Extracellular vesicle release inhibition identification using drug detection machine learning modeling

Abstract

Systems and methods include a machine learning model comprising an input for receiving a set of training data; a processor that applies a set of fingerprints to the training data; and a set of machine learning algorithms that predict bioactivity of a set of drugs using the training data applied to a stored listing of drugs.

Description

Attorney Docket No.: 03284.0363WO01 EXTRACELLULAR VESICLE RELEASE INHIBITION IDENTIFICATION USING DRUG DETECTION MACHINE LEARNING MODELING RELATED APPLICATIONS

[0001] This application claims priority to U.S. Provisional Application Serial No. 63 / 587,215 filed October 2, 2023, entitled “Extracellular Vesicle Release Inhibition Drug Identification for Preventing Cancer Progression using a novel Machine Learning Model,” the entirety of which is incorporated by reference herein. STATEMENT REGARDING FEDERALLY FUNDED RESEARCH FOR DEVELOPMENT.

[0002] This invention was made with government support under grant numbers 5SC3GM113782-07 and 41925-00-06 awarded by the National Institute of Health. The government has certain rights in the invention. FIELD

[0003] The present concepts relate generally to a machine learning model and software platform that identifies new molecules, compounds, and drugs and a software platform to identify drugs for neuron invasion, photoreceptor migration, and synapse formation. BACKGROUND

[0004] Extracellular vesicles (EVs) are lipid-enclosed particles that are generated by most types of cells and have been identified as an important mechanism for cell communication aside from cell direct contact. EVs contain a variety of functional macromolecules, proteins, lipids, RNA, and these materials can be transferred from the host cells to the target cells. The molecular pathways of EV biogenesis have been studied during the past decades and the process involves a number of genes and proteins. EVsplay a role in tumor progression and metastasis. For example, some cancers are known to undergo metastasis using vesicles released by donor cells and taken up by recipient cells in a cell-to-cell communication, referred to as EV release and uptake. Blocking the EV release from cancer cells with targeted drugs may slow disease pathogenesis and prevent a recurrence. SUMMARY

[0005] In one aspect, a method for drug identification, comprises receiving, by a machine learning model stored at and executed by a special-purpose computer, a set of input data; arranging the input data into active compounds, inactive compounds, and decoy compounds; applying a plurality of chemical fingerprint processes to the active compounds, inactive compounds, and decoy compounds to generate a training dataset; receiving the training dataset by a machine learning model, comprising a plurality of different machine learning algorithms; performing a validation process on an output of each of the different machine learning algorithms; and predicting bioactivity levels of a plurality of drugs by applying a listing of the drugs to an output of the validation process.

[0006] In another aspect, a machine learning model, comprises an input for receiving a set of training data; a processor that applies a set of fingerprints to the training data; and a set of machine learning algorithms that predict bioactivity of a set of drugs using the training data applied to a stored listing of drugs.

[0007] In another aspect, a method for detecting drugs for treating a medical condition comprises administering a drug to an eye of a human being, wherein the drug is selected from a group consisting of midostaurin, amiloride, ponatinib, selumetinib and belumosudi.

[0008] In another aspect, a machine learning model, comprises an input for receiving a set of training data; a processor that applies a set of fingerprints to the training data; and a set of machine learning algorithms that predict bioactivity of a set of drugs using the training data applied to a stored listing of drugs.

[0009] In another aspect, a drug library is generated according to a process, comprising: arranging input data into active compounds, inactive compounds, and decoycompounds; applying a plurality of chemical fingerprint processes to the active compounds, inactive compounds, and decoy compounds to generate a training dataset; receiving the training dataset by a machine learning model, comprising a plurality of different machine learning algorithms; performing a validation process on an output of each of the different machine learning algorithms; and predicting bioactivity levels of a plurality of drugs by applying a listing of the drugs to an output of the validation process. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The above and further advantages of this invention may be better understood by referring to the following description in conjunction with the accompanying drawings, in which like numerals indicate like structural elements and features in various figures. The drawings are not necessarily to scale, emphasis instead being placed upon illustrating the principles of the invention. In the drawings:

[0011] FIG.1 is a flow diagram of a method for drug identification, in accordance with some embodiments.

[0012] FIG.2A, FIG.2B and FIG.2C are flow diagrams of a method that incorporates a machine learning model to recognize the chemical structure of novel compounds and their potential to inhibit EV genesis, in accordance with some embodiments.

[0013] FIG.3 is a diagram illustrating a training dataset for a machine learning model targeting a Protein kinase C (3.&Į) isoform for reducing an EV release in cancer cells, in accordance with some embodiments.

[0014] FIG.4 is a table that includes examples of EV release inhibitors and corresponding target proteins, in accordance with some embodiments.

[0015] FIG.5 is a table that includes example candidate compounds that inhibit PKCĮ, in accordance with some embodiments.

[0016] FIG.6 is a table including a set of decoys incorporated into a training dataset, in accordance with some embodiments.

[0017] FIG.7A and FIG.7B show tables that include example including a setof molecular access system (MACCS) fingerprints of training compounds that inhibit 3.&Į, in accordance with some embodiments.

[0018] FIG.8 is a diagram illustrating a workflow of a Quantitative Structure- Activity Relationship machine learning (QSAR-ML) model for predicting FDA drug bioactivity, in accordance with some embodiments.

[0019] FIG.9A is a table that includes a list of FDA-approved drugs identified by a machine learning model, in accordance with some embodiments.

[0020] FIG.9B are illustrations of compounds of the inhibitors listed in the table of FIG.9A.

[0021] FIGs.10A-10F are graphs of experimental results of retinoblastoma cell related EV size and concentration distributions.

[0022] FIG.11 is a block diagram of a computer platform in which embodiments of the present inventive concept can be practiced.

[0023] FIGs.12A-12C are illustrations of isolated photoreceptor precursor cells (PPCs) transplanted into a retina ex vivo using a drug library, in accordance with some embodiments. DETAILED DESCRIPTION

[0024] In brief overview, the systems and methods in accordance with embodiments of the present inventive concept include a novel machine learning model to identify new molecules and FDA-approved drugs to target EV release from a cell and uptake where EVs enter recipient cells. In addition, the machine learning model can be used to identify novel drugs with functional properties similar to existing EV inhibition drugs. The novel drugs are validated in vitro for inhibition of EV genesis in cells. With the help of the novel machine learning model and related computational tools, embodiments of the present inventive concept identify novel drugs to modify extracellular vesicle release and uptake. Other embodiments include a drug library to drive photoreceptor migration and synapse formation, but not limited thereto. In other embodiments, a first library is for collection is based on a gene expression level for the collection of drugs of interest developed for new uses, and a second library is based on chemical structural data generated in response to a machine learning model.

[0025] In some embodiments, the machine learning models include Quantitative Structure-Activity Relationship (QSAR)- machine learning models, which represent a solid approach in the realm of drug discovery and repurposing, especially in medicine of EV release inhibition. The QSAR models are augmented with machine learning algorithms to establish and discern the complex and intricate relationships between the structural descriptors and bioactivity levels of chemical compounds. Employing a multi- algorithmic strategy, the models are trained on a dataset of experimental compounds and decoys. Thus, unlike traditional QSAR models, the machine learning models in accordance with embodiments of the present inventive concept incorporate decoys that structurally resemble active compounds but lack biological activity, thereby enhancing the model’s predictive accuracy. Multiple chemical fingerprint sequences serve as machine-readable codes that encapsulate the structural information from these chemical compounds, and act as the input variables, while inhibitory concentration values against specific target proteins or bioactivity metrics, such as IC values, function as response variables within the models. The training dataset is assembled as described in FIG.1. Here, outlining the workflow of the QSAR-ML model, a variety of machine learning algorithms are employed to train on disparate types of chemical fingerprints. Once validated, these models hold the capability to predict the bioactivity of FDA-approved drugs, offering a pathway to swiftly and cost-effectively identify new therapeutic applications where existing EV targeting drugs can be used with a novel machine learning model to generate novel drugs. The integrated workflow epitomizes a transformative approach to drug identification, offering a robust, accurate, and streamlined methodology for discovering viable drug candidates for inhibition of EV release, which can prevent various diseases, such as ocular cancer metastasis.

[0026] The embodiments of the inventive concept include methodologies, both data-driven and experimental, for identifying viable candidates for drug repurposing. Drug repurposing technologies hold significant promise, particularly with respect to FDA-approved drugs designed to treat a variety of illnesses. The advantages of repurposing are manifold. It offers a pathway that may bypass the high failure rates, exorbitant costs, and protracted timelines typically associated with novel drug discovery. Utilizing 'de-risked' compounds in this way can result in reduced development costs andexpedited time-to-market. Also, the novel or re-purposed drugs to improve migration or synaptogenesis may be of interest to pharmaceutical companies to improve neural cell replacement across neurodegenerative or neurotraumatic conditions to improve transplanted neural stem cell survival and integration into central nervous system (CNS) tissue.

[0027] FIG.1 is a flow diagram of a method 100 for drug identification, in accordance with some embodiments.

[0028] At block 102, an input dataset of target proteins related to EV is collected. In some embodiments, as shown in FIG.4, a dataset of EV release inhibitors and the corresponding target proteins can be obtained from publications, web-based tools, datasets, and repositories, but not limited thereto. The input can be derived from a database of genes and drugs that have been used by researchers to inhibit EV release. In some embodiments, the dataset can be bifurcated into active and inactive classes to facilitate effective model training. The active compounds are experimentally validated inhibitors of the target protein, sourced from a ChEMBL database, or related curated database of bioactive molecules with drug-like properties. On the other hand, the inactive compounds in the dataset are a combination of two types. The first type includes compounds that have been experimentally verified as insensitive to inhibiting the target protein, with IC50 values above a specific threshold. The second type encompasses decoy compounds generated using the Directory of Useful Decoys, enhanced (DUD-E) to serve as negative controls in the study. This comprehensive and balanced training and testing dataset is designed to fortify the predictive accuracy of a QSAR-machine learning model, which is well-known for providing an approach in the realm of drug discovery and repurposing, especially with respect to EV release inhibition.

[0029] At block 104, a machine learning model is built from the target proteins related to EV biogenesis and release and their corresponding chemical inhibitors collected from multiple databases, repositories, and publications of block 102. In some embodiments, a collection of chemical compounds are used to build one or more QSAR models to identify the relationship between the chemical fingerprint of compounds and biological activity via ML methods. However, the machine learning model built in block 104 is different from traditional QSAR models, in that the machine learning model hereuniquely incorporates decoys that structurally resemble active compounds but lack biological activity, significantly enhancing the model’s predictive accuracy. In addition, as described in FIGs.2A-2C, each input dataset is interpreted by multiple chemical fingerprints and additional level of training is accomplished by a plurality of different machine learning algorithms. In some embodiments, the models can further apply to the FDA-approached drug dataset and bioactive molecules to identify the FDA drugs with repurposing and novel chemicals that can inhibit the EV release. The novel druggable compounds will be validated by preform by the knockdown of the target gene and quantification of EV release inhibition and uptake.

[0030] At block 106, the machine learning model is trained and validated.

[0031] At block 108, once validated, the model is employed to forecast the bioactivity against designated targeted proteins.

[0032] At block 109, a database or related storage device can store data related to drugs of interest or the like from the model output of block 108.

[0033] At block 110, the predicted biological activities of the drugs of interest are collected for further analysis. For example, the half-maximal inhibitory concentration (IC50) is a used and informative measure of a drug's efficacy, and indicates how much drug is needed to inhibit a biological process by half, thus providing a measure of potency of an antagonist drug in pharmacological research. In this example, the IC50 of FDA-approved drugs can be predicted. In doing so, the drugs are prioritized then prioritize these drugs based on their predicted bioactivity levels, extracting the top candidates from each model's output. From this refined list, drugs are identified that appear consistently across all top-ranking categories, earmarking these as potential EV inhibitors. These selected compounds will undergo further biological validation to confirm their effectiveness in inhibiting EV release.

[0034] FIGs.2A-2C show a flow diagram of a method 200 that incorporates a machine learning model to recognize the chemical structure of novel compounds and their potential to inhibit EV genesis, in accordance with some embodiments.

[0035] Blocks 202-210 are part of a target identification process. An EV inhibitor 202 and corresponding target protein 204 are matched (208) against a drug discovery repository that houses medicinal chemistry information derived from researchand development processes, for example, a ChEMBL database. The EV inhibitor 202 and target protein 204 can be an EV inhibitor obtained from a dataset of collected target proteins related to EV biogenesis and release and their corresponding chemical inhibitors from multiple databases, repositories, and publications. If at decision diamond 208 there is no match record in the ChEMBL database, the method 200 proceeds to block 206 where an alternate target protein is identified. If there is a match, then the method proceeds to block 210, where experimental compounds inhibit the target protein 204. In blocks 212-216, a dataset of the experimental compounds and decoys are provided on which the subsequent models are trained.

[0036] Shown in block 212 are active compounds that are experimentally validated inhibitors of the target protein, for example, sourced from the well-established ChEMBL database. In some embodiments, bisindolylmaleimide I (BIM I, GF-109203) is used as a highly selective inhibitor of protein kinase C (PKC), specifically targeting the 3.&Į^LVRIRUP^^7KLV^LQKLELWRU^KDV^EHHQ^SURYHQ^HIIHFWLYH^LQ^UHGXFLQJ^(9^UHOHDVH^LQ^FDQFHU^ cells. For the QSAR-ML machine learning model, as shown in FIG.3, ^^^^3.&Į- inhibiting compounds are initially identified from the ChEMBL database. After eliminating duplicates, 621 unique chemical entities are left (see also the table in FIG.5). 7KHVH^ZHUH^WKHQ^FDWHJRUL]HG^EDVHG^RQ^WKHLU^,&^^^YDOXHV^DJDLQVW^3.&Į compounds with IC50 values below 1,000 nM (IC50<1000nM) in block 212 were classified as ‘active’, while those with values in block 216 above 10,000 nM (IC50>10,000nM) were deemed ‘inactive’. This resulted in a dataset comprising 360 active and 162 inactive compounds. Values between 1000nM and 10,000 nM may be discarded as being in an intermediate class to assist with the training of the model, which processes compounds outside this range.

[0037] To further balance the training data and enhance the model’s predictive accuracy, decoys can be incorporated into the training dataset, shown in block 214. The decoys are molecules that are presumed to be inactive against a target, i.e., no binding to the target. The decoys may be structurally similar to active ligands or the like but lack the biological activity compounds. Utilizing the DUD-E database, by way of example, 975 decoy compounds (see, for example, FIG.6) are used that are structurally similar to the DFWLYH^FRPSRXQGV^EXW^ELRORJLFDOO\^LQHUW^DJDLQVW^3.&Į^^The decoys may be manuallyassigned significantly low IC50 values to signify their inactivity. In this experiment, as shown in FIG.3, the training dataset 302 for the QSAR-0 / ^PRGHO^WDUJHWLQJ^3.&Į^ consisted of 1,335 compounds (see FIG.3), thereby providing a robust foundation for generating accurate and reliable predictions on the compound bioactivity. The training dataset 302 may be similar to the dataset 210-216 shown in FIGs.2A-2C. In some embodiments, the training dataset 302 is bifurcated into active and inactive bioactivity classes inhibiting 3.&Į identified from the ChEMBL database. Details of example candidate compounds that inhibit 3.&Į identified from ChEMBL database are illustrated in FIG.5. Shown is a set of relevant compounds undergoing processing by the machine learning model, as well as a chemical structure language (Smiles) in a sequence for each compound for each class (i.e., active, intermediate, and inactive).

[0038] One or more or a combination of the active compounds in block 212, decoys in block 214, and inactive compound in 216 are processed for chemical descriptor generation. Each input dataset is interpreted by a plurality of different chemical fingerprints. In some embodiments, the collection of training data 210 consisting of active compounds 212 and inactive compounds 216 are encoded into a chemical fingerprint sequence 222, which serve as input variables to the machine learning model.

[0039] A chemical fingerprint is a high-dimensional vector generated from the molecular structure of a chemical compound, acting as one of the key descriptors in cheminformatics analytics. Specifically designed for the machine learning training regimen, embodiments of the model may include a QSAR-ML model that leverages the fingerprints to encode essential chemical and molecular traits, represented as bitstrings, bit-vectors, or arrays. Each bit in this computational representation symbolizes a predefined molecular characteristic; for instance, a “1” would indicate the presence of a specific feature like an H-donor, whereas a “0” would signify its absence.

[0040] As shown in FIGs.2A-2C, for comprehensive model training, three distinct types of chemical fingerprints may be employed: MACSS (Molecular ACCess System) keys 222A, Morgan fingerprint 222B, and PubChem fingerprints 222C, or a combination 222D of the fingerprints 222A-222C. To generate MACCS 222A and Morgan 222B fingerprints, a rdkit Python package is utilized in some embodiments, which efficiently encodes the chemical structures of the compounds into the desiredfingerprint formats. On the other hand, the padelpy package is employed for encoding the chemical structures into PubChem fingerprints 222C. In some embodiments, MACCS keys 222A consist of 166 pre-established structural fragments, individually crafted by medicinal chemists, to indicate the presence or absence of specific molecular features. Examples of MAACS fingerprints of training compounds that inhibit 3.&Į are shown in FIGs.7A and 7B. In other embodiments, Morgan fingerprints 222B employ the Morgan algorithm to allocate bits to atoms based on their circular molecular environments, with the radius determining the scope of neighboring bonds and atoms considered. PubChem fingerprints 222C provide a binary substructure, principally used for similarity searches and neighborhood calculations with the PubChem database. These three fingerprints may be combined (222D) as new types of fingerprints to represent the compounds for the QSAR-ML model training. The multiplicity of fingerprint types facilitates a nuanced understanding of EV inhibitor activity, allowing for comparative performance analyses within the QSAR model. This multifaceted approach enhances our model’s ability to accurately predict bioactivity based on structural features, thereby capturing intricate details of chemical structure.

[0041] In some embodiments, four distinct machine-learning algorithms 224A- 224D, generally, 224) are strategically deployed to facilitate the training of the QSAR models: Lasso Regression (Lasso) 224A, Decision Trees (DT) 224B , Random Forest (RF) 224C, and Artificial Neural Networks (ANN) 224D. Each algorithm has been chosen for its unique capabilities in uncovering the relationship between structural features and bioactivity in chemical compounds.

[0042] Lasso Regression 224A not only performs linear regression but also feature selection by applying penalty terms to the coefficients. This is invaluable for QASR models where the feature space is often high-dimensional and full of multicollinearity. The Lasso algorithm effectively eliminates irrelevant or redundant features, focusing solely on the structural descriptors that have a significant influence on bioactivity.

[0043] Decision Trees 224B are hierarchical models that partition the feature space into distinct regions, enabling straightforward interpretability. In the context of QSAR, each node 226 in the tree represents a decision criterion based on a chemicalfingerprint, and the leaf nodes correspond to the predicted bioactivity. This model is particularly useful for capturing local patterns and is easy to understand. Its interpretability is crucial for debugging and fine-tuning, especially when requiring transparent decision-making processes to understand how particular structural features impact bioactivity.

[0044] An ensemble learning method, Random Forest 224C, build multiple Decision Trees during training and output the mode of the classes or mean prediction of individual trees for classification or regression tasks, respectively. By leveraging the collective wisdom of multiple Decision Trees, Random Forests tend to deliver highly accurate predictions. Their inherent feature-importance metrics provide an added layer of interpretability, helping us to identify which structural aspects are most indicative of a compound’s bioactivity.

[0045] Artificial Neural Networks (ANNs) 224D are powerful non-linear models inspired by biological neural networks. They are particularly adept at capturing complex, intricate relationships in the data, which linear models may fail to do. Given the highly complex nature of the relationship between chemical structure and bioactivity, ANNs can unearth patterns that might be too intricate for other algorithms to detect.

[0046] The selection of these diverse algorithms serves a dual purpose. First, it allows us to capture different facets of the relationship between chemical structure and bioactivity, shown by leaf nodes 226, from the linear associations focused on by Lasso to the complex non-linear relationships modeled by ANNs. Second, the ensemble approach effectively mitigates the risk of overfitting and algorithmic bias. By evaluating and comparing the performance of each algorithm rigorously, only the most reliable and validated models are utilized for subsequent stages, thereby enhancing the robustness and accuracy of the machine learning models.

[0047] In some embodiments, the training set includes 80% of the initial dataset and the rest as the testing set. When collecting the data for the machine learning training, the input data set is split according, i.e., 80% for training and 20% for testing to improve accuracy of the model. This splitting can be processed at random by using the sample function from a built-in model in a Python program. Each time the data is randomly selected for training, and after multiple passes, i.e., tenfold of the training and testing, theaverage of the model’s performance can be determined. If the average is high for performance, then validation of the model can establish high performance, for example, after the chemical descriptor generation step shown in FIG.2A, or between fingerprint preparation and beginning of the training process is where the training and testing portions are split. In one embodiment, Pearson’s correlation coefficient (r) and the root means squared error (RMSE) are used to evaluate the performance of the QSAR-ML models. The r value measures the correlation of two variables and ranges from -1 to 1, which r > 0 implies that positive correlation of variables and vice versa for the negative value. The value of r equals zero means there is no correlation between the variables. The RMSE is a parameter that is used to measure the differences between the predicted values from the QSAR-ML model and the experimental values.

[0048] In a cross-validation step 228, the ability of the QSAR-ML models built in the previous steps is validated using the tenfold cross-validation and y-randomization test. Leaf nodes 230 are similar to the leaf nodes 226 except that they are cross-validated under step 228. For example, for the tenfold cross-validation of the leaf nodes, 9 subsets of the entire dataset may be used as training data and the reserving subset as the testing data. For tenfold cross-validation, the data set is divided randomly into a plurality of sections or subsets of data, where one of the subsets is not considered to avoid overfeeding of the machine learning algorithm, where the other subsets, e.g., 9 subsets, are provided for the model. By repeating this procedure ten times and averaging the model evaluation score, the predicted accuracies are obtained by statistical analysis. Y- randomization is a validation of the robustness of model performance by shuffling the data. The coefficient of determination (R2) (testing set) and cross-validated correlation coefficient (Q2) (cross-validation set) are used to quantify the models' performance for all validation tests. For example, R2>0.65 shown in FIG.2B, but not limited thereto, can be set as a threshold, which serves as a benchmark to assess model performance. It quantifies how well the variability in the observed data can be explained by the model’s input variables, and indicates that performance of the model is better than a randomization. Although the threshold of 0.65 is provided in some embodiments, it is not an absolute standard but a chosen value indicative of satisfactory performance for this specific application. It reflects a level of predictive accuracy deemed sufficient for ourpurposes. This threshold can be adjusted by the system if required, based on further validation or feedback from model testing.

[0049] To avoid bias, the QSAR-ML model randomly selects data and repeats 50 times. As the evaluation of the model, the Pean’s correlation coefficient (r) and the root means squared error (RMSE) of each test run are collected and used to calculate the mean and standard deviation. All the validation tests are the external validation for the QSAR model. The validation tests generate the R2and Q2as the measurement of model prediction and obtain the mean and standard deviation for comparison.

[0050] At blocks 232A-232D (generally, 232), the output of the leaf nodes 230 is combined with FDA drugs applied to each validated model for prediction, where top ranking drugs are identified. For example, in the context of inhibiting target proteins and subsequently reducing Extracellular Vesicle (EV) release, the output of the validated machine learning models that are fine-tuned during the training and testing phases is employed. Leveraging the same fingerprint generation techniques that were applied in constructing the training dataset, the chemical fingerprints 234 of FDA drugs 233 are output into the validated models to forecast the bioactivity (specifically, IC50 values) against designated target proteins. When the model is trained, it is used for prediction purposes. The chemical fingerprints 234 of FDA drugs 233 are compared to the output of the model to develop a list of inhibitors or the like of interest. These drugs are prioritized based on their predicted bioactivity levels, extracting the top 100 candidates from each model's output 230. From this refined list, drugs are identified in step 232 that appear consistently across all top-ranking categories, earmarking these as potential EV inhibitors 240. These selected compounds may undergo further biological validation to confirm their effectiveness in inhibiting EV release. In some embodiments, a screening step may be performed on the top ranking drugs identified in step 232 that may repeatedly appear. The screening step includes an application of an FDA-approved drug 234 to the output of the model, and the new dataset is applied to the model which repeats the abovementioned steps in FIGs.2A-2C.

[0051] FIG.8 is a diagram illustrating a workflow 800 of a Quantitative Structure-Activity Relationship machine learning (QSAR-ML) model for predicting FDA drug bioactivity, in accordance with some embodiments. In describing the workflow800, reference is made to elements of FIGs.1-7.

[0052] At block 802 of the workflow, training data is collected including active and inactive compounds, in this example, ChEMBL388978 and ChEMBL3350558. The training data may include a dataset of EV release inhibitors stored in a data repository, for example, a CHEMBL database.

[0053] At block 804, these compounds are encoded into chemical fingerprints sequence, serving as the input variables X. Corresponding bioactivity data quantified as ,&^^^YDOXHV^DJDLQVW^3.&Į^^DUH^REWDLQHG^IURP^WKH^&+(0% / ^GDWDEDVH^DQG^XVHG^DV^WKH^ response variables Y. The model is trained to identify important chemical features that contribute to fluctuation in IC50 values. At block 806, once validated, the model is employed to predict the IC50 of FDA-approved drugs, illustrated here by the midostaurin, which is a protein kinase inhibitor considered for the treatment of various cancers such as acute myeloid leukemia.

[0054] The FDA drugs as the data to predict are applied to the validated models. The top 5 drugs and their detailed drug facts are listed in the table 900 shown in FIG.9A. For identified drugs, the new method of treatment includes novel routes of administration including eye drops, ocular injection, and ocular tumor antibody-drug conjugate delivery. The predicted biological activities of all FDA drugs are collected for further analysis. In particular, shown are FDA-approved drugs identified via machine learning for novel ocular delivery to inhibit EV release as a method of childhood retinoblastoma treatment. The top drugs to be delivered using intraocular injection alone, or conjugated to antibodies for retinoblastoma tumor targeting in the eye include Midostaurin, Amiloride, Ponatinib, Selumetinib and Belumosudi.

[0055] FIGs.10A-10F are graphs of experimental results of retinoblastoma cell related EV size and concentration distributions. In describing the graphs in FIGs.10A- 10F, reference is made to features of the systems and methods described herein.

[0056] Retinoblastoma is a rare and fast-growing cancer that develops in the retina. Retinoblastoma are known to undergo metastasis using vesicles released by donor cells and taken up by recipient cells in an EV release and uptake. A machine learning model in accordance with embodiments herein can generate a drug library with functional properties similar to existing EV inhibition drugs. The novel drugs can be validated invitro for inhibition of EV genesis and uptake using childhood ocular cancer retinoblastoma cells. With the help of the computer platform 100 of FIG.11, novel drugs can be generated that may be useful in therapeutic applications to reduce retinoblastoma pathogenesis. FIGs.10A-10F illustrate an application of existing EV targeting drugs which have been used with the machine learning model to generate novel drugs. More specifically, FIGs.10A-10F illustrate retinoblastoma cell released EV size and concentration distributions An MA-treated sample shown in FIGs.10B, 10E have an unreasonably high concentration and a low average size of EV, so the result is not trustable. Other than that, the average concentration of control (CT) EV (FIGs.10A, 10D) is 4*E7 particles / ml, which is higher than the average of Y27632 treated Y79 EV (FIGs.10C, 10F) 1.7*E7 particles / ml. FIGs.10C and 10F illustrate results of another drug Y27632 (Y). The graphical results support the ability to reduce the EV release of Y27632 (Y). As shown in the comparative drawings, MA has the potential to inhibit EV release, but more experiments need to be repeated to correct the abnormal results of sample shown in FIG.10B.

[0057] In some embodiments, drugs identified from the method 200 of FIGs.2A- 2C may be administered via intravitreal injection. In other embodiments, administering occurs via intravitreal injection and the drug is conjugated to an antibody. In other embodiments, the administering occurs via ophthalmic arterial injection. In other embodiments, the administering occurs via ophthalmic arterial injection and the drug is conjugated to an antibody. Although the foregoing may be ocular cancer retinoblastoma, identified drugs may be used for other medical conditions.

[0058] FIG.11 is a block diagram of a computer platform 1100 in which embodiments of the present inventive concept can be practiced. In some embodiments, the computer platform 1100 can identify drugs for neuron invasion and synapse formation. In other embodiments, the computer platform 1100 can be part of a system to develop a novel drug library to drive PPC migration into the retina and functional synapse formation or to target cancer extracellular vesicle release and uptake mechanisms to prevent metastasis. More specifically, the computer platform 1100 can be used to build a drug library and improve new photoreceptors transplanted into a damaged retina to migrate toward functional synapse formation, for example, shown in FIGs.12A-12C.In some embodiments, the drug library can be implemented to inhibit EV uptake. In some embodiments, the drug library can be implemented to inhibit EV release. In some embodiments, the drug library can be implemented to drive transplantable neuron migration. In some embodiments, the drug library can be implemented to drive transplantable neuron invasion. In some embodiments, the drug library can be implemented to drive neuron synaptogenesis. It is well-known that blinding diseases destroy photoreceptors leading to blindness. To restore vision, new photoreceptors may be transplanted and have to migrate into the host retina and form new synapses. As described in FIG.11, the computer platform 1100 can provide a drug library in a database 1112 used to target mechanisms to enhance photoreceptor replacement cell migration and synaptogenesis. The computer platform 1100 is not limited to providing these drug libraries described herein, and other applications may equally apply.

[0059] The computer platform 1100 that stores and executes software, for example, implementing machine learning models 1114, e.g., similar to or the same as machine learning models described above, to identify drugs for particular purposes, such as neuron invasion and synapse formation, where the drug database can target extracellular vesicle release and uptake inhibition and to drive neuron migration and synaptogenesis. In other embodiments, the computer platform 1100 may include a database for discovering drugs useful in therapeutic applications to address diseases related to vision loss. Vision loss due to retinal degenerative diseases like Retinitis Pigmentosa is overwhelmingly attributed to the death of retinal photoreceptor cells, which initiate vision by phototransduction and signaling to the visual cortex. Photoreceptors develop from photoreceptor precursor cells (PPCs) during embryonic and early postnatal development. Recent studies show that transplanted early postnatal PPCs can replace degenerated photoreceptors in the adult retinas of mice. The level of restored visual response depends on the migration of transplanted cells from the injection site in the sub-retinal space, through the interphotoreceptor matrix (IPM) and into the retina.

[0060] As shown in FIG.11, the computer platform can include one or more processors 1102, non-transitory computer readable medium or memory 1104, I / O interface devices 1106 (e.g., wireless communications, etc.) and a network interface (not shown). The computer readable medium 1104 may include an operating system 1108,one or more applications 1110 for performing some or all of the method 200 of FIGs.2A- 2C, a database 1112 in accordance with the systems and methods described herein.

[0061] In operation, the processor 1102 may execute the application 1110 stored in the computer readable medium 1104. The application 1110 may include software instructions that, when executed by the processor 1102, cause the processor 1102 to perform operations for identifying new molecules and drugs and a software platform to identify drugs for neuron invasion, photoreceptor migration, and synapse formation, as described and shown in FIGS.1 – 10, with particular reference to the steps of the methodology shown in FIG. 2. The database 1112 is constructed and arranged to store one or more drug libraries.

[0062] The application program 1110 may operate in conjunction with the database 1112 and the operating system 1108. The device 1100 may communicate with other devices, systems, or the like, for example, an artificial intelligence system, quantum computing system, or the like, via the I / O interface 1106.

[0063] With further regard to FIGs.12A-12C, shown at FIG.12A at a scale of ^^^P are isolated PPCs and their transplantation into a retina ex vivo. In some embodiments, the PPCs are isolated from crx-gfp + / + mouse retina and FACS sorted for migration analysis. FIG.12B is a diagram of IPM ensheathing photoreceptor outer segments and filling the subretinal space (blue cross-hatching) where PPCs are transplanted. FIG.12C is a view shown at a scale of ^^^P of a cryosection of GFP+ cells migrated into wild-type retina 5 days after transplantation into the IPM region (host retina nuclei blue) leads to migration into the retina.

[0064] While the invention has been described with reference to certain embodiments, it will be understood by those skilled in the art that various changes may be made and equivalents may be substituted for elements thereof to adapt to particular situations without departing from the scope of the disclosure. Therefore, it is intended that the claims are not limited to the particular embodiments disclosed, but that the claims will include all embodiments falling within the scope and spirit of the appended claims.

Claims

What is claimed is:

1. A method for drug identification, comprising: receiving, by a machine learning model stored at and executed by a special- purpose computer, a set of input data; arranging the input data into active compounds, inactive compounds, and decoy compounds; applying a plurality of chemical fingerprint processes to the active compounds, inactive compounds, and decoy compounds to generate a training dataset; receiving the training dataset by a machine learning model, comprising a plurality of different machine learning algorithms; performing a validation process on an output of each of the different machine learning algorithms; and predicting bioactivity levels of a plurality of drugs by applying a listing of the drugs to an output of the validation process.

2. The method of claim 1, wherein the training data is output for generating a plurality of EV release inhibitors.

3. The method of claim 1, wherein the validating step includes using the tenfold cross- validation and y-randomization test.

4. The method of claim 1, further comprising selecting potential EV inhibitors from a result of a ranking of the plurality of drugs in response to applying the drugs to the output of the validation process.

5. The method of claim 1, wherein predicting bioactivity levels of a plurality of drugs comprises comparing at least one food and drug administration (FDA) drug to anoutput the machine learning model, and outputting a result to the machine learning model to determine at least one inhibitor of interest.

6. The method of claim 1, further comprising: prioritizing drugs of interest from a result of a comparison of the predicted bioactivity levels of the plurality of drugs and at least one food and drug administration (FDA) drug of the listing of the drugs.

7. The method of claim 1, further comprising: generating a list of potential EV inhibitors identified from the output of the machine learning model.

8. The method of claim 1, further comprising: generating a first library based on a gene expression level for a collection of drugs of interest developed for new uses; and generating a second library based on chemical structural data generated in response to the machine learning model.

9. The method of claim 1, wherein the identified drug is for treating a medical condition including ocular cancer retinoblastoma.

10. A machine learning model, comprising: an input for receiving a set of training data; a processor that applies a set of fingerprints to the training data; and a set of machine learning algorithms that predict bioactivity of a set of drugs using the training data applied to a stored listing of drugs.

11. The machine learning model of claim 10, wherein the input is arranged into active compounds, inactive compounds, and decoy compounds.

12. The machine learning model of claim 11, wherein the fingerprints include a plurality of chemical fingerprint processes applied to the active compounds, inactive compounds, and decoy compounds to generate the training data.

13. The machine learning model of claim 10, further comprising a special purpose computer processor that performs at least one validation process on an output of the machine learning algorithms 14. The machine learning model of claim 10, wherein the training data is output for generating a plurality of EV release inhibitors.

15. The machine learning model of claim 14, wherein the EV release inhibitors are generated from a result of a ranking of a plurality of drugs determined from the output of the machine learning model.

16. The machine learning model of claim 10, wherein predicting bioactivity levels of a plurality of drugs comprises comparing at least one food and drug administration (FDA) drug to an output the machine learning model, and outputting a result to the machine learning model to determine at least one inhibitor of interest.

17. A drug library generated according to a process, comprising: receiving, by a machine learning model stored at and executed by a special- purpose computer, a set of input data; arranging the input data into active compounds, inactive compounds, and decoy compounds; applying a plurality of chemical fingerprint processes to the active compounds, inactive compounds, and decoy compounds to generate a training dataset; receiving the training dataset by a machine learning model, comprising a plurality of different machine learning algorithms;performing a validation process on an output of each of the different machine learning algorithms; and predicting bioactivity levels of a plurality of drugs by applying a listing of the drugs to an output of the validation process.

18. The drug library of claim 17, method of claim 1, further comprising generating chemical structural data in response to the machine learning model.

Images