Methods for identifying drugs for patient-specific treatment

Abstract

The present invention provides a computer-implemented method for identifying a drug for treating a patient, comprising the steps of: a. providing a dataset of expression values ​​of biological markers from samples taken from said patient; b. selecting a plurality of drugs; n For each of the multiple drug response distance values ​​D n wherein D for each drug d is the difference between the distribution of expression of biological markers of susceptibility to drug d relative to the distribution of expression of biological markers of resistance to drug d; c. providing one or more trained predictive models; wherein the one or more trained predictive models are used for at least as many drugs d as in step b. n Multiple drug response distance values ​​D n and one or more trained predictive models are trained on the plurality of drugs d in order of predicted efficacy in the sample taken from the patient. n d. the plurality of drug response distance values ​​D obtained in step b. n into the one or more trained predictive models; and e. using one or more of the trained predictive models to rank the plurality of drugs d in order of predicted efficacy in the patient. n providing a drug ranking from the

Description

[Technical Field]

[0001] The present invention relates to a computer implementation method for identifying drugs to treat a patient. The present invention provides a method for ranking drugs in order of their predicted efficacy in a particular patient, enabling treatment selection for that patient. The present invention is particularly useful in the field of cancer treatment. [Background technology]

[0002] Even cancers originating from the same histological and pathological classification exhibit a high degree of genetic and phenotypic variability among individuals. In fact, this heterogeneity is said to be the reason why cancer patients have a wide range of responses to treatment. To address this issue, the field of personalized medicine aims to identify personalized therapeutic interventions by measuring the amount of biomarkers that indicate the effectiveness of specific drugs or drug combinations.

[0003] Current methods for stratifying patients for treatment primarily rely on the association between genetic markers and drug responses, although protein markers are also employed; for example, HER2 and estrogen receptor expression directly influence breast cancer treatment. While current genetic indicators used in companion diagnostics can increase the proportion of patients for whom a given treatment is beneficial, mutations and other genetic abnormalities often lead to inaccurate stratification, resulting in high frequencies of false positives and false negatives. Consequently, while patient selection using genetic biomarkers may increase the overall efficacy of a given treatment, these markers often exhibit low precision in accurately stratifying individual patients for treatment. These findings can be explained by the complex biological context of cancer, where multiple biochemical pathways interact and contribute to the oncogenic phenotype (Casado et al., 2013b; Klempner et al., 2013).

[0004] The application of machine learning (ML) to biomedicine has the potential to revolutionize how cancer is diagnosed and treated in the future. Projects such as Cancer Target Discovery and Development (CTD2), DepMap, and Genomics of Drug Sensitivity in Cancer (GDSC) are evaluating ML as a means of predicting drug responses by associating genomic features, gene expression patterns, and copy number variations with drug sensitivity. However, despite anecdotal evidence suggesting that proteomics-derived features may predict drug responses more accurately than genome substitutes, this approach has not been systematically applied using large-scale proteomics and phosphorylated proteomics data (Casado et al., 2013a; Casado et al., 2018; Frejno et al., 2017; Paulitschke et al., 2019; van Alphen et al., 2020). Compared to other omics techniques, a limitation of proteomics and phosphorylated proteomics by liquid chromatography combined with tandem mass spectrometry (LC-MS / MS) is low. Most proteomics methods also involve comparing proteins after chemical or metabolic labeling, thus limiting the number of samples that can be directly compared and used as input for ML model generation (Roumeliotis et al., 2017). Furthermore, because labeling methods measure the amount of protein or phosphorylation site as a ratio rather than an absolute amount, it may be difficult to validate drug response models constructed using labeled proteomics data and subsequently implement them in validation datasets and clinical settings.The emergence of label-free and moderate-sample-throughput proteomics techniques (Cutillas, 2017; Leutert et al., 2019; Montoya et al., 2011; Rudolph et al., 2016), along with the recent availability of systematic drug response profiles for numerous cell lines and drugs (Basu et al., 2013; Smirnov et al., 2018; Yang et al., 2013), now enables the use of proteomics and phosphorylated proteomics data as input to predictive models of drug responses. Thus, evaluating the performance of ML models constructed using proteomics data as input is timely and essential for assessing the accuracy and potential of proteomics to advance the field of precision medicine. [Overview of the Initiative]

[0005] The inventors investigated an approach called Drug Ranking Using ML (DRUML) to construct and unify ML models. DRUML uses a set of proteomics, phosphorylated proteomics, and transcriptomics features to create a ranked list based on the efficacy of drugs, for example, in reducing cancer cell proliferation. DRUML's ability to predict drug rankings within a cancer cell population without the need for comparison with standard samples is crucial for the clinical implementation of ML and fulfills a core objective of precision medicine.

[0006] Therefore, in a first aspect, the present invention provides a computer implementation method for identifying a drug for treating a patient, the method comprising the following steps: a. To provide a dataset of expression levels of biological markers from samples taken from the aforementioned patients; b. Multiple drugs d n Multiple drug reaction distance values ​​D for each of them nusing the dataset to calculate; where D for each drug d is the difference between the distribution of the expression of the biological marker of sensitivity to drug d and the distribution of the expression of the biological marker of resistance to drug d; c. providing one or more trained prediction models; where the one or more trained prediction models are for a plurality of drugs d equal to at least the same number as in step b n for a plurality of drug response distance values D n for which they are trained; and the one or more trained prediction models are trained to provide a ranking of the drugs from the plurality of drugs d in order of the predicted efficacy in the sample taken from the patient; n and are trained to provide a ranking of the drugs from the plurality of drugs d in order of the predicted efficacy in the patient; d. inputting the plurality of drug response distance values D obtained in step b n into the one or more trained prediction models; and e. using one or more of the trained prediction models to provide a ranking of the drugs from the plurality of drugs d in order of the predicted efficacy in the patient. Detailed description of the invention n

[0007] In a first aspect, the present invention relates to a computer-implemented method for identifying a drug for treating a patient. The method of the first aspect may alternatively be represented as a method for ranking drugs in order of efficacy for treating a particular patient, particularly as a computer-implemented method. The method includes calculating a plurality of drug response distance values D for each of a plurality of drugs d n where D for each drug d is the difference between the distribution of the expression of the biological marker of sensitivity to drug d and the distribution of the expression of the biological marker of resistance to (the same) drug d. The first step a. of the method of the present invention is to provide a dataset of the expression values of biological markers from a sample taken from the patient. n where D for each drug d is the difference between the distribution of the expression of the biological marker of sensitivity to drug d and the distribution of the expression of the biological marker of resistance to (the same) drug d.

[0008] The first step a. of the method of the present invention is to provide a dataset of the expression values of biological markers from a sample taken from the patient.

[0009] Any large-scale omics dataset can be used as an input to the method of the present invention, which may also be referred to herein as DRUML. The dataset of expression values used in the method of the first aspect of the present invention can be obtained, for example, from experiments in phosphoproteomics, proteomics or transcriptomics, and thus can correspond to counts of phosphorylated proteins or peptides, proteins, peptides, or gene transcripts. Transcriptomics data can be obtained, for example, using RNA sequencing (RNA-seq). RNA-Seq data can be obtained from repositories such as the DepMap repository (Corsello et al., 2020).

[0010] For example, phosphoproteomics data can be obtained using the inventors' TIQUAS (targeted and in-depth quantification of s ignalling) technology as described in International Publication No. WO 2010 / 119261 pamphlet (International Patent Application No. PCT / GB2010 / 000770), the entirety of which is incorporated herein by reference. This technology enables sensitive, rapid and comprehensive quantification of modified peptides. This method can simultaneously measure the amounts of thousands of phosphorylation sites and other modifications on proteins in one simple quantification. Other computer programs and workflows, such as MaxQuant (Nature Biotechnology 26, 1367-1372 (2008)), may be used for peptide quantification, and these are compatible with the present invention. Other sources of proteomics and phosphoproteomics data available to the public include Jarnuczak et al., 2019 and Piersma et al., 2015.

[0011] The method of the present invention may include the step of obtaining a dataset of expression levels of biological markers from a sample taken from the patient. This can be carried out using phosphoproteomics, proteomics, or transcriptomics techniques, as described herein. The step of taking a sample from the patient typically does not constitute part of the method.

[0012] Next, the dataset of biological marker expression values ​​from the sample taken from the patient is used in step b. for multiple drugs d. n Multiple drug reaction distance values ​​D for each of them n This is used to calculate the difference between the distribution of expression of biological markers for susceptibility to drug d and the distribution of expression of biological markers for resistance to drug d for each drug d.

[0013] The D metric is essentially a measure of the distribution of susceptibility markers relative to resistance markers. D can be defined as the difference in the overall expression of markers positively associated with drug susceptibility compared to markers positively associated with drug resistance in a sample. Calculating D involves analyzing the expression of biological markers that increase in expression in biological samples susceptible to drug d, and the expression of biological markers that increase in expression in biological samples resistant to the same drug d.

[0014] As used herein, “biological marker of susceptibility to drug d” means a biological marker whose expression is consistently increased in a biological sample (typically cells) susceptible to drug d compared to its expression in a biological sample (typically cells) resistant to drug d. As used herein, “biological marker of resistance to drug d” means a biological marker whose expression is consistently increased in a biological sample (typically cells) resistant to drug d compared to its expression in a biological sample (typically cells) susceptible to drug d.

[0015] Such biological markers of susceptibility and tolerance to a particular drug d are collectively referred herein to as empirical markers of drug responses (EMDRs). Such biological markers may be identified using a computer program such as the R package Limma (http: / / bioconductor.org / packages / release / bioc / html / limma.html) described in the examples herein. The biological markers of susceptibility and tolerance are typically identified before carrying out the method of the first aspect of the present invention, and are typically identified using multiple samples. A dataset of expression values ​​of biological markers from patient samples typically includes expression values ​​for any of the biological markers already identified as biological markers of susceptibility and tolerance to drug d that are present in the sample.

[0016] The examples herein describe the identification of various biological markers of susceptibility and resistance to a particular drug, referred to herein as EMDR. Thus, in one embodiment, the biological marker of susceptibility and / or resistance may be any one or more of those identified in Figure 3. Figure 3C shows the phosphorylation sites, proteins, and transcripts most frequently identified as biological markers of susceptibility or resistance (see charts titled “Phosphorylated Proteomics,” “Proteomics,” and “Transcriptomics,” respectively), and the present invention encompasses the use of any one or more of these biological markers.

[0017] The biological sample referred to in relation to the calculation of D is a cell line, which may optionally be a cancer cell line. Alternatively, such a biological sample may be primary tissue, which may optionally be a cancer biopsy, obtained directly from a patient or from in vivo or ex vivo experiments. There may be several advantages to training machine learning models using primary cells, such as primary cancer cells, isolated directly from tissue. This is because primary cells are not immortalized, and therefore, compared to cell lines, primary cells are more similar to cells obtained directly from a patient in terms of cell morphology and biological function.

[0018] The present invention has the advantage that D values ​​do not require comparison with other samples. Once EMDRs are identified, they can be quantified in the sample and used to calculate D values ​​in biological samples (e.g., tumor biopsies in clinical testing when other samples for comparison are not available).

[0019] One exemplary method for calculating D is shown in the examples herein and described in relation to Figure 1A. Thus, in one embodiment, D is calculated for each drug d using: D d =[S Q2 -R Q2 ]+[S Q3 -R Q3 ],

[0020] Here, S Q2 and S Q3 These are the median and third-quartile expression values ​​of biological markers of sensitivity to drug d; and R Q2 and R Q3 These are the median and third-quartile expression values ​​of biological markers of resistance to drug d.

[0021] Any suitable alternative method, including the z-value, the Kolmogorov-Smirnov test, and related methods such as the following, can also be used to calculate D: D d =[Median(M S )+Q3(MS )]-[Median(M R )+Q3(M R )]

[0022] Here, M S = Expression of biological markers of susceptibility to drug d (increased protein, phosphorylation site, or RNA transcript in a biological sample (typically a cell) susceptible to drug d), and M R = Expression of a biological marker of resistance to drug d (an increased protein, phosphorylation site, or RNA transcript in a biological sample (typically a cell) of drug d resistance).

[0023] Regarding determining which biological samples (e.g., cell lines) are resistant to or susceptible to a particular drug d, this can be carried out using any suitable means. For example, the median cutoff of the area on the curve (AAC) can be used, as described in the examples herein. Alternatively, susceptibility data can be obtained using databases such as PharmacoDB (Smirnov et al., 2018).

[0024] The next step c of the method of the present invention is to provide one or more trained predictive models, wherein the one or more trained predictive models are the same number of drug response distance values ​​D as in step b. n They are trained in this.

[0025] One or more trained predictive models (such as machine learning models) can predict multiple drugs d n Multiple drug reaction distances D n The model is trained on process b. and at least the same number of drugs d. n Multiple drug reaction distances D n Training can be provided on multiple drugs d nThe group of drugs can be any number of drugs, but the method of the present invention is particularly useful in providing a ranking of drugs in order of expected efficacy in a particular patient when there are many drugs to choose from. For example, the group of drugs may consist of 100 to 500 drugs, for example, 150 to 450, 200 to 400, or 300 to 350 drugs.

[0026] Typically, multiple drug-to-drug reaction distance values ​​D n This includes some of the D values ​​that are most positively correlated and most negatively correlated with the D value for drug d. For example, an equal or different number of D values ​​that are positively correlated and D values ​​that are negatively correlated with the D value for drug d may be used. In the examples herein, seven D values ​​that are most positively correlated and seven D values ​​that are most negatively correlated with the D value for drug d were used, but any suitable number of D values ​​that are most positively correlated and most negatively correlated with the D value for drug d may be used, for example, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 or 20, each of which may be the D values ​​that are most positively correlated and most negatively correlated with the D value for drug d.

[0027] The trained predictive model may be a machine learning model and may be obtained using statistical learning methods. Any suitable method can be used to train the predictive model. For example, one or more trained predictive models may be trained using a learning algorithm selected from Random Forest (RF), Cubist, Bayesian estimation of generalized linear models (BGLM), Partial Least Squares (PLS), Principal Component Regression (PCR), Deep Learning (DL), and Neural Network (NNET) learning algorithms. The inventors found that PCR and RF had lower errors on verification datasets obtained from independent laboratories, as described in the examples herein, but DL was the best performer when using in-house training and verification datasets.

[0028] One or more trained predictive models (such as machine learning models) predict the efficacy of the multiple drugs d in the order of their predicted efficacy in the sample taken from the patient. n It is trained to provide a ranking of drugs from a plurality of drugs d in order of expected potency in the patient. n By providing a ranking of drugs from a set of drugs, it enables the prediction of drug efficacy in a patient. In other words, one or more trained predictive models rank the multiple drugs d in order of their predicted efficacy in the patient. n The method is trained to provide drug rankings from a set of trained predictive models. This means that the method of the present invention can be used to inform treatment decisions for a particular patient. The drug rankings provided by one or more trained predictive models may take the form of a report ranking drugs in order of their predicted efficacy in a patient, for example. Thus, the output of the method of the first aspect of the present invention may be a report ranking drugs in order of their predicted efficacy in a patient.

[0029] A first aspect of the present invention is a method for identifying a drug for treating a patient. This method is particularly useful in identifying a drug for treating a patient who has been diagnosed with cancer or is suspected of having cancer.

[0030] Cancer can be any type of cancer, such as lymphoma, leukemia (e.g., acute myeloid leukemia), or solid tumor (e.g., esophageal cancer or hepatocellular carcinoma). Therefore, the sample taken from the patient is typically a biopsy from the tumor.

[0031] Therefore, each drug d is typically an anticancer agent (i.e., a drug specifically developed to treat cancer) or a different type of drug repurposed for the treatment of cancer. Anticancer agents include kinase inhibitors, such as AGC kinases, such as protein kinase A (PKA), protein kinase B (PKB) (also known as Akt), protein kinase C (PKC), and protein kinase G (PKG); tyrosine kinases; tyrosine kinase-like kinases; calcium / calmodulin-dependent protein kinases; casein kinase 1 group; CMGC group, such as CDK, MAPK, GSK3, and CLK kinases; and inhibitors of human protein kinases selected from the group consisting of STEs, which are homologs of yeast sterile 7, sterile 11, and sterile 20 kinases. Kinase inhibitors suitable for use according to the present invention include AZD-5438 (CDK2i;), GF-109203X (PKCαi; Tocris), PF-3758309 (PAKi; Calbiochem), trametinib (MEKi; Selleckchem), MK-2206 (AKTi; Selleckchem), KU-0063794 (mTORi; Chemdea), TAK 715 (P38αi;), PKC-412 (PKC / Flt3i; Tocris), TBB (CK2i;), PF-3758309 (PAKi), and C4945 (CK2i;). However, the DRUML method is not particularly limited to any specific type of anticancer agent and has been demonstrated to be effective in ranking drugs with diverse mechanisms of action.

[0032] The method according to the first aspect of the present invention involves the plurality of drugs d nThe method may include an additional step f. to identify a drug for treating a patient by selecting one of the highest-ranked drugs according to a ranking of drugs in order of predicted potency. This could be the highest-ranked drug. However, in practice, the method of the first embodiment is not typically used alone to identify a drug for treating a patient. Typically, DRUML can assist in prioritizing drugs by complementing information obtained from clinicopathological parameters and mutational analysis. Thus, the method may be used in conjunction with considering other factors for identifying a drug for treating a patient, such as cost, safety, and / or regulatory issues (e.g., recommendations from NICE in the UK, EMA in Europe, and FDA in the US). Therefore, drug ranking using the method of the present invention may inform treatment decisions made by physicians or using other computer-implemented methods such as ML methods. Thus, the drug selected to treat a patient may be the second-highest ranked drug, or, for example, the third, fourth, fifth, sixth, seventh, eighth, ninth, or tenth-highest ranked drug, according to the ranking obtained by the method of the first embodiment of the present invention.

[0033] The present invention also extends to a method for training one or more predictive models to carry out a method according to a first aspect of the present invention.

[0034] Therefore, in a second aspect, the present invention provides a plurality of drugs d in the order of expected efficacy in a sample taken from a patient. n This provides a computer implementation method for training one or more predictive models to provide drug rankings, the method including: i. Multiple drugs d n Multiple drug reaction distance values ​​D for each of them n Provide training data including; where D for each drug d is the difference between the distribution of expression of biological markers of susceptibility to drug d and the distribution of expression of biological markers of susceptibility to drug d; ii. Train one or more predictive models using training data, and classify the multiple drugs d in order of their predicted efficacy in samples taken from patients. n To provide a ranking of drugs from [source].

[0035] Embodiments of a second aspect of the present invention are as described above with respect to the first aspect of the present invention.

[0036] As described above with respect to the first aspect of the present invention, one or more predictive models represent the plurality of drugs d in order of predicted efficacy in the patient. n By providing a ranking of drugs from, the model is trained to predict the effectiveness of drugs in a patient. In other words, step ii of the method of a second aspect of the present invention is to train one or more predictive models using training data and rank the plurality of drugs d in order of their predicted effectiveness in the patient. n This includes providing a ranking of drugs from a selection of drugs.

[0037] A method according to a second aspect of the present invention may also be described as a method for providing a trained predictive model for use according to the first aspect of the present invention. A method according to a second aspect of the present invention may be used to train multiple predictive models.

[0038] Therefore, the method of the first aspect of the present invention may include a step of training one or more predictive models, as described in relation to the second aspect of the present invention. In this embodiment, the method of the first aspect of the present invention may be expressed as follows: a. To provide a dataset of expression levels of biological markers from samples taken from the aforementioned patients; b. Multiple drug reaction distance values ​​D for each of the multiple drugs d. n The aforementioned dataset is used to calculate; where D for each drug d is the difference between the distribution of expression of biological markers of susceptibility to drug d and the distribution of expression of biological markers of susceptibility to drug d; c. Provide one or more trained predictive models, where one or more trained predictive models are trained using a method according to a second aspect of the present invention; d. The plurality of drug reaction distances D obtained in step b. n Inputting the above into one or more trained predictive models; and e. Using one or more trained predictive models, the multiple drugs d are ordered in order of their predicted efficacy in the patient. n To provide a ranking of drugs from [source].

[0039] As described above with respect to the first aspect of the present invention, the drug ranking provided by one or more trained predictive models may take the form of a report ranking drugs in order of their predicted efficacy in patients. Thus, the output of the method of the first aspect of the present invention may be a report ranking drugs in order of their predicted efficacy in patients.

[0040] In a third aspect, the present invention provides a computer-readable storage medium, or a medium storing instructions for implementing a method according to the first or second aspect of the present invention.

[0041] In a fourth aspect, the present invention provides a system for identifying a drug for treating a patient, the system comprising memory and one or more processors, the one or more processors configured to cause the method of the first aspect of the present invention.

[0042] In a fifth aspect, the present invention provides a method for treating a patient who has cancer or is suspected of having cancer, the method comprising identifying a drug for treating the patient by a method of a first aspect of the present invention, or using a system of a third aspect of the present invention, and administering the drug to the patient. In this aspect, the drug is an anticancer agent as defined herein.

[0043] The treatment method may be for human or animal subjects, and the present invention extends equally to use in both human medicine and / or veterinary medicine. The drug is preferably administered to the individual in a “therapeutically effective dose,” which is sufficient to benefit the individual and / or to improve, eliminate or prevent one or more symptoms of a disease. As used herein, “treatment” includes any regimen that may benefit a human or non-human animal, preferably a mammal, such as economically important mammals such as cattle, sheep, goats and pigs. The treatment may be for an existing condition or it may be preventive (preventive treatment).

[0044] The inventors have designed an approach called Drug Ranking Using Machine Learning (DRUML) that ranks drugs based on their predicted efficacy within a cancer model. In contrast to the inventors' previous “K-score” approach described in PCT application number PCT / EP2016 / 077845 (published as WO 2017 / 085116), which is limited to kinase inhibitors, DRUML predicts and ranks drug efficacy within a given cancer cell model. This is an important distinction because the ability to predict the best drug to treat a given patient without the need to compare it with samples from other patients is crucial for the clinical implementation of ML in selecting the best treatment for a given patient. Thus, the invention serves a core objective of precision medicine.

[0045] The present invention also has advantages over the aforementioned approaches, including the ability to predict whether a patient is sensitive to or tolerant of a particular drug, as well as to rank drugs according to their predicted efficacy in a particular patient.

[0046] Preferred features relating to second and subsequent aspects of the present invention are applied mutatis mutandis to the first aspect. It will be understood that all embodiments described herein are broadly applicable and can be combined as appropriate with any and all other consistent embodiments. Such combinations are considered to fall within the scope of the present invention. [Brief explanation of the drawing]

[0047] The present invention will now be further described with reference to the following examples, which exist for illustrative purposes only. The following figures are referenced in the examples: [Figure 1]An overview of Drug Ranking Using Machine Learning (DRUML). A) Drug response (AAC) values ​​were modeled for 659 drugs using different DL / ML methods. Of these, 466 models were generated using at least one learning algorithm. The input for DL / ML model generation is the mean values ​​of empirical markers of drug responses (EMDR), which are combined to derive the distance metric D. For each drug d and each biological sample b, Dd, b = [SQ2 - RQ2] + [SQ3 - RQ3], where SQ2 and SQ3 are the median and third-quartile expression values ​​of the increased empirical markers in cells sensitive to a given drug, respectively; RQ2 and RQ3 are the median and third-quartile expression values ​​of the increased empirical markers in cells resistant to the same drug. B) LC-MS / MS workflow for generating proteomics and phosphateproteomics datasets used to train DRUML. C) Principal component analysis of proteomics, phosphorylated proteomics, and drug response datasets for each cell line used to train DRUML. D, E) Measurement of empirical markers of drug sensitivity and resistance to barasertib. Cell lines are divided into sensitive (dark gray) and resistant (light gray) groups based on barasertib response (AAC value) and limma (E), which is used to identify response markers by resampling. F) Distribution of empirical markers of resistance and sensitivity is shown for resistant cell line (OCI-M1) and sensitive cell line (P31-FUJ) to barasertib or intermediate phenotype (GDM-1). Median (Q2) and third quartile (Q3) are marked for the EMDR distribution of barasertib in P31-FUJ. G) Barasertib D value for each cell line calculated from the EMDR distribution shown in F. [Figure 2]Dimensionality reduction using empirical markers of drug response. A, B, C) Correlation between cell line response to valacertib and distance (D_valasertib) metrics obtained from phosphorylated proteomics (A), proteomics (B), and RNA-seq (C) data. As shown in Figure 1A and herein, distance values ​​were calculated by combining the expression of empirical markers of sensitivity and resistance. R and p values ​​were obtained by Pearson's test. D) Expression of the top 14 drug marker distance values ​​that correlate both positively and negatively with drug sensitivity to valacertib. Rows are organized in order of valacertib sensitivity (AAC). The intensity and size of the dot color are proportional to the distance value normalized from 0 to 1. E) Overall correlation between each distance marker and valacertib sensitivity. The size of the dot is proportional to the -log10 Spearman p value. [Figure 3] Summary of systematic empirical markers of responses to over 400 drugs. A) Number of empirical susceptibility and resistance markers identified per drug. Not all drugs were profiled in all cell lines; markers were successfully identified for 445–466 drugs with sufficient data points. B) Frequency of phosphorylation sites (top), proteins (middle), and transcripts (bottom) identified as empirical drug response markers. C) Phosphorylation sites, proteins, and transcripts most frequently identified as markers of susceptibility or resistance are shown. D) Principal component analysis using empirical response markers as input. Representative drug classes are annotated. [Figure 4] Performance of predictive models for response to valacertib. To demonstrate the DRUML for a single drug, the performance of predictive models for valacertib using D values ​​from different datasets was compared. A) Comparison of measured and predicted responses readjusted by eight different learning methods using phosphorylated proteomics data as input. Solid, dashed, and dotted lines indicate the 0%, 10%, and 20% error boundaries, respectively. B) Similar to (A), but with D values ​​obtained from proteomics data. C) Comparison of standard errors (SE) in a validation set from the data in (A), (B), and Figure 11. P values ​​were calculated using the Kruskal-Wallis test. [Figure 5] Performance and accuracy of DRUML for ranking drugs based on efficacy. A) Total number of models obtained from each data input. B) Training and validation errors of each model binned by the ML method and input dataset. C) Comparison of measured and predicted drug response values, generated by DL analysis of phosphorylated proteomics distance values ​​in the validation dataset. Each data point represents a drug prediction, and the shape of the data point is coded by the mechanism of action of the drug. Each cell line was analyzed three times. Dotted lines indicate 1 with a slope or 0 intercept. D) Absolute validation error of the learned model binned by the ML method and input dataset. [Figure 6] Accuracy evaluation of DRUML for ranking drugs based on efficacy using independent phosphoproteomics datasets. DRUML was used to predict drug response in colorectal cancer (CRC) cell lines represented by phosphoproteomics data obtained from Piersma et al (Jimenez lab, PRIDE PXD001550). A) Comparison of measured and predicted drug response within the cell model. Each data point represents a drug prediction. B) Comparison of measured and predicted drug response binned by drug developmental stage. C) Distribution of absolute error validation by the ML model. D, E) Number (D) and percentage (E) of accurate predictions within absolute errors of 0.05, 0.1, 0.15, and 0.25. [Figure 7]Accuracy evaluation of DRUML for ranking drugs based on efficacy using independent proteomics datasets derived from 47 tumor models and 8 pathologies. Drug responses in cell lines, shown using proteomics data from 12 different laboratories and collected by Jarnuczak et al (Vizcaino lab, PRIDE PXD013455), were predicted using DRUML. A) Comparison of measured and predicted drug responses within cell models using a random forest method. Each data point represents a drug prediction. B) Distribution of validation absolute errors in each tumor cell model from DRUML analysis using a random forest method. C) Comparison of the percentage of accurate predictions at error cutoffs of 0.05, 0.1, 0.15, and 0.25 returned by different ML methods. [Figure 8] Qualitative evaluation of proteomics and phosphoproteomics data. A) Number of peptides, phosphorylated peptides, and proteins identified from proteomic analysis of 48 cell lines of acute myeloid leukemia, esophageal cancer, and hepatocellular carcinoma shown in Figure 1. B) Number of phosphorylated peptides (left) and unmodified proteins (right) quantified in each replica from all 48 cell lines analyzed. [Figure 9] The number of empirical markers of drug responses (EMDRs) identified for each drug. Markers of resistance and sensitivity refer to increased or decreased phosphorylation sites, proteins, or transcripts in cells resistant to a given drug, respectively. [Figure 10] PCA is based on empirical markers of drug responses common to different drugs, identified from AML cell lines. [Figure 11] Performance of a parasertib reaction model using transcriptome data. [Figure 12] Comparison of drug response values ​​between measurements and predictions generated by DL analysis of phosphorylated proteomics distance values ​​in the training dataset. [Figure 13] Qualitative evaluation of phosphorylated proteomics data from eight colorectal cancer cell lines obtained from Piersma et al. [Figure 14] Evaluation of the accuracy of DRUML for ranking drugs based on efficacy using independent proteomics datasets derived from 47 tumor models and 8 pathologies. Results of DRUML-based drug response prediction using PCR (left) and DL (right) models are shown. [Examples]

[0048] Details of the experimental model and subjects Cell lines were obtained from various repositories and cultured according to the recommended cell culture conditions provided by each source.

[0049] AML cell line AML cell lines AML-193, CMK, K-052, Kasumi-1, KG-1, HEL, ME-1, ML-2, MOLM-13, MONO-MAC-6, MV4-11, OCI-AML2, OCI-AML3, OCI-AML5, P31 / FUJ, PL-21, SIG-M5, SKM-1, and THP-1 were derived from male patients, while GMD-1, KMOE-2, HL-60, M-07e, NB-4, and NOMO-1 were derived from female patients. The sex of patients from whom OCI-M1 cells were derived was not identified in the DSMZ-German Collection of Microorganisms and Cell Cultures GmbH database.

[0050] In short, SIG-M5 and M-07e cell lines were maintained in IMDM supplemented with 20% FBS and 1% penicillin / streptomycin (P / S). M-07e cells were supplemented with 10 ng / mL GM-CSF. OCI-M1 and AML-193 cells were grown in IMDM supplemented with 10% FBS and 1% P / S. AML-193 cells were supplemented with 5 ng / mL GM-CSF. OCI-AML2, OCI-AML3, and OCI-AML5 cell lines were grown in α-MEM supplemented with 20% FBS and 1% P / S. OCI-AML5 cells were also supplemented with 5 ng / mL GM-CSF. K-052 cells were maintained in α-MEM supplemented with 10% FBS and 1% P / S. GDM-1 and SKM-1 cells were grown in RPMI-1640 supplemented with 20% FBS and 1% P / S. SKM-1 cells were also supplemented with 1 ng / mL GM-CSF. All other AML cell lines were maintained in RPMI-1640 supplemented with 10% heat-inactivated FBS and 1% (RPMI / FBS medium). All cell lines were maintained under humidified conditions at 37°C and 5% CO2. Cell density was 0.5–1.5 × 10⁶. 6 The concentration was maintained at cells / mL.

[0051] For proteomics and phosphorylated proteomics analysis, AML cell lines were seeded in IMDM medium supplemented with 10% FBS and 1% P / S in T75 flasks (20 × 10 in 10 mL). 6 Cells were incubated at 37°C and 5% CO2 under humid conditions for 3 hours. The cell suspension was then transferred to a 15 mL Falcon tube and centrifuged at 5°C and 1500 rpm for 5 minutes. The supernatant was removed and the cells were washed twice with ice-cold DPBS supplemented with phosphatase inhibitors (1 mM NaF, 1 mM Na3VO4). During washing, the cells were centrifuged at 5°C and 1500 rpm for 5 minutes. The cell pellet was transferred to a 1.5 mL Protein LoBind tube, rapidly frozen in dry ice, and stored at -80°C. Biologically independent replication of each cell line (n=3) was performed on different days.

[0052] liver cancer cell line The hepatocellular carcinoma cell lines HEP 3B2.1-7, HEP G2, JHH2, JHH4, SK-HEP-1, SNU182, SNU-398, SNU-423, SNU-449, and SNU-475 were derived from male patients, while the cell line SNU-387 was derived from female patients. The sex of the patients from whom the PLC / PRF / 5 cells were derived was not identified in the American Type Culture Collection (ATCC) database.

[0053] Cell lines SNU-387, SNU-423, SNU-182, SNU-398, SNU-475, and SNU-449 were maintained in RPMI-1640 supplemented with 1 mM sodium pyruvate, 10% FBS, and 1% P / S. Cell lines HEP 3B2.1-7, HEP G2, JHH2, JHH4, and PLC / PRF / 5 were grown in MEM supplemented with 1 mM sodium pyruvate, 2 mM L-glutamine, 1X NEAA solution, 10% FBS, and 1% P / S, and the SK-HEP-1 cell line was maintained in MEM supplemented with 1 mM sodium pyruvate, 2 mM L-glutamine, 1X NEAA solution, 20% FBS, and 1% P / S. All cell lines were maintained under humidified conditions at 37°C and 5% CO2. Cell density is 0.2-0.4 × 10⁶ per mL 6 The cells were maintained between each other, and the culture medium was changed 2-3 times per week.

[0054] For proteomics and phosphoproteomics analysis, hepatocyte lines were seeded in Petri dishes (0.3-3.74 × 10¹³ cells in 20 mL). 6Cells were kept in an incubator for 3–8 days under humid conditions at 37°C and 5% CO2 until cell confluence reached approximately 80%. 1.5 hours prior to cell harvesting, the medium was replaced with fresh complete medium. Subsequently, the cells were washed three times with cold DPBS supplemented with 1 mM NaF and 1 mM Na3VO4, and lysed in 500 μL of urea buffer (8 M urea in 20 mM HEPES at pH 8.0 supplemented with 1 mM NaF, 1 mM Na3VO4, 1 mM Na4P2O7, and 1 mM β-glycerophosphate). After cell collection with a scraper, the lysate was rapidly frozen in a protein LoBind tube and stored at -80°C for further sample preparation.

[0055] Esophageal cancer cell line The esophageal cancer cell lines KYSE-70, KYSE-140, KYSE-410, KYSE-450, and OE-19 were derived from male patients, while the cell lines COLO-680N, KYSE-150, KYSE-510, KYSE-520, and EO-33 were derived from female patients.

[0056] The KYSE-150 cell line was maintained in RPMI-1640 supplemented with 49% F12, 2% FBS, and 1% P / S, and the KYSE-450 cell line was grown in RPMI-1640 supplemented with 45% F12, 10% FBS, and 1% P / S. All other esophageal cell lines were maintained in RPMI-1640 supplemented with 10% FBS and 1% P / S. All cell lines were maintained under humidified conditions at 37°C and 5% CO2. Cell density was 0.1–0.25 × 10⁶. 6 The concentration was maintained at cells / mL.

[0057] For proteomics and phosphoproteomics tests, esophageal cell lines were seeded in Petri dishes (1.5-3.5 × 10¹³ cells per 10 mL). 6The cells were kept overnight in an incubator at 37°C and 5% CO2 under humid conditions. The cells were then washed twice with cold DPBS supplemented with 1 mM NaF and 1 mM Na3VO4, dissolved in 500 μL of urea buffer (8M urea in 20 mM HEPES at pH 8.0 supplemented with 1 mM NaF, 1 mM Na3VO4, 1 mM Na4P2O7, and 1 mM β-glycerophosphate), rapidly frozen, and stored at -80°C until further processing.

[0058] Method details Sample preparation for phosphoproteomics and proteomics analysis Phosphorylated proteomics and proteomics analyses were performed as previously described (Casado et al., 2018; Montoya et al., 2011; Wilkes and Cutillas, 2017). AML cell pellets were lysed in 320 μL of urea buffer (8M urea in 20 mM HEPES at pH 8.0, supplemented with 1 mM Na3VO4, 1 mM NaF, 1 mM Na4P2O7, and 1 mM β-glycerophosphate). AML cell lysates were homogenized by sonication for 90 cycles (30 seconds on, 30 seconds off), while thawed esophageal and hepatocyte cell lines were homogenized in Diagenode Bioruptor® for 15 cycles (30 seconds on, 40 seconds off), and insoluble material was removed by centrifugation.

[0059] Proteins were quantified using a BCA protein assay. Next, 300 μg of protein was subjected to cysteine ​​reduction and alkylation by continuous incubation with 10 mM dithiothreitol (DDT) and 16.6 mM iodoacetamide (IAM) for 1 hour and 30 minutes, respectively, at 25°C with stirring. Trypsin beads (50% slurry of TLCK-trypsin) were equilibrated by washing three times with 20 mM HEPES (pH 8.0). The urea concentration in the protein suspension was reduced to 2 M by adding 900 μL of 20 mM HEPES (pH 8.0), and 100 μL of the equilibrated trypsin beads were added. The sample was incubated overnight at 37°C. The trypsin beads were removed by centrifugation (2000 xg, 5°C for 5 minutes), and the sample was divided into 250 μg for phosphoproteomics analysis and 50 μg (200 μL) for proteomics analysis.

[0060] For phosphoproteomics analysis, the peptide solution was desalted using an Oasis HLB cartridge (Waters) as directed by the manufacturer. Briefly, the cartridge was placed in a vacuum manifold device and the pressure was adjusted to 5 mmHg. The cartridge was then prepared with 1 mL of acetonitrile (ACN) and equilibrated with 1.5 mL of washing solution (0.1% trifluoroacetic acid (TFA), 2% ACN). The peptide was loaded into the cartridge and washed twice with 1 mL of washing solution. Finally, the peptide was eluted with 500 μL of glycolic acid buffer 1 (1 M glycolic acid, 5% TFA, 50% ACN). Phosphorylated peptides were concentrated using TiO2 beads. The desalted eluate was normalized to 1 mL with glycolic acid buffer 2 (1 M glycolic acid, 5% TFA, 80% ACN) and incubated with 25 μL of TiO2 buffer (containing 500 mg TiO2 beads in 500 μL of 1% TFA) at room temperature for 5 minutes. TiO2 beads were packed into an empty spin column pre-washed with ACN by centrifugation. The TiO2 beads were washed three times by centrifugation (1500 xg, 3 minutes) with 100 μL of glycolic acid buffer 2, ammonium acetate solution (containing 100 mM ammonium acetate in 25% ACN), and neutral solution (10% ACN). For phosphorylated peptide elution, the spin tip was transferred to a new tube, 50 μL of elution 1 (5% NH4OH, 7.5% ACN) was added, and the tip was centrifuged at 1500 xg for 3 minutes. The elution process was repeated a total of four times. Finally, the samples were rapidly frozen, dried in a SpeedVac vacuum concentrator, and the phosphorylated peptide pellet was stored at -80°C.

[0061] For the proteomics experiment, the peptide solution was desalted using a C18+ carbon top tip (Glygen Corporation). The tip was prepared twice with 200 μL of elution solution 2 (70 / 30 ACN / H2O + 0.1% TFA) and equilibrated twice with 200 μL of washing solution. The sample was loaded onto the top tip and washed twice with 200 μL of washing solution. For peptide elution, the tip was transferred to a new tube and the peptide was eluted three times with 250 μL of elution solution 2. In all desalting steps, the tip was centrifuged at 1500 xg at 5°C for 5 minutes. The eluted peptide solution was dried in a SpeedVac vacuum concentrator, and the peptide pellet was stored at -80°C.

[0062] mass spectrometry Mass spectrometry for the identification and quantification of proteins and phosphorylated peptides was performed by LC-MS / MS as previously described (Montoya et al., 2011; Wilkes and Cutillas, 2017). For phosphoproteomics analysis, the peptide pellet was reconstituted in 13 μL of reconstitution buffer (20 fmol / μL enolase in 3% ACN, 0.1% TFA), and 5 μL was loaded into the LC-MS / MS system. For proteomics analysis, the peptide pellet was reconstituted in 10 μL of 0.1% TFA, 2 μL of this solution was further diluted in 18 μL of reconstitution buffer, and 2 μL was injected into the LC-MS / MS system.

[0063] The LC-MS / MS platform consisted of a Dionex UltiMate 3000 RSLC coupled to a Q Exactive® Plus Orbitrap Mass Spectrometer (Thermo Fisher Scientific) via an EASY-Spray source. The mobile phase for peptide chromatographic separation consisted of solvent A (3% ACN; 0.1% FA) and solvent B (99.9% ACN; 0.1% FA). Peptides were packed into a micropre-column and separated on the analytical column using a 3%–23% B gradient for 60 minutes (for phosphorylated proteomics) or 120 minutes (for proteomics). The UPLC system supplied flow rates of 2 μL / min (loading) and 250 nL / min (gradient elution). The Q Exactive Plus was operated on a 2.1-second duty cycle. Therefore, a full-scan survey spectrum (m / z 375–1500) with a resolution of 70,000 FWHM was acquired, followed by selecting 15 of the strongest ions and acquiring HCD (high-energy collision dissociation) and MS / MS scans (200–2000 m / z) with a resolution of 17,500 FWHM. A 30-second dynamic exclusion period was allowed within a ±10 ppm m / z window.

[0064] Phosphorylated peptide and protein identification Peptide identification from MS data was automated using the Mascot Daemon 2.6.0 workflow, where Mascot Distiller v2.6.1.0 generated a peak list file (MGF) from the RAW data, and the Mascot search engine (v2.6) matched the MS / MS data stored in the MGF file to the peptides using the SwissProt database (SwissProt_2016Oct.fasta). The search allowed for peptides with an FDR of less than 1%, two trypsin misscleavages, a mass tolerance of ±10 ppm for MS scans and ±25 mmu for MS / MS scans, and allowed for carbamide methyl Cys as a fixed modification, and PyroGlu on the N-terminal Gln, oxidation of Met, and phosphorylation on Ser, Thr, and Tyr as variable modifications (phosphorylation was included only for searches performed using phosphorylated proteomics data).

[0065] Quantification and statistical analysis Quantification of phosphorylated peptides and proteins The identified peptides were quantified using a label-free procedure based on extracted ion chromatograms (XICs). Missing data points were minimized by constructing XICs across all LC-MS / MS runs for all peptides identified in at least one LC-MS / MS run (Cutillas, 2017). The XIC mass window and retention time window were ±7 ppm and ±2 minutes, respectively. Peptide quantification was performed by measuring the area under the peaks of the XICs. Individual peptide intensity values ​​in each sample were normalized to the sum of the intensity values ​​of all peptides quantified in that sample. Data points that were not quantified for a particular peptide were given a peptide intensity value equal to the minimum intensity value quantified in that sample divided by 10. In phosphorylated proteomics experiments, the phosphorylation index (ppIndex) was obtained by summing the signals of all peptide ions containing the same modification site. In proteomics experiments, the protein intensity value was calculated by adding the intensities of all peptides derived from the protein. The protein score value was expressed as the maximum Mascot protein score value obtained for the entire sample.

[0066] Data source and processing Susceptibility data were obtained from PharmacoDB (Smirnov et al., 2018). RNA-Seq data were obtained from the DepMap repository (Corsello et al., 2020). Proteomics and phosphorylated proteomics data were generated in-house or obtained from (Jarnuczak et al., 2019; Piersma et al., 2015) for 26 AML cell lines, 10 esophageal cell lines, and 12 HCC cell lines (see above). Drug response, proteomics, and phosphorylated proteomics datasets were normalized by sigmoid, and proteomics and phosphorylated proteomics data were further normalized by centering and scaling. RNA-seq data were obtained as quantile-normalized values.

[0067] Methods for empirical markers of susceptibility and tolerance To reduce the dimensionality of the input dataset, empirical markers of drug responses (EMDRs) were obtained using a collection of statistical differences between cells in resistance or susceptibility to a given drug. For each drug, cell lines were divided into relatively resistant or susceptible populations using the median drug response (AAC) as a cutoff. The resistant and susceptible populations were divided into 10 groups using the createMultiFolds caret function (https: / / cran.r-project.org / package=caret). The AAC values ​​in each susceptible group were compared to each in the resistant group, resulting in 100 comparisons. Markers were obtained using resistant and susceptible samples in repeats with p-values ​​less than 0.05 between AAC values. A linear model was constructed, and contrasts were calculated for phosphorylation sites, proteins, or transcripts across the sampled cell lines using the Limma package (http: / / bioconductor.org / packages / release / bioc / html / Limma.html). The significance of these comparisons was assessed by p-values ​​adjusted for multiple tests using empirical Bayesian statistics and the Benjamini-Hochberg method. EMDR was defined as significant if a fold value of ±0.8 and a p-value <0.05 were achieved in at least 80% of the repeats. Increases found in susceptible cells compared to resistant cells were considered susceptibility markers, while increases in resistant cells were considered resistance markers.

[0068] DRUML method Drug enrichment values ​​were calculated using a custom in-house script from the DRUMLR package. For each drug and each cell line, the distance (D) value was calculated by subtracting the median and third quartile values ​​of resistance marker expression from the median and third quartile values ​​of sensitivity marker expression. The D values ​​for each drug were correlated with the total drug response data using Spearman ranking, and the top 14 D values ​​with the highest correlations (7 positive correlations and 7 negative correlations) were used as input to the machine learning model. The model was built using the caret (https: / / cran.r-project.org / package=caret) and h2o (https: / / cran.r-project.org / package=h2o) packages. The data was divided into training and validation populations with a distribution ratio of 0.8 using the createDataPartion function of caret. Subsequently, before being used to construct regression models, the D-values ​​were sigmoid-normalized using deep learning (DL), neural networks (nnet), Bayesian estimation in general linear models (bglm), random forests, partial least squares (pls), principal component regression (pcr), support vector machines (svm), and Cubist ML models. The h2o R package was used for DL ​​model generation, and the caret R package was used for all other models. Each model underwent hyperparameter tuning with 10-fold cross-validation using RMSE as the loss function, and was then validated using validation data and MS data from other labs. The code used to create and evaluate different learning models is provided in the DRUMLR package.

[0069] Approach Overview DRUML consists of a collection of ML models trained on cell responses to over 400 drugs, allowing these drugs to be ranked based on their predicted efficacy within a sample (Figure 1A). In principle, any large omics dataset can be used as input to DRUML. While the use of gene copy number and RNA-seq for generating trained models is well known, the usefulness and relative performance of large-scale proteomics and phosphorylated proteomics data have not been sufficiently explored. Here, we used phosphorylated proteomics and proteomics datasets obtained from 48 AML, esophageal cancer, and hepatocellular carcinoma cell lines as input to DRUML to build models applicable to leukemia and solid tumors (Figures 1B, 1C). To reduce the impact of data noise on model performance, we first reduced the dimensionality of the omics datasets obtained as part of this investigation to obtain empirical markers of drug response (EMDR, Figure 1A) and calculated a comprehensive index of drug response (named Drug Response Distance, D). The D metric is the difference in the overall expression of markers positively associated with drug sensitivity compared to markers positively associated with drug resistance in a sample. The D value was then used as an input feature for a predictive model of drug response, which was trained and tested using our dataset and validated using independent datasets from other laboratories (described below). This important feature of DRUML is due to two reasons: firstly, since D can be calculated using incomplete omics data, the use of averaged marker values ​​avoids the problem of lacking predictors when making validation or predictions on future datasets; secondly, D is an internally normalized metric obtained by subtracting the averaged signals from two sets of phosphorylation sites, proteins, or transcripts in a given sample; therefore, once the model is built, the application of DRUML to predict drug response in novel cancer-derived samples does not require comparison to a control or reference sample set.

[0070] Input dataset To develop DRUML, the inventors, similar to previous studies (Casado et al., 2018; Hijazi et al., 2020), analyzed the proteome and phosphorylated proteome of a panel of 26 AML, 10 esophageal cancer, and 12 hepatocellular carcinoma cell lines in triplicate (three independent media per cell line) by LC-MS / MS (Figure 1B). This analysis required 288 LC-MS / MS runs, generating a sufficiently large ground-state phosphorylated proteomics and proteomics dataset containing 22,804 phosphorylated peptides and 6,455 proteins, resulting in 3,283,776 and 929,520 quantitative data points, respectively (Figure 8). To the inventors' knowledge, these represent the largest set of phosphorylated proteomics available to date with matched proteomics data for AML, esophageal cancer, and hepatocellular carcinoma cell lines.

[0071] Drug response data (AAC values) for the morphology of regions on the curve were obtained from PharmacoDB (Smirnov et al., 2018) for the same cell lines from which the inventors created the phosphoproteomics and proteomics data. The AAC values ​​were sigmoid-normalized so that the values ​​ranged from 0 (no drug effect) to 1 (maximum cell death) within a given cell line. To ensure that the dataset had a sufficient range of sensitivity, drugs were filtered by interquartile range, reducing the number of profiled drugs from 659 to 466. For comparison, RNA-seq data obtained from the DepMap portal (Corsello et al., 2020) was also used as input to the model. Figure 1C shows that principal component analysis of the proteomics, phosphoproteomics, and drug response data grouped the cell lines by cancer type. Therefore, in order to ensure that the generated models elucidate the biological mechanisms of susceptibility without the influence of cancer type, the inventors constructed separate DRUML models for solid and AML tumor samples, as described below.

[0072] Dimensionality Reduction To illustrate the approach used to reduce dimensionality, Figures 1D-G show the measurement of EMDR for barasertib. For each drug, cell lines were split into those resistant or susceptible to those drugs (using a median AAC cutoff), and repeated resamplings were compared (Figures 1D, 1E). The R package Limma was then used to identify phosphorylation sites, proteins, and transcripts frequently associated with drug response across the resampling groups. Markers consistently found to be increased or decreased in susceptible cells by Limma were saved as EMDRs and provided in the DRUMLR package. We refer to markers increased in susceptible cells as susceptibility markers and markers decreased as resistance markers. As outlined above, our approach involves combining the identified EMDRs into a distance metric (D), which is essentially a measure of the distribution of susceptibility markers against resistance markers; D is formally defined in Figure 1A. Figure 1F shows the distribution of phosphorylation site markers associated with resistance and sensitivity to valacertib in drug-resistant cells (OCI-M1), sensitive cells (P31-FUJ), or intermediate response cells (GDM-1). These distributions were then measured to derive D values, which correlate with drug response across these three cell lines (Figure 1G) and across all models tested (Figures 2A, B, C). As expected, the correlation was statistically significant when markers derived from AML or solid tumors were applied to their respective cancer types, but not significant when comparing AML cell responses using solid tumor-derived markers or vice versa (Figures 2A, B, C).

[0073] To further reduce dimensionality, for each drug, DRUML selected the top 14 D-values ​​(7 positively correlated and 7 negatively correlated) from EMDRs of phosphorylated proteomics, proteomics, and transcriptome for all 466 drugs. Figure 2D shows that the parasertib response correlates with the D-values ​​of several drugs. As a positive control for this approach, we found that parasertib D-values ​​obtained from three different marker datasets consistently correlated with the response to this drug. The D-values ​​of AT7867 and CR1.31B also correlated with the parasertib response, while the D-values ​​of rigosertib, SL.0101.1, FK866, and FH535 were inversely correlated (Figure 2E). Thus, D-values ​​from different drugs and datasets consistently showed association with the parasertib response, highlighting that D-value reproducibility is associated with drug response regardless of the omics dataset from which they were obtained.

[0074] When this approach was systematically applied to 466 drugs in AML and solid tumor cell lines (Figure 3A), 1,232 and 1,139 phosphorylation sites, 542 and 480 proteins, and 3,046 and 3,699 transcripts were identified as response markers for AML and solid models, respectively (Figure 3B). On average, each drug was annotated with 128 ± 37 (mean ± SD, range 53–278) and 97.6 ± 43 (15–269) phosphorylation site markers for drug response in AML and solid models, respectively, as well as an average of 40–50 (range 10–131) resistance or sensitivity protein markers in solid and AML models (Figure 9). The number of RNA transcripts associated with drug response was large due to the size of the input data. As shown in Figure 3C, some of these phosphorylation sites, proteins, and transcripts were found to be markers of response to several drugs. Interestingly, while phosphorylation sites on FAM129B, SRRM2, lamin (LMNA), and mTOR substrate 4EBP1 were found to be sensitivity markers for over 200 drugs, phosphorylation sites and the entire protein of NPM1 (protein name nucleophosmin, NPM) were frequently markers of resistance (Figure 3C).

[0075] The inventors also hypothesized that these proteins and phosphorylation sites would group marker drugs based on the known targets and mechanisms of action of the compounds, in order to explore the biological relevance of their EMDRs. In general, drugs that target a common enzyme or have similar mechanisms of action were found to be grouped in the PCA space (Figure 3D and Figure 10). For example, barasertib grouped with other AurB inhibitors, as well as with the microtubule destabilizers vinplatin and vinorelbine (Figure 3D), consistent with the role of AurB in microtubule stabilization (Haase et al., 2017). Similarly, EGFR inhibitors were grouped together in the PCA space and separated from IGFR inhibitors and IR inhibitors (Figure 3D). Another example is provided by ERK MAPK pathway inhibitors grouped by target in the PCA space (Figure 3D). Therefore, drug markers in drug response groups group drugs according to their mechanisms of action, which is consistent with the concept that these markers serve as indicators of the biological mechanisms that determine the response to the profiled drug.

[0076] ML model of response to varacertib Next, the inventors generated a drug response learning model for a given drug using the top 14 correlation D values, as described above for varacertib (Figures 1 and 2). Because there was no prior knowledge of potentially more appropriate learning methods for predicting drug responses from proteomics datasets, the inventors first evaluated the performance of various ML methods based on the learning algorithms of Random Forest (RF), Cubist, Generalized Linear Model Bayesian Estimation (BGLM), Partial Least Squares (PLS), Principal Component Regression (PCR), Deep Learning (DL), and Neural Network (NNET). As mentioned above, the inventors' primary objective was to compare models constructed from D values ​​obtained from phosphorylated proteomics and proteomics data, but as a benchmark, models were constructed using D values ​​obtained from RNA-seq data as input. The RNA-seq dataset was obtained from a publicly available resource (Corsello et al., 2020), making it difficult to directly compare the results obtained using RNA-seq with those obtained from in-house generated proteomics and phosphorylated proteomics data. Therefore, this embodiment does not draw any conclusions regarding the relative performance of RNA-seq-derived models. For model generation, the inventors split the data in an 80:20 ratio (different training and validation sets for each drug to generate splits with similar response value distributions in the two datasets), and trained regression models on normalized drug response (AAC) data using 10-fold cross-validation with the root mean square error (RMSE) metric as the loss function. The DL / ML models were then evaluated on the validation set by comparing predicted responses with actual responses using absolute error or standard error (SE) and RMSE (for individual data points and overall model performance, respectively).

[0077] As a first example, we used barasertib to evaluate the performance of different models generated using D values ​​derived from phosphoproteomics and proteomics datasets as predictors (shown in Figure 2D). Figures 4 and 11 show that DL and NNET using D values ​​from phosphoproteomics data predicted all response values ​​from all cell lines with an absolute error of less than 0.2 AAC units, resulting in smaller validation errors (Figures 4A, 4B, and 4C). In the DL model, 12 out of 12 validation data points were predicted within 0.1 AAC units from the phosphoproteomics D data (Figure 4A), compared to 7 out of 12 data points from the protein D data being predicted within 0.1 AAC units (Figure 4B). Direct comparison of SEs obtained from different ML methods (Figure 4C) confirmed that for barasertib, the DL model constructed using phosphoD data yielded the lowest validation error.

[0078] A DRUML model ensemble for ranking drug responses Next, using phosphorylated proteomics, proteomics, and RNA-seq distance D data obtained from AML and solid tumors as input, we systematically constructed predictive models for 466 drugs (of which 412 could be modeled) using the approach described above (using varacertib as an example). A total of 17,064 trained models were constructed (Figure 5A), consuming 4.31 gigabytes of disk space. Approximately the same number of models were created from the phosphorylated proteomics and proteomics datasets (Figure 5A). Similar to the analysis of varacertib-sensitivity models, the DL algorithm produced smaller validation errors, with standard errors less than 0.1 in all cases, for proteomics and phosphorylated proteomics data derived from solid tumor types and AML tumor types (Figure 5B).

[0079] Next, we examined whether the ML model could rank drugs within cell lines based on predicted efficacy. Figure 5C shows the ranking of drugs in AML cells used to validate the DL algorithm (the ranking of training data is shown in Figure 12). We found a remarkably high correlation between predicted and actual responses within the cell model across drugs with diverse mechanisms of action. For DL ​​models derived from phosphoproteomics, proteomics, and RNA-seq datasets, the mean absolute validation errors between predicted and actual responses were 0.04, 0.051, and 0.11, respectively. Examining the validation errors of models derived from training methods (resampling) and other algorithms (Figure 5D), we confirmed that the error with DL was smaller for all datasets. Therefore, our results suggest that DRUML can be used to accurately rank drugs with diverse mechanisms of action within tumors based on predicted efficacy.

[0080] Validation in independent datasets Ultimately, for a drug response prediction model to be useful, it should be able to accurately predict treatment outcomes regardless of the laboratory from which the data was obtained. Therefore, to validate DRUML using data collected in independent laboratories, we attempted to test whether a model generated using our training dataset could predict drug responses from publicly available, label-free proteomics and phosphoproteomics datasets generated by other groups. Label-free phosphoproteomics data was downloaded from eight colorectal cancer cell lines from Piersma et al.'s PRIDE [(Piersma et al., 2015), pride id: PXD001550], reprocessed using our mass spectrometry informatics pipeline (Cutillas, 2017; Hijazi et al., 2020), and 12,197 phosphorylated peptides were identified and quantified label-free (Figure 13). Next, this dataset was used with DRUML to predict drug responses from previously generated models using phosphorylated proteomics data of solid tumors for six of these cell lines (for which drug response data was available). Figure 6 shows a significant correlation between DRUML-derived drug response predictions and actual responses to these six cell lines, and the predictions were accurate for drugs with diverse mechanisms of action (Figure 6A) and developmental phases (Figure 6B). For all DL / ML models tested, the mean absolute error was less than 0.15 AAC units (Figure 6C). On this dataset, PCR and RF learning algorithms performed comparably to or better than DL, with over 80% of all responses predicted with an absolute error of less than 0.15 AAC units (Figures 6D, E).

[0081] Furthermore, DRUML was applied to predict drug responses in a panel of 47 cell lines derived from diverse solid tumor species. The input for this analysis was proteomics data obtained from Jarnuczak [(Jarnuck et al., 2019), pride id: PXD013455], compiled from 11 separate studies. Using a DRUML model generated with our solid tumor proteomics training dataset, responses were predicted without further processing (except averaging of iBAQ values ​​from replicated cell line measurements) using iBAQ values ​​provided by Jarnuczak et al. Figures 7A and 14 show that DRUML-predicted and actual drug responses were highly correlated across the 47 cell lines, regardless of their assigned disease state. The median SE of predictions was less than 0.1 for all cell lines (Figure 7B), and similar to predictions from phosphorylated proteomics data, over 80% of drug responses were predicted with an absolute error of less than 0.15 AAC units, and 95% of those with an error of less than 0.25 AAC units, indicating that predictions from RF and PCR were as accurate as those provided by DL and NNET models (Figure 7C). Overall, these data demonstrate that proteomics data collected using conventional LC-MS / MS can be used as input for DRUML to accurately predict and rank the efficacy of drugs with diverse mechanisms of action in cancer cells originating from different pathological conditions.

[0082] Consideration This example demonstrates the usefulness of proteomics and phosphorylated proteomics data as inputs to DL / ML to rank drugs based on their predicted efficacy in reducing the proliferation of a given population of cancer cells. This example led to the development of DRUML, a collection of predictive models trained on 412 drugs with different mechanisms of action and developmental stages. This example is made possible because methods for systematic and relatively high-throughput label-free analysis of the proteome and phosphateproteome are now available (Aasebo et al., 2020; Aebersold and Mann, 2016; Bodenmiller et al., 2010; Cutillas, 2017; Leutert et al., 2019; van Alphen et al., 2020; Vowinckel et al., 2018; Wilkes et al., 2015), and drug response data for a vast number of compounds have become publicly available in recent years (Chiu et al., 2019; Corsello et al., 2020; Menden et al., 2019; Smirnov et al., 2018). However, since the use of large-scale LC-MS / MS proteomics data for DL / ML model generation has not been systematically investigated as part of DRUML development, we evaluated the suitability of such large datasets as input to predictive drug response models. Initial evaluations showed that the difference in accuracy between proteomics and phosphorylated proteomics-based models was small, but the training and validation errors with phosphorylated proteomics data were consistently minimal. This is consistent with previous findings from our laboratory and other laboratories that have shown that phosphorylated proteomics and proteomics data reflect the mechanisms of drug response (Casado et al., 2013a; Casado et al., 2018; Frejno et al., 2017; Roumeliotis et al., 2017; van Alphen et al., 2020).

[0083] To limit the impact of noise and missing values ​​in omics datasets on DL / ML model performance and to make this approach practical, instead of individual phosphorylation sites, proteins, or transcripts, DRUML uses a distance metric (denoted as D) as input, which measures the difference in distribution levels between sensitivity and resistance markers for a given drug. This feature contributes to DRUML's robustness because D is an internally normalized value, takes into account the marker scores (and thus dilutes the contribution of outliers), and addresses the potential problem of missing values ​​in validation and verification datasets. In fact, each D value is calculated by averaging over hundreds of proteins, phosphorylation sites, or transcripts (Figures 3 and 9), so the D metric can be calculated even if not all individual EMDRs are identified in the sample being analyzed. DRUML uses D values ​​for each drug and those calculated for other drugs. For example, the D values ​​selected to build a DL / ML model for varacertib include the D_varasertib value and D values ​​for rigosertib, AT7867, etc. (Figure 2). In this embodiment, to avoid overfitting, DL / ML models for each drug were constructed using the top 14 D values. The performance of the learning model can be adjusted by controlling the number of D values ​​included in the model.

[0084] The calculation of the D-value relied on the measurement of EMDR, and as input to DRUML, we obtained over 2000 phosphorylation sites and over 800 such EMDR proteins (Figures 3 and 9). Previous studies have generally indicated that drug efflux pump expression is a major variable influencing drug tolerance to various drugs (Roumeliotis et al., 2017), and in particular, suggest that kinase activation (detected as phosphorylation of kinase activity markers) underlies the response to kinase inhibitors (Casado et al., 2018; van Alphen et al., 2020). In this embodiment, we did not examine the biochemical function of our set of predictive biomarkers, as this was outside the scope of this embodiment. Further investigation of the ontology to which these markers belong may provide insights into biochemical pathways mediating drug sensitivity or tolerance to different treatments. Nevertheless, analysis by unsupervised classification revealed that our set of EMDRs group drugs based on their mechanism of action (Figure 3). Therefore, it was suggested that EMDR, used as input by DRUML, reflects the biological mechanisms of responses to different drugs.

[0085] Of the various learning algorithms tested, DL showed the best performance using in-house training and validation datasets (Figures 4 and 5), while PCR and RF were found to have lower errors on validation datasets obtained from independent laboratories (Figures 6 and 7). This is in contrast to findings from previous studies using transcriptome data as input, and we found DL to be superior to other ML methods (Sakellaropoulos et al., 2019). This difference may be explained by the fact that when applied to large datasets, the performance of DL is only significantly greater than the performance provided by other ML methods (LeCun et al., 2015). Therefore, the DL method may be better suited for training predictive models from phosphorylated proteomics and proteomics as larger datasets become available.

[0086] Evaluation of DRUML using an external validation dataset from 53 cell lines analyzed by independent laboratories (Jarnuczak et al., 2019; Piersma et al., 2015) revealed that approximately 80% of drugs could be ranked with an absolute error of less than 0.15 AAC units across all cancer types, and 95% of those showed an error of less than 0.25 AAC units (Figures 6 and 7). This represents remarkably high accuracy, considering that DRUML was trained using esophageal and liver cancer, while the validation dataset consisted of data from cell lines derived from bone cancer, brain cancer, breast cancer, cervical cancer, colorectal cancer, ovarian cancer, and prostate cancer.

[0087] In summary, this embodiment evaluated the accuracy of DRUML for creating a list of drugs ranked by their predicted efficacy in reducing the proliferation of a given cancer cell population. The approach was trained and validated using analysis of 48 cell lines profiled in our laboratory and validated using a set of 53 cancer cell models profiled by 12 other groups. The results of this embodiment demonstrate that DRUML ranks drugs with different mechanisms of action with low error based on their predicted efficacy across different cancer types. Ultimately, DRUML can assist in drug prioritization by complementing information obtained from clinicopathological parameters and mutational analysis. References

[0088] Aasebo, E., Berven, FS, Bartaula-Brevik, S., Stokowy, T., Hovland, R., Vaudel, M., Doskeland, SO, McCormack, E., Batth, TS, Olsen, JV, et al. (2020). Proteome and Phosphoproteome Changes Associated with Prognosis in Acute Myeloid Leukemia. Cancers (Basel) 12. Aebersold, R., and Mann, M. (2016). Mass-spectrometric exploration of proteome structure and function. Nature 537, 347-355. Basu, A., Bodycombe, N.E., Cheah, J.H., Price, E.V., Liu, K., Schaefer, G.I., Ebright, R.Y., Stewart, M.L., Ito, D., Wang, S., et al. (2013). An interactive resource to identify cancer genetic and lineage dependencies targeted by small molecules. Cell 154, 1151-1161. Bodenmiller, B., Wanka, S., Kraft, C., Urban, J., Campbell, D., Pedrioli, P.G., Gerrits, B., Picotti, P., Lam, H., Vitek, O., et al. (2010). Phosphoproteomic analysis reveals interconnected system-wide responses to perturbations of kinases and phosphatases in yeast. Science signaling 3, rs4. Casado, P., Alcolea, M.P., Iorio, F., Rodriguez-Prados, J.C., Vanhaesebroeck, B., Saez-Rodriguez, J., Joel, S., and Cutillas, P.R. (2013a). Phosphoproteomics data classify hematological cancer cell lines according to tumor type and sensitivity to kinase inhibitors. Genome biology 14, R37. Casado, P., Rodriguez-Prados, J.C., Cosulich, S.C., Guichard, S., Vanhaesebroeck, B., Joel, S., and Cutillas, P.R. (2013b). Kinase-substrate enrichment analysis provides insights into the heterogeneity of signaling pathway activation in leukemia cells. Sci Signal 6, rs6. Casado, P., Wilkes, E.H., Miraki-Moud, F., Hadi, M.M., Rio-Machin, A., Rajeeve, V., Pike, R., Iqbal, S., Marfa, S., Lea, N., et al. (2018). Proteomic and genomic integration identifies kinase and differentiation determinants of kinase inhibitor sensitivity in leukemia cells. Leukemia 32, 1818-1822. Chiu, Y.C., Chen, H.H., Zhang, T., Zhang, S., Gorthi, A., Wang, L.J., Huang, Y., and Chen, Y. (2019). Predicting drug response of tumors from integrated genomic profiles by deep neural networks. BMC Med Genomics 12, 18. Corsello, S.M., Nagari, R.T., Spangler, R.D., Rossen, J., Kocak, M., Bryan, J.G., Humeidi, R., Peck, D., Wu, X., Tang, A.A., et al. (2020). Discovering the anticancer potential of non-oncology drugs by systematic viability profiling. Nature Cancer 1, 235-248. Cutillas, P.R. (2017). Targeted In-Depth Quantification of Signaling Using Label-Free Mass Spectrometry. Methods Enzymol 585, 245-268. Frejno, M., Zenezini Chiozzi, R., Wilhelm, M., Koch, H., Zheng, R., Klaeger, S., Ruprecht, B., Meng, C., Kramer, K., Jarzab, A., et al. (2017). Pharmacoproteomic characterisation of human colon and rectal cancer. Mol Syst Biol 13, 951. Haase, J., Bonner, M.K., Halas, H., and Kelly, A.E. (2017). Distinct Roles of the Chromosomal Passenger Complex in the Detection of and Response to Errors in Kinetochore-Microtubule Attachment. Dev Cell 42, 640-654 e645. Hijazi, M., Smith, R., Rajeeve, V., Bessant, C., and Cutillas, P.R. (2020). Reconstructing kinase network topologies from phosphoproteomics data reveals cancer-associated rewiring. Nat Biotechnol 38, 493-502. Jarnuczak, A.F., Najgebauer, H., Barzine, M., Kundu, D.J., Ghavidel, F., Perez-Riverol, Y., Papatheodorou, I., Brazma, A., and Vizcaino, J.A. (2019). An integrated landscape of protein expression in human cancer. bioRxiv, 665968. Klempner, S.J., Myers, A.P., and Cantley, L.C. (2013). What a tangled web we weave: emerging resistance mechanisms to inhibition of the phosphoinositide 3-kinase pathway. Cancer Discov 3, 1345-1354. LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature 521, 436-444. Leutert, M., Rodriguez-Mias, R.A., Fukuda, N.K., and Villen, J. (2019). R2-P2 rapid-robotic phosphoproteomics enables multidimensional cell signaling studies. Mol Syst Biol 15, e9021. Menden , MP , Wang , D , Mason , MJ , Szalai , B , Bulusu , KC , Guan , Y , Yu , T , Kang , J , Jeon , M , Wolfinger , R , et al. (2019). Community assessment to advance computational prediction of cancer drug combinations in a pharmacogenomic screen. Nat Commun 10 , 2674 . Montoya , A. , Beltran , L. , Casado , P. , Rodriguez-Prados , JC , and Cutillas , PR (2011). Characterization of a TiO(2) enrichment method for label-free quantitative phosphoproteomics. Methods (San Diego, Calif) 54, 370–378. Paulitschke , V. , Eichhoff , O. , Gerner , C. , Paulitschke , P. , Bileck , A. , Mohr , T. , Cheng , PF , Leitner , A. , Guenova , E. , Saulite , I. , et al. (2019). Proteomic identification of a marker signature for MAPKi resistance in melanoma. EDUCATION J 38, e95874. Piersma, S.R., Knol, J.C., de Reus, I., Labots, M., Sampadi, B.K., Pham, T.V., Ishihama, Y., Verheul, H.M., and Jimenez, C.R. (2015). Feasibility of label-free phosphoproteomics and application to base-line signaling of colorectal cancer cell lines. Journal of proteomics 127, 247-258. Roumeliotis, T.I., Williams, S.P., Goncalves, E., Alsinet, C., Del Castillo Velasco-Herrera, M., Aben, N., Ghavidel, F.Z., Michaut, M., Schubert, M., Price, S., et al. (2017). Genomic Determinants of Protein Abundance Variation in Colorectal Cancer Cells. Cell Rep 20, 2201-2214. Rudolph, J.D., de Graauw, M., van de Water, B., Geiger, T., and Sharan, R. (2016). Elucidation of Signaling Pathways from Large-Scale Phosphoproteomic Data Using Protein Interaction Networks. Cell Syst 3, 585-593 e583. Sakellaropoulos, T., Vougas, K., Narang, S., Koinis, F., Kotsinas, A., Polyzos, A., Moss, T.J., Piha-Paul, S., Zhou, H., Kardala, E., et al. (2019). A Deep Learning Framework for Predicting Response to Therapy in Cancer. Cell Rep 29, 3367-3373 e3364. Smirnov, P., Kofia, V., Maru, A., Freeman, M., Ho, C., El-Hachem, N., Adam, G.A., Ba-Alawi, W., Safikhani, Z., and Haibe-Kains, B. (2018). PharmacoDB: an integrative database for mining in vitro anticancer drug screening studies. Nucleic Acids Res 46, D994-D1002. van Alphen, C., Cloos, J., Beekhof, R., Cucchi, D.G.J., Piersma, S.R., Knol, J.C., Henneman, A.A., Pham, T.V., van Meerloo, J., Ossenkoppele, G.J., et al. (2020). Phosphotyrosine-based Phosphoproteomics for Target Identification and Drug Response Prediction in AML Cell Lines. Mol Cell Proteomics 19, 884-899. Vowinckel, J., Zelezniak, A., Bruderer, R., Mulleder, M., Reiter, L., and Ralser, M. (2018). Cost-effective generation of precise label-free quantitative proteomes in high-throughput by microLC and data-independent acquisition. Sci Rep 8, 4346. Wilkes, E., and Cutillas, P.R. (2017). Label-Free Phosphoproteomic Approach for Kinase Signaling Analysis. Methods Mol Biol 1636, 199-217. Wilkes, E.H., Terfve, C., Gribben, J.G., Saez-Rodriguez, J., and Cutillas, P.R. (2015). Empirical inference of circuitry and plasticity in a kinase signaling network. Proc Natl Acad Sci U S A 112, 7719-7724. Yang, W., Soares, J., Greninger, P., Edelman, E.J., Lightfoot, H., Forbes, S., Bindal, N., Beare, D., Smith, J.A., Thompson, I.R., et al. (2013). Genomics of Drug Sensitivity in Cancer (GDSC): a resource for therapeutic biomarker discovery in cancer cells. Nucleic Acids Res 41, D955-961.

Claims

1. A computer implementation method for identifying drugs to treat a patient, comprising the following steps: a. To provide a dataset of expression levels of biological markers from samples taken from the patient; b. Multiple drugs d n Multiple drug reaction distance values ​​D for each of them n To calculate the above dataset, D is used; where D for each drug d is the difference between the distribution of expression of biological markers for susceptibility to drug d and the distribution of expression of biological markers for susceptibility to drug d, and D is calculated for each drug d using the following formula: D d = [S Q2 - R Q2 ] + [S Q3 - R Q3 ], Here, S Q2 and S Q3 are the median and third-quartile expression values ​​of the biological markers of susceptibility to drug d; and R Q2 and R Q3 are the median and third-quartile expression values ​​of the biological markers of resistance to drug d; c. Provide one or more trained predictive models, where one or more trained predictive models include at least the same number of drugs d as in step b. n Multiple drug reaction distance values ​​D n They are trained in this; Furthermore, one or more trained predictive models rank the multiple drugs d in order of their predicted efficacy in the sample taken from the patient. n They are trained to provide drug rankings from; d. The plurality of drug reaction distance values ​​D obtained in step b. n Inputting the above into one or more trained predictive models; and e. Using one or more trained predictive models, the multiple drugs d are ordered in order of their predicted efficacy in the patient. n To provide a ranking of drugs from [source].

2. A method according to claim 1, further comprising: f. Identifying a drug to treat a patient by selecting one of the highest-ranking drugs.

3. A method according to claim 1 or 2, wherein the plurality of drug reaction distance values ​​D n A method in which each of the D values ​​includes any number of most positively correlated D values ​​and any number of most negatively correlated D values ​​for the D value of drug d.

4. The method according to claim 3, wherein each of the plurality of drug reaction distance values D n includes an equal number of D values that are positively correlated and D values that are negatively correlated with the D value for drug d.

5. The method according to claim 4, wherein the plurality of drug reaction distance values ​​D n Each of these methods includes the seven most positively correlated D values ​​and the seven most negatively correlated D values ​​for the D value of drug d.

6. A method according to any one of claims 1 to 5, wherein the biological marker for susceptibility to drug d is a biological marker whose expression is consistently increased in a biological sample for susceptibility to drug d compared to its expression in a biological sample for resistance to drug d, and the biological marker for resistance to drug d is a biological marker whose expression is consistently increased in a biological sample for resistance to drug d compared to its expression in a biological sample for susceptibility to drug d.

7. A method according to claim 6, wherein the biological sample is a cell line, and optionally a cancer cell line.

8. The method according to claim 6, wherein the biological sample is primary cells obtained from a patient, and optionally primary cancer cells obtained from a patient.

9. A method according to any one of claims 1 to 8, wherein the biological markers of susceptibility to drug d and / or the biological markers of resistance to drug d are identified using a computer program.

10. A method according to any one of claims 1 to 9, wherein the biological marker for sensitivity to drug d is FAM129B(S691); SRSF6(S301); HSP90AA1(S252); SRRM2(S2102); SRRM2(T2104); DBNL(K288); DBNL(T291); EIF4EBP1(S83); EIF4EBP1(S85); EIF4EBP1(S44); LMNA(S392); TRA2A(S262); TRA2B(S266); TRA2A(S260); TRA2B(S264); LMNA(S390); ACIN1(S1004); BCLAF1(S512); MCM2(T25); HSP90AB1(S255); HSP90AB2P(S177); HSP90AB1(S255); SRRM2(M1168); SRRM2(T1177); IRF2BP2(S175); ADD1(S726); ADD2(S713); ADD3(S693); FAM129B(S691); FAM129B(S692); EIF4EBP1(T37); EIF4EBP1(T41); EIF4EBP1(S35);EIF4EBP1(T50); HNRNPK(S121); G3BP1(S230); PLEC_HUMAN; ANXA2_HUMAN; FAS_HUMAN; LMNA_HUMAN; PDIA1_HUMAN; H4_HUMAN; 1433Z_HUMAN; PRKDC_HUMAN; AHNK_HUMAN; MYH9_HUMAN; EF2_HUMAN; P3_HUMAN; SPB1_HUMAN; HSPB1_HUMAN; FLNB_HUMAN; KPYM_HUMAN; LEG1_HUMAN; H2A1A_HUMAN; CH10_HUMAN; CH60_HUMAN; HS90B_HUMAN; Selected from the group consisting of CSTA; SDC2; BEX3; SPOCK1; BEX1; RPS4Y1; VCAN; EMP1; VIM; HGD; CHI3L1; TGFBI; RTL8C; CDH1; FAM20C; GTSF1; ALOX5AP; TUBB6; HLA; PPP1R27; LGALS1; CCL2; SPARC; LYZ and RAB34, and / or The biological marker of resistance to drug d is, WRNIP1(S151); CCDC86(S18); NCL(T69); NPM1(S125); HMGA1(S36); PML(S518); HMGA1(T53); NOLC1(S623); PML(S527); SRRM1(S773); SNW1(S234); SNW1(S224); PGM2L1(M171); PGM2L1(T173); SRRM1(S775); SRRM1(S597); IGF2BP1(S181); SRRM1(T614); NUCKS1(S181); ASXL3(S76); DVL2(S480); SRRM1(S769); TPI1(S58); SRP72(S630); SRP72(T633); AKT1S1(T90); NPM_HUMAN; A4_HUMAN; AAAS_HUMAN; ACTBL_HUMAN; PRDX1_HUMAN; TOPK_HUMAN; H2B1B_HUMAN; 2B1B_HUMAN; G3P_HUMAN; ACTB_HUMAN; KPYM_HUMAN; HSP7C_HUMAN; K1614_HUMAN; H2A1B_HUMAN; ACTA_HUMAN; PROF1_HUMAN; PPIA_HUMAN; ANXA1_HUMAN; EF1A1_HUMAN; TBA1C_HUMAN; NUCL_HUMAN; CH60_HUMAN; A method selected from the group consisting of HLA; TSPAN1; AS1; JUP; GYPC; TFPI; SLPI; ITM2A; GSTP1; LCN2; ALDH1A1; ABCB1; CDH1; CA2; CLEC2B; TSPAN13; IL1B; DPA1; RAB38; CAVIN2; AC093840.1; LGALS3; BEX4; PRAME and DDX60.

11. A method according to any one of claims 1 to 10, wherein the expression value is obtained from a phosphoproteomics, proteomics, or transcriptomics experiment.

12. A method according to any one of claims 1 to 11, wherein the one or more trained predictive models are obtained using a machine learning or statistical learning method.

13. The method according to claim 12, wherein the one or more trained predictive models are trained using a learning algorithm selected from random forest (RF), Cubist, Bayesian estimation of generalized linear models (BGLM), partial least squares (PLS), principal component regression (PCR), deep learning (DL), and neural network (NNET) learning algorithms.

14. A method according to any one of claims 1 to 13, wherein the patient has been diagnosed with cancer or is suspected of having cancer.

15. A method according to claim 14, wherein the cancer is leukemia or a solid tumor.

16. A method according to claim 15, wherein the leukemia is acute myeloid leukemia, or the solid tumor is esophageal cancer or hepatocellular carcinoma.

17. A method according to any one of claims 1 to 16, wherein each drug d is an anticancer agent.

18. A method according to any one of claims 1 to 17, wherein the sample taken from the patient is a biopsy sample obtained from a tumor.

19. In samples taken from patients, multiple drugs were ranked in order of predicted efficacy. n A computer implementation method for training one or more predictive models to provide a ranking of drugs from a given list, the method comprising: i. Multiple drug reaction distance values ​​D for each of multiple drugs dn n To provide training data including, where D for each drug d is the difference between the distribution of biological markers for susceptibility to drug d and the distribution of biological markers for resistance to drug d, and D is calculated for each drug d using the following formula: D d = [S Q2 - R Q2 ] + [S Q3 - R Q3 ], Here, S Q2 and S Q3 are the median and third-quartile expression values ​​of the biological markers of susceptibility to drug d; and R Q2 and R Q3 are the median and third-quartile expression values ​​of the biological markers of resistance to drug d; ii. Train one or more predictive models using training data, and rank the multiple drugs d in order of their predicted efficacy in samples taken from patients. n To provide a ranking of drugs from [source].

20. A method according to claim 19, wherein the one or more predictive models are trained using a learning algorithm selected from random forest (RF), Cubist, Bayesian estimation of generalized linear models (BGLM), partial least squares (PLS), principal component regression (PCR), deep learning (DL), and neural network (NNET) learning algorithms.

21. A computer-readable storage medium storing instructions for implementing the method described in any one of claims 1 to 20.

22. A system for identifying a drug to treat a patient, wherein the system includes memory and one or more processors, the one or more processors being configured to cause the method according to any one of claims 1 to 18.