may include at least one feature vector associated with a small molecule known to be bioactive against at least one human target miRNA, and at least one feature vector associated with a small molecule known to be bioactive against at least one mouse or rat target miRNA. In some embodiments, the first training set may include at least one feature vector associated with a small molecule with known bioactivity for each target miRNA in a first species and each target miRNA in a second species. Returning to the example, the first training set 910 may include at least one feature vector of a bioactive small molecule for each of the 131 human target miRNAs, and at least one feature vector of a bioactive small molecule for each of the mouse or rat target miRNAs.

[0142] As mentioned above, the first training set can be combined with the second training set to form an extended training set. The second training set may contain multiple feature vectors associated with multiple small molecules that have unknown biological activity against target miRNAs in other species. The extended training set can be used to train a neural network model that predicts the biological activity of small molecules against target miRNAs in other species.

[0143] Figures 10A and 10B show the performance of neural network model 1002 (sChemNET) trained to predict the bioactivity of small molecules against target miRNAs in different species. Figure 10A shows the performance of model 1002 in predicting the bioactivity of small molecules in mice, and Figure 10B shows the performance of model 1002 in predicting the bioactivity of small molecules in rats. In some embodiments, the neural network may perform best without the use of sequence similarity information. Here, the mouse model shows that the neural network can retrieve small molecules with more than 43% bioactivity on average within the top 25% of retrieved predictions.

[0144] The neural network model 1002 can be trained using the aforementioned extended training set in conjunction with Figures 9A and 9B. Similar to Figures 4A and 4B, the performance of the sChemNET model in predicting the bioactivity of small molecules in mice and rats is shown for the modified sChemNET model, sChemNET. Suv=1 1004 (Here, s in the sChemNET model uv(is set to 1) and can be compared with other existing models. Existing models include the XGBoost model 1006, the logistic regression model 1008, the random forest model 1010, the feedforward neural network ("FNN") model 1012, the chemical similarity model 1014, and the random model 1016. The chemical similarity model 1014 predicts the bioactivity of small molecules using the highest Tanimoto chemical similarity with a group of active small molecules in the training set. The random model randomly assigns a bioactivity of 0 or 1 to small molecules. As in the figure above, the models are, from left to right in each figure, sChemNET model 1002, sChemNET Suv=1 The models are grouped in the following order: Model 1004, XGBoost model 1006, logistic regression model 1008, random forest model 1010, FNN model 1012, chemical similarity model 1014, and random model 1016.

[0145] Figure 10A shows the proportion of small molecules known to have biological activity against target miRNAs in mice, which are chemically dissimilar (Tanimoto chemical dissimilarity < 0.6), and which were ranked in the top 2.5%, top 7.5%, top 12.5%, and top 25% of the small molecules with the highest biological activity by each model. The sChemNET model 1002 performed best without using sequence similarity information, and it can be observed that it can read small molecules with activity exceeding 43% on average within the top 25% of read predictions.

[0146] The sChemNET model 1002 significantly outperformed the baseline, recovering over 50% of active small molecules within the top 25% of read predictions. Considering all instances in our evaluation, chemical similarity has good predictive performance and may also exhibit a bias toward chemically similar compounds in the small chemical dataset available for mice.

[0147] Similarly, Figure 10B shows the proportion of small molecules that are known to have biological activity against target miRNAs in brown rats, are chemically dissimilar (Tanimoto chemical dissimilarity < 0.6), and are ranked in the top 100, top 300, top 500, and top 1,000 for the most biologically active small molecules by each model. Similarly, Figure 10B shows the predictive performance of different methods for chemically dissimilar instances of small molecules active against miRNA targets in brown rats. In this dataset, sChemNET model 1002 outperformed the comparison model by 6.18–24.67% for the top 300 (7.5%) of the read predictions, and by 2.74–20.50% for the top 1,000 (12.5%). Logistic regression performs 0.726% better than sChemNET in the top 100 (2.5%), however, this result is not sustained and reverses as more small molecules are read. For 25% of the read small molecules, the neural network can read over 50% of the known small molecules in the training set.

[0148] Expanding the size of the training set The example describes a computer-implemented method for increasing the size of a training set of data for an artificial intelligence engine. In one embodiment, the method may include providing a first training set comprising a first plurality of training samples, each training sample comprising an associated input data label and a corresponding output label. The method may also include providing a second training set comprising a second plurality of partially labeled samples, each partially labeled sample comprising an associated input data label without a corresponding data output label, wherein the second training set is larger than the first training set. The method may then optionally apply a data augmentation technique to the initial training set, the data augmentation technique being used to generate augmented training samples, the data augmentation technique introducing variation into the input data while preserving the corresponding output labels.

[0149] This method can be continued by combining the first and second training sets to form an extended training set. The technique for combining datasets may depend on the size and shape of the datasets. Combining datasets of different sizes may involve merging the data in a way that preserves information from both datasets. One approach is to use a technique called concatenation, in which case the smaller dataset is appended to the larger dataset. This can be done by aligning the columns or features of the two datasets and stacking the rows of the smaller dataset below the rows of the larger dataset. Using this method, the combined dataset will have a larger total sample size, incorporating data from both sources. In general, datasets can be useful if they are compatible in terms of their data type, column names, and feature representations.

[0150] When joining two datasets of different sizes into a single matrix with zero values ​​for unlabeled entries, a common approach is to create a sparse matrix representation. The larger dataset forms the main structure of the matrix, while the smaller dataset is aligned and inserted into the appropriate rows and columns of the larger dataset. Any entries in the larger dataset that do not have corresponding data in the smaller dataset can be padded with zeros to represent unlabeled entries.

[0151] Once the combined dataset is generated, it can be used to train a neural network that generates predictive scores of the bioactivity of candidate drug compounds using an expanded training set. To combine the data in a meaningful way, the contribution of a second set of unlabeled samples in training the neural network can be scaled by a parameter α, where α is greater than zero.

[0152] Improving the performance of neural networks using loss functions The example describes a computer-based method for training a neural network to generate candidate drug compounds that affect one or more biological targets. In some embodiments, the computer-based method may include collecting a set of known drug compounds from a database. The method may continue by creating a first training set, which includes chemical and biological data associated with the set of known drug compounds, wherein the chemical data includes structural information of the drug compounds, and the biological data includes information on the bioactivity of the drug compounds against one or more biological targets, each of which has associated sequence information. The method may continue by creating a second training set, which includes chemical data associated with drug compounds having unknown bioactivity against one or more biological targets. The method may then combine the first and second training sets to form an extended training set.

[0153] This method may optionally include calculating sequence similarity scores between biological targets based on the sequence information of the biological targets. Next, the method can train a neural network using an extended training set to generate predictive scores of the biological activity of candidate drug compounds against the biological targets, wherein the contribution of each biological target to other biological targets is weighted by the sequence similarity score, and the contribution of a second set of unlabeled samples is reduced in the training of the neural network compared to the first set of labeled data. For example, a loss function similar to the loss function in Equation 2 may be used. Finally, the method may output one or more improved neural network models capable of generating candidate drug compounds with predicted activity.

[0154] Integration of sequence similarity information The example describes a computer-based method for training a neural network engine that generates candidate drug compounds affecting one or more biological targets. This method may include: collecting a set of known drug compounds from a database; then creating a training set containing chemical and biological data associated with the known drug compounds, wherein the chemical data includes structural information of the drug compounds, and the biological data includes information on the bioactivity of the drug compounds against one or more biological targets, with each biological target having associated sequence information. The method may calculate sequence similarity scores between biological targets based on the sequence information of the biological targets. The method may then train a neural network using the expanded training set to generate predictive scores for the bioactivity of candidate drug compounds against the biological targets, where the contribution of each biological target to other biological targets is weighted by the sequence similarity scores. Finally, the method may output one or more improved neural network models capable of generating candidate drug compounds with predicted activity.

[0155] Prediction of active small molecules and novel small molecules targeting miRNAs in model organisms through interspecies data integration. The example describes a computer-based method in which the first training set further includes chemical data, and the biological data includes biological data derived from non-human animals. In some embodiments, the computer-based method may include calculating chemical similarity scores from the chemical data and then identifying novel candidate drug compounds by generating candidate drug compounds by inputting candidate drug compounds with low chemical similarity scores into a trained neural network model. The computer-based method may or may not include weighting the contribution of each biological target to other biological targets.

[0156] For example, a deep learning approach can predict miRNA targets targeted by small molecules. For example, these small molecules and / or targets may influence miRNA function. For example, a nonlinear relationship between the chemical characteristics of small molecules and their miRNA targets can be learned. In some examples, learning from small bioactive miRNA-chemical datasets can be done by using information from labeled and unlabeled chemicals. For example, it can be effective to predict molecule-miRNA associations obtained from Homo sapiens and other mammalian model organisms. In some examples, this model provides a novel predictive understanding of the chemical principles by which small molecules have bioactivity against specific miRNA targets. In some examples, this knowledge can be used as a hypothesis generator for experimental design.

[0157] This disclosure provides experimental validation in zebrafish embryos and human cells demonstrating that small molecules predicted by sChemNET can directly act on miRNAs to influence their processing or expression, or to regulate gene expression in miRNA-target networks. Both mechanisms of action are plausible, as both pathways enable desired outputs that complement miRNA activity.

[0158] One example is α-calcidol, which does not directly affect the levels of miR-451 or its cluster partner miR-144, but promotes hematopoietic production. The reason α-calcidol does not affect the levels of red blood cell-specific miR-144 is that Dicer and miR-144 are involved in a negative feedback loop in red blood cells. Dicer processes miR-144, but is also a target of miR-144 (PMID:32191872), which substantially offsets the potential for drug-induced increase in miR-144 output.

[0159] In several cases, experimental validation in zebrafish embryos demonstrated the success of drugs predicted to modulate miR-451 activity or the expression of its targets. Zebrafish embryos were incubated with different drug candidates predicted by sChemNET, in combination with phenylthiourea (PTU), a chemical known to induce anemia via oxidative stress when miR-451 activity is reduced, but not in wild-type embryos. At 48 hours post-fertilization, the embryos exhibit robust blood circulation. At this stage, the accumulation of mature red blood cells in clear embryos can be easily assessed using O-dianisidine, a hemoglobin-specific staining agent. Drugs that reduce miR-451 activity induce anemia, while drugs that promote miR-451 increase red blood cell production (blood circulation).

[0160] Images of the ventral side of 2-day-old embryos stained with O-dianisidine show hemoglobinized cells (brown staining) in wild-type embryos as well as those treated with docetaxel, β-elemen, and α-calcidol. Blood accumulated in the ventral region (Cuvier's duct). A lateral view of another embryo aggregate shows excessive blood accumulation in the tail region during drug treatment.

[0161] definition The various exemplary logic, logic blocks, modules, circuits, and algorithmic steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or a combination of both. Hardware-software compatibility is generally described in terms of functionality and is shown in the various exemplary components, blocks, modules, circuits, and steps described above. Whether such functionality is implemented in hardware or software depends on the specific application and the design constraints imposed on the overall system.

[0162] Hardware and data processing devices used to implement the various exemplary logic, logic blocks, modules, and circuits described in relation to the embodiments disclosed herein may be implemented using or performed with general-purpose single-chip processors or multi-chip processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general-purpose processor may be a microprocessor, or any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors in use with a DSP core, or any other such configuration. In some embodiments, specific steps and methods may be performed by circuits specific to a given function.

[0163] In one or more embodiments, the functions described may be implemented in hardware, digital electronic circuits, computer software, firmware, or any combination thereof, including structures disclosed herein and their structural equivalents. Embodiments of the subject matter described herein may also be implemented as modules of one or more computer programs, i.e., one or more computer program instructions encoded on a computer storage medium for execution by a data processing device or for controlling the operation of a data processing device.

[0164] When implemented in software, the functionality may be stored as one or more instructions or codes on a computer-readable medium, or transmitted via a computer-readable medium. Steps of the methods or algorithms disclosed herein may be implemented in a processor-executable software module that may reside on a tangible, non-temporary computer-readable medium. Computer-readable mediums include both computer storage media and communication media, which include any media that can enable the transfer of computer programs from one location to another. Storage media may be any available media that can be accessed by a computer.

[0165] Software modules may reside in random access memory (RAM), flash memory, read-only memory (ROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. The storage medium is coupled to the processor so that the processor can read information from and write information to the storage medium. Alternatively, the storage medium may be integrated with the processor. As used herein, disks and discs include compact discs (CDs), laser discs, optical discs, digital multipurpose discs (DVDs), floppy disks, and Blu-ray discs, where disks typically reproduce data magnetically, while discs reproduce data optically using a laser. The above combinations also fall within the scope of computer-readable media. Processors and storage mediums may reside in ASICs. ASICs may reside in user terminals. Alternatively, processors and storage mediums may reside as discrete components in user terminals.

[0166] In addition, methods or algorithmic operations may exist as one or any combination of codes and instructions, or sets thereof, on machine-readable and computer-readable media, and these may be incorporated into computer program products.

[0167] Various modifications to the embodiments described herein may be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other embodiments without departing from the spirit or scope of this disclosure. Accordingly, the claims are not intended to be limited to the embodiments shown herein, but should be granted the broadest scope according to this disclosure, the principles disclosed herein, and the novel features. In addition, those skilled in the art will readily understand that the terms “top” and “bottom” may be used for the convenience of describing the figures and indicate relative positions corresponding to the orientation of the figures on a properly placed page, and may not necessarily reflect the proper orientation of the components when implemented.

[0168] While certain embodiments of this disclosure are described, these embodiments are presented for illustrative purposes only and are not intended to limit the scope of this disclosure. In fact, the novel methods and systems described herein may be embodied in various other forms. Furthermore, various omissions, substitutions, and modifications in the systems and methods described herein may be made without departing from the spirit of this disclosure. The appended claims and their equivalents are intended to cover any such forms or modifications that may fall within the scope and spirit of this disclosure. Accordingly, the scope of this disclosure is defined solely by reference to the appended claims.

[0169] Features, materials, properties, or sets described in conjunction with a particular aspect, embodiment, or example should be understood to be applicable to any other aspect, embodiment, or example described in this section or elsewhere in this specification, unless incompatible with any other aspect, embodiment, or example. All features disclosed herein (including any appended claims, abstract, and drawings) and / or all steps of any method or process disclosed herein may be combined in any combination, except for any combination in which at least some of such features and / or steps are mutually exclusive. The protection is not limited to the details of any of the embodiments described herein. The protection extends to any novel features or any novel combination of features disclosed herein (including any appended claims, abstract, and drawings), or any novel steps or any novel combination of steps of any method or process disclosed herein.

[0170] Furthermore, certain features described in this disclosure in the context of separate embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any preferred partial combination in multiple embodiments. Even if features can be described above as acting in a particular combination, one or more features may be removed from the claimed combination, and this combination may be claimed as a partial combination or a variation of a partial combination.

[0171] The features and attributes of the specific embodiments disclosed above may be combined in different ways to form additional embodiments, all of which are included within the scope of this disclosure. Furthermore, the separation of various system components in the embodiments described above should not be understood as requiring such separation in all embodiments, and the described components and systems can generally be integrated together into a single product or packaged into multiple products.

[0172] Furthermore, even if operations can be depicted in a specific order in the drawings or described in the specification, such operations do not need to be performed in the specific order or sequential order shown to achieve the desired result, nor do all operations need to be performed. Other operations not depicted or described can be incorporated into the exemplary methods and processes. For example, one or more additional operations can be performed before, after, simultaneously with, or between any of the described operations. Moreover, operations can be rearranged or reordered in other embodiments. Those skilled in the art will understand that in some embodiments, the steps actually taken in the illustrated and / or disclosed processes may differ from those shown in the drawings. Depending on the embodiment, certain steps described above may be omitted, and others may be added. Furthermore, the features and attributes of the specific embodiments disclosed above can be combined in different ways to form additional embodiments, all of which are included within the scope of this disclosure.

[0173] For the purposes of this disclosure, certain aspects, advantages, and novel features are described herein. Not all of these advantages can necessarily be achieved according to any particular embodiment. Therefore, for example, a person skilled in the art will recognize that this disclosure can be embodied or performed in a manner that achieves one advantage or set of advantages taught herein without necessarily achieving other advantages that can be taught or suggested herein.

[0174] Conditional expressions such as "can," "could," "might," or "may," unless otherwise explicitly stated or otherwise understood differently in the context in which they are used, are generally intended to convey that a particular embodiment includes certain features, elements, and / or steps, while other embodiments do not. Therefore, such conditional expressions are not generally intended to suggest that features, elements, and / or steps are essential in any sense in one or more embodiments, or that one or more embodiments necessarily include logic for determining, with or without user input or instruction, whether these features, elements, and / or steps are included in any particular embodiment or should be performed in any particular embodiment.

[0175] Conjunctions such as "at least one of X, Y, and Z" are generally understood, in the context in which they are used, to indicate that an item, term, etc., may be one of X, Y, or Z, unless otherwise explicitly stated. Therefore, such conjunctions do not generally suggest that a particular embodiment requires the presence of at least one of X, at least one of Y, and at least one of Z. Accordingly, as used herein, the phrase "at least one of X, Y, and Z" is intended to encompass X, Y, Z, X and Y, X and Z, Y and Z, and X, Y and Z.

[0176] The headings provided herein, where present, are for convenience only and do not necessarily affect the scope or meaning of the devices and methods disclosed herein.

[0177] As used herein, expressions indicating degree, such as “approximately,” “about,” “generally,” and “substantially,” represent values, quantities, or characteristics that are close to the stated values, quantities, or characteristics that perform the desired function or achieve the desired result. For example, “approximately,” “about,” “generally,” and “substantially” may refer to quantities that are within the range of less than 10%, less than 5%, less than 1%, less than 0.1%, and less than 0.01% of the stated quantity.

[0178] The scope of this disclosure is not intended to be limited by the specific disclosure of embodiments in this section or elsewhere in this specification, but may be defined by the claims presented or to be presented in this section or elsewhere in this specification. The wording of the claims should be interpreted broadly based on the wording adopted in the claims, and not limited to the examples described herein or in the examination process of the application, and these examples should be interpreted non-exclusively.

[0179] The term "training data" is used herein to mean the dataset used to train a machine learning model. This consists of input features and corresponding target labels or outcomes. The model learns patterns and relationships in the training data to make predictions on new, unseen data.

[0180] The term “feature engineering” is used herein to mean the process of selecting, transforming, and extracting relevant features from raw data to improve the performance of machine learning models. This includes techniques such as normalization, scaling, coding categorical variables, dimensionality reduction, and creating new derived features.

[0181] The term "model architecture" is used herein to mean the structure and arrangement of layers, nodes, and connections in a machine learning model. This determines how the model processes and learns from the input data, thereby enabling it to make predictions or decisions.

[0182] In this specification, the term "loss function" is used to mean a function that quantifies the difference between the predicted output of a machine learning model and the actual target label. This serves as the basis for model optimization during the training process.

[0183] In this specification, the term "hyperparameters" refers to parameters that are set before the training process, rather than being learned directly from the data. These control the behavior of a machine learning model, such as the learning rate, regularization strength, and architectural configuration.

[0184] The term “exemplary validation dataset” is used herein to mean a subset of training data used to evaluate the performance of a machine learning model during training. This provides an estimate of how well the model will generalize to new, unseen data.

[0185] The term “exemplary overfitting” is used herein to refer to a phenomenon in machine learning where a model performs exceptionally well on the training data but fails to generalize to new, unseen data. This can occur when the model becomes excessively complex or when it incorporates noise in the training data.

[0186] The term "feature selection" is used herein to mean the process of selecting a subset of relevant features from the available input data. This aims to reduce dimensionality, improve model interpretability, and enhance model performance by focusing on the most useful features.

[0187] The term "cross-validation" is used herein to mean a technique used to evaluate the performance and generalization ability of a machine learning model. This involves dividing a dataset into multiple subsets, iteratively training and evaluating the model on different subsets, and obtaining a robust performance estimate.

[0188] Embodiment The following are non-limiting examples of certain embodiments of systems and methods for characterizing coronary artery plaque. Other examples and embodiments may include one or more other or different features discussed herein.

[0189] In some embodiments, a computer-based method for training a neural network to generate candidate drug compounds may include creating a first training set that includes structural information of the drug compounds and information on the bioactivity of the drug compounds against biological targets.

[0190] In the example, a first training set is created, which includes structural information and bioactivity information of the drug compound against a biological target; and a second training set is created, which includes structural information of the drug compound having unknown bioactivity against the biological target.

[0191] In the example, the first training set and the second training set are combined to form an extended training set. In the example, the neural network is trained using the extended training set.

[0192] In an example, output one or more improved neural network models capable of generating predictive scores for the biological activity of candidate drug compounds against biological targets.

[0193] In the example, the method performed by the computer may include training a neural network using an expanded training set, which includes scaling the relative contribution of the second training set to the first training set by a parameter greater than 0 and less than 1.

[0194] In this example, the computer-based method may include generating candidate drug compounds by inputting chemical data associated with those candidate drug compounds into a trained neural network.

[0195] In this example, the computer-based approach may include using a trained neural network model to evaluate the predicted activity of a candidate drug compound.

[0196] In the example, the computer-based method may include predicting small, bioactive molecules that are chemically dissimilar to those available for training neural network models.

[0197] In the example, in a computer-based method, any chemically dissimilar small molecule may have a Tanimoto similarity of less than 0.6 compared to a small molecule available for training a neural network model.

[0198] In this example, predictive performance can be greater than 40% when evaluated using the one-miss cross-validation (LOOCV) method and reported as the recall rate of active small molecules out of 1000 small molecules read from the test set.

[0199] In the example, the predictive performance is greater than at least one of 50%, 55%, 60%, and 65%.

[0200] In the example, the first training set is smaller than the second training set.

[0201] In the example, the first training set is one-tenth the size of the second training set.

[0202] In the example, the first training set is between 1 / 100 and 1 / 5000 of the second training set.

[0203] In the example, the first training set contains fewer than 10 known bioactive drug compounds.

[0204] In the example, the first training set includes drug compounds with fewer than 5 known biological activities.

[0205] In this example, a neural network is applied to evaluate candidate drug compounds that modulate at least one of the following: miRNA, mRNA, and protein targets.

[0206] In the example, the neural network is applied to at least one of the following targets: FLT3, ALK, IGF1R, and EGFR.

[0207] In the example, the neural network is applied to at least one of the following targets: miR-10300, miR-155, miR-10b, and miR-181.

[0208] In the examples, each disclosed example, when executed by one or more processors, can be implemented by a computer-readable storage medium that stores instructions causing one or more processors to perform the disclosed example.

[0209] In the examples, each disclosed example may be carried out by a system for training a neural network to generate candidate drug compounds, the system comprising a memory device and modules stored on the memory device and configured to cause one or more processors to execute the examples.

[0210] In the example, the disclosed method may expand the size of the training set. For example, a computer-implemented method for increasing the size of training set data for an artificial intelligence engine may include providing a first training set comprising a first set of training samples, each training sample comprising an associated input data label and a corresponding output label.

[0211] In some examples, the method performed by the computer may include providing a second training set comprising a second set of partially labeled samples, each partially labeled sample having associated input data labels without corresponding data output labels. In some examples, the second training set is larger than the first training set.

[0212] In the example, the method performed by the computer may include combining the first training set and the second training set to form an extended training set.

[0213] In the examples, the computer-based approach may involve training a neural network that generates predictive scores for the bioactivity of candidate drug compounds using an expanded training set. In some examples, the contribution of a second set of unlabeled samples in training the neural network is scaled by a parameter α, where α is greater than 0. In some examples, the input data labels are the chemical fingerprints of the chemical compounds, and the data output labels are the bioactivity of the chemical compounds against a biological target.

[0214] In some cases, the chemical fingerprint is determined to be at least one of the MACCS fingerprint, Daylight fingerprint, and RDKit fingerprint.

[0215] In some examples, the group of chemical compounds includes compounds found in at least one of the SM2miR database and the drug repositioning hub.

[0216] In some examples, the second training set includes partially labeled samples that represent a broader range of chemical diversity than the first training set, which contains multiple training samples.

[0217] In some examples, chemical diversity corresponds to a vector space of chemical feature vectors based on chemical fingerprints.

[0218] In some examples, the second training set includes partially labeled samples that map to lower probability scores than the first training set, which contains multiple training samples.

[0219] In some examples, the second training set includes partially labeled samples, the mean probability score of which is less than the probability score of the first training set, which includes multiple first training samples.

[0220] In some examples, the second training set includes partially labeled samples having associated input data labels and corresponding output labels containing a set of non-zero latent features that encode the biological interaction between the drug and its bioactivity.

[0221] In some cases, non-zero latent features are determined to exist by improved predictive performance for neural networks trained without an extended training set.

[0222] In the example, the method performed by the computer may include training a neural network using an extended training set by employing a non-negative matrix decomposition model.

[0223] In the example, the method performed by the computer may include training a neural network with an extended training set by employing zero-driven regularization in a non-negative matrix decomposition model.

[0224] In the disclosed example, when the method is executed by one or more processors, it may be stored on a computer-readable storage medium that stores instructions causing one or more processors to perform the disclosed method.

[0225] In the examples, any of the disclosed examples may be implemented by a system for scaling the size of the training set, which comprises a memory device and modules stored on the memory device and configured to cause one or more processors to perform any of the steps of a preceding example.

[0226] Some examples may include improving the performance of a neural network by using a loss function that includes disclosure features, such as any combination of chemical similarity, unlabeled data, augmented datasets, and sequence similarity features.

[0227] Computer-based methods for training a neural network engine to generate candidate drug compounds that affect one or more biological targets may include collecting a set of known drug compounds from a database.

[0228] In the example, the method performed by the computer may include creating a first training set containing chemical and biological data associated with a known group of drug compounds.

[0229] In some examples, chemical data includes structural information of the drug compound, and biological data includes information on the biological activity of the drug compound against one or more biological targets.

[0230] In some cases, each biological target has associated sequence information,

[0231] In the example, the method performed by computer may include creating a second training set containing chemical data associated with drug compounds having unknown biological activity against one or more biological targets.

[0232] In the example, the method performed by the computer may include combining a first training set and a second training set to form an extended training set.

[0233] In the example, the method performed by the computer may include calculating a sequence similarity score between biological targets based on the sequence information of the biological targets.

[0234] In the example, the computer-based method may include training a neural network to generate predictive scores of the biological activity of candidate drug compounds against biological targets using an expanded training set.

[0235] In some examples, the contribution of each biological target to other biological targets is weighted by a sequence similarity score.

[0236] In some examples, when training a neural network, the contribution of a second set of unlabeled samples is reduced compared to the first set of labeled data, and

[0237] In some examples, the computer-based methods may include outputting one or more improved neural network models capable of generating candidate drug compounds with predicted activity.

[0238] In this example, the computer-based method may include generating candidate drug compounds by inputting chemical data associated with the drug compounds into a trained neural network model, and then using the trained neural network model to evaluate the predicted activity of the candidate drug compounds.

[0239] In some cases, biological data for all biological targets applies to each individual biological target.

[0240] In some cases, unlabeled drug compounds are assigned near-zero initial prediction scores for each biological target during the training of a neural network.

[0241] In the example, the method performed by the computer may include training a neural network using an expanded training set, wherein training the neural network involves using a Bayesian optimization approach for hyperparameters relating to the contributions of a second set of unlabeled samples, based on one-out cross-validation of drug compounds known to target biological targets not included in any of the training sets.

[0242] In the disclosed example, when the method is executed by one or more processors, it may be stored on a computer-readable storage medium that stores instructions causing one or more processors to perform the disclosed method.

[0243] In the example, any of the disclosed examples may be implemented by a system comprising a memory device and modules stored on the memory device and configured to cause one or more processors to perform any of the steps of the preceding example.

[0244] For example, a computer-based method for training a neural network engine to generate candidate drug compounds that affect one or more biological targets, the computer-based method may include collecting a set of known drug compounds from a database.

[0245] In the example, the computer-based method may include creating a training set containing chemical and biological data associated with a known group of drug compounds.

[0246] In the example, the method performed by a computer may include creating a training set containing chemical and biological data associated with a known group of drug compounds, wherein the chemical data contains structural information of the drug compounds and the biological data contains information on the biological activity of the drug compounds against one or more biological targets.

[0247] In some cases, each biological target has associated sequence information.

[0248] In the example, the method performed by the computer may include calculating a sequence similarity score between biological targets based on the sequence information of the biological targets.

[0249] In this example, the computer-based method may include training a neural network to generate predictive scores of the biological activity of candidate drug compounds against biological targets using an expanded training set.

[0250] In some cases, the contribution of each biological target to other biological targets is weighted by a sequence similarity score.

[0251] In the example, the method performed by the computer may include outputting one or more improved neural network models capable of generating candidate drug compounds with predicted activity.

[0252] In the disclosed example, when the method is executed by one or more processors, it may be stored on a computer-readable storage medium that stores instructions causing one or more processors to perform the disclosed method.

[0253] In the examples, any of the disclosed examples may be implemented by a system for scaling the size of the training set, which comprises a memory device and modules stored on the memory device and configured to cause one or more processors to perform any of the steps of a preceding example. Further examples of embodiments are listed in the following set of claims #1 to #4.

Claims

1. A computer-based method for training a neural network to generate candidate drug compounds, wherein the method is: To create a first training set that includes structural information of a drug compound and information on the biological activity of the drug compound against a biological target, To create a second training set containing structural information of drug compounds having unknown biological activity against the aforementioned biological target, The first training set and the second training set are combined to form an extended training set, The neural network is trained using the aforementioned expanded training set. A method comprising outputting one or more improved neural network models capable of generating predictive scores for the biological activity of candidate drug compounds against the biological target.

2. A computer-based method according to claim 1, wherein training the neural network using the expanded training set includes scaling the relative contribution of the second training set to the first training set by a parameter greater than 0 and less than 1.

3. The computer-based method according to claim 1, further comprising generating candidate drug compounds by inputting chemical data associated with the candidate drug compounds into the trained neural network.

4. The computer-based method according to claim 1, further comprising evaluating the predicted activity of the candidate drug compound using the trained neural network model.

5. The computer-based method according to claim 1, further comprising predicting a small bioactive molecule that is chemically dissimilar to those available for training the aforementioned neural network model.

6. The computer-based method according to claim 5, wherein the chemically dissimilar small molecule has a Tanimoto similarity of less than 0.6 compared to the small molecule available for training the neural network model.

7. A computer-based method according to claim 2, wherein the predictive performance exceeds 40%, is evaluated by a one-out cross-validation (LOOCV) method, and is reported as the recall rate of active small molecules among 1000 small molecules read from the test set.

8. The computer-based method according to claim 7, wherein the predictive performance exceeds at least one of 50%, 55%, 60%, and 65%.

9. The computer-based method according to claim 1, wherein the first training set is smaller than the second training set.

10. The computer-based method according to claim 9, wherein the first training set is one-tenth the size of the second training set.

11. The computer-based method according to claim 9, wherein the first training set is 1 / 100 to 1 / 5000 of the second training set.

12. The computer-based method according to claim 1, wherein the first training set comprises fewer than 10 known bioactive drug compounds.

13. The computer-based method according to claim 1, wherein the first training set comprises drug compounds having fewer than 5 known biological activities.

14. The computer-based method according to claim 1, wherein the neural network is applied to evaluate candidate drug compounds for modulating at least one of miRNA, mRNA, and protein targets.

15. The computer-based method according to claim 11, wherein the neural network is applied to at least one of the following targets: FLT3, ALK, IGF1R, and EGFR.

16. The computer-based method according to claim 11, wherein the neural network is applied to at least one of the following targets: miR-10300, miR-155, miR-10b, and miR-181.

17. A computer-readable storage medium that, when executed by one or more processors, stores instructions causing the one or more processors to perform the method according to any one of claims 1 to 16.

18. A system for training a neural network to generate candidate drug compounds, wherein the system is Memory device and A system comprising: a module stored on the memory device and configured to cause one or more processors to perform the method described in any one of claims 1 to 16.

19. A computer-based method for training a neural network engine to generate candidate drug compounds that affect one or more biological targets, wherein the method is: Collecting known drug compounds from a database, Creating a first training set that includes chemical and biological data associated with the known group of drug compounds, The chemical data includes structural information of the drug compound, and the biological data includes information on the biological activity of the drug compound against one or more biological targets. To create a first training set containing sequence information associated with each biological target, To create a second training set containing chemical data associated with drug compounds that have unknown biological activity against one or more biological targets, The first training set and the second training set are combined to form an extended training set, Based on the sequence information of the biological targets, a sequence similarity score is calculated between the said biological targets, The method involves training a neural network that generates predictive scores for the biological activity of candidate drug compounds against biological targets using the aforementioned expanded training set, The contribution of each biological target to other biological targets is weighted by the sequence similarity score. The neural network is trained such that the contribution of the second set of unlabeled samples is reduced compared to the first set of labeled data. A method comprising outputting one or more improved neural network models capable of generating candidate drug compounds having predicted activity.

20. A computer-based method according to claim 19, further comprising generating candidate drug compounds by inputting chemical data associated with drug compounds into the trained neural network model, and evaluating the predicted activity of the candidate drug compounds using the trained neural network model.

21. The computer-based method according to claim 19, wherein biological data for all biological targets are applied to each biological target.

22. The computer-based method according to claim 19, wherein during the training of the neural network, unlabeled drug compounds are assigned an initial prediction score close to zero for each biological target.

23. The computer-based method according to claim 19, wherein training the neural network using the expanded training set comprises using a Bayesian optimization approach to the hyperparameters relating to the contributions of the second plurality of unlabeled samples, based on one-out cross-validation of drug compounds known to target biological targets not included in any of the training sets.

24. A computer-readable storage medium that, when executed by one or more processors, stores instructions causing the one or more processors to perform the method described in any one of claims 19 to 23.

25. A system for training a neural network to generate candidate drug compounds, wherein the system is Memory device and A system comprising: a module stored on the memory device and configured to cause one or more processors to perform the method described in any one of claims 19 to 23.

Images