Drug combinations screening
Patent Information
- Application Number
- HK62026125825
- Authority / Receiving Office
- HK · HK
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-10-06
- Filing Date
- 2026-07-07
- Publication Date
- 2026-09-11
Smart Images

Figure 00000000_0000_ABST
Abstract
Description
(19) State Intellectual Property Office (12) Invention Patent Application (10) Application Publication Number (43) Application Publication Date (21) Application Number 202480063566.0 (22) Application Date 2024.10.04 (30) Priority Data 23202094.1 2023.10.06 EP (85) PCT International Application Entering National Phase Date 2026.04.01 (86) PCT International Application Application Data PCT / EP2024 / 078058 2024.10.04 (87) PCT International Application Publication Data WO2025 / 073970 EN 2025.04.10 (71) Applicant: Hofmeister Roche Ltd. Address: Basel, Switzerland (72) Inventors: M. Beck, N. Gazzarra, H.K. Hayward-Konek, J. Ofrael, A. Sadat Varnossafadrani, G.A. Tanei (74) Patent Agency: Beijing Kunrui Law Firm, 11494 Patent Attorney: Feng Xinqin (51) Int.Cl. G16C 20 / 30 (2006.01) (54) Invention Title: Drug Combination Screening (57) Abstract: A method for screening drug combinations is described. The method includes: obtaining values of a plurality of features associated with the drug combination for a first drug and a second drug; and using the values of the plurality of features, using a machine learning model trained to predict whether the drug combination is likely to be synergistic to predict the effect of the drug combination. The plurality of features includes at least one feature that characterizes the similarity between one or more indications or diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug. Related methods, products, and systems are also described. Claims 3 pages, Description 37 pages, Drawings 27 pages, CN 122342008 A 2026.07.03 CN 1 22 34 20 08 A 1. A computer-implemented method for predicting the effect of a drug combination, the method comprising: obtaining values of a plurality of features associated with the drug combination for a first drug and a second drug; and using the values of the plurality of features, using a machine learning model trained with training data to predict whether the drug combination is likely to be synergistic to predict the effect of the drug combination, the training data including for each of a plurality of training drug combinations: (i) the values of the plurality of features; and (ii) an indication of whether the drug combination is synergistic; wherein the plurality of features includes at least one feature characterizing one or more indications or diseases associated with one or more targets of the first drug and associated with one or more targets of the second drug.1. Similarity between one or more indications or diseases associated with the drug. 2. The method of claim 1, wherein the machine learning model has been trained to classify drug combinations among multiple categories including synergistic and non-synergistic categories, and wherein predicting the effect of the drug combination includes determining whether the drug combination is likely to belong to the synergistic category. 3. The method of any preceding claim, wherein the plurality of features comprises: at least one feature characterizing the similarity between one or more indications associated with one or more targets of the first drug and one or more indications associated with one or more targets of the second drug, wherein the indication is provided as a Medical Subject Headings term associated with the one or more targets of the drug; and / or at least one feature characterizing the similarity between one or more diseases associated with one or more targets of the first drug and one or more diseases associated with one or more targets of the second drug, wherein the disease is provided as a disease ontology term associated with the one or more targets of the drug. 4. The method according to any preceding claim, wherein the similarity between one or more indications or diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug is obtained as a semantic similarity measure between indication or disease terms associated with the one or more targets of the first drug and indication or disease terms associated with the one or more targets of the second drug, or as a summary measure aggregating multiple such semantic similarity measures, each such semantic similarity measure corresponding to a pair of indication or disease terms, the pair of indication or disease terms including indication or disease terms associated with the target of the first drug and indication or disease terms associated with the target of the second drug. 5. The method according to any preceding claim, wherein the plurality of features comprises: one or more drug similarity features, one or more target similarity features, and one or more indication similarity features, optionally wherein the one or more drug similarity features are selected from: side effect similarity, mode of action similarity, and molecular structure similarity. 6. The method of claim 5, wherein the one or more target similarity features are selected from: pathway similarity; disease association similarity, which characterizes the similarity between one or more diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug; and a measure of the proximity of the targets of the first drug and the targets of the second drug on a protein-protein interaction network. 7. The method of claim 5 or 6.The method of any one of the following, wherein the one or more indication similarity features are selected from: an indication similarity measure based on the target of the drug, which characterizes the similarity between one or more indications associated with one or more targets of the first drug and one or more indications associated with one or more targets of the second drug; and an indication similarity measure based on the drug, which characterizes the similarity between indication annotations associated with the first drug and the second drug. 8. The method of claim 7, wherein the indication annotation is provided as Anatomical Therapeutic Chemistry (ATC) coding, and / or wherein the indication similarity measure based on the drug is selected from: a similarity measure that quantifies the similarity between multiple levels of the ATC codes for the first drug and the second drug; and a similarity measure that compares a first level of the ATC codes for the first drug and the second drug. 9. The method of claim 7 or claim 8, wherein the similarity metric for comparing the first level of the ATC codes for the first drug and the second drug is a binary feature indicating whether the first drug and the second drug are associated with ATC codes belonging to the same first level category or to a particular pair of ATC categories, and / or wherein the plurality of features comprises a set of binary features indicating whether the first drug and the second drug belong to each of a plurality of pairs of first level ATC categories, optionally wherein the plurality of pairs of first level ATC categories includes LL, NN, JJ, and CC. 10. The method of any preceding claim, wherein the machine learning model is selected from: decision trees and variants thereof, including regularized and / or gradient boosting decision trees and random forest models; regularized discriminant analysis; logistic regression models; artificial neural networks; Naive Bayes classifiers; support vector machines; and multivariate adaptive regression splines, and / or wherein the machine learning model is a nonlinear model, and / or wherein the machine learning model is an ensemble model, and / or wherein the machine learning model is a tree-based model. 11. The method according to any preceding claim, wherein the plurality of features comprises: at least one feature characterizing the similarity between one or more indications associated with one or more targets of the first drug and one or more indications associated with one or more targets of the second drug, wherein the indication is provided as a medical subject term associated with the one or more targets of the drug; side effect similarity, mode of action similarity, molecular structure similarity, pathway similarity,Disease association similarity, characterizing the similarity between one or more diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug; and a measure of the proximity of the targets of the first drug and the targets of the second drug on a protein-protein interaction network; and a drug-based indication similarity measure, characterizing the similarity between indication annotations associated with the first drug and the second drug, wherein the indication annotations are provided as Anatomical Therapeutic Chemistry (ATC) codes, and wherein the drug-based indication similarity measure includes a similarity measure that quantifies the similarity between multiple levels of the ATC codes for the first drug and the second drug and optionally a similarity measure that compares a first level of the ATC codes for the first drug and the second drug, wherein the similarity measure that compares the first level of the ATC codes for the first drug and the second drug is a set of binary features indicating whether the first drug and the second drug belong to each of multiple pairs of first-level ATC categories, optionally wherein the multiple pairs of first-level ATC categories include LL, NN and optionally JJ, CC, One or more of JS, DR, DS, CJ, AA, AM, AN, and BN. Claims 2 / 3 pages 3 CN 122342008 A 12. The method according to any of the preceding claims, wherein the machine learning model has been trained using a training dataset comprising at least 1000, 2000, or 3000 unique drug combinations, and / or wherein, when evaluated on unseen data, the trained machine learning model has at least 90% specificity and 70% sensitivity in predicting synergistic combinations, and / or wherein the method further comprises, prior to providing the values of the plurality of features as input to the machine learning model, applying one or more steps selected from: missing value imputation, filtering, standardization, and normalization. 13. A method for providing a tool for predicting the effects of drug combinations, the method comprising: for each of a plurality of training drug combinations comprising a first drug and a second drug, obtaining: (i) values of a plurality of features associated with the drug combination; and (ii) information indicating whether the drug combination is synergistic; and using the training data to train a machine learning model to predict whether the drug combination is likely to be synergistic, wherein the machine learning model is trained to take the values of the plurality of features for the drug combination as input and produce an indication as output of whether the drug combination is likely to be synergistic;The plurality of features includes at least one feature characterizing the similarity between one or more indications or diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug. 14. A method for identifying synergistic drug combinations, the method comprising: obtaining a plurality of candidate drug combinations; predicting the effect of each of the plurality of candidate drug combinations using the method of any one of claims 1 to 13; and using the results of the prediction to select a subset of the plurality of candidate drug combinations for experimental testing; and optionally further comprising testing the selected subset of the plurality of candidate drug combinations using one or more in vitro assays. 15. A system comprising: at least one processor; and at least one non-transitory computer-readable medium containing instructions that, when executed by the at least one processor, cause the at least one processor to perform the method of any one of claims 1 to 13. Claims 3 / 3 Page 4 CN 122342008 A Drug Combination Screening Technology
[0001] This disclosure relates to methods for predicting the effects of drug combinations, and more particularly to using machine learning to predict the effects of treatment utilizing the pair against a specific condition from the properties of individual drugs in a candidate drug pair. Related methods, systems, and products are described. Background Art
[0002] Despite advances, effective treatments for the progression and disability of multiple sclerosis (MS) remain limited. Drug combinations have shown promise in fields such as oncology and diabetes. However, screening new drug combinations is costly, time-consuming, and often impossible on a near-exhaustive scale, even when limited to approved therapies. For diseases such as MS (for which there are few known disease-modifying therapies), the real potential lies in exploring combinations with drugs that have not yet been indicated for treating the disease, a combination space that is simply too vast to be explored experimentally.
[0003] Computational methods exist for predicting the effects of drug combinations, and these have been reviewed in Paltun et al. 2021. These include machine learning-based methods, among others, including methods designed to identify new synergistic combinations (“drug synergistic classification” methods). Previous research on such methods indicates that drug target proteins, anatomical therapeutic chemistry (ATC) encoding, and drug indications are highly informative in making such predictions (Paltun et al., 2021). In particular, ATC drug encoding information has been identified as the most useful information for such predictions. However, even with this information, predicting new combinations (especially those involving drugs with no known drug combinations) has been found difficult.
[0004] For example, Ji et al. (2019) proposed a five-feature approach...The XGBoost model is used to classify drug combinations as synergistic or antagonistic. This model uses: a disease overlap score, calculated as the proportion of the two drugs sharing the same indication; an adverse reaction overlap score, calculated as the Jaccard similarity between adverse drug reactions recorded for the two drugs; a biological process similarity score, measured as the semantic similarity between gene ontology annotations of the targets, enzymes, and transporters for the two drugs; an action mode similarity score, obtained by integrating drug-target interaction information (i.e., promoting / inhibiting interactions) and directed and signed protein interaction networks (i.e., interactions between proteins indicating promoting or inhibiting interactions); and a separation score, which quantifies the network-based separation of the two drugs on a protein-protein interaction network. Chemical structure similarity and ATC similarity were also tested but excluded from the final model after feature selection.
[0005] Existing methods still have predictive accuracy that can be improved and rely heavily on predictive features, such as ATC encodings unavailable for novel drug candidate compounds, further limiting the predictive accuracy of these methods for any combination involving such drugs.
[0006] Therefore, there is a need for improved systems and methods for predicting the effects of drug combinations. Summary of the Invention
[0007] Here, the inventors have set out to construct a predictive model to use machine learning to discover “high-potential drug combinations” for improving the treatment of MS. In doing so, they have developed a new machine learning method for predicting the effects of drug combinations, which utilizes new predictive features found to contribute to high predictive accuracy. In particular, the inventors have defined a novel predictive feature for characterizing drug combinations based on the similarity of the disease-related annotations of the drug target specification (page 1 / 37, CN 122342008 A). Existing methods make extensive use of disease indication information associated with the drug itself, and particularly extensive use of ATC codes. However, such disease indication information is only available for drugs that have been assigned an ATC code. ATC codes are assigned by the WHO based on applications from drug owners and require readiness to submit a market authorization application in at least one country. Therefore, drugs still in clinical trials typically do not have an ATC code. In contrast, the inventors propose novel predictive features that quantify the similarity of disease indications and / or associated information by measuring: (a) overlapping medical subject terms (MTTs) in the database for each drug target; and / or (b) the similarity between disease ontologies associated with the drug target. These effectively quantify target-disease similarity based on information about the drug target rather than the drug itself. In particular, the use of MTT information has been found to be more effective than based on ATCs.The metrics are highly informative, and these metrics were previously considered to be the most informative features for such models.
[0008] The inventors demonstrate that the newly developed model has extremely high predictive performance (specificity 94%, sensitivity 70%, accuracy 90%) on hold-out test data including approved drugs. They further demonstrate that the model is applicable to disease situations poorly captured by existing models due to the scarcity of information such as the availability of drugs with ATC codes, because the most informative features of this model are newly proposed features that do not rely on this information and whose missing data problem is greatly reduced.
[0009] The newly developed method and model are applicable to any disease situation and any type of drug, and have been tested and validated on datasets including thousands of drug combinations tested in clinical trials.
[0010] According to a first aspect of this disclosure, a method for predicting the effect of a drug combination is provided, the method comprising the steps of: obtaining, for a first drug and a second drug, values of a plurality of features associated with the drug combination; and using the values of the plurality of features, predicting the effect of the drug combination using a machine learning model trained with training data to predict whether the drug combination is likely to be synergistic, the training data for each of a plurality of training drug combinations including: (i) the values of the plurality of features and (ii) an indication of whether the drug combination is synergistic; wherein the plurality of features includes at least one feature characterizing the similarity between one or more indications or diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug.
[0011] The method of the first aspect may have any of the following optional features or any combination thereof.
[0012] The machine learning model may have been trained to classify the drug combination among a plurality of categories including synergistic and non-synergistic categories, and predicting the effect of the drug combination may include determining whether the drug combination is likely to belong to a synergistic category. Thus, the training data may include a tag indicating one of the plurality of categories. In other words, an indication of whether a drug combination is synergistic can be presented as a tag indicating one of a plurality of categories.
[0013] The plurality of features may include at least one feature characterizing the similarity between one or more indications associated with one or more targets of a first drug and one or more indications associated with one or more targets of a second drug, wherein the indications are provided as medical subject heading terms associated with one or more targets of the drug. The plurality of features may include at least one feature characterizing the similarity between one or more diseases associated with one or more targets of a first drug and one or more diseases associated with one or more targets of a second drug, wherein the diseases are provided as disease ontology terms associated with one or more targets of the drug.
[0014] The similarity between one or more indications or diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug can be obtained as a semantic similarity measure between indication or disease terms associated with one or more targets of the first drug and indication or disease terms associated with one or more targets of the second drug, or as a summary measure aggregating multiple such semantic similarity measures, each such semantic similarity measure corresponding to a pair of indication or disease terms, which includes indication or disease terms associated with the targets of the first drug and indication or disease terms associated with the targets of the second drug.
[0015] Multiple features may include: one or more drug similarity features, one or more target similarity features, and one or more indication similarity features. One or more drug similarity features may be selected from: side effect similarity, mode of action similarity, and molecular structure similarity. One or more target similarity features may be selected from: pathway similarity; disease association similarity, which characterizes the similarity between one or more diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug; and a measure of the proximity of the targets of the first drug and the targets of the second drug on a protein-protein interaction network. One or more indication similarity features may be selected from: drug-based indication similarity measures, which characterize the similarity between one or more indications associated with one or more targets of the first drug and one or more indications associated with one or more targets of the second drug; and drug-based indication similarity measures, which characterize the similarity between indication annotations associated with the first drug and the second drug.
[0016] Indication annotations may be provided as Anatomical Therapeutic Chemistry (ATC) codes. Drug-based indication similarity measures may be selected from: similarity measures that quantify the similarity between multiple levels of ATC codes for the first drug and the second drug; and similarity measures that compare a first level of ATC codes for the first drug and the second drug. One or more indication similarity features may include similarity measures that quantify the similarity between multiple (optionally all) levels of the ATC codes for the first and second drugs, and one or more similarity measures that compare the first level of the ATC codes for the first and second drugs. The similarity measures comparing the first level of the ATC codes for the first and second drugs may be ATCs indicating whether the first and second drugs belong to the same first level category or to a specific pair of ATC categories.Encoding associated binary features. Multiple features may include a set of binary features indicating whether the first and second drugs belong to each of multiple pairs of first-level ATC categories. Multiple pairs of first-level ATC categories may include LL, NN, JJ, and CC.
[0017] The machine learning model may be selected from: decision trees and their variants, including regularized and / or gradient boosting decision trees and random forest models; regularized discriminant analysis; logistic regression models; artificial neural networks; Naive Bayes classifiers; support vector machines; and multivariate adaptive regression splines. The machine learning model may be a nonlinear model. The machine learning model may be an ensemble model. The machine learning model may be a tree-based model. In this embodiment, the machine learning model is a gradient boosting tree model.
[0018] The multiple features may include: (i) at least one feature characterizing the similarity between one or more indications associated with one or more targets of the first drug and one or more indications associated with one or more targets of the second drug, wherein the indications are provided as medical subject terminology terms associated with one or more targets of the drug; (ii) side effect similarity; (iii) mode of action similarity; (iv) molecular structure similarity; (v) pathway similarity; (vi) disease association similarity characterizing the similarity between one or more diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug; and a measure of the proximity of the targets of the first drug and the targets of the second drug on a protein-protein interaction network; and (vii) a drug-based indication similarity measure characterizing the similarity between indication annotations associated with the first drug and the second drug, wherein the indication annotations are provided as Anatomical Therapeutic Chemistry (ATC) codes, and wherein the drug-based indication similarity measure includes ATCs for the first drug and the second drug. A similarity measure that quantifies the similarity between multiple levels of coding. The indication-based similarity measure for a drug may further include a similarity measure comparing the first level of the ATC codes for the first drug and the second drug, wherein the similarity measure comparing the first level of the ATC codes for the first drug and the second drug (page 3 / 37, CN 122342008 A) is a set of binary features indicating whether the first drug and the second drug belong to each of multiple pairs of first-level ATC categories, optionally wherein the multiple pairs of first-level ATC categories include LL, NN, and optionally one or more of JJ, CC, JS, DR, DS, CJ, AA, AM, AN, and BN.
[0019] The machine learning model may have used at least 1000, 2000, or 3000The training dataset is used to train the machine learning model on a unique dataset of drug combinations. When evaluated on unseen data, the trained machine learning model has at least 90% specificity and 70% sensitivity in predicting synergistic combinations. The method may further include applying one or more steps selected from the following before providing the values of the plurality of features as input to the machine learning model: missing value imputation, filtering, standardization, and normalization.
[0020] Embodiments of any aspect may include providing the results of any step to a user, for example, through a user interface.
[0021] According to a second aspect, a method is provided for providing a tool for predicting the effects of drug combinations, the method comprising obtaining, for each of a plurality of training drug combinations comprising a first drug and a second drug: (i) values of a plurality of features associated with the drug combination; and (ii) information indicating whether the drug combination is synergistic; and using the training data to train a machine learning model to predict whether the drug combination is likely to be synergistic, wherein the machine learning model is trained to take the values of the plurality of features for the drug combination as input and produce an indication as output of whether the drug combination is likely to be synergistic; wherein the plurality of features includes at least one feature characterizing the similarity between one or more indications or diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug.
[0022] The method according to this aspect may have any one or more features described in connection with the first aspect.
[0023] According to a third aspect, a method is provided for identifying synergistic drug combinations, the method comprising: obtaining a plurality of candidate drug combinations; predicting the effects of each of the plurality of candidate drug combinations using the method of any embodiment of the first aspect; and using the prediction results to select a subset of the plurality of candidate drug combinations for experimental testing. The method may further include testing a selected subset of a plurality of candidate drug combinations using one or more in vitro assays.
[0024] The candidate drug combinations may each comprise a drug that is approved or used in clinical or preclinical trials for the treatment of a specific disease or condition (such as, for example, MS).
[0025] According to a fourth aspect, a system is provided comprising: at least one processor; and at least one non-transitory computer-readable medium containing instructions that, when executed by at least one processor, cause at least one processor to perform the method of any embodiment of any of the first, second, or third aspects or any method described herein.
[0026] According to yet another aspect, a non-transitory computer-readable medium is provided containing instructions that, when executed by at least one processor, cause at least one processor to perform the method of any embodiment of any of the first, second, or third aspects or any method described herein.
[0027] According to another aspect, a computer program is provided, the computer program including code that, when executed on a computer, causes the computer to perform the methods of any embodiment of the first, second, or third aspect or any methods described herein. Brief Description of the Drawings
[0028] Embodiments of the present disclosure will now be described by way of example with reference to the accompanying drawings, in which:
[0029] FIG1 is a flowchart illustrating a method for predicting the effect of a drug combination according to a general embodiment of the present disclosure (page 4 / 37, 8 CN 122342008 A);
[0030] FIG2 is a flowchart illustrating a method for providing a method for predicting the effect of a drug combination according to an embodiment of the present disclosure;
[0031] FIG3 schematically illustrates an exemplary system according to the present disclosure;
[0032] FIG4 schematically illustrates a process for predicting the effect of an example drug combination of the present disclosure.
[0033] FIG5 schematically illustrates the concept of a measure of proximity of drug targets in a protein-protein interaction (PPI) network. A. Example of two overlapping drugs. B. Example of two separate drugs.
[0034] Figure 6 schematically illustrates the data structure used in the examples of this disclosure.
[0035] Figure 7 shows the missing data in the predicted features for the training data used in the examples of this disclosure. Rows are drug pairs in the training dataset, and columns are the predicted features for the drug pairs. The percentage in the column labels is the percentage of missing data for a specific feature. Se_sim = side effect similarity; kegg_sim = pathway similarity (kegg pathway database); atc_sim = ATC encoding similarity; mesh_sim = MESH similarity; ppi_mean_dist = average distance between drug target clusters in the PPI network; do_sym = disease ontology similarity; moa_sim = mode of action similarity; smiles_sim = molecular structure similarity.
[0036] Figure 8 shows the results of the feature selection process used in the examples of this disclosure. A. F1 score with progressively smaller predicted feature sets. B. Tolerance (loss of F1 performance compared to the best model, %). Triangles indicate the selected model (8 features). Solid dots indicate the optimal model (11 features). A 5% tolerance (dashed horizontal line) is applied to the model selection.
[0037] Figure 9 shows the results of a survey on the importance of predicted features (1-AUC for 10 random permutations of the corresponding features) in the model according to this disclosure. A. Individual predicted features, ATC aggregated (i.e., all ATC-related features are considered together, including ATC coding similarity and ATC...A. Pairwise probabilities of combinations of 14 different Level 1 values encoded. B. Individually predicted features, showing the selected ATC binary categories. C. Importance aggregated by category of predicted features.
[0038] Figure 10 shows the correlation between predicted features in an example of this disclosure. A. Correlation between the 8 features used in the model after feature selection (atc_sim only shows the full encoded similarity). Feature plot example is shown in Figure 7. B. Correlation between all features considered, except that only the ATC Level 1 binary categories of their most important categories are shown. Feature plot example is shown in Figure 7, except that the ATC Level 1 binary categories with high importance are each indicated as “atc_xx”, where “xx” refers to the first left of the ATC code for each drug in the corresponding drug (e.g., “atc_ll” is the value of the binary variable, which is “1” if both drugs belong to ATC Level 1 category L, and “0” otherwise). On both plots, the color intensity and size of the dots indicate the magnitude of the correlation.
[0039] Figure 11 shows the distribution of predicted feature values separated by antagonistic and synergistic examples for each of the multiple predicted features. A. Individual predicted features, ATC pooled (i.e., all ATC-related features are considered together, including ATC encoding similarity and the pairwise probability of combinations of 14 different first-level values of ATC encoding). B. Individual predicted features, showing the selected ATC binary categories. MESH = Mesh Similarity, Molecular Structure = Molecular Structure Similarity, Network Separation = PPI Proximity, MOA = Mode of Action Similarity, Pathway = Pathway Similarity (kegg), ATC = Pooled ATC Similarity, Disease Ontology = Disease Ontology Similarity, Side Effect = Side Effect Similarity.
[0040] Figure 12 shows the results of a comparison of different model architectures and data preprocessing pipelines. A. F1 scores (top) and balanced accuracy obtained on the hold-out test dataset using different model architectures and data preprocessing pipelines. B. Post-test distribution of average balance accuracy obtained on the holdout test dataset using different model architectures and data preprocessing pipelines. (Page 5 / 37, CN 122342008 A) C. Post-test distribution of average balance accuracy on the holdout test data between the Random Forest model on data processed only by imputation and the XGBoost model on data processed by imputation and normalization. boost_tree = Gradient Boosting Tree (XGBoost), decision_tree = Decision Tree Ensemble, discrim_regularized = RegularizationDiscriminant analysis, logistic_reg = logistic regression, mars = multivariate adaptive regression splines, mlip = multilayer perceptron, rand_forest = random forest, svm_rbf = support vector machine with radial basis function kernel, rec_impute = data processed only by imputation, rec_impute_norm = data processed by imputation followed by standardization and normalization, rec_impute_norm_pca = data processed by imputation followed by standardization and normalization, followed by dimensionality reduction by principal component analysis (PCA), rec_impute_norm_pls = data processed by imputation followed by standardization and normalization, followed by dimensionality reduction by partial least squares regression (PLS). D. Practical equivalence of the workflow set: This proportion of the model was obtained on resampled data (from cross-validation), and the model has an effect size of up to 2% (average loss of balanced accuracy compared to the best model). This indicates that the XGboost model performs best on both the raw and normalized data, and the Random Forest model is approximately 68% practically equivalent to the XGboost model.
[0041] Figure 13 shows the results of the threshold selection and optimization process used to obtain the model of this disclosure (XGboost model). A. F1 score, precision (increasing data sequence), and recall (decreasing data sequence) obtained by the XGBoost model as the threshold for identifying predictions as "positive examples" (cooperative) for the training data increases. B. Balanced accuracy, sensitivity (decreasing data sequence), and specificity (increasing data sequence) obtained by the XGBoost model as the F1 threshold increases (same model as in Figure 13B, but evaluated on different performance metrics). Dashed lines indicate the threshold selected based on the F1 score.
[0042] Figure 14 shows the predictive performance of the model of this disclosure (XGBoost) evaluated on the hold-out test dataset. A. Confusion matrix obtained by evaluating the trained XGBoost model on the held-out test data. B. Receiver operating characteristic (ROC) curves obtained by applying different predicted probability thresholds to the model results.
[0043] Figure 15 shows calibration curves (x-axis indicating the midpoint of the interval) of the observed co-log (event rate) for each of the multiple intervals (bins) of the predicted probabilities from the XGBoost model. The 0-1 probability range is divided into 10 equal-sized intervals [0, 0.1) [0.1, 0.2)……[0.9, 1], with the indicated midpoint, i.e.0.05, 0.15, ..., 0.95. Plot the proportion (solid dots) of the pairs that are predicted to fall within each interval, and the range of the predicted probabilities for these observations.
[0044] Figure 16 shows a partial dependency plot for the predicted feature pairs. Mesh_sim = MESH similarity, ATC_sim = ATC similarity (full ATC coding similarity), do_sim = disease ontology similarity, smiles_sim = structural similarity, kegg_sim = pathway similarity, se_sim = side effect similarity, ppi_separation = proximity of drug targets on the PPI network, moa_sim = mode of action similarity. The partial dependency plot shows how the expected values predicted by the model (here, the co-probabilities on the y-axis) are represented by the selected explanatory variables (indicated on the x-axis). Each gray line corresponds to a single drug combination, and the bold blue line is the average of all gray lines. The PD curve at the value z of the explanatory variable represents the empirically estimated expected value predicted by the model when the joint distribution of that explanatory variable relative to all other explanatory variables is fixed at the value z.
[0045] Figure 17 shows the missing values for each drug pair containing a known disease-modifying therapy for MS. Each row represents the drug combination, and black indicates missing values. Mesh_sim = MESH similarity, ATC_sim = ATC similarity, do_sim = disease ontology similarity, smiles_sim = structural similarity, kegg_sim = pathway similarity, se_sim = side effect similarity, ppi_separation = proximity of drug targets on the PPI network, moa_sim = mode of action similarity. The percentage of missing data is indicated on the x-axis for each feature. Instruction manual 6 / 37 pages 10 CN 122342008 A
[0046] Figure 18 shows a bar chart of the distribution of the number of missing features for each combination in the data of Figures 17 to 18 (i.e., each bar represents the number of combinations in the MS dataset with the number of missing features). There are 33,313 combinations with a maximum number of missing features of 2 (i.e., 0, 1, or 2 missing features).
[0047] Figure 19 is a bar chart showing the predicted synergistic probability for each of the 33,313 combinations highlighted in Figure 18, as well as the cutoff point on the predicted probability (dashed vertical line) and the percentage of combinations predicted as synergistic. The preceding predicted pairs are listed in the illustration, while MS Disease Modifying Therapy (DMT) is listed in the first column and in the second column.Additional therapies are listed, and the probabilities predicted by the model are listed in the third column.
[0048] Figure 20 shows the feature importance for one of the previously predicted combinations in the data on Figure 19. Each row represents a feature, and the y-axis labels indicate the feature and the value of the feature for a particular drug combination. Shaded values are the average contribution of the feature to the probability predicted by the model for that pair (Shapley additive interpretation), and the box plot shows the distribution of contributions obtained across multiple possible rankings used to calculate the Shapley attribution.
[0049] Where the accompanying drawings illustrate embodiments of the invention, these drawings should not be construed as limiting the scope of the invention. Where appropriate, similar reference numerals will be used in different drawings to refer to the same structural features of the illustrated embodiments. Detailed Description
[0050] Specific embodiments of the invention will now be described with reference to the accompanying drawings.
[0051] This disclosure describes a method for using a machine learning model to predict the effects of drug combinations, which takes as input values of multiple predictive features of the drug combination.
[0052] In the context of this disclosure, "drug" means any active substance used in medicine. A drug may also be referred to as a "therapeutic agent". This specifically includes small molecules and macromolecules (e.g., antibodies). A drug can advantageously be any active substance that modulates the activity of one or more targets. A target can be a gene or its expression product (e.g., a protein). In other words, a drug can be any active substance that can be associated with one or more targets, wherein the drug modifies the activity of the target(s). The term is not limited in terms of how the activity of the target is modulated, whether the modulation is direct or indirect, or whether the modulation is positive or negative. Furthermore, drugs without known targets can also be included if sufficient predictive features about the drug can be determined. As explained below, predictive features can include features associated with drug targets, features associated with drug indications, and features associated with drug structure. Therefore, for drugs without known targets, predictive features can still be available, and such drugs can still be used for training and / or prediction using machine learning as described herein when more than half or more two-thirds of the predictive features used (e.g., five or more of eight or nine predictive features, preferably six or more of eight or nine predictive features) can be determined. To train a machine learning model as described herein, it is advantageous to use a training dataset that includes most drug combinations (e.g., at least 50%, at least 60%, at least 70%, at least 80%, or at least 90%), where both drugs have at least one known target.
[0053] A drug combination refers to the administration of a pair of drugs to a subject. Therefore, predicting the role of drug combinations refers to predicting the drug...The therapeutic effect that the combination will have when administered to a subject. This may include classifying the drug combination among multiple categories, including one or more synergistic categories and one or more non-synergistic categories, as will be further described below. Simply distinguishing between synergistic and non-synergistic drug combinations may be particularly advantageous. In fact, this produces a simple classification that is easier to adapt and easier to assign labels for training purposes. This further reflects the clinical reality that drug combinations are more likely to be considered promising when they have a synergistic effect, as this advantageously balances the additional risks and side effects associated with using a pair of drugs rather than a single drug. This disclosure relates primarily to methods for predicting the effect that a drug combination will have when administered to a subject, but generally does not include any steps of administering the drug or drug combination to a subject. However, the prediction is not limited to any mode of administration, nor is it limited to concurrent or subsequent administration of the drugs. In fact, the prediction is based purely on the properties inherent in the function and structure of the drugs. The determination of formulations and treatment regimens for use in administering drug combinations is beyond the scope of this disclosure. These drug combinations have been selected as potentially synergistic using the methods described herein and have been experimentally validated.
[0054] The term "machine learning model" refers to a mathematical model that has been trained to predict one or more output values (the predicted variables) based on input data (the values of the predictor variables, also known as predictor features). Training refers to the process of learning the parameters of a mathematical model using training data, which produce a model that can predict output values that satisfy one or more optimality criteria. In the case of supervised learning, training typically refers to the process of learning the parameters of a mathematical model using training data, which produce a model that can predict output values with minimal error compared to (known) values associated with the training data (where these comparison values are often referred to as "labels" or "benchmark realities"). There are two main types of supervised learning models: classification models and regression models. The purpose of a classification model is to classify observations among multiple categories. Classification models are generally suitable when the output to be predicted is a category label in the training data (e.g., whether the drug combination is synergistic or not). The purpose of a regression model is to predict the value of a continuous variable associated with an observation (e.g., a drug combination has a continuous, quantifiable effect on the outcome variable). Regression models are generally suitable when the output to be predicted is a continuous value in the training data. Classification models can provide a classification label and / or one or more probabilities as outputs for an observation belonging to one or more corresponding categories. Binary classification models can provide a single probability as output, indicating the probability that the observation belongs to the positive category (e.g., a synergistic drug combination). A predetermined threshold can be applied to one or more categories.The rate is assigned to a category label. This threshold can be determined based on the expected characteristics of the classification, such as specificity at the expected level, sensitivity, or any accuracy performance that combines aspects of specificity (precision) and sensitivity (recall) (such as accuracy and F1 score, or a balanced form thereof, taking into account the proportion of observations in the training data belonging to each category in the category).
[0055] Embodiments of this disclosure utilize a classification model trained using training data, which includes drug pairs and associated labels that identify drug combinations as belonging to one of a plurality of categories, including at least “synergistic” categories. The plurality of categories may include synergistic drug combinations of a first category and non-synergistic drug combinations of a second category. When using a binary classification model (i.e., the plurality of categories consists of the synergistic and non-synergistic categories), the latter may also be referred to as “antagonistic” drug combinations, even if they may not be strictly antagonistic (i.e., their combined effect may not be worse than their individual effects, but their combined effect is not better than the sum of their individual effects). A synergistic drug combination is a combination of drugs in which the combination is better than the sum of the individual effects of each of the drugs on at least one clinical endpoint. The term "better" refers to an improvement in the endpoint, which can be a reduction or an increase, depending on the circumstances (i.e., whether the endpoint is reduced by treatment (e.g., the presence and / or extent of lesions) or increased by treatment (e.g., motor function score)). A non-synergistic drug combination is a combination of drugs in which the combination is no better than the sum of the individual effects of each drug in the drug for at least one clinical endpoint, and / or the combination's side effects are as bad as or worse than the sum of the individual side effects of each drug in the drug. Additional categories and / or subsets of these categories may be used. For example, multiple categories may include synergistic drug combinations of category one, antagonistic drug combinations of category two, and additional drug combinations of category three. In such embodiments, an antagonistic drug combination may refer to a combination of drugs in which the combination is worse than the sum of the individual effects of each drug in the drug for at least one clinical endpoint, and / or the combination's side effects are worse than the sum of the individual side effects of each drug in the drug. Additional drug combinations can refer to drug combinations in which the effect of the combination on at least one clinical endpoint is equal to the sum of the individual effects of each drug in the drug, and the side effects of the drug combination are not worse than the sum of the individual side effects of each drug in the drug. Training data including drug pairs and associated markers are available from multiple online resources such as, for example, the Drug Comb database (Liu et al., 2020;The association (drugcombdb.denglab.org / ) identifies drug combinations as belonging to one of several categories that include at least the "synergistic" category. Synergistic drug combinations are documented in the FDA publication "Approved Drug Products with Therapeutic Equivalence Evaluations," commonly known as the "Orange Book" (FDA, 2023). Therefore, synergistic drug combinations used to train machine learning models as described herein are available from the FDA Orange Book. Such data can be downloaded from the DrugComb database (Liu et al. 2020). Non-synergistic drug combination data are available in DrugBank (Wishart et al. 2017) and can be downloaded from the DrugComb database.
[0056] The term "machine learning algorithm" or "machine learning method" refers to an algorithm or method for training and / or deploying a machine learning model. The machine learning model of this disclosure is trained through supervised learning. The optimality criterion can be the minimization of a loss function that quantifies the model prediction error based on the observed (benchmark true) values and predicted values of the predicted variables. Suitable loss functions for training machine learning models are known in the art and include mean squared error (MSE) and mean absolute error (MAE). Either of these can be used according to this disclosure. The mean squared error (MSE) can be expressed as , where and are the predicted and observed (benchmark true) values of the data instance (in this example, the candidate drug combination in the training data), respectively. The mean absolute error (MAE) can be expressed as: , where and are the predicted and observed (benchmark true) values of the data instance (in this example, the candidate drug combination in the training data), respectively. MAE is considered more robust to outlier observations than MSE. MAE can also be referred to as the "L1 loss function." However, MSE remains a very commonly used loss function, especially when strong effects from outliers are not expected, as it can make the optimization problem easier to solve. A regularized loss function is a function that includes the loss function described above plus one or more terms that penalize model complexity to reduce the risk of overfitting. Overfitting occurs when a machine learning model is trained to reproduce the features of the training dataset very closely, resulting in poor performance (i.e., poor generalization) on other datasets that do not have the same characteristics. L1 regularization (also known as "Lasso" in the context of regression) adds a regularization term to the loss function, penalizing the model based on the sum of the absolute values of its coefficients. L2 regularization (also known as "ridge" in the context of regression) adds a regularization term to the loss function, penalizing the model based on the sum of the squared values of its coefficients.Penalized models. L1 regularization can be used as a feature selection method because it minimizes the coefficients associated with predictive features that have little information. In this embodiment, the machine learning model is a regularized model. In this embodiment, the machine learning model is a regularized tree-based model. In this embodiment, the machine learning model is a regularized gradient boosting decision tree model. Examples of such models are available in the XGBoost software library (xgboost.ai / ). Such models can be referred to as "XGBoost" models, although any other implementation of a regularized gradient boosting model can be used equivalently.
[0057] The machine learning models described herein may be selected from: decision trees and their variants, including regularized and / or gradient boosting decision trees and random forest models; regularized discriminant analysis; logistic regression models; artificial neural networks (ANNs), including multilayer perceptrons (with linear or nonlinear activation functions) and deep learning models (e.g., long short-term memory networks (LSTM), recurrent neural networks (RNN), generative adversarial networks (GAN), etc.); Naive Bayes classifiers; support vector machines (SVMs, using linear or nonlinear kernels, such as radial basis functions); and multivariate adaptive regression splines (MARS). The inventors have found it particularly advantageous to use nonlinear models, such as decision trees and their variants (particularly including random forests and gradient boosting trees), SVMs with nonlinear kernels, and ANNs with nonlinear activation functions (e.g., multilayer perceptrons, page 9 / 37 of the specification, CN 122342008 A). In this particular study, relatively simple ANNs (such as MLPs) are preferred because more complex deep learning models (such as LSTM, RNN, and GAN) typically have more trainable parameters and therefore perform best when trained with large amounts of training data.
[0058] In embodiments, the machine learning model includes an ensemble of models whose predictions are combined. Alternatively, the machine learning model may include a single model. Random forest models and gradient boosting tree models (such as XGBoost) are ensemble models. Any model of ensemble form can be constructed. Ensemble models are expected to result in better prediction performance than single models and are therefore preferred in the context of the methods described herein. For example, the machine learning model may be a random forest classifier or a gradient boosting decision tree model. A random forest classifier is a model that includes an ensemble of decision trees and outputs the class as the average prediction of the individual trees. A decision tree recursively partitions the feature space until each leaf (the final set of partitions) is associated with a single value of the target. Gradient boosting is a machine learning method that forms an ensemble of weak prediction models (such as decision trees) from which a combined strong prediction is obtained. The algorithm iteratively adds new weak predictors to improve the output by combining the weak predictors.The predictions obtained. In contrast, random forests use a random subset of training data to iteratively train a set number of trees.
[0059] The method described herein predicts the effect of a drug combination based on the values of one or more predictive features. The predictive features (also simply referred to herein as “features”) are properties of the drug combination that can be quantified and characterize the properties of the drug combination other than its effect. The effect of the drug combination is the predicted feature of the machine learning model according to the present disclosure. The predictive features according to the present disclosure may include one or more features (indication similarity features) indicating the similarity between indications of the drugs in the combination. The predictive features according to the present disclosure may include one or more features (target similarity features) indicating the similarity between targets of the drugs in the combination. The predictive features according to the present disclosure may include one or more features (drug similarity features) indicating the similarity between the drugs in the combination.
[0060] Indication similarity may include one or more indication similarity measures based on the drug(s) targets(s) and / or one or more indication similarity measures based on the drug(s). In embodiments, one or more predictive features used by the machine learning model include at least one indication similarity measure based on the drug(s) targets(s). Drug target-based indication similarity metrics can be a similarity measure that quantifies the similarity between medical annotations associated with a drug target. For example, medical subject headings (MeSH or MESH – see www.nlm.nih.gov / mesh / meshhome.html) annotations can be associated with genes / proteins, as described in Yu 2018. Specifically, genes can be associated with MESH terms through text mining, for example by pairing entries from the online Human Mendelian Inheritance (OMIM) database with MESH terms, as described in Nakazato et al. 2009 (with gene-level MESH annotations available at gendoo.dbcls.jp), and / or by using annotations between Entrez gene identifiers and PubMed articles in the gene2pubmed data (Maglott et al. 2007; available in the “Related articles in PubMed” section of any entry in the Entrez gene database www.ncbi.nlm.nih.gov / gene / and in bulk downloads on the NGBI website), where PubMed articles are manually selected and associated with MESH terms (since the MESH vocabulary is the primary means of categorizing PubMed content).(Developed for this purpose). Sequence similarity tools can also be used to provide annotations for genes lacking annotations; for example, bidirectional Blast Hits (RBBH) can be used to assign annotations to the most similar genes across different genomes. This is particularly useful for annotating genes from one species to orthologous genes in another. The annotations used in this example utilize both annotations obtained by pairing entries from OMIM with MESH terms and annotations obtained from gene2pubmed data. Therefore, in embodiments, a drug target-based indication similarity measure can be a similarity measure that quantifies the similarity between medical annotations associated with the drug target, provided as MESH term annotations associated with the gene identifier of the target, where the MESH term annotations are annotations obtained by pairing entries from OMIM with MESH terms and / or annotations obtained from gene2pubmed data.
[0061] Drug-based indication similarity metrics can be similarity metrics that quantify the similarity between indication annotations directly associated with a drug. For example, an ATC code is an indication annotation associated with a drug. Similarity between ATC codes can be a quantification of the similarity between complete ATC codes (or at least multiple levels of ATC codes) or between first levels of ATC classification. The similarity between complete ATC codes can be obtained using any similarity metric suitable for comparing discrete sets (e.g., Jaccard similarity, as described below) or any similarity metric suitable for comparing discrete sets organized as an ontology (e.g., the Wang, Resnik, Relevance, Jiang, and Lin methods, as described below). It should be noted that each drug is assigned a unique ATC code, and therefore when determining the similarity between complete ATC codes, it is not necessary to aggregate similarity scores for multiple annotations. In other words, each drug is associated with a single code, which fully defines the drug's position in the ATC ontology. The similarity between complete ATC codes can be obtained using a similarity metric between discrete sets (e.g., Jaccard similarity), where the discrete set for a drug is the set of all possible subsets of the drug's ATC code (starting from the first-level code and adding one character for each subset). For example, the code L04AA36 can be represented by the set [L, L0, L04, L04A, L04AA, L04AA3, L04AA36]. The similarity between the first levels of ATC classification can be obtained as a binary feature.This indicates whether drugs forming part of a combination are associated with ATC codes belonging to the same first-level category or to a specific pair of ATC categories. Table 1 lists the ATC classifications for first-level categories. For example, a pair of drugs both belonging to first-level category L can be associated with a similarity of "1," while a pair of drugs belonging to different first-level categories (e.g., L and N, respectively) can be associated with a similarity of "0." As another example, a pair of drugs both belonging to first-level category L (a more informative category pair) can be associated with a similarity of "1," while a pair of drugs belonging to any other category or combination of categories can be associated with a similarity of "0." It should be noted that a more informative category pair can be any first-level category pair. For example, it might be relevant if the combination includes drugs belonging to category J and drugs belonging to category S. The inventors have found that identical category pairs are more likely to be a more informative feature, and therefore, advantageously, this similarity feature indicates whether both drugs in the combination fall into the same specific first-level category. Furthermore, the inventors have discovered that certain class pairs are particularly informative, such as LL, NN, JJ, and CC. Each combination of first-level classes that contributes to the prediction can be referred to as an "informative class pair." The similarity between the first levels of ATC classification can be obtained as a set of binary features (i.e., binary vectors) indicating whether a drug forming part of a combination is associated with each of the multiple categories defined by a specific combination of first-level classes. For example, each of the aforementioned informative class pairs (or any class pair found to be informative for a particular training dataset) can be represented. Alternatively, each possible class pair can be represented. Thus, the similarity between the first levels of ATC classification can be obtained as a set of binary features including "0" or "1" at each of multiple positions in the binary vector corresponding to a particular class pair. Note that since a drug is assigned a single ATC code, such a binary vector can only contain a single "1". Class pairs can include one or more of LL, NN, JJ, CC, JS, DR, DS, CJ, AA, AM, AN, and BN. Categories including at least LL and NN are advantageous.
[0062] Drug similarity features may include one or more features selected from: side effect similarity, mode of action similarity, and molecular structure similarity.
[0063] Side effect similarity is the similarity of side effect terms compared to those associated with drugs in a drug combination. Side effects are documented using recognized terms from databases such as the FDA Orange Book. A list of side effects associated with the first drug and those associated with the second drug can be compared using any similarity metric known in the art for comparing discrete sets.The list of side effects is compared. For example, the similarity between side effects associated with two drugs can be obtained as a Jaccard similarity (page 11 / 37 of the specification, CN 122342008 A). Note that in embodiments where side effects are specified according to a controlled vocabulary organized as an ontology, any similarity measure designed for the ontology can be used, as further explained below.
[0064] The mode of action similarity is the similarity of the mode of action terms associated with the drugs in the drug combination. Mode of action terms are recorded using recognized terms from databases such as DrugBank. The list of mode of action terms associated with the first drug and the list of mode of action terms associated with the second drug can be compared using any similarity measure known in the art for comparing discrete sets. For example, the similarity between mode of action terms associated with two drugs can be obtained as a Jaccard similarity. Note that in embodiments where the mode of action of the drugs is specified according to a controlled vocabulary organized as an ontology, any similarity measure designed for the ontology can be used, as further explained below.
[0065] Molecular structural similarity is the similarity of the 2D and / or 3D structures of drugs in a drug combination. The 2D structure of a molecule can be represented as a fingerprint. This can be a numerical vector (also known as a binary fingerprint) indicating the presence of a specific substructure in the molecule. Any method known in the art for obtaining molecular fingerprints can be used, such as, for example, the extended connectivity fingerprint ECFP4 (described in Rogers and Hahn (2010)). Any similarity metric known in the art for comparing molecular fingerprints can be used, including, for example: Tanimoto similarity (i.e., Jaccard similarity), Dice similarity, modified Tanimoto, cosine similarity (see Equation (0), Ochiai (the ratio of the square root of the product of the number of features shared by two molecules and the number of features in the two molecules), Kulczynski2 (the ratio of the product of the number of features shared by two molecules multiplied by the average size of the total number of features present divided by the size of the total number of features present (i.e., ((A+B) * 0.5(A+B)) / (A*B))).
[0066] Target similarity features may include one or more of the following: pathway similarity, target protein similarity, disease association similarity, co-expression, and proximity on a protein-protein interaction network (PPI). In embodiments, target similarity features include one or more of the following: pathway similarity, disease association similarity, and proximity on a protein-protein interaction network (PPI). It should be noted that indication similarity is based on the drug's (one or more) targets.Degree metrics can also be viewed as target similarity features, but are classified as indication similarity features because their information content is found to be more similar to drug-related indication similarity metrics.
[0067] Disease association similarity is the similarity of comparisons between disease-related annotations associated with any known target of the drug. Disease-related annotations may be in the form of disease ontology terms. Disease ontology terminology annotations for genes / proteins may be available, for example, from disease and gene annotation databases (DGA, dga.nubic.northwestern.edu; Peng et al. 2013). Any similarity metric described below may be used. Preferably, the similarity metric is an ontology-designed metric (also referred to as semantic similarity in this context), as further described below.
[0068] Pathway similarity is the similarity of comparisons between pathway annotations associated with any known target of the drug. Pathway annotations may come from any pathway annotation database, including Kegg (www.genome.jp / kegg / pathway.html), Reactome (reactome.org / ), and gene ontology biological process annotations (see geneontology.org / ). Reactome and Kegg provide pathway maps indicating genes involved in known biological pathways. The inventors tested each of these pathway information sources and found they produce highly correlated predictive features, indicating that any of them can be used interchangeably. In this embodiment, pathway annotations are obtained from Kegg. Each drug may be associated with one or more drug targets. Each drug target may be associated with one or more pathways. Therefore, each drug may be associated with a set of pathways (including all pathways associated with any of the drug's targets). Pathways associated with the first drug and pathways associated with the second drug can be compared by calculating one or more pairwise similarities (including the similarity between each pair of pathways associated with the first drug and the pathway associated with the second drug). The similarity between two pathways can be quantified by comparing the sets of genes involved in the two pathways, using any metric suitable for discrete sets. The similarity between two pathway sets can be quantified by combining such similarities using any method known in the art for combining multiple pairwise similarities. Examples of such methods are further provided below.
[0069] Target protein similarity is a similarity measure that compares the targets of two drugs in a pair. In this embodiment, it is a score between 0 and 1, which is applied to two lists (each including the targets of the first drug and the second drug, respectively).The similarity between the targets is quantified. Any similarity measure applicable to discrete sets, such as, for example, Jaccard similarity, can be used for this purpose.
[0070] Co-expression similarity is the comparison of the similarity of expression profiles across multiple samples or conditions (including, for example, cell lines, organs, etc.) for each pair of genes in a combination of drugs, including drug targets of the first drug and gene targets of the second drug. For example, all pairwise comparisons between the targets of the first drug and the targets of the second drug can be performed (i.e., a similarity measure can be obtained for each such comparison), and one or more similarities obtained can be combined using any of the methods described herein (where the total number depends on the number of drug targets). The similarity between expression profiles can be calculated using any measure known in the art, such as, for example, Pearson correlation, Spearman rank correlation, cosine similarity, etc. The similarity measure can compare the average expression of all targets of drug 1 in each of multiple tissues and the average expression of all targets of drug 2 in each of multiple tissues.
[0071] Proximity on a protein-protein interaction network is a comparison of the similarity of subnetworks associated with the drug targets of the first and second drugs in a drug combination, wherein the subnetworks are subsets of a protein-protein interaction (PPI) network. A PPI network is a network comprising nodes corresponding to genes / proteins and edges corresponding to functional and / or physical interactions between genes / proteins associated with nodes connected to the edges. Any PPI network known in the art can be used, including any PPI network that combines information from multiple PPI networks. For example, a PPI network can be selected from STRING (e.g., STRING v11, which is available at string-db.org / and described in Szklarcyk et al. 2019) or any subset thereof, including interactions of pathways from Kegg (Kanehisa et al. 2017), Reactome (Fabregat et al. 2016), BioCyc (Caspi et al. 2016), and / or Gene Ontology (The Gene Ontology, 2017). PPIs can be functional interactions. Functional interactions can include direct interactions within pathways and / or protein complex membership. A drug-associated subnetwork can be obtained as a minimal subnetwork encompassing all targets of the drug. This can include all nodes of the PPI that are drug targets, as well as all nodes on the shortest paths (or multiple shortest paths, if multiple shortest paths exist) between nodes that are drug targets. Other criteria can be used to construct the subnetwork, such as including, for example, nodes corresponding to drug targets.All nodes and all nodes within a certain number of edges from the drug target. A drug-associated subnetwork can be a set of nodes that are targets of the drug. Proximity between such subnetworks can be obtained as a score quantifying the distance and / or overlap between subnetworks. In cases where multiple disjointed subnetworks exist for at least one drug in the drug combination, a pairwise proximity metric can be calculated for all subnetwork pairs, including the subnetwork associated with the first drug in the drug combination and the subnetwork associated with the second drug. The resulting multiple proximity metrics can be combined using any method known in the art for combining similarities or using any aggregation metric known in the art. Proximity between subnetworks that are sets of target nodes for the corresponding drugs can be the average (or maximum or minimum) distance between the drug targets of the two drugs. The distance between two drug targets can be the length of the minimum length path identified between the two drug targets. The distance between two drugs can be the average, minimum, or maximum of such distances across all target pairs including targets of drug 1 and targets of drug 2. The average can conveniently be used as a robust metric. Specification 13 / 37 pages 17 CN 122342008 A
[0072] Any annotation, whether associated with a drug or a drug target, may take the form of one or more terms in a controlled vocabulary. The controlled vocabulary may be organized as an ontology. In this context, an ontology is a representation of terms that form part of a controlled vocabulary and the relationships between them. In the context of this disclosure, an ontology may be a hierarchical ontology. This may include multiple hierarchical levels, where each level includes multiple categories, each of which may each include one or more subcategories at the next level. Thus, a controlled vocabulary may be represented as a directed acyclic graph, where each node of the graph corresponds to a term in the controlled vocabulary, and edges indicate parent-child relationships between nodes.
[0073] Examples of controlled vocabularies organized as hierarchical ontologies include medical subject terms, gene ontology (GO, geneontology.org / ), and disease ontology (DO, disease-ontology.org / ). Additionally, the Anatomical Therapeutic Chemistry (ATC) classification system (see www.who.int / tools / atc-ddd-toolkit / atc-classification) is also organized as a hierarchical classification system and is therefore a hierarchical ontology that can be analyzed using any method designed to compare annotations organized as hierarchical ontology.
[0074] The terms “similarity measure” and “similarity score” (or sometimes simply “score”) are used interchangeably to refer to a value that quantifies the degree to which two values, annotations, or sets of values / annotations are similar to each other. The terms “annotation” and “the term” are used interchangeably, andAnd refers to words / vocabularies related to drugs or drug targets. These are typically selected from controlled vocabularies, such as MESH, GO, DO as described herein.
[0075] The measure of similarity between annotations may include any similarity measure known in the art that is applicable to discrete sets in general or to discrete sets organized as ontologies. The measure of similarity between annotations may be a score between 0 and 1. Similarity measures applicable to discrete sets include the Jaccard index (also known as Jaccard similarity), Dice similarity, cosine similarity, Ochiai similarity, and the Kulczynski2 measure. Dice similarity is a score between 0 and 1, which is calculated as twice the number of elements shared between the two sets divided by the number of elements in the two sets (the number of unique elements across the two sets), see Equation (16). Jaccard similarity is the ratio of the number of features shared by the two sets to the total number of features in the two sets, i.e., (A intersects B) / (A + B - (A intersects B)). The cosine similarity between two vectors is provided as equation (0), where, when comparing two sets of annotations, each set can be expressed as a vector of length equal to the total number of unique elements across the two sets, where a "1" or "0" for each element depends on whether the element exists in or not in a particular set. The Ochiai similarity measure (also known as Carbo similarity or cosine coefficient) is calculated as the ratio of the square root of the product of the number of features shared by the two molecules to the number of features in the two molecules. The Kulczynski2 metric is calculated as the ratio of the product of (the number of features shared by the two molecules multiplied by the average size of the total number of features present) divided by the size of the total number of features present (i.e., ((A * B) * 0.5(A + B)) / (A * B)).
[0076] Similarity measures designed for ontology take into account the fact that two terms to be compared may be different but can be related to each other through the ontology. In other words, similarity measures designed for ontology compare terms by taking into account the relation of the terms in the ontology. Similarity measures designed for ontology (where the ontology specifies a controlled vocabulary, as is the case for ontology such as DO, GO, and MESH) can also be called “semantic similarity measures” because they take into account the meaning of the terms to be compared (in the form of their relationship within the ontology). Similarity measures applicable to annotations that form parts of a hierarchical ontology include the Wang, Resnik, Relevance, Jiang, and Lin similarity measures. The Wang similarity measure is described in Wang et al. 2007 and is computed using equations (1), (2), (3), (4a), (4b), (5a), and (5b) provided below.The method is described in Resnick 1999. It uses equation (6) to calculate the semantic similarity between two concepts. The Lin method is described in Lin 1998. The Lin similarity measure is a node-based semantic similarity measure based on the information content of the least common subsumer. This method is calculated using equations (7), (8), and (9). The Relevance method is described in the specification, page 14 / 37, 18 CN 122342008 A Schlicker et al. 2006. This method combines the methods of Resnick and Lin and is calculated using equation (10). The Jiang method is described in Jiang and Conrath 1997. This method is calculated using equation (11). Each of the above measures compares pairs of ontology terms. In cases where multiple ontology terms are associated with one or both sets of two sets to be compared, any of the measures described above can be used to compare multiple pairs of ontology terms, thereby generating multiple similarities between pairs of annotations for the corresponding drugs in the combination.
[0077] Multiple similarities between multiple pairs for a drug combination (each pair comprising an entry associated with one drug in the combination and an entry associated with another drug, where the entry can be an annotation, a pathway can be represented as a set of genes, etc.) can be combined using any method known in the art. Methods known in the art for combining multiple pairwise similarities obtained by comparing two sets of entries include any method selected from the maximum value method, the average value method, the rcmax method, and the best-match average value method. The maximum value method calculates the maximum similarity score relative to all term pairs (see Equation 12). The average value (avg) method uses the average of the similarity scores relative to all term pairs (see Equation 13). The rcmax method calculates the maximum of the following: (i) the average of the maximum similarity on each row; and (ii) the average of the maximum similarity on each column (see Equation 14), where rows and columns refer to the rows and columns of a matrix, where terms annotated to one drug correspond to rows and terms annotated to another drug correspond to columns, and the values in the matrix are the similarity between the corresponding terms. The Best Match Average (BMA) method measures the average of the maximum similarity scores on each row and each column (see Equation 15), where rows and columns are as above. In embodiments, any of the methods described above, such as the BMA method, is used to combine multiple similarity measures obtained for corresponding pairs of target and / or ontology terms for a pair of drugs.
[0078] Multiple predicted features for a drug combination can also be combined by obtaining a summary value of multiple predicted features.Throughout this disclosure, a “summary value” can be any statistic that summarizes a set of values relative to multiple values, such as, for example, a sum, mean, median, or predetermined percentile (e.g., the 1st, 2nd, 5th, 10th, 15th, 85th, 90th, 95th, 98th, or 99th percentile). A summary value can include multiple values, provided that the dimension of the summary value is lower than the dimension of the values it summarizes. For example, a summary value can be obtained by dimensionality reduction (e.g., using PCA). In an embodiment, there are no summarized predictive features, and each predictive feature (optionally preprocessed, such as by normalization) is used individually as input to a machine learning model.
[0079] Any set of predictive features can be preprocessed before being provided as input to a machine learning model. Preprocessing can include one or more of the following: filtering, logarithmic transformation, imputation, standardization, and normalization. Imputation refers to the process of providing values for missing data. Any imputation method known in the art can be used, such as, for example, mean substitution (replacing missing values with the mean of the values of features in a dataset (e.g., a training dataset), regression imputation (where a regression model is estimated to predict observed values of a variable based on other variables, and once trained, the model can be used to predict missing values of the variable), etc. Filtering can be based on, and may include, one or more of, the following: removing duplicate entries (e.g., removing from the training or test dataset), removing entries with a certain number or proportion of missing values above a predetermined threshold (e.g., removing any entry that is missing 3 or more of the predicted features before any imputation), etc. Standardization refers to scaling the values of a feature based on a measure of the variability of the features in a reference dataset (typically the standard deviation). Normalization can be performed using any method known in the art, including mean centering, obtaining z-scores (also known as standard scores), minimum-maximum scaling (also known as cell-based normalization, in which a transformer value for each value of a feature is obtained in the form of (x-xmin) / (xmax-xmin), where x is the value of the feature, and xmin and xmax are the minimum and maximum values observed for the feature in a reference dataset). The reference dataset can be a training dataset or dataset to which the data points to be normalized belong.
[0080] Drug combinations selected using the methods described herein can be validated using one or more assays. Specification 15 / 37 pages 19 CN 122342008 A Assays can be in vitro assays or in vivo assays. Assays can be in vitro assays. In vitro assays can be cell-based assays or tissue-based assays. Cell-based assays can be cytotoxicity assays, cell proliferation assays, or measurements of target cells.Any assay of a phenotype (motility, viability, proliferation, expression, etc.). Tissue-based assays (including tissue culture and organoid-based assays) can be any assay for toxicity or to measure a target tissue phenotype (such as, for example, tissue composition, structure, lesions, etc.). Readouts of in vitro cell- or tissue-based assays can be obtained using any bioanalytical instrument known in the art (such as, for example, microscopy, flow cytometry, etc.) or any omics protocol known in the art. In vivo assays can characterize any target phenotype or physiological characteristic, such as, for example, tumor growth, viability, physiological function (e.g., function of an organ or system, etc.). Readouts of in vivo assays can include physiological measurements, phenotypic measurements (including, for example, behavioral measurements), and omics measurements (e.g., genomics, transcriptomics, proteomics, microbiome, metabolomics characterization of a subject).
[0081] Figure 1 is a flowchart illustrating a method for predicting the effects of a drug combination according to a general embodiment of the present disclosure. At step 110, a list of drugs or drug combinations comprising a first drug and a second drug is obtained. When a list of drugs is obtained, step 110 may further include defining drug pairs from the list. At step 112, values for multiple features associated with each drug in the combination are obtained. Obtaining values for features associated with the drug combination may include, for each of the multiple features, receiving the value from a database, user interface, or memory, or calculating the value based on the characteristics of the first and second drugs. Therefore, step 110 may further include, for each drug in the drug combination to be evaluated, obtaining multiple features from which values for the multiple features may be calculated. The multiple features may include one or more of the following: indication annotation associated with the drug, one or more drug targets of the drug, indication annotation associated with one or more targets of the drug, disease annotation associated with one or more targets of the drug, molecular structure of the drug or information representing the molecular structure of the drug (e.g., molecular fingerprint), side effect annotation associated with the drug, and mode of action annotation associated with the drug. Alternatively, when the drug combination is obtained at step 110, values for features associated with the respective drug combination may also be obtained at step 110. The multiple features include at least one feature that characterizes the similarity between one or more indications or diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug. The multiple features may advantageously include at least one feature that characterizes the similarity between one or more indications and / or diseases associated with one or more targets of the first drug and one or more indications and / or diseases associated with one or more targets of the second drug. Indications may be provided as medical subject heading terms associated with one or more targets of the drug.A disease can be provided as a disease ontology term associated with one or more targets of a drug. Multiple features may include: one or more drug similarity features, one or more target similarity features, and one or more indication similarity features, as described herein. For example, multiple features may include: at least one feature characterizing the similarity between one or more indications associated with one or more targets of the first drug and one or more indications associated with one or more targets of the second drug, wherein the indications are provided as medical subject heading terms associated with one or more targets of the drug; side effect similarity; mode of action similarity; molecular structure similarity; pathway similarity; disease association similarity characterizing the similarity between one or more diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug; and a measure of the proximity of the targets of the first drug and the targets of the second drug on a protein-protein interaction network; and a drug-based indication similarity measure characterizing the similarity between indication annotations associated with the first drug and the second drug, wherein the indication annotations are provided as Anatomical Therapeutic Chemistry (ATC) codes, and wherein the drug-based indication similarity measure includes a similarity measure that quantifies the similarity between multiple levels of ATC codes for the first drug and the second drug. The indication-based similarity measure for drugs may further include a similarity measure comparing the first level of the ATC codes for the first drug and the second drug at page 16 / 37 of the instruction manual, 20 CN 122342008 A, wherein the similarity measure comparing the first level of the ATC codes for the first drug and the second drug is a set of binary features indicating whether the first drug and the second drug belong to each of multiple pairs of first-level ATC categories, optionally wherein the multiple pairs of first-level ATC categories include LL, NN and optionally one or more of JJ, CC, JS, DR, DS, CJ, AA, AM, AN and BN.
[0082] At step 114, the values obtained at step 11 may be preprocessed. This may include one or more of the following: missing value imputation, filtering, standardization and normalization. Missing value imputation may use any algorithm known in the art, such as, for example, KNN. The missing value imputation process may use information from training and / or testing datasets that have been used to train the model. Filtering can exclude combinations with a certain proportion of missing features exceeding a predetermined threshold (e.g., more than 2 missing features out of 8 or 9 predicted features). Standardization and normalization can utilize information from the training and / or test datasets already used to train the model. For example, the mean and standard deviation of features across the training and / or test datasets can be used to...The values of features in the combinations whose effects are to be predicted are normalized and standardized.
[0083] At step 116, the values obtained at step 112 or step 114 are used to predict the effect of at least one drug combination obtained at step 110. This uses a machine learning model trained with training data to predict whether a drug combination is likely to be synergistic, the training data for each of a plurality of training drug combinations including: (i) the values of the plurality of features; and (ii) an indication of whether the drug combination is synergistic. The machine learning model may have been trained to classify drug combinations among a plurality of categories including synergistic and non-synergistic categories, and step 116 may include determining whether a drug combination is likely to belong to the synergistic category.
[0084] At step 118, one or more of the plurality of combinations whose effects were predicted at step 116 are selected for further validation. For example, the top x predicted combinations (the top x combinations with the highest probability of synergy) may be selected. Alternatively, one or more combinations may be selected from the top x combinations. For example, factors such as prior knowledge and availability may be used to select drug combinations from the preceding predicted combinations. The value of x can depend on practical factors such as testing capability. Similarly, predicted combinations or subsets of such combinations with a synergistic probability higher than a threshold y can be selected. For example, factors such as prior knowledge and availability can be used to select drug combinations from the preceding predicted combinations with probabilities higher than the threshold y. Again, the value of the threshold y can depend on practical factors and the probability distribution predicted for a selected set of drugs.
[0085] At step 120, the selected combination is validated using one or more in vitro and / or in vivo assays (e.g., preclinical and clinical trials). At step 122, the results of any of the preceding steps can be provided to the user.
[0086] Figure 2 is a flowchart illustrating a method for providing a prediction of the effect of a drug combination according to an embodiment of the present disclosure. Step 210 includes obtaining for each of a plurality of training drug combinations comprising a first drug and a second drug: (i) values of a plurality of features associated with the drug combination or properties of the drugs from which such values can be obtained; and (ii) information indicating whether the drug combination is synergistic. At step 212, if values for multiple features associated with the drug combination have not yet been obtained at step 210, these values are calculated. The multiple features include at least one feature characterizing the similarity between one or more indications or diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug.
[0087] At step 214, the training data obtained at steps 210 and 212 is optionally preprocessed. This may include...This includes one or more of the following: filtering, missing value imputation, normalization, and standardization. For example, data can be filtered to exclude any drug combination with more than a predetermined number or proportion of missing values. For example, drug combinations missing more than two features can be removed. The optionally filtered data can then be subjected to data imputation to obtain a complete dataset (i.e., without missing values). For example, this can be done using the KNN algorithm. The data can then be normalized (subtracting the mean across observations for each feature) and / or standardized (dividing each feature by the standard deviation across observations). Specification 17 / 37 pages 21 CN 122342008 A
[0088] At step 216, the values obtained at step 212 or 214 are used to train a machine learning model using the training data obtained at steps 210 and 212 to predict whether the drug combination is likely to be synergistic. The machine learning model is trained to take the values of the multiple features for the drug combination as input and produce an indication as output of whether the drug combination is likely to be synergistic. This may include performing feature selection (e.g., by determining the performance of each of several models trained with corresponding subsets of multiple features) and selecting the smallest model whose performance is within a predetermined tolerance of the best model. Performance can be evaluated using any metrics known in the art, such as F1 score, accuracy, balanced accuracy, precision, recall, etc. In the present context, metrics such as F1 score and balanced accuracy are considered advantageous because drug combination training data may be imbalanced (since more non-synergistic combinations are known compared to synergistic combinations).
[0089] At step 218, the trained model is provided to a user, data storage area, or computing device. This can be used in the method described with reference to FIG1.
[0090] Unless the context otherwise indicates (e.g., in the case of measurement steps and / or wet steps), the methods described herein are computer-implemented. Therefore, the methods described herein are generally performed using a computer system or computer device. Thus, any reference to an action (such as “acquire,” “process,” “determine”) may refer to a processor performing the action or a processor executing instructions that cause the processor to perform the action. In practice, the method of the present invention involves at least processing drug annotation to a large ontological or drug structural similarity, particularly on a scale of tens or hundreds of thousands of drug combinations, making it impossible to perform in human thought. As used herein, the term "computer device" as used in "computer system" includes hardware, software, and data storage devices for implementing the system or carrying out computer-implemented methods. For example, a computer system may include one or more processing units (such as a central processing unit (CPU) and / or a graphics processing unit (GPU)), input structures, etc.The computer system includes components, output components, and data storage devices, which can be implemented as one or more connected computing devices. Preferably, the computer system has a display or includes a computing device having a display to provide visual output display (e.g., in the design of a business process). The data storage may include RAM, disk drives, or other computer-readable media. The computer system may include multiple computing devices connected via a network and capable of communicating with each other on the network. For example, the computer system may be implemented as a cloud computer. The term "computer-readable media" includes, but is not limited to, any non-transitory medium or media that can be directly read and accessed by a computer or computer system. The medium may include, but is not limited to, magnetic storage media such as floppy disks, hard disk storage media, and magnetic tape; optical storage media such as optical discs or CD-ROMs; electronic storage media such as memory including RAM, ROM, and flash memory; and mixtures and combinations of the above, such as magnetic / optical storage media.
[0091] The methods described herein are applicable to a variety of situations. For example, the methods described herein can be used to determine which of a plurality of candidate drug combinations (e.g., a plurality of drug combinations comprising known disease-modifying drugs for a specific disease and additional drugs not yet clinically tested in combination with known disease-modifying drugs and / or not yet clinically tested for the treatment of a specific disease) are likely to produce synergistic effects, and / or to select one or more candidate drug combinations from a plurality of candidate drug combinations to be included in a set of drug combinations for further characterization (e.g., based on experimental validation) for the treatment of a specific disease. For example, candidate combinations predicted to produce synergistic effects can be selected. For example, a predetermined number of drug combinations can be selected as the drug combinations with the highest predicted synergistic probability. As another example, a predetermined number of drug combinations can be selected as the drug combinations from the top x (e.g., top 10, top 20) drug combinations with the highest predicted synergistic probability, for example, based on additional literature support, expert knowledge, availability for testing, etc. As another example, a predetermined number of drug combinations can be selected as drug combinations from a set of drug combinations whose predicted synergistic probability is higher than a predetermined threshold (e.g., 0.20 or higher, 0.25 or higher, 0.30 or higher, etc.), for example, based on additional literature support, expert knowledge, availability for testing, etc. As another example, all drug combinations whose predicted synergistic probability is higher than a predetermined threshold (e.g., 0.20 or higher, 0.25 or higher, 0.30 or higher, etc., as per specification page 18 / 37, 22 CN 122342008 A) can be selected for further characterization.
[0092] Because the methods can prioritize promising drug combinations, these methods can significantly reduce the cost and time associated with screening drug combinations. This is very valuable in practice because drug combinations are cumbersome and can...The number of combinations tested is exponentially increasing. However, if effective combinations can be identified, this can be very valuable because it can provide patients with improved treatments without the need to develop and characterize entirely new compounds.
[0093] Therefore, a method for screening multiple candidate drug combinations is also described herein, comprising: obtaining multiple candidate drug combinations; predicting the effect of each candidate drug combination among the multiple candidate drug combinations using methods as described herein; and using the prediction results to select a subset of the multiple candidate drug combinations for experimental testing. Similarly, a method for providing candidate drug combinations for treating a specific disease is also described herein, comprising: obtaining multiple candidate drug combinations, each candidate drug combination containing at least one drug as a known disease-modifying drug for the specific disease; and identifying at least one drug combination among the multiple candidate drug combinations as potentially having a synergistic effect in treating the specific disease using methods as described herein.
[0094] Figure 3 illustrates embodiments of a tool for predicting the effect of a drug combination, a tool for providing the effect of a drug combination, and / or a system for providing a drug combination according to the present disclosure. The system includes a computing device 1, which includes a processor 101 and a computer-readable storage device 102. In the illustrated embodiment, computing device 1 also includes a user interface 103, exemplified as a screen, but may include any other means such as conveying information to the user via sound or visual signals. Computing device 1 is communicatively connected, for example via a network, to one or more databases 2 storing drug data (values of multiple predictive features associated with the drug and, optionally, real-label data of drug combination effects). The one or more databases 2 may further store one or more of the following: one or more machine learning algorithms, training data, parameters (such as, for example, parameters of a machine learning model, feature selection algorithms, data preprocessing methods, etc.). The computing device may be a smartphone, tablet, personal computer, or other computing device. The computing device is configured to implement the methods as described herein. In an alternative embodiment, computing device 1 is configured to communicate with a remote computing device (not shown), which itself is configured to implement the methods as described herein. In such cases, the remote computing device may also be configured to send the results of the method to the computing device. Furthermore, the various steps of the methods described herein may be decoupled between computing device 1 and the remote computing device. Communication between computing device 1 and a remote computing device can be made via a wired or wireless connection, and can occur via a local or public network 6 (such as, for example, via the public Internet). Computing device 1 can also be connected to drug combination testing component 3 (such as, for example, an assay system for in vitro testing of drug combinations). The drug combination testing component can be connected to computing device 1.Wired connection, or communication may be possible via wireless connection (such as, for example, via WiFi and / or via the public Internet), as shown. The connection between computing device 1 and drug combination testing component 3 may be direct or indirect (such as, for example, via a remote computer). Drug combination testing component 3 may include any assay system configured to acquire biological data indicative of the effect of candidate drug combinations.
[0095] The following examples are presented by way of example and should not be construed as limiting the scope of the claims.
[0096] Examples
[0097] Exemplary methods of the present disclosure will now be described. In particular, the following examples demonstrate the use of the methods of the present invention to discover “high-potential drug combinations” for improving the treatment of MS. The general workflow followed in these examples is shown in Figure 4.
[0098] Materials and Methods
[0099] Data. Publicly available databases (FDA Orange Book and DrugBank) provide data on 474 synergistic drug combination products (19 / 37 pages, 23 CN 122342008 A) and 2,650 antagonistic drug combination products (Figure 4, Step 1). The FDA Orange Book and DrugBank contain data from 3,133 drug combinations, of which 474 were found to have synergistic effects and 2,659 were found to have antagonistic effects through clinical and / or preclinical / study methods. Data were downloaded from the DrugComb database (Liu et al., 2020), where it was annotated as synergistic (combinations from the FDA Orange Book) or non-synergistic (combinations from DrugBank). Synergistic drugs are those in which effect 1+2 > effect 1 + effect 2. Additional drugs (effect 1 + 2 = effect 1 + effect 2) are not considered here. Antagonistic drugs are those in which action 1 + 2 < action 1 + action 2, or side effects 1 + 2 > side effects 1 + side effects 2. The drugs in a test pair are considered synergistic if the combined effect is better than the sum of the effects of the two isolated drugs. An example of this is the combination of chlorpromazine and pentanemidine. Neither drug showed any antitumor activity when tested individually, but together they prevented tumor growth even more effectively than paclitaxel (a known cytotoxic drug). Drugs are considered antagonistic if the combined effect is not better than the sum of their isolated effects and / or their combined side effects are worse than the sum of their isolated side effects (i.e., action 1 + 2 < action 1 + action 2, or side effects 1 + 2 > side effects 1 + side effects 2). An example of this is sorafenib and pegylated interferon.α2-β, when used together, produce more skin side effects compared to using the drugs alone. Among these, five drug combinations contain an existing MS disease-modifying therapy as one member of the pair. All five pairs, along with a randomly selected set of pairs representing 20% of the data, were included in the unseen data used for model testing. For each drug pair, a feature vector containing 113 features is constructed (Figure 4, Step 2):
[0100] - Cooperative Combination: This is a binary value of 0 (no cooperation) or 1 (cooperation) (the baseline true label used for model training);
[0101] - Pathway Similarity: Score = Dice similarity calculated using the R package BioCor (Revilla Sancho L, 2023);
[0102] - MESH Similarity: Score = Semantic similarity calculated using the R package Mesh (Yu G, 2018);
[0103] - ATC Similarity: Score = Jaccard similarity;
[0104] - MOA Similarity: Score = Jaccard similarity;
[0105] - Disease Ontology Similarity: Score = Semantic similarity calculated using the R package DOSE (Yu G, 2015);
[0106] - Network Separation: Score = Minimum distance between clusters, calculated using the R package igraph To calculate (Csárdi G et al. 2023);
[0107] - Side effect similarity: overlapping side effects (intersection / union);
[0108] - Molecular structure similarity: score = Tanimoto similarity calculated using R package fingerprint (Guha R, 2018);
[0109] - Level 1 ATC: 105 features describing all combinations of Level 1 ATC encoding, one-hot encoded as binary values.
[0110] Each of these is described in more detail below.
[0111] Predictive features – Drug similarity features. Three different drug similarity features are used: molecular structure similarity, mode of action similarity, and side effect similarity.
[0112] Side effect similarity is calculated as the number of common side effects (intersection) divided by the number of unique side effects listed for either of the two drugs (union) (Jaccard similarity). Side effects were extracted from OnSIDES (tagged with side effects resource, available at tatonettilab.org / onsides-side-effects-extracted-from-fda-structured-product-labels / , this database lists FDA structured product labels from DailyMed).Adverse reactions and boxed warnings extracted from product labels; data downloaded in tabular form), and compared by exact matching to the instructions 20 / 37 pages 24 CN 122342008 A (DailyMed uses a uniform vocabulary). For example, if drug 1 has side effects “drowsiness, diarrhea, vomiting” and drug 2 has side effects “drowsiness, thirst”, the intersection = 1 (vomiting) and the union = 4 (drowsiness, diarrhea, vomiting, thirst), so the side effect similarity is ¼ (i.e. 0.25). Side effect similarity is a score between 0 and 1.
[0113] The action pattern similarity is calculated as the number of common action pattern terms (intersection) divided by the number of unique action pattern terms listed for either drug (union). The action pattern terms were extracted from DGIdb (Pharmacogene Interaction Database, Freshour et al. 2020) and compared by exact matching (action patterns are specified in DGIdb using a predefined vocabulary, see www.dgidb.org / interaction_types; each drug can be associated with one or more action pattern terms, e.g., each action pattern term is associated with a corresponding target). For example, in the case where drug 1 has the action pattern terms “inhibitor, blocker” and drug 2 has the action pattern terms “antibody, inhibitor”, the intersection = 1 (inhibitor) and the union = 3 (inhibitor, blocker, antibody), so the side effect similarity is 1 / 3 (i.e., 0.33). The action pattern similarity is a score between 0 and 1.
[0114] Molecular structure similarity is calculated as a Tanimoto similarity as implemented by R-pack fingerprinting. The Tanimoto similarity is a score between 0 and 1, which quantifies the degree to which two small molecules are similar to each other. Any molecular structure similarity measure of this type known in the art can be used instead. For example, fingerprint packages also include implementations of Dice, Modified Tanimoto, Simple, Jaccard, Cosine, and Achai similarity, among others. The similarity between two molecules can be calculated based on their corresponding fingerprints. A fingerprint can be specified as a binary vector (a vector of 0s and 1s). A fingerprint can characterize the 2D structure of a molecule (2D fingerprint). Various methods exist for determining binary molecular fingerprints. For example, the extended connectivity fingerprint ECFP4 is commonly used and described by Rogers and Hahn (2010). Any method known in the art for calculating molecular fingerprints can be used. In particular, any method known in the art for calculating binary 2D fingerprints can be used. This focuses on the presence of specific circular substructures around each atom in the molecule. In this example, the Rcpi R package (Cao et al. 2015) is used.To obtain a fingerprint, the `getSmiFromChEMBL` function in this package takes a CHEMBL identifier as input and produces the string "SMILES" as output. The Simplified Molecular Linear Input System (SMILES) is a specification for describing the structure of chemical species in linear notation. This can then be converted into an FP4 fingerprint using the function `extractDrugOBFP4`. Any similarity measure known in the art for comparing molecular fingerprints can be used (including any of the molecular similarity measures reviewed in Maggiora and Shanmugasundaram, 2011), such as, for example: Tanimoto, Dice, modified Tanimoto, cosine similarity, Ochiai, and Kulczynski2. Tanimoto similarity (also known as the Tanimoto coefficient or Jaccard similarity) is the most commonly used measure of similarity or chemical distance between two compounds; it is the ratio of the number of features shared by the two molecules to the total number of features, i.e., (A crosses B) / (A+B-(A crosses B)). The Dice coefficient is the number of features shared by two molecules relative to the average size of the total number of features present, i.e., (A * B) / 0.5(A + B). The modified Tanimoto method is a modification of the Tanimoto method, which aims to be more accurate for larger molecules. The cosine similarity between any two vectors A and B of size n is calculated as:
[0115]
[0116] The Ochiai similarity measure (also known as Carbo similarity or cosine coefficient) is calculated as the ratio of the square root of the product of the number of features shared by the two molecules and the number of features in the two molecules. The Kulczynski2 metric is calculated as the ratio of (the number of features shared by the two molecules multiplied by the average size of the total number of features present) divided by the product of the size of the total number of features present (i.e., ((A * B) * 0.5(A + B)) / (A * B)).
[0117] Macromolecules (such as antibodies) are assigned a missing value (NA) for this predicted feature. However, similarity measures for macromolecules (such as antibodies, see page 21 / 37 of specification 25 CN 122342008 A) can be calculated using measures based on one or more of primary sequence, secondary structure (e.g., Musnier et al. 2022), and 3D structure (e.g., RMSD of all or part of the structure).
[0118] Predictive features – target similarity features. Five different target similarity features are used: pathway similarity, target protein similarity, disease association similarity, co-expression, and proximity on protein-protein interaction networks (PPIs).
[0119] Target protein similarity is between 0 and 1 based on known targets of the combination drugs.The score between the targets is calculated as a Jaccard similarity, i.e., target_protein_similarity = intersection(drug1_target, drug2_target) / union(drug1_target, drug2_target). This feature was not used in the final model.
[0120] Co-expression is a score between 0 and 1 based on the average expression of the drug in all corresponding targets across all tissues. The average expression of each gene in all tissues was downloaded from www.proteinatlas.org. The score is calculated as correlation(drug1_target_expression, drug2_target_expression), where correlation is the Pearson correlation coefficient, and drug1_target_expression is a vector of the average expression of the drug1 target, including one value for each tissue (and similarly for drug2_target_expression). This feature was not used in the final model.
[0121] Disease association similarity is a score between 0 and 1 based on the disease ontology of the drug(s)(one or more) targets. For example, drug 1, which targets CD40, can be associated with the ontology term "CD40-related disease," which can be associated with "cancer" and "MS." Drug 2, which targets CD20, can be associated with the ontology term "CD20-related disease," which can be associated with "viral infection" and "MS." The similarity between drug 1 and 2 will reflect the semantic similarity between these two ontology terms. Disease association similarity is calculated using the R package DOSE (Yu G, 2015). DOSE is an R / Bioconductor package used for semantic and enrichment analysis of disease ontology. Disease ontology is a consistent scheme for annotating human genes from a disease perspective. Disease ontology is available in the DO database (disease-ontology.org; Schriml et al., 2012), and gene DO annotations are available in the disease and gene annotation database (DGA, dga.nubic.northwestern.edu; Peng et al., 2013). As an ontology, DOs are organized as directed acyclic graphs, and therefore annotations can be compared using semantic similarity algorithms. The DOSE package provides an algorithm (geneSim function) for calculating semantic similarity between genes based on annotated DO terms. The geneSim function can use one of five similarity measures (Wang, Resnik, Relevance, Jiang, and Lin) and one of four methods (maximum, avg, rcmax, BMA) to aggregate semantic similarity scores for multiple DO terms associated with a gene to calculate the semantic similarity as an entrez.The gene identifier provides the similarity between gene pairs. In this example, the Wang method is used to obtain pairwise similarities between DO terms, but alternatively, any of the methods described above can be used. In this example, the BMA method is used to aggregate multiple pairwise similarities, but any of the aggregation methods described above can also be used. Any of these aggregation methods can also be used to combine similarities for multiple pairs of targets (i.e., calculating similarities for each pair including targets of drug 1 and drug 2, and aggregating the resulting similarities using any of the methods described above). In this example, the BMA method is used for this purpose, but alternatively, any of the aggregation methods described above can be used. The Wang method is described in Wang et al. 2007. The method calculates the similarity between genes G1 and G2 as follows:
[0122]
[0123] where GO1 = {go11, go12, …, go1m} and GO2 = {go21, go22, …, go2n} are the terms set of the ontology (e.g., GO, DO, etc., depending on the case) annotated to genes G1 and G2, respectively, and Sim(go,GO) is the semantic similarity between a term go and the term set GO, which is defined as: Specification 22 / 37 Page 26 CN 122342008 A
[0124]
[0125] where goi is a term in the set GO, and k is the size of the set GO, and the similarity between two terms A and B (e.g., A is go and B is goi in the above equation) defined by the corresponding directed acyclic graphs DAGA=(A,TA,EA) and DAGB=(B,TB,EB) is defined as:
[0126]
[0127] Where terms t are terms related to terms A and B respectively, TA is the set of GO terms in DAGA, including term A in the figure and all its ancestor terms, and EA is the set of edges (semantic relations) connecting GO terms in DAGA (and similarly, connecting corresponding terms in DAGB); SV(A) is the semantic value of GO term A, and SV(B) is the semantic value of GO term B, which is provided by:
[0128]
[0129] Where the S value of each term in each DAG is defined as:
[0130]
[0131] Where we is the semantic contribution factor for the edge e ∈ EA (correspondingly EB) connecting term t and its sub-term t′, where 0 < we < 1. The semantic value of a GO term is the sum of all terms in the DAG representing this GO term.Semantic contribution. Depending on the relation type, the weight we (semantic contribution factor) is set to a fixed value. For example, “is-a” can be associated with a weight of 0.8, and “part-of” can be associated with a weight of 0.6.
[0132] The Resnick method is described in Resnick 1999. This method calculates the semantic similarity between two concepts as:
[0133]
[0134] where S(c1,c2) is the set of concepts containing both c1 and c2, and p is a function p:C→[0,1], such that for any c ∈ C, p(c) is the probability of encountering an instance of concept c, and the probability of the top node is 1. This can be calculated as the relative frequency p(c) = freq(c) / N, where N is the total number of observed nouns, and where word(c) is the set of words contained in concept c.
[0135] The Lin method is described in Lin 1998. The Lin similarity measure is a node-based semantic similarity measure, which is based on the information content of the least common subsumer. The similarity measure is defined as:
[0136]
[0137] where sim(c1,c2) is the Resnick similarity as defined in equation (6), and IC(c1) and IC(c2) are the information content of concepts c1 and c2, which is defined as:
[0138] IC(c)=-log p(c) (8)
[0139] where p(c) is defined as explained above: Specification 23 / 37 Page 27 CN 122342008 A
[0140] p(c)=nc' / N | c' ϵ {c, c's sub} (9)
[0141] where nc' is the number of terms c', and N is the total number of terms in the GO corpus. It should be noted that Resnick similarity is equal to the information content of the most informative common ancestor (MICA).
[0142] The Relevance method is described in Schlicker et al. 2006. This method combines the Resnick and Lin methods and is defined as:
[0143]
[0144] where sim(c1,c2) is the Resnick similarity as defined in equation (6), and IC(c1) and IC(c2) are the information content of concepts c1 and c2 as defined in equations (8) to (9), and MICA is the most informative common ancestor, which is the closest common ancestor term compared between two terms.
[0145] The Jiang method is described in Jiang and Conrath 1997. The similarity measure is defined as:
[0146]
[0147] where MICA is the common ancestor with the largest information content mentioned above.
[0148] The maximum value method calculates the maximum similarity score relative to all term pairs (Equation 12), the avg method uses the average of the similarity scores relative to all term pairs (Equation 13), the rcmax method calculates the maximum of the following: (i) the average of the maximum similarity on each row and (ii) the average of the maximum similarity on each column (Equation 14), and the Best Match Average (BMA) method measures the average of the maximum similarity scores on each row and each column (Equation 15). Therefore, any of the following equations can be used to combine the semantic similarity scores between two term sets relative to term pairs (where the pairwise similarity scores are calculated using any of the metrics described above):
[0149]
[0150] Pathway similarity is a score between 0 and 1 calculated as Dice similarity based on KEGG pathways (www.genome.jp / kegg / pathway.html). Note that any other pathway database annotations can be used for this. In particular, both Gene Ontology (GO) and Reactome annotations were tried. For example, considering drug 1 associated with pathways “immune system” and “cell proliferation” and drug 2 associated with pathways “interferon 1 activation” and “apoptosis”, the score would reflect the similarity between pathways “cell proliferation” and “apoptosis” and pathways “interferon 1 activation” and “immune system”. Pathway associations are based on drug targets, i.e., drugs are considered to be associated with any pathway annotated by any of their targets. Use the R package BioCor to calculate Kegg or Reactome pathway similarity scores (Revilla Sancho 2023). This is done by calculating the similarity between two genes specified by their entrez gene identifiers by calculating the Dice similarity between each pair of pathways involved in the genes and combining multiple similarity values. Combining multiple similarities between pathway pairs in two pathway sets can be done using one of the following: the maximum value method, which calculates the maximum similarity score relative to all pathway pairs (see Equation 12, but where “go” now refers to the pathway and sim is the Dice similarity between pathway pairs); the avg method, which uses the average of the similarity scores relative to all pathway pairs (see Equation 13, but where “go” now refers to the pathway and sim is the Dice similarity between pathway pairs); Specification 24 / 37 pages 28 CN 122342008The Arcmax method calculates the maximum of the following: (i) the average of the maximum similarity on each row and (ii) the average of the maximum similarity on each column (see Equation 14, but where “go” now refers to the pathway and sim is the Dice similarity between path pairs); the Best Match Average (BMA) method measures the average of the maximum similarity scores on each row and each column (see Equation 15, but where “go” now refers to the pathway and sim is the Dice similarity between pathway pairs); the bidirectional method (described in Tao et al. 2007, Equation (3)) calculates twice the sum of the bidirectional maximum similarities above a threshold divided by the number of elements in the two pathway sets being compared (i.e., , where t is the threshold set to remove pairs with low similarity, A and B are two pathway sets with pathways ai and bi, containing the number of pathways |A| and |B|, P is all bidirectional pathway pairs (ai, bi), and sim is the Dice similarity between pathway pairs); or rcmax.avg The method calculates the sum of the maximum values by row and column divided by the number of columns and rows. The BMA method is used in this example, but any other method can be used. Dice similarity is a score between 0 and 1, calculated as twice the number of elements common between the two groups (i.e., the number of genes common between the two pathways) divided by the number of elements in the two groups (the total number of unique genes across the two pathways), i.e.:
[0151]
[0152] where X and Y are the two sets to be compared (in this case, the two pathways defined by a set of genes involved in the pathway). It should be noted that Dice and Jaccard similarity measures are related to each other, and any Dice similarity value can be converted to a Jaccard similarity value. In fact, the corresponding Jaccard index is J=D / (2-D). Conversely, D=2J / (1+J). For GO annotation, only pathway annotation is used, and the GOSemSim software is used to calculate similarity (Yu et al. 2010). This uses one of five similarity measures (Wang, Resnik, Relevance, Jiang, and Lin), and one of four methods (maximum, avg, rcmax, and BMA) for pooling semantic similarity scores for multiple GO terms associated with a gene or a pair of genes. In this example, the Wang and BMA methods are used.
[0153] Proximity on the protein-protein interaction network (PPI) is calculated as a separation score, which is applied to the PPI.The distances between clusters of proteins targeted by the drugs are quantified. In these examples, the STRING network (string-db.org) is used as the PPI. STRING (Szklarczyk et al. 2018) is a database that collects protein-protein functional interaction information from multiple data sources, including interactions between direct pathway members and within protein complexes. This is illustrated in Figure 5, which shows two drugs (A) with overlapping target clusters and two drugs (B) with separate target clusters, along with their corresponding proximity scores. The drug targets were obtained from DGIdb (www.dgidb.org) and located on the STRING network. A “cluster” is simply defined as a set of drug targets on the network. The proximity score is calculated as the average distance between the drug targets of the two drugs using the igraph package, which is used to calculate graph-based distances. The distance function in iGraph calculates the length of all shortest paths from or to the selected vertex in the network. In other words, all pairwise distances between the target of drug 1 and the target of drug 2 are calculated (the length of the minimum length path is identified), and the average of these distances is used as a proximity metric. Note that the minimum or maximum distance can be used alternatively, as the average, minimum, and maximum distances have been found to be correlated. Alternatively, any score that can be used to quantify the distance between subnets (clusters) in a network can be used.
[0154] Predictive Features – Indication Similarity Features. Three different indication similarity features are used: Medical Subject Headings (MESH) similarity and ATC similarity, where ATC similarity is calculated in two ways: Jaccard similarity and Level A similarity (page 25 / 37 of the first specification, CN 122342008).
[0155] Medical Subject Headings (MESH) similarity is a score between 0 and 1 that quantifies the semantic similarity between indications associated with one or more targets of the drug. This is calculated using the R package Mesh. MeSH is a controlled vocabulary used by the U.S. National Library of Medicine to manually index articles on MEDLINE / PubMed (see www.nlm.nih.gov / mesh / meshhome.html). Mesh packages associate MeSH terms with Entrez gene identifiers and provide semantic similarity measures between gene groups. These are used to obtain drug similarity measures based on the Entrez gene identifiers of drug targets. Any of the measures described above for comparing ontologies can be used in this context; these measures include, for example, the Resnick method and Lin...The Mesh package includes the following methods: Relevance, Jiang, and Wang. All of these are implemented in the Mesh package. The Wang method is used in this example, but any of the methods mentioned above can be used. Furthermore, any of the above aggregation methods can be used to combine the similarity of multiple pairs including targets of drug 1 and targets of drug 2. The BMA method is used in this example. The Mesh package provides the following three methods to associate genes with MeSH terms: gendoo, gene2pubmed, and RBBH. The gendoo method is a text mining-based method described in Nakazato et al. 2009. The inventors chose this method in these examples because it is considered more comprehensive and may return results for more drugs than other methods. The gene2pubmed method uses manually selected results from the NCBI team. The Bidirectional Blast Best Hit (RBBH) method is based on sequence homology (E value < 10-50) with the BLASTP search.
[0156] Anatomical Therapeutic Chemistry (ATC) codes are unique codes associated with drugs according to a classification system that indicates the organ or system in which the drug acts and how it acts. The ATC system is maintained by the World Health Organization (WHO). According to this classification system, each active substance is classified in a hierarchical structure with five different levels. At the first level, the classification distinguishes 14 main anatomical / pharmacological groups as defined in Table 1. The second level indicates a therapeutic or pharmacological subgroup and is provided with two numbers. The third and fourth levels indicate chemical, pharmacological, or therapeutic subgroups. They are each designated as a letter. The fifth level uniquely identifies the chemical substances in the group.
[0157] For example, the active substance ocrelizumab is assigned the code L04AA36, which corresponds to antitumor and immunomodulatory agents (L), immunosuppressants (L04), and selective immunosuppressants (L04AA). The active substance rituximab is assigned the code L01FA01, which corresponds to antitumor and immunomodulatory agent (L), antitumor agent (L01), and CD20 inhibitor (L01FA). Specification 26 / 37 pages 30 CN 122342008 A
[0158]
[0159] Table 1. ATC Level 1 Group.
[0160] Two distinct features are derived from the ATC codes. The first feature is the similarity between complete ATC codes, where each ATC code is represented by all possible subsets of the ATC code (e.g., ATC code L04AA36).The ATC code L01FA01 is represented by the set [L, L0, L04, L04A, L04AA, L04AA3, L04AA36], and the ATC code L01FA01 is represented by the set [L, L0, L01, L01F, L01FA, L01FA0, L01FA01]. Note that the last number is always unique for the drug and can therefore be included in or excluded from this representation; in this example, the number is included. This representation is considered to capture more information in terms of ATC code similarity than full encoding or level-to-level comparison. Any similarity metric that can be used to compare discrete sets as described above, such as, for example, Jaccard or Dice similarity, can be used. In this example, the inventors used Jaccard similarity:
[0161]
[0162] In the above example, ozoglucon and rituximab were assigned a similarity of 0.17 (the intersection of the two ATC representations is 2 elements (L and L0), and the union is 12 elements (L, L0, L04, L04A, L04AA, L04AA3, L04AA36, L01, L01F, L01FA, L01FA0, L01FA01), making the Jaccard similarity 2 / 12 = 0.167, rounded to 0.17).
[0163] The second feature is a feature encoded as a binary vector that assigns a 0 or 1 value to each of the 105 possible combinations of the 14 categories in the first ATC level. In other words, this feature indicates which of the two drugs fall into the first level categories. In the example above, both ozoleizumab and rituximab belong to category L, and therefore the corresponding features are vectors for all categories except "L_L" including 0. In other words, each drug in the first level of a pair is one-hot encoded to produce a binary vector.
[0164] Imputation and data preprocessing. Various preprocessing pipelines were evaluated. In all cases, missing data were imputed using a KNN (k nearest neighbor) with k=10. This imputed the missing features as the average across the 10 most similar drugs. The data is then: (i) used as is (labeled "rec_impute"); (ii) scaled and normalized (labeled "rec_impute_norm" on page 27 / 37 of the specification, CN 122342008 A); (iii) scaled, normalized, and subjected to PCA (principal component analysis; labeled "rec_impute_nrom_pca"); or (iv) scaled, normalized, and subjected to partial least squares (PLS; labeled "rec_impute_norm_").pls”). In this example, imputed, scaled, and normalized data (labeled “rec_impute_norm”) is used. This is because in this example, no significant improvement in the model was found from performing PCA or PLS, and since these methods are computationally intensive, it is more efficient to use rec_impute_norm for inference. Scaling and normalization are always performed separately for each predictive feature. These steps help address potential issues with the data to avoid some features dominating the predictions. In this example, scaling and normalization show a modest improvement since all features are between 0 and 1, but this step is maintained because it is computationally efficient and safe to apply. PCA and PLS are used only for non-binary features. The number of non-binary features is reduced to 5 using these methods (the first 5 principal components or the first 5 latent variables of PLS), which yield a summary predictor variable as a linear combination of the original variables.
[0165] Model Architecture and Training. A binary prediction XGBoost algorithm is trained on this data to predict synergies. Via parsnip The package (parsnip.tidymodels.org / ; cran.r-project.org / web / packages / parsnip / index.html) uses the implementation in the R package xgboost (cran.r-project.org / web / packages / xgboost / index.html). Independent features used for prediction include molecular structure, mode of action, side effects, pathways, disease associations, protein-protein interactions, networks, and drug indications. As explained above, the model was trained on 80% of the data (excluding previous MS disease-modifying therapy combinations) and tested on 20% of unseen data (Figure 4, step 3). The final tuning parameters were found by training the model by randomly partitioning the data into 10 partitions (i.e., 10-fold cross-validation). The final model has 1237 trees, a depth of 6, and mtry (the number of randomly selected predictors at each split) of 76. All these hyperparameters were tuned using grid search (e.g., www.tmwr.org / grid-search). As described in [the text], this grid search attempts a range of different parameters and uses the parameters that yield the best performance. Because the data used for training and testing is highly imbalanced (few positive examples compared to the number of negative examples), a threshold of 0.29 is chosen to identify predictions as cooperative based on the F1 score (i.e., a predicted probability > 0.29 is considered cooperative). The F1 score is calculated as F1 = (2 * Precision * Recall) / (Precision + Recall).The precision and recall are combined to measure prediction accuracy, where precision (also known as specificity) is the ratio of true predictions to positive predictions (TP / (TP+FP), where TP is the true positive and FP is the false positive), and recall (also known as sensitivity) is the ratio of true predictions to positive positives (TP / (TP+FN), where FN is the false positive). A high F1 score indicates good precision and recall (i.e., the metric aims to maximize both competing objectives). Balanced accuracy (the average of recall and precision) is also calculated for all models. Feature selection is performed using recursive feature elimination implemented in the R package Caret (cran.r-project.org / web / packages / caret / index.html; topepo.github.io / caret / recursive-feature-elimination.html#rfe). Feature importance analysis was performed using DALEX (ema.drwhy.ai / featureImportance.html) and the VIP package (cran.r-project.org / web / packages / vip / index.html). The model was further evaluated by performing a Bayesian analysis to compare the models using resampling statistics with the tidyposterior package (cran.r-project.org / web / packages / tidyposterior / index.html) and obtaining calibration curves using the Probably package, as described in www.tidyverse.org / blog / 2022 / 11 / model-calibration.
[0166] Alternative model architectures. Various alternative architectures were also used, and each model was trained as described above. The tested model architectures include: a decision tree classifier (an ensemble of 1237 trees with a depth of 6 and mtry (randomly selected predictors) of 76), a random forest classifier, a logistic regression model, multivariate adaptive regression splines (MARS), regularized discriminant analysis (RDA), a multilayer perceptron (MLP), and a support vector machine (SVM) with radial basis function (RBF) kernels (see page 28 / 37, CN 122342008 A). All models are accessed via the parsnip R package (cran.r-project.org / web / packages / parsnip / index.html), which accesses the following R...Packages: For decision trees: rpart (cran.r-project.org / web / packages / rpart / rpart.pdf); For random forests: ranger (cran.r-project.org / web / packages / ranger / ranger.pdf); For logistic regression: the glm function in the stats package (stat.ethz.ch / R-manual / R-devel / library / stats / html / glm.html); For MARS: earth (cran.r-project.org / web / packages / earth / index.html); For RDA: klaR (cran.r-project.org / web / packages / klaR / index.html); For MLP: nnet (cran.r-project.org / web / packages / nnet / index.html); For SVM: kernlab (cran.r-project.org / web / packages / kernlab / index.html).
[0167] Validation of MS disease-modifying drugs. The model was applied to 33,313 drug pairs, including combinations of 17 existing disease-modifying therapies for MS and 7,678 approved or investigational drugs for various diseases (Figure 4, Step 4). These were selected as any DMT for MS in Phase 1 or above from OpenTargets (www.opentargets.org) using ChEMBL and were documented. These were filtered from the total number of possible combinations to exclude combinations with more than 3 missing values for predictive features. Disease-modifying therapies include: alemtuzumab (CHEMBL1 201587), cladribine (CHEMBL161 9), ozenafil (CHEMBL2108041), natalizumab (CHEMBL1201607), ofatumumab (CHEMBL1201836), fingolimod (CHEMBL314854), ponesimod (CHEMBL1096146), interferon beta-1b (CHEMBL1201563), and interferon beta-1a.(CHEMBL1201562), Mitoxantrone (CHEMBL58), Siponimod (CHEMBL2336071), Dimethyl fumarate (CHEMBL2107333), Glatiramer acetate (CHEMBL1201507), Teriflunomide (CHEMBL973), Diroximel fulmarate (CHEMBL3989944), Azathioprine (CHEMBL1542), and Ozanimod (CHEMBL3707247).
[0168] Results
[0169] Figure 6 schematically illustrates the structure of the data used for prediction. For each drug combination for which clinical trial data was available, all candidate predictive features described above were calculated. Each drug combination was further labeled as “synergistic” or “antagonistic”. The data was then analyzed to determine the extent of missing values in the data. This is shown in Figure 7, which also shows the overall proportion of missing values (12.5%) and the proportion of missing values for each feature (from left to right: side effect similarity, kegg pathway similarity, ATC similarity, MESH similarity, proximity on PPI, disease ontology similarity, mode of action similarity, and structural similarity (labeled "Smiles similarity")). This indicates that data imputation may be necessary in this dataset, as removing any entries from the missing data would remove too much data. The data in Figure 7 illustrates the overall pattern of missing values, demonstrating the challenges of addressing real-world data and the potential for significant benefits associated with data engineering and cleaning. For example, the data shows that the side effect feature has many missing values. This is because commonly approved drugs (approximately 2000 drugs) have their side effect tags. Therefore, the inventors tested two approaches: imputing this missing data or using a model that does not have this feature. The latter resulted in lower model performance, so the feature was preserved and the missing data was imputed. However, to limit the impact of missing data and imputation, the inventors decided to remove all rows with >2 missing values (i.e., any drug combination). They then imputed the missing values as part of preprocessing (rec_impute_norm), as described above. Feature selection was then performed using recursive feature elimination, i.e., starting with the entire set of variables being tested, the model was tested with progressively fewer variables, and an F1 score was calculated. Each time, the variable with the least impact on the performance specification (pages 29 / 37, 33 CN 122342008 A) was removed. The results are shown in Figure 8.As shown, this figure illustrates the results as the average of 10-fold cross-validation (Figure 8A shows the average F1 score, and Figure 8B shows the average tolerance %, i.e., the percentage of error accepted). Using a 5% tolerance threshold, models with F1 scores up to 5% lower than the best model's score were accepted (indicated by dashed lines in Figure 8B, leading to the selection of the 8-feature model whose F1 score is highlighted as a solid triangle in Figure 8A, representing the smallest model whose performance differs from the best model by less than 5%, indicated by a solid dot in Figure 8A). Based on this, the inventors decided to remove the following predictive features: target similarity, pathway similarity based on gene ontology and Reactome (maintaining pathway similarity based on Kegg annotation), and co-expression. Therefore, the following features are maintained in the final model: molecular structure similarity, mode of action similarity, side effect similarity, pathway similarity (Kegg annotation), disease association similarity, proximity on protein-protein interaction networks, MESH similarity, and ATC similarity (both full-coding similarity and first-level binary similarity).
[0170] Using these features to train the XGBoost model, when evaluated on unseen data (20% reserved for test data), the model achieved 94% specificity and 70% sensitivity in predicting synergistic combinations.
[0171] Figure 9 shows the results of the feature importance analysis. First, the relative importance of the predictive features in the model was investigated by randomly permuting the variables in the test data and calculating the average (1-AUC) obtained by applying the trained model to 10 random permutations of each of the indicated features (i.e., the value of the shuffled feature, predicted from the model, and measuring how much worse the model fits the data compared to before shuffling). The results for this are shown in Figure 9A, where the black line represents the median across 10 permutations, the box shape shows the interquartile range, and the results must be extended to the minimum and maximum values. For this analysis, the binary ATC feature was removed for computational efficiency. Figure 9A shows that indication similarity features (MESH and ATC) are most important for the model's predictions, followed by mode of action similarity. The dashed line shows the performance of the complete model; therefore, the further the distribution of (1-AUC) values is from this line, the more important the feature is (i.e., the more performance the model loses when the feature is perturbed). Figure 9B shows the results of a survey on feature importance when the first-level ATC feature is treated as a separate feature. For this analysis, all features (8 features + binary ATC encoding) were used without permutation, as this would be computationally too large to include all binary features. Therefore, the method described in Greenwell et al. (2018) using the R package...The variance-based approach implemented in Vip again demonstrates that MESH similarity features are most important for prediction, but also shows that the importance of ACT similarity features is actually separated between ATC similarity (fully encoded similarity) and first-level binary similarity, especially for some combinations (e.g., if a pair of drugs belongs to L_L, N_N, J_J, or C_C categories, this is very informative about whether the drugs will synergize, while other categories are less informative). Some of these first-level binary similarity categories (especially L_L and N_N) are even more informative than the mode of action features. Figure 9C shows the results of the same analysis as in Figure 9A, but features are grouped by category (i.e., indication similarity, target similarity, and drug similarity). Features are grouped into each of these groups and arranged together, and then the model performance is quantified as (1-AUC) across 10 permutations, as described above. This indicates that, to date, the indication similarity feature (combining MESH similarity and both types of ATC similarity) is the most important for the predictions made by the model.
[0172] Figure 10 shows the Pearson correlations between features of the final predictive model, across the training and test datasets (A) for all eight features in the 8-feature model, and across all features (B) for the initial survey. For ATC similarity, only the full coded similarity measure was used to produce the graph on Figure 10A. The color intensity and size of the dots indicate the correlation between variables, and the color (red or blue) indicates the sign of the correlation. The data on Figure 10A shows that, as expected, ATC and MESH similarities are highly correlated, as are disease ontology and pathway similarity (kegg_sim). However, the data on Figure 9A indicates that, at least when it comes to ATC and MESH, the information captured is not entirely redundant. Figure 30 / 37, page 34, CN 122342008 A 10B, shows that kegg, gene ontology, and Reactome-based pathway similarities are highly correlated (serving as the basis for the decision to include only one of these similarities in the set of predictive features used; in this example, only Kegg-based similarities are included, although any of these can be used equally).
[0173] Figure 11 shows the distribution of each predictive feature in the training data separated by combined categories (left for antagonism, right for synergy). Figure 11A shows the results for the variables included in the final model. All differences were found to be statistically significant (test: Wilcoxon test, **: p <= 0.01; ***: p <= 0.001; ****: p <= 0.001).<= 0.0001). The data shows that the importance of the predicted features is not necessarily related to the magnitude of the difference between the two distributions. For example, the action pattern similarity distribution is more different from the MESH and ATC similarity (fully encoded similarity) distributions because the values in the co-category are more widely distributed. Figure 11B shows the same data for all variables in the initial survey. The data shows that the removed features (co-expression, go_sim, and reactome_sim) either have small differences between categories or have similar distributions to other predicted features.
[0174] Figure 12 shows the results of the study on the impact of the survey model selection and data preprocessing methods. Figure 12A shows the impact of model selection and data preprocessing on F1 score and balance accuracy. F1 and balance accuracy were used as performance metrics because the dataset is very imbalanced (containing more adversarial pairs compared to co-category pairs). For each model and data preprocessing method tried, the F1 accuracy ranked in descending order of F1 (top) and the balance accuracy ranked in descending order of balance accuracy (bottom) are shown. The data shows that logistic regression performs worse than all other models, while boosting trees perform particularly well. However, all models except logistic regression exhibit good performance. Nonlinear models tend to perform better than linear models (logistic regression, regularized discriminant analysis). All data preprocessing methods are appropriate, but normalized data performs slightly better overall, and dimensionality reduction with PCA / PLS does not appear to be beneficial. Figure 12B shows the posterior distribution of the average balance accuracy associated with each of the indicated combinations of models and preprocessing methods. This compares multiple models by resampling, thus evaluating the performance of sub-models on cross-validation resampling (in this case, balance accuracy). Using these observations, a Bayesian method is used to compute the posterior distribution of the balance accuracy for each model. This confirms that the model using normalized data performs better than its unnormalized counterpart and shows that the XGBoost model performs significantly better than all other models, followed by random forests and multilayer perceptrons. Figure 12C more closely compares the posterior distribution of average balance accuracy between the Random Forest and XGBoost models. The top plot shows the difference in average balance accuracy between the XGBoost model on normalized data and the Random Forest on imputed but unnormalized data (note that for the Random Forest model, there is no significant difference compared to using normalized data; see Figure 12B). The bottom plot shows a scaled-down version of the corresponding posterior distribution of average balance accuracy associated with these models shown in Figure 12B. This indicates that the XGBoost model does indeed perform slightly better than the Random Forest model. Figure 12DThe proportion of models with an effect size of 2% is shown, where the effect size is the amount of difference in accuracy that we are willing to accept. This again shows that the XGBoost model has the best performance, followed by RDA and Random Forest, and normalization has a small impact on the performance of these models.
[0175] Figure 13 shows the results of the threshold selection and optimization process used to obtain the final XGBoost model. The data on Figure 13A shows the F1 score, precision (increasing data sequence), and recall (decreasing data sequence) obtained by the XGBoost model as the threshold for identifying predictions as "positive examples" (cooperative) for the training data increases. This shows that a threshold of 0.29 was found to maximize the F1 score. The F1 score was chosen as the performance metric because the dataset is imbalanced. However, the same approach can be applied to maximize any other metric such as precision or recall. Figure 13B shows the balanced accuracy, sensitivity (decreasing data sequence), and specificity (increasing data sequence) obtained by the XGBoost model as the F1 threshold increases (the same model as in Figure 13B, but evaluated on different performance metrics). This figure shows that, as per specification page 31 / 37, 35 CN 122342008 A, maximizing the F1 score (rather than bal_accuracy) allows us to also maximize bal_accuracy with a threshold of 0.29. Therefore, by choosing 0.29, we obtain higher F1 and bal_accuracy.
[0176] Figure 14A shows the confusion matrix obtained by evaluating the trained XGBoost model (trained with the features described above and selected based on the F1 score = 0.29 threshold as described above) on the held-out test data. This shows that when evaluated on unseen data (20% held-out test data), the model achieves 94% specificity and 70% sensitivity in predicting co-op combinations. This is exceptionally good performance. The classification accuracy is 90%. This is higher than, for example, the best model described in Ji et al. 2019, which is based on a similar but different predictive features (78.7% accuracy). Figure 14B shows the associated receiver operating characteristic (ROC) curves, again demonstrating very high predictive performance.
[0177] Figure 15 is a calibration curve showing the observed co-log (event rate) for each of several intervals of predicted probabilities from the XGboost model (x-axis indicates the midpoint of the interval). The idea behind the calibration plot is that if we group the predictions based on the predicted probabilities, we should see a certain percentage that matches such probabilities.The events. For example, if we collect a set of predictions whose probability is estimated to be about 10%, then we should expect that about 10% of those predictions in that set are indeed events. The plot shown here can be used as a diagnostic to see if the model’s predictions are consistent with the observed event rate. In this plot, the event rate is the ratio of synergistic to antagonistic. The top dashed line shows the probability of synergistic combinations, and the bottom dashed line shows the probability of antagonistic combinations. Typically, the challenging part is the tail (close to 0 or 1), but the data suggests that the model performs well (close to the x=y line). The plots show that the model’s predictions are consistent with the observed event rate, indicating that the model performs well.
[0178] Figure 16 shows a partial dependency plot for the predicted feature pairs. The partial dependency plot shows how the expected value of the model’s prediction (synergistic) changes with the features (mesh_sim, atc_sim, ...). Each gray line here corresponds to a row in the test dataset (i.e., a single drug combination, its features, and its synergistic effect). The blue line corresponds to the average of all the gray lines. This plot shows the overall behavior of the output when we change each feature. For example, when we increase the mesh_sim score from 0 to 1, the synergistic probability increases. This makes biological sense. Therefore, this figure is a diagnostic plot to ensure that the model's predictive features are represented in a biologically meaningful way – this is the case here.
[0179] Next, the trained model was applied to combinations of 17 validated disease-modifying therapies (DMTs) for the treatment of MS with drugs currently in Phase 2 or later clinical trials (7,678 investigational drugs, resulting in approximately 130,000 combinations). Figure 17 shows the missing values for this data. Each row represents a drug combination, and black indicates missing values. This indicates that the data contains a high proportion of missing values, primarily related to side effects, kegg, and ATC similarity. This is not surprising, as many of these drugs may not yet have fully characterized side effects and ATC encoding, and some of the investigational drugs do not even have known targets (resulting in a high proportion of missing values for features that rely on known targets, such as kegg pathway similarity). However, the data also show that the Mesh similarity feature has the lowest percentage of missing values. This is particularly significant given that this feature was found to be highly predictive. Therefore, this indicates that predictions from this model may be more effective than existing models for disease areas where few commercially available DMTs are available, such as MS, because the predictions can utilize realistic, information-rich data rather than imputing these drug combinations.
[0180] Figure 18 shows a bar chart of the distribution of the number of missing features for each drug combination in the MS drug combination data.Given the large amount of missing data in this dataset, further analysis focused on 33,313 combinations with a maximum of 2 missing features (i.e., 0, 1, or 2 missing features). The resulting predictions are shown in a bar chart in Figure 19. Of the 33,313 combinations, 17% were predicted to be synergistic when an optimized probability threshold of 0.29 was applied. The inset shows the predicted pairs, with the MS disease-modifying therapy (DMT) listed in the first column, the adjunctive therapy listed on pages 32 / 37 of the instruction manual (CN 122342008 A) in the second column, and the probability predicted by the model in the third column. This was repeated for specific single-purpose DMTs, where the top 3 predictions were extracted and the scientific literature was searched for any evidence that the adjunctive therapy involved MS or MS-related pathways. The synergistic effect of all the top 3 predictions is supported by the literature, regardless of whether there is direct evidence of synergy (in individual cases where two therapies are administered simultaneously for unrelated conditions or at the preclinical level, since all these combinations are novel and have not yet been the subject of clinical trials). Figure 20 shows the feature importance for one of these top 3 predictions, and the actual value of the feature for that combination (labeled on the y-axis).
[0181] This figure shows why this combination receives the probability assigned to it (i.e., 0.87). The x-axis shows the value of the predictive feature for this combination (do_sim=0.285, mesh_sim=0.222, ...). The x-axis shows how much each feature contributes to the predicted probability. For example, do_sim contributes 0.3 out of 0.87. This is obtained by calculating the contribution attributable to individual features as a Shapley additive explanation for the average model prediction for a particular observation. The box plot shows the distribution of contributions across all rankings used to calculate the Shapley attribution. The figure shows that, for this particular combination, disease ontology and mesh similarity are the largest contributors to the obtained probabilities.
[0182] Conclusion
[0183] This work demonstrates that machine learning provides valuable tools for understanding drug properties and successful combinations. This study designed a high-performance model that identifies potentially successful combinations for further evaluation in experimental trials. This is the first model specifically validated to predict drug combinations for neurological diseases such as MS, where combination trials for MS were saved as unseen data and used for validation. The model not only identifies new combinations but also accurately predicts the outcomes of all previous MS combination trials based on biological evidence and expert opinion. Further refinement can select the most promising predicted synergistic candidates that are consistent with specific development goals.
[0184] This study demonstrates the potential of machine learning in finding effective drug combinations for the treatment of MS. It bridges the gap in available therapies and holds promise for enhancing MS management. Further exploration and validation in experimental trials can be applied to evaluate the identified combinations and advance MS treatment options.
[0185] Furthermore, the data demonstrate the ability of a number of new methods to characterize drug combinations to predict the likelihood of synergistic effects. The first newly developed high-importance feature is drug target MESH similarity. As shown in Figure 9A, this new feature contributes extremely valuable information for synergistic prediction. The second newly developed importance feature is the ATC Level 1 combination feature. As shown in Figure 9B, this method of characterizing drug pairs contributes information that simple ATC similarity measures do not capture, and at least some of these ATC pair categories are highly informative in making predictions. These features together constitute the most important part of the predictive model, as shown in Figure 9C (where the indication similarity feature importance is a combination of the importance of MESH similarity, ATC similarity, and the ATC binary category variable). The third importance feature is the disease ontology similarity measure, which, as shown in Figures 9 (9B and 9A), is more important for prediction than side effects or even ATC similarity when considered separately from the ATC binary category variable. Importantly, both MESH similarity and disease ontology similarity rely on the availability of drug target annotations, and are therefore available even for drugs that have not yet been assigned an ATC or whose side effects are not registered in the FDA Orange Book. This means that the values of these predictive features can be used for a wider range of drugs, thereby increasing the applicability of the model and the amount of training data available for accuracy. Therefore, it is believed that the use of the above three features (each individually and collectively) can form the basis for the benefits of this model, including its very high predictive performance.
[0186] References
[0187] All documents mentioned in this specification are incorporated herein by reference in their entirety.
[0188] Revilla Sancho L (2023). BioCor: Functional similarities. bioconductor.org / packages / BioCor, llrs.github.io / BioCor / . Instruction manual, pages 33 / 37, CN 122342008 A
[0189] Guha Rajarshi, fingerprint: Functions to Operate on Binary Fingerprint Data, package version 3.5.7, released on2018‑01‑07, CRAN.R‑project.org / package= fingerprint
[0190] Csárdi G, Nepusz T, Traag V, Horvát S, Zanini F, Noom D, Müller K (2023). igraph for R: R interface of the igraph library for graph theory and network analysis doi:10.5281 / zenodo.7682609, R package version 1.5.1, CRAN.R‑project.org / package=igraph.
[0191] G Yu, LG Wang, GR Yan, QY He. DOSE: an R / Bioconductor package for Disease Ontology Semantic and Enrichment analysis. Bioinformatics 2015, 31(4):608‑609. dx.doi.org / 10.1093 / bioinformatics / btu684.
[0192] Yu G (2018). "Using meshes for MeSH term enrichment and semantic analyses." Bioinformatics, 34(21), 3766‑3767. doi:10.1093 / bioinformatics / bty410.
[0193] Schriml LM et al., Disease Ontology: a backbone for disease semantic integration Nucleic Acids Res. 2012 40 D940 D946
[0194] Peng K, Xu W, Zheng J, Huang K, Wang H, Tong J, Lin Z, Liu J, Cheng W, Fu D, Du P, Kibbe WA, Lin SM, Xia T. The Disease and Gene Annotations (DGA): an annotation resource for human disease. Nucleic Acids Res. January 2013;41(Database issue): D553‑60.
[0195] Wang JZ, Du Z, Payattakool R, Yu PS, Chen CF. A new method to measure the semantic similarity of GO terms. Bioinformatics. 2007 May 15;23(10):1274‑81.
[0196] Resnik P. Semantic similarity in a taxonomy: an information‑based measure and its application to problems of ambiguity in natural language, J. Artificial Intelligence Res., 1999, Volume 11 (95‑130)
[0197] Jiang J, Conrath D. Semantic Similarity Based on Corpus Statistics and Lexical Taxonomy, 1997 tenth International Conference on Research on Computational Linguistics (ROCLING X) Taiwan, China
[0198] Lin D. An information‑theoretic definition of similarity, Semantic Similarity Based on Corpus Statistics and Lexical Taxonomy, fifteenth International Conference on Machine Learning, 1998 (296‑304)
[0199] Schlicker, Andreas, Francisco S Domingues, Jörg Rahnenführer and Thomas Lengauer. 2006. "A New Measure for Functional Similarity of Gene Products Based on Gene Ontology." BMC Bioinformatics 7: 302.
[0200] YingTao, Lee Sam, Jianrong Li, Carol Friedman, Yves A. Lussier, Information theory applied to the sparse gene ontology annotation network to predict novel gene function, Bioinformatics, Vol. 23, No. 13, July 2007, pp. i529–i538
[0201] Rogers D, Hahn M. Extended‑connectivity fingerprints. J Chem Inf Model. 24 May 2010; 50(5): 742‑54. DESCRIPTION 34 / 37 PAGE 38 CN 122342008 A
[0202] Ji X et al. Five‑Feature model for developing the classifier for synergistic vs. antagonistic drug combinations built by XGBoost. Frontiers in Genetics (2019) 10: 600.
[0203] Paltun B G, Kaski S., Mamitsuka H. Machine learning approaches for drug combination therapies. Briefings in Bioinformatics, 22(6), 2021, 1‑16.
[0204] Nakazato T, Bono H, Matsuda H, Takagi T. Gendoo: functional profiling of gene and disease features using MeSH vocabulary. Nucleic Acids Res. July 2009; 37(Web Server issue): W166‑9.
[0205] Maglott D, Ostell J, Pruitt KD, Tatusova T. Entrez Gene: gene‑centered information at NCBI. Nucleic Acids Res. January 2007; 35(Databaseissue):D26‐31。
[0206] FDA, 2023: Approved Drug Products with Therapeutic Equivalence Evaluations.Section 43,2023, Products approved under section 505 of the Federal Food, Drug and Cosmetic Act Act, December 31, 2022
[0207] Liu H, Zhang W, Zou B, Wang J, Deng Y, Deng L. DrugCombDB: a comprehensive database of drug combinations toward the discovery of combinatorial therapy.Nucleic Acids Res.
[0208] Wishart DS, Feunang YD, Guo AC, Lo EJ, Marcu A, Grant JR, Sajed T, Johnson D, Li C, Sayeeda Z. Assempour N. Iynkkaran I. Liu Y. Maciejewski A. Gale N. Wilson A. Chin L. Cummings R. Le D. Pon A. Knox C. Wilson M. DrugBank 5.0: a major update to the DrugBank database for 2018.Nucleic Acids Res. Apr 8, 2017.
[0209] Sharon L Freshour, Susanna Kiwala, Kelsy C Cotto, Adam C Coffman, Joshua F McMichael, Jonathan J Song, Malachi Griffith, Obi L Griffith, Alex H Wagner, Integration of the Drug–Gene Interaction Database (DGIdb 4)..0) with open crowdsource efforts, Nucleic Acids Research, Volume 49, Issue D1, 8 January 2021, pp. D1144–D1151.
[0210] Cao, D .‑S . et al. (2015) Rcpi: R / Bioconductor package to generate various descriptors of proteins, compounds and their interactions. Bioinformatics, 31, 279–281.
[0211] Maggiora, G .M., Shanmugasundaram, V. (2011). Molecular Similarity Measures. In: Bajorath, J. (Ed.) Chemoinformatics and Computational Chemical Biology. Methods in Molecular Biology, Volume 672. Humana Press, Totowa, NJ.
[0212] Musnier, A.; Bourquard, T.; Vallet, A.; Mathias, L.; Bruneau, G.; Ayoub, M .A .; Travert, O.; Corde, Y .; Gallay, N.; Boulo, T . et al. A New in Silico Antibody Similarity Measure Both Identifies Large Sets of Epitope Binders with Distinct CDRs and Accurately Predicts Off‑Target Reactivity. Int. J. Mol. Sci. 2022, 23, 9765.
[0213] G Yu, F Li, Y Qin, X Bo*, Y Wu, S Wang*. GOSemSim: an R package for DESCRIPTION 35 / 37 pages 39 CN 122342008 A measuring semantic similarity among GO terms and gene products .Bioinformatics. 2010, 26(7): 976‑978.
[0214] Szklarczyk et al. Nucleic acids research 47.D1 (2018): D607‑D613.2
[0215] Nakazato T, Bono H, Matsuda H, Takagi T. Gendoo: functional profiling of gene and disease features using MeSH vocabulary. Nucleic Acids Res. July 2009; 37(Web Server issue): W166‑9.
[0216] Brandon M. Greenwell, Bradley C. Boehmke, Andrew J. McCarthy. A Simple and Effective Model‑Based Variable Importance Measure . arXiv:1805 .04755v1 [stat.ML]. May 12, 2018.
[0217] Kanehisa M., Furumichi M., Tanabe M., Sato Y., Morishima K.. KEGG: new perspectives on genomes , pathways , diseases and drugs . Nucleic Acids Res . 2017; 45:D353–D361.
[0218] Fabregat A., Sidiropoulos K., Garapati P., Gillespie M., Hausmann K., Haw R., Jassal B., Jupe S., Korninger F., McKay S. et al. The reactome pathway Knowledgebase. Nucleic Acids Res. 2016; 44:D481–D487.
[0219] Caspi R., Billington R., Ferrer L., Foerster H., Fulcher C .A., Keseler I.M., Kothari A., Krummenacker M., Latendresse M., Mueller L.A. et al. TheMetaCyc database of metabolic pathways and enzymes and the BioCyc collection of pathway / genome databases. Nucleic Acids Res. 2016; 44:D471–D480.
[0220] The Gene Ontology, C Expansion of the Gene Ontology knowledgebase and resources. Nucleic Acids Res. 2017; 45:D331–D338.
[0221] Equivalents and Scope
[0222] Unless the context otherwise requires, the description and definition of the above features are not limited to any particular aspect or embodiment of the invention and are equally applicable to all aspects and embodiments described.
[0223] The term “and / or” as used herein will be considered as a specific disclosure of each of two particular features or components, which may or may not include the other. For example, “A and / or B” will be considered as a specific disclosure of each of (i) A, (ii) B, and (iii) A and B, as if each were listed separately herein.
[0224] It must be noted that, as used in this specification and the appended claims, the singular forms “a,” “an,” and “the” include plural referents unless the context explicitly specifies otherwise. A range may be expressed herein as from “about” one particular value and / or to “about” another particular value. When expressing such a range, another embodiment includes from one particular value and / or to another particular value. Similarly, when a value is expressed as an approximation, the particular value will be understood to form another embodiment by using the antecedent “about” or “approximately.” The terms “about” or “approximately” in relation to numerical values are optional and mean, for example, + / - 10%.
[0225] Throughout this specification, including the following claims, unless the context requires otherwise, the words “comprising” and “including,” and variations such as “comprising” and “including,” should be understood to imply inclusion of the stated integers or steps or groups of integers or steps, but do not exclude any other integers or steps or groups of integers or steps.
[0226] Other aspects and embodiments of the invention provide for the above aspects and embodiments, wherein the term "comprising" is replaced by the terms "consisting of" or "substantially consisting of," unless the context otherwise requires.
[0227] Features disclosed in the foregoing specification, or expressed or implemented in a particular form in the following claims, are described in the specification on pages 36 / 37, 40 CN 122342008 ARegarding the manner in which these features are used, or the methods or processes for obtaining the disclosed results, these features may be used individually or in any combination of their various forms to achieve the invention.
[0228] Although the invention has been described in conjunction with the exemplary embodiments described above, many equivalent modifications and variations will be apparent to those skilled in the art when this disclosure is given. Therefore, the exemplary embodiments of the invention described above are to be considered illustrative rather than restrictive. Various changes may be made to the described embodiments without departing from the spirit and scope of the invention.
[0229] To avoid any doubt, any theoretical explanations provided herein are intended to improve the reader's understanding. The inventors do not wish to be bound by any of these theoretical explanations.
[0230] Any section headings used herein are for organizational purposes only and should not be construed as limiting the subject matter described. Instruction manual page 37 / 37, page 41, CN 122342008 A, Figure 1; Instruction manual figure 1 / 27, page 42, CN 122342008 A, Figure 2; Instruction manual figure 2 / 27, page 43, CN 122342008 A, Figure 4; Instruction manual figure 3 / 27, page 44, CN 122342008 A, Figure 5A; Instruction manual figure 5B; Instruction manual figure 4 / 27, page 45, CN 122342008 A, Figure 6; Instruction manual figure 5 / 27, page 46, CN 122342008 A, Figure 7; Instruction manual figure 6 / 27, page 47, CN 122342008 A, Figure 8A; Instruction manual figure 7 / 27, page 48, CN 122342008 A, Figure 9A; Instruction manual figure 8 / 27, page 49, CN 122342008 A, Figure 9B; Instruction manual figure 9 / 27, page 50, CN 122342008 A Figure 9C Instruction Manual Appendix 10 / 27 Page 51 CN 122342008 A Figure 10A Instruction Manual Appendix 11 / 27 Page 52 CN 122342008 A Figure 10B Instruction Manual Appendix 12 / 27 Page 53 CN 122342008 A Figure 11A Instruction Manual Appendix 13 / 27 Page 54 CN 122342008 A Figure 11B Instruction Manual Appendix 14 / 27 Page 55 CN 122342008 A Figure 12A Instruction Manual Appendix 15 / 27 Page 56 CN 122342008 A Figure 12B Instruction Manual Appendix 16 / 27 Page 57 CN 122342008 A Figure 12C Instruction Manual Appendix 17 / 27 Page 58CN 122342008 A Figure 12D Instruction Manual Drawing, Page 18 / 27, 59 CN 122342008 A Figure 13A Instruction Manual Drawing, Page 19 / 27, 60 CN 122342008 A Figure 13B Instruction Manual Drawing, Page 20 / 27, 61 CN 122342008 A Figure 14A Figure 14B Instruction Manual Drawing, Page 21 / 27, 62 CN 122342008 A Figure 15 Instruction Manual Drawing, Page 22 / 27, 63 CN 122342008 A Figure 16 Instruction Manual Drawing, Page 23 / 27, 64 CN 122342008 A Figure 17 Instruction Manual Drawing, Page 24 / 27, 65 CN 122342008 A Figure 18 Instruction Manual Drawing, Page 25 / 27, 66 CN 122342008 A Figure 19 Instruction Manual Drawing, Page 26 / 27, 67 CN 122342008 A Figure 20 Instruction Manual Appendix 27 / 27 Page 68 CN 122342008 A
Claims
1. A computer-implemented method for predicting the effects of drug combinations, the method comprising: For the first drug and the second drug, values of multiple features associated with the drug combination are obtained; as well as Using the values of the plurality of features, the effect of the drug combination is predicted using a machine learning model trained with training data to predict whether the drug combination is likely to be synergistic, the training data for each of the plurality of training drug combinations including: (i) the values of the plurality of features; and (ii) an indication of whether the drug combination is synergistic; The plurality of features includes at least one feature that characterizes the similarity between one or more indications or diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug.
2. The method of claim 1, wherein the machine learning model has been trained to classify drug combinations among multiple categories including synergistic and non-synergistic categories, and wherein predicting the effect of the drug combination includes determining whether the drug combination is likely to belong to the synergistic category.
3. The method according to any of the preceding claims, wherein the plurality of features comprises: At least one feature characterizing the similarity between one or more indications associated with one or more targets of the first drug and one or more indications associated with one or more targets of the second drug, wherein the indications are provided as Medical Subject Headings terms associated with the one or more targets of the drug; and / or At least one feature characterizes the similarity between one or more diseases associated with one or more targets of the first drug and one or more diseases associated with one or more targets of the second drug, wherein the disease is provided as a disease ontology term associated with the one or more targets of the drug.
4. The method according to any preceding claim, wherein the similarity between one or more indications or diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug is obtained as a semantic similarity measure between indication or disease terms associated with the one or more targets of the first drug and indication or disease terms associated with the one or more targets of the second drug, or as a summary measure aggregating multiple such semantic similarity measures, each such semantic similarity measure corresponding to a pair of indication or disease terms, the pair of indication or disease terms including indication or disease terms associated with the target of the first drug and indication or disease terms associated with the target of the second drug.
5. The method of any preceding claim, wherein the plurality of features comprises: One or more drug similarity features, one or more target similarity features, and one or more indication similarity features, optionally wherein said one or more drug similarity features are selected from: Similarity in side effects, similarity in modes of action, and similarity in molecular structure.
6. The method of claim 5, wherein the one or more target similarity features are selected from: pathway similarity; Disease association similarity, which characterizes the similarity between one or more diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug; And a measure of the proximity of the target of the first drug and the target of the second drug on a protein-protein interaction network.
7. The method of any one of claims 5 or 6, wherein the one or more indication similarity features are selected from: indication similarity measures based on the targets of the drug, which characterize the similarity between one or more indications associated with one or more targets of the first drug and one or more indications associated with one or more targets of the second drug; and indication similarity measures based on the drug, which characterize the similarity between indication annotations associated with the first drug and the second drug.
8. The method of claim 7, wherein the indication annotation is provided as an Anatomical Therapeutic Chemistry (ATC) code, and / or wherein the indication similarity measure based on the drug is selected from: a similarity measure that quantifies the similarity between multiple levels of the ATC codes for the first drug and the second drug; and a similarity measure that compares a first level of the ATC codes for the first drug and the second drug.
9. The method of claim 7 or claim 8, wherein the similarity metric for comparing the first level of the ATC codes for the first drug and the second drug is a binary feature indicating whether the first drug and the second drug are associated with ATC codes belonging to the same first level category or to a particular pair of ATC categories, and / or wherein the plurality of features comprises a set of binary features indicating whether the first drug and the second drug belong to each of a plurality of pairs of first level ATC categories, optionally wherein the plurality of pairs of first level ATC categories include LL, NN, JJ, and CC.
10. The method according to any of the preceding claims, wherein the machine learning model is selected from: decision trees and variants thereof, including regularized and / or gradient boosting decision trees and random forest models; regularized discriminant analysis; logistic regression models; artificial neural networks; Naive Bayes classifiers; support vector machines; and multivariate adaptive regression splines, and / or wherein the machine learning model is a nonlinear model, and / or wherein the machine learning model is an ensemble model, and / or wherein the machine learning model is a tree-based model.
11. The method according to any preceding claim, wherein the plurality of features comprises: At least one feature characterizing the similarity between one or more indications associated with one or more targets of the first drug and one or more indications associated with one or more targets of the second drug, wherein the indications are provided as medical subject terminology terms associated with the one or more targets of the drug; Side effect similarity Similarity of action patterns Molecular structural similarity Pathway similarity Disease association similarity, which characterizes the similarity between one or more diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug; And a measure of the proximity of the target of the first drug and the target of the second drug on a protein-protein interaction network; as well as The indication similarity measure based on the drugs characterizes the similarity between indication annotations associated with the first drug and the second drug, wherein the indication annotations are provided as Anatomical Therapeutic Chemistry (ATC) codes, and wherein the indication similarity measure based on the drugs includes a similarity measure that quantifies the similarity between multiple levels of the ATC codes for the first drug and the second drug, and optionally a similarity measure that compares a first level of the ATC codes for the first drug and the second drug, wherein the similarity measure that compares the first level of the ATC codes for the first drug and the second drug is a set of binary features indicating whether the first drug and the second drug belong to each of multiple pairs of first-level ATC categories, optionally wherein the multiple pairs of first-level ATC categories include LL, NN, and optionally one or more of JJ, CC, JS, DR, DS, CJ, AA, AM, AN, and BN.
12. The method according to any preceding claim, wherein the machine learning model has been trained using a training dataset comprising at least 1,000, 2,000, or 3,000 unique drug combinations, and / or wherein, when evaluated on unseen data, the trained machine learning model has at least 90% specificity and 70% sensitivity in predicting synergistic combinations, and / or wherein the method further comprises applying one or more steps selected from: missing value imputation, filtering, standardization, and normalization before providing the values of the plurality of features as input to the machine learning model.
13. A method for providing a tool for predicting the effects of drug combinations, the method comprising: For each of a plurality of training drug combinations comprising a first drug and a second drug, obtain: (i) values of a plurality of features associated with the drug combination; and (ii) information indicating whether the drug combination is synergistic; as well as The training data is used to train a machine learning model to predict whether a drug combination is likely to be synergistic, wherein the machine learning model is trained to take the values of the plurality of features for the drug combination as input and produce an indication as output of whether the drug combination is likely to be synergistic. The plurality of features includes at least one feature that characterizes the similarity between one or more indications or diseases associated with one or more targets of the first drug and one or more indications or diseases associated with one or more targets of the second drug.
14. A method for identifying synergistic drug combinations, the method comprising: Multiple drug candidate combinations were obtained; The effect of each of the plurality of candidate drug combinations is predicted using the method according to any one of claims 1 to 13; as well as The results of the predictions are used to select a subset of the plurality of candidate drug combinations for experimental testing; and optionally further include testing the selected subset of the plurality of candidate drug combinations using one or more in vitro assays.
15. A system comprising: At least one processor; as well as At least one non-transitory computer-readable medium containing instructions that, when executed by the at least one processor, cause the at least one processor to perform the method according to any one of claims 1 to 13.
