Method for predicting transcriptional response to novel drug perturbation and virtual screening method and system

Abstract

The invention provides a method for predicting a transcription response to disturbance of a novel drug, a virtual screening method based on the prediction method, and a virtual screening system based on the above methods. Wherein the prediction model comprises a disturbance adapter, a disturbance encoder and a disturbance decoder. The disturbance adapter is used for encoding the input chemical structures of one or more novel drugs and the dose information of the novel drugs into disturbance embedding; the disturbance encoder is used for mapping the influence of the disturbance of the novel medicine on the input undisturbed transcriptional spectrum to an interpretable potential space to obtain a potential representation of an undisturbed state; and a perturbation decoder for predicting a distribution of a transcriptional response of the novel drug perturbation based on the potential representation and perturbation embedding. And performing virtual screening according to the predicted transcription response of the novel drug, calculating the enrichment score of the candidate compound to the disease-sensitive drug, and recommending the compound according to the enrichment score. The flexibility and expansibility of the prediction model do not need priori knowledge or annotation.

Description

Technical Field

[0001] The present invention belongs to the fields of computational science and biomedicine, and in particular relates to a method for predicting transcriptional response to novel drug disturbance, a virtual screening method based on the prediction method, and a virtual screening system based on the above method. Background Art

[0002] The statements in this section are merely intended to provide background information related to the technical solution of the present application to aid understanding, and they do not necessarily constitute prior art for the technical solution of the present application.

[0003] Virtual screening is one of the common methods for drug discovery and screening. It uses computer-assisted drug screening. Virtual screening can select active compounds from a massive database of small molecules, thereby greatly reducing the number of experimental screening compounds, shortening the research cycle, and reducing the cost of drug development. Virtual screening is divided into structure-based virtual screening and gene marker-based screening. Traditional virtual screening methods are mainly structure-based virtual screening, which uses small molecule structures and drug target structures for molecular docking operations, including receptor structure-based virtual screening and ligand-based virtual screening. In recent years, some studies have performed virtual screening by predicting gene expression after drug perturbation. Gene markers can be calculated by comparing the expression profiles before and after drug perturbation, and matching disease gene markers and small molecule gene markers can be used to screen potential drugs. The change in expression profile before and after drug perturbation is the change in gene expression levels in cells or tissues before and after drug perturbation. This change can reflect the effect of the drug on cell function, including the drug's mechanism of action, possible side effects, and the efficacy of the drug. Gene markers are single or combined sets of genes with unique characteristic gene expression patterns.

[0004] There are currently some methods based on deep learning to predict expression profiles after drug perturbation. For example, scGen (Lotfollahi M, Wolf FA, Theis F J. scGen predicts single-cell perturbation responses. Nature methods, 2019, 16 (8): 715-721.) and CPA (Lotfollahi M, Klimovskaia Susmelj A, De Donno C, et al. Predicting cellular responses to complex perturbations in high-throughput screens. Molecular systems biology, 2023, 19 (6): e11517.) map drug-induced transcriptome effects to latent space based on autoencoder models to model drug perturbation responses. scGen and CPA are used to model single-cell drug perturbation expression profiles. Among them, the scGen model reconstructs the expression spectrum after drug perturbation through a variational encoder, and learns cell type and species-specific responses in the latent space; CPA reconstructs the expression spectrum after drug perturbation based on the encoder and decoder structure, and classifies the embedded vectors in the latent space through the discriminator to identify cell types and perturbed drugs. chemCPA (Hetzel L, Boehm S, Kilbertus N, et al. Predicting cellular responses to novel drug perturbations at a single-cell resolution. Advances in Neural Information Processing Systems, 2022, 35: 26711-26722.) models drug perturbations through a perturbation network, and predicts the expression spectrum after drug perturbation through an encoder and decoder structure. Summary of the invention

[0005] The inventors found in their research and practice that both scGen and CPA decouple expression profile-related properties by reconstructing the expression profile after drug perturbation, and cannot predict the expression profile of new cell types or new drugs after perturbation, and cannot predict the expression profile after multi-cell data perturbation, which has certain limitations for drug virtual screening and recommendation. ChemCPA can only predict the expression profile after a new single drug perturbation, and cannot match the disease gene markers and small molecule gene markers for potential drug or drug combination screening.

[0006] Therefore, the object of the present invention is to overcome the defects of the above-mentioned prior art and provide a method for predicting the transcriptional response to novel drug perturbations, a method for drug virtual screening based on the prediction method, and a drug virtual screening system based on the above method.

[0007] According to a first aspect, an embodiment of the present invention provides a method for predicting transcriptional responses to novel drug perturbations, comprising: inputting the chemical structure of the novel drug and its dosage information as well as an unperturbed transcriptional spectrum into a prediction model; the prediction model outputs the distribution of transcriptional responses perturbed by the input novel drug; wherein the prediction model comprises: a perturbation adapter, for encoding the input chemical structure of the novel drug and its dosage information into a perturbation embedding; a perturbation encoder, for generating a latent embedding according to the perturbation embedding and the unperturbated transcriptional spectrum; and a perturbation decoder, for predicting the distribution of transcriptional responses to the input novel drug perturbations based on the latent embedding and the perturbation embedding.

[0008] Preferably, after the prediction model outputs the distribution of transcriptional responses perturbed by the novel drug, it further comprises: sampling from the distribution of transcriptional responses perturbed by the novel drug to generate an expression spectrum in response to the novel drug perturbation.

[0009] Preferably, the perturbation adapter is also used to: use the input SMILES format of the new drug and the dosage information of the new drug as a drug perturbation attribute set; convert the SMILES format of each of the new drugs in the drug perturbation attribute set into a corresponding first fingerprint embedding; calculate the second fingerprint embedding of the drug perturbation attribute set based on the first fingerprint embedding of each of the new drugs in the drug perturbation attribute set and its dosage information; and generate the perturbation embedding based on the second fingerprint embedding.

[0010] Preferably, the perturbation adapter, the perturbation encoder and the perturbation decoder are respectively multi-layer perceptrons with 2 or more layers.

[0011] Preferably, the training objective of the prediction model is to minimize Gaussian negative log-likelihood loss.

[0012] According to the second aspect, an embodiment of the present invention provides a method for virtual screening of compounds to be screened based on disease-sensitive drugs, comprising: using the prediction method described in the first aspect to respectively predict the expression profiles of sensitive drugs and compounds to be screened after perturbing the cell line; based on the expression profiles before and after the cell line is perturbed by the sensitive drug, calculating the first up-regulation information or the first down-regulation information of the gene level in the cell line, and genetically marking the genes of the cell line perturbed by the sensitive drug according to the first up-regulation information or the first down-regulation information; based on the expression profiles before and after the cell line is perturbed by the compound to be screened, calculating the second up-regulation information or the second down-regulation information of the gene level in the cell line, and genetically marking the genes of the cell line perturbed by the compound to be screened according to the second up-regulation information or the second down-regulation information; according to the gene markers corresponding to the sensitive drugs and the gene markers corresponding to the compounds to be screened, calculating the enrichment scores of the compounds to be screened for the sensitive drugs of the cell line, and screening the compounds to be screened according to the enrichment scores.

[0013] Preferably, based on the expression profiles of the cell line before and after the perturbation of the cell line by the sensitive drug, the first up-regulation information or the first down-regulation information of the gene level in the cell line is calculated, and the genes of the cell line perturbed by the sensitive drug are genetically marked according to the first up-regulation information or the first down-regulation information, including: comparing the expression profiles before and after the perturbation of the sensitive drug, calculating the fold change value of the gene expression data of each gene in the cell line by the sensitive drug as the first fold change value; marking the genes perturbed by the sensitive drug based on the first fold change value; and / or, based on the expression profiles of the cell line before and after the perturbation of the cell line by the compound to be screened, calculating the second up-regulation information or the second down-regulation information of the gene level in the cell line, and genetically marking the genes of the cell line perturbed by the compound to be screened according to the second up-regulation information or the second down-regulation information, including: for each compound to be screened: comparing the expression profiles before and after the perturbation of the compound to be screened, calculating the fold change value of the gene expression data of each gene in the cell line by the compound to be screened as the second fold change value; marking the genes perturbed by the compound to be screened based on the second fold change value.

[0014] Preferably, the enrichment score of the compounds to be screened for sensitive drugs of the cell line is calculated based on the gene markers corresponding to the sensitive drugs and the gene markers corresponding to the compounds to be screened, including: sorting the genes of the cell line according to the first fold change value as a first sorted list; determining the up-regulated gene set and the down-regulated gene set corresponding to the sensitive drug according to the gene marker results perturbed by the sensitive drug; for each compound to be screened: sorting the genes of the cell line according to the second fold change value as a second sorted list; and calculating the enrichment score of the compounds to be screened for sensitive drugs of the cell line based on the first sorted list, the second sorted list, the up-regulated gene set, the down-regulated gene set and the number of cell line genes.

[0015] According to the third aspect, an embodiment of the present invention provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are loaded and executed by a processor, the method as described in any one of the first and second aspects is implemented.

[0016] According to the fourth aspect, an embodiment of the present invention provides a drug virtual screening system, comprising: a data uploading module for uploading an undisturbed expression spectrum and a compound structure of a drug, or selecting an undisturbed expression spectrum and a compound structure of a drug provided by a system; a model prediction module for executing any method described in the second aspect; a data storage module for storing uploaded data and prediction results, and at the same time storing predicted expression spectrum drug features after drug administration; and a result visualization module for analyzing and statistically analyzing the prediction results and visually displaying the screening results.

[0017] Compared with the prior art, the prediction method proposed in this application can effectively screen large compound libraries for various diseases and cell types without the need for prior knowledge or annotations, and has good flexibility and scalability. This scheme uses the chemical structure and dosage of the drug and the unperturbed transcriptional spectrum as input to predict the transcriptional response after drug perturbation. It is not limited to fixed sequencing technology, and can predict transcriptional responses at both single-cell and multi-cell levels. The prediction does not require prior knowledge of the drug, and can achieve efficient virtual high-throughput transcriptional response prediction. In addition, the drug virtual screening scheme provided in this application can perform efficient drug screening, and a large-scale compound library can be screened for a given disease-sensitive drug. The screening results can be visualized by uploading and selecting the system, which is easy to operate and suitable for large-scale screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The embodiments of the present invention are further described below with reference to the accompanying drawings, in which:

[0019] Figure 1 A schematic diagram of a model structure for predicting transcriptional response to a novel drug disturbance according to one embodiment of the present invention;

[0020] Figure 2 A schematic diagram of a method for virtual drug screening based on disease-sensitive drugs according to an embodiment of the present invention;

[0021] Figure 3 Schematic diagram of the structure of a drug virtual screening system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail by specific embodiments in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0023] The response of the transcriptome to drug perturbation refers to the changes in the gene expression pattern of cells or tissues after drug perturbation. This change can be detected by high-throughput sequencing technology such as RNA sequencing (RNA-seq), thereby revealing the effect of drugs on gene expression from the cellular level to the individual level. The study of the response of the transcriptome to drug perturbation helps to understand the mechanism of action of drugs, discover new drug targets, and redirect drugs. In drug development, this analysis is particularly important because it can provide a comprehensive understanding of the mechanism of action of drugs and help discover new drug mechanisms and treatment opportunities. Although deep learning-based methods have become an important tool for modeling the response of transcriptomes to drug perturbation, existing methods have their own limitations. The inventors found in the study of virtual screening of small molecule drugs that scGen and CPA both decouple expression profile-related attributes including cell type, drug, drug dosage, etc. by reconstructing the expression profile after drug perturbation, so they cannot predict the expression profile after new cell types or new drugs perturbation. scGen and CPA decouple perturbation attributes based on single-cell data, and cannot predict the expression profile after perturbation of multi-cell data, and have certain limitations for virtual screening and recommendation of drugs. chemCPA models drug perturbations through perturbation networks and can only predict the expression profile after a new single drug perturbation. It cannot match disease gene markers and small molecule gene markers for potential drug or drug combination screening.

[0024] In view of the defects that only single-cell or multi-cell sequencing methods are considered in the prior art, and there are certain limitations in scalability, some methods can only model experimentally perturbed data, cannot predict new drugs, and some methods cannot perform cell type-specific drug screening. In the embodiments of this application, a new solution for predicting transcriptional responses to new drug perturbations is provided, and the post-perturbation transcriptional response prediction problem is defined as a distribution generation problem conditioned on perturbation. The pre-perturbation expression spectrum and drug structure of single cells or multi-cells are mapped to the hidden space as perturbation conditions, and then the expression spectrum distribution after drug perturbation is generated through a deep generative model. Such a solution can be generalized to new cell types and new drug perturbation problems for small molecule drug virtual screening.

[0025] In one embodiment of the present application, a method for predicting the transcriptional response to a novel drug disturbance is provided. The chemical structure, dosage, and unperturbed transcriptional profile of the drug are provided as input to a pre-trained prediction model, through which the prediction model is used to predict the distribution of the transcriptional response after the given drug dose disturbance. Among them, the unperturbed transcriptional profile refers to the gene expression in the cells or tissues of the subject without external intervention or experimental treatment. In other words, it reflects the gene expression level of the cell in a natural state or under experimental control conditions. In biomedical research, scientists often perturb cells through experimental means (such as drug treatment, gene knockout, or environmental changes, etc.) to study the effects of these perturbations on cell function and gene expression. In contrast, the unperturbed transcriptional profile provides a baseline or reference point for comparing and understanding the changes caused by experimental perturbations. In the process of drug development, researchers may apply drugs to cells and then compare the differences in transcriptional profiles before and after drug treatment to understand how drugs affect cell gene expression and biological pathways. The unperturbed transcriptional profile is the "before" state in this comparison, while the transcriptional profile after drug treatment reflects the "after" state, that is, the result of drug perturbation. Unperturbed transcriptional profiles are very important for understanding the normal physiological functions of cells, the occurrence and development of diseases, and the mechanism of action of drugs. By analyzing these transcriptional profiles, researchers can identify key biomarkers, potential drug targets, and new strategies for disease treatment.

[0026] In one embodiment, Figure 1As shown, the prediction model may include a perturbation adapter, a perturbation encoder, and a perturbation decoder. The perturbation adapter encodes the chemical structure of the input drug and its dosage information into a perturbation embedding, and the perturbation encoder maps the effect of the input drug perturbation on the unperturbed transcriptional state into an interpretable latent space, that is, the perturbation encoder generates a latent embedding based on the perturbation embedding and the unperturbed transcriptional spectrum, and the perturbation decoder predicts the distribution of the transcriptional response after the input drug perturbation based on the latent representation of the unperturbed state (i.e., the latent embedding) and the perturbation embedding. The perturbation adapter of the prediction model of the embodiment of the present invention encodes the drug chemical structure and dosage information into the perturbation embedding, which only requires the chemical structure of the drug and can be expanded to a large-scale compound library without prior knowledge or annotation. The perturbation encoder maps the effect of the drug perturbation on the unperturbed transcriptional state into an interpretable latent space, and the perturbation encoder can encode two different levels of perturbations, single cell and multi-cell, and separate different cell types and drugs in the hidden space; the perturbation decoder predicts the distribution of the transcriptional response based on the latent representation of the unperturbed state, the perturbation embedding, and the noise.

[0027] The structure and training process of the prediction model of the embodiment of the present invention will be specifically introduced below.

[0028] In this application, the drug perturbation transcriptional response prediction problem is treated as a conditional probability generation problem. The prediction model is trained to learn a function f for predicting the transcriptional response under a new drug perturbation. In one embodiment, Gaussian negative log-likelihood loss can be selected as the loss function. The construction process of the training data set is as follows:

[0029] Given a training dataset represents the sample of the ith n-dimensional gene expression data (i.e., the gene expression data obtained by sequencing before and after the disturbance). In a training data set, the sequencing technology of the gene expression data samples is the same, that is, the total number of genes measured by the sequencing method is the same. i is the set of drug perturbation attributes, where s i =(s i,1 ,s i,2 ,s i,j …,si, M ) represents the canonical simplified molecular linear input specification (SMILES, a standardized format for representing chemical molecular structures using ASCII strings) format of M drugs that perturb the i-th sample of n-dimensional gene expression data (which can be referred to as the i-th perturbation). i =(d i,1 ,d i,2 ,d i,j ,…,d i,M ) represents the given dose of these M drugs. i,j= 0, it means that the jth drug is not applied to the i-th disturbance. If M = 0, it means that no disturbance is applied to the i-th sample. At this time, the i-th sample is in an undisturbed state, otherwise it is in a disturbed state. i Represents other covariates of the ith perturbation, such as which cell line the ith sample comes from, and can be expanded to cells, cell lines, patients, species, etc. according to the data set. The drug perturbation attribute set includes the SMILES format of all drugs input into the prediction model and the dosage information of the drug. The drug input into the prediction model can be a single drug or a combination of multiple drugs. When a drug combination is input, the prediction model outputs the distribution of transcriptional responses of the drug combination to the perturbation of cell gene expression. The training data also includes the gene expression profile x for each perturbed gene. i Assign an unperturbed gene expression profile The prediction model is used to predict the transcriptional response of new drug perturbations by randomly selecting from the unperturbed expression profiles of the same cell line in the dataset. Given an unperturbed gene expression The prediction model maps the unperturbed state to a distribution The transcriptional response after perturbation is sampled.

[0030] The prediction model of the embodiment of the present invention is a conditional generative model that aims to predict gene expression profiles under different perturbation conditions. The prediction model consists of three components: 1) a perturbation adapter, an extensible adapter The complex disturbance (s i ,d i ) is encoded as a k-dimensional perturbation embedding 2) Disturbance encoder The perturbation can be embedded in and unperturbed n-dimensional gene expression Encode into the latent space and get the latent state 3) Perturbation decoder Embed the perturbation and potential state Decoded into the Gaussian likelihood distribution of the perturbation state And through from Sampling the transcriptional response (i.e. expression profile) after the perturbation These three components are described in detail below.

[0031] 1) Perturbation Adapter

[0032] The disturbance adapter will P i Encoded as a fixed-size perturbation embedding Specifically, firstly, P iThe jth drug (in the embodiment of the present invention, the drug is the main compound component of the drug) is represented as the standard SMILES format s i,j , then through the RDKit encoder Convert it into a fixed-size first fingerprint embedding FCFP i,j (FCFP is a Morgan fingerprint with a fingerprint radius of 2, focusing on functional topological pharmacophores). Among them, RDKit encoder is a tool that uses the RDKit library to generate and process different types of encodings of chemical molecules. RDKit is an open source chemical informatics software library that provides a variety of functions to process chemical structures, including but not limited to SMILES encoding, SMARTS encoding, molecular fingerprints (such as Morgan fingerprints, including ECFP, FCFP), etc.

[0033] By all P i The first fingerprint embedding of the compound in is weighted summed up to calculate P as shown in formula (1) i The second fingerprint is embedded in rFCFP i :

[0034]

[0035] where φ is the j-th compound dose (d i,j ), j∈[0,M], H(s i,j ) represents the first fingerprint embedding after being encoded by the RDKit encoder.

[0036] As shown in formula (2), by perturbation adapter Generate the final perturbation embedding

[0037]

[0038] Among them, θ p are the parameters of the perturbation adapter, which can be a multi-layer perceptron with 2 or more layers.

[0039] Due to the extensibility of the SMILES format and the RDKit encoder, any compound can be encoded as a potential embedding, and thus the design of perturbation adapters is scalable to virtually screen new compounds for perturbations.

[0040] 2) Perturbation Encoder and Perturbation Decoder

[0041] Inspired by the variational autoencoder (VAE) framework, the prediction model consists of an encoder and a decoder estimating a Gaussian distribution for each gene, which is represented by the mean (μ i ) and variance Parameterization. Due to experimental technical limitations, only unpaired perturbed and unperturbed transcriptional profile data can be observed. In order to match the expression profiles before and after the perturbation, for single-cell RNA-seq datasets, each perturbed cell will be randomly assigned an unperturbed cell of the same cell type; for batch RNA-seq datasets, the perturbed state will be assigned the observed value of the same cell line.

[0042] Given the perturbation embedding generated by the perturbation adapter According to formula (3), the disturbance encoder Will and unperturbed expression profiles Mapping to latent embeddings Then according to formula (4), Input to the perturbation decoder Estimating the Gaussian likelihood distribution of the perturbed expression profile

[0043]

[0044] in, and They are all multi-layer perceptrons with 2 or more layers, θ e and θ d They are and Parameters of perturbation embedding It is used as the input of both the disturbance encoder and the disturbance decoder, and is multiplexed in the disturbance decoder. Combining chemical and biological background information helps generate more precise outputs.

[0045] In some embodiments, in order to increase the robustness of the model, a perturbation decoder can be used to embed the perturbation into Potential Embedding and noise Decoded into the Gaussian likelihood distribution of the perturbation state That is, formula (4) can be

[0046] According to formula (5) to formula (7), from the estimated Gaussian likelihood distribution The predicted perturbation expression profile is obtained by sampling

[0047]

[0048] in, Represents the nonlinear softplus function. μ i and are the mean and variance of each gene after perturbation, both of which are n-dimensional vectors. is the predicted transcriptional response after the n-dimensional perturbation.

[0049] 3) Training and testing

[0050] The prediction model training objective of the embodiment of the present invention is to minimize the Gaussian negative log-likelihood loss, which is defined as shown in formula (8):

[0051]

[0052] Among them, eps is used to limit To maintain stability, the default setting is 1e-6. By minimizing the loss of all cells / samples, the prediction model is able to learn the most appropriate parameters.

[0053] In order to train and test the prediction model, all data sets are strictly divided into training set, validation set and test set according to the disturbance attribute in a ratio of 6:2:2. Strict data division can prevent data leakage and bring better generalization performance. The prediction model of the embodiment of the present invention is trained on each divided category using a 5-fold cross-validation method.

[0054] like Figure 2As shown, according to one embodiment of the present invention, a method for virtual screening of new drugs based on disease-sensitive drugs is provided, comprising: in step S201, using the prediction method of the above embodiment, respectively predicting the expression spectrum after the disturbance of the sensitive drug and the compound to be screened in the compound library to the cell line (i.e., the cell line affected by a specific disease) (the expression spectrum after the disturbance is the gene expression data after the disturbance, which can also be called the transcription response after the disturbance). In step S202, based on the expression spectrum before and after the sensitive drug disturbs the cell line, the first up-regulation (increased expression) information or the first down-regulation (decreased expression) information of the gene level in the cell line is calculated, and the gene of the cell line disturbed by the sensitive drug is genetically marked according to the first up-regulation information or the first down-regulation information. In step S203, based on the expression spectrum before and after the compound to be screened disturbs the cell line, the second up-regulation information or the second down-regulation information of the gene level in the cell line is calculated, and the gene of the cell line disturbed by the compound to be screened is genetically marked according to the second up-regulation information or the second down-regulation information. In step S204, the enrichment score of the compound to be screened for the sensitive drug of the cell line is calculated according to the gene markers corresponding to the sensitive drug and the gene markers corresponding to the compound to be screened, and the compound to be screened is ranked according to the enrichment score. This ranking reflects the degree of influence of the compound to be screened on the expression of genes related to a specific disease. The higher the enrichment score, the more similar the gene expression effect pattern of the compound and the known sensitive drug on the cell line is, and therefore it may have a therapeutic effect or mechanism similar to that of the sensitive drug. Among them, sensitive drugs refer to drugs known to show efficacy on cell lines. New drugs, i.e. compounds to be screened, refer to drugs whose therapeutic effects are to be evaluated other than sensitive drugs.

[0055] In some embodiments, step S201 includes predicting the expression profiles of sensitive drugs and compounds to be screened in the compound library after perturbing the cell line at multiple concentration (i.e., dosage) gradients by a prediction method. In order to ensure the robustness of the calculation, the prediction step can be repeated multiple times to obtain multiple perturbed expression profiles, preferably 3 times.

[0056] In some embodiments, after step S201, the following step may be performed: the expression profile of the reference gene in the cell line is transformed and expanded into n u In this example, 978 genes were selected to represent the diversity of biological pathways and processes in human cells, and these representative genes are called benchmark genes.

[0057] In some embodiments, step S202 includes comparing the gene expression data before and after the sensitive drug perturbation, and calculating the gene expression data of each gene in the cell line by the sensitive drug (if the reference gene is linearly transformed, then here is the expanded n uThe gene expression data of each gene is obtained by calculating the fold change value of each gene (if not linearly transformed, it is calculated according to the reference gene). The first up-regulation information or the first down-regulation information of each gene is determined according to the fold change value. In some embodiments, a gene with a fold change value greater than 1 is an up-regulated gene, and the fold change value is also used as the up-regulation information (i.e., gene marker) of the gene. A gene with a fold change value less than 1 is a down-regulated gene, and the fold change value is also used as the down-regulation information of the gene. In some embodiments, if step S201 is repeated multiple times to obtain gene expression data after multiple disturbances, the average of these multiple fold change values ​​can be used as the up-regulation information or down-regulation information. For the sensitive drug, the genes of the cell line are sorted according to the fold change value corresponding to the gene (the first up-regulation information or the first down-regulation information) as the first sorted list.

[0058] Among them, up-regulation information and down-regulation information describe the changes in gene expression levels. Up-regulation means that under certain conditions (such as drug treatment, environmental changes, disease states, etc.), the expression level of a specific gene increases relative to a control group or a baseline state. This means that the amount of mRNA and / or protein transcribed from the gene increases, which is usually associated with the enhancement or activation of the gene function. Down-regulation means that under specific conditions, the expression level of a gene decreases relative to a control group or a baseline state. This usually means that the amount of mRNA and / or protein transcribed from the gene decreases, which may be associated with the weakening or inhibition of the gene function.

[0059] In some embodiments, step S203 includes comparing the gene expression data before and after the disturbance of the compound to be screened, and calculating the second up-regulation and second down-regulation information of the compound to be screened on the gene expression in the specific disease cell line. The gene after the disturbance of the compound to be screened is marked according to the second up-regulation information or the second down-regulation information. And for each compound to be screened, the genes of the cell line are sorted according to the fold change value (second up-regulation information or second down-regulation information) corresponding to the gene, as a second sorting list. The specific operation process here is similar to that in step 202, and will not be repeated here.

[0060] In some embodiments, step S204 includes using the set of up-regulated genes and the set of down-regulated genes corresponding to the sensitive drug to be screened, and comparing them with the gene markers after the disturbance of the compound to be screened. Through Gene Set Enrichment Analysis (GSEA), the enrichment score of the screening compound for the sensitive drug of the cell line can be calculated. The compounds to be screened are ranked according to the enrichment score, and the compounds to be screened are screened according to the ranking results. Wherein, the enrichment score is a statistic used to measure the similarity between the compounds to be screened and the known sensitive drugs in terms of affecting the expression of disease-related genes. The score is calculated based on the correlation between the gene expression changes caused by the compounds to be screened and the genes affected by the sensitive drugs, and the distribution of the up-regulated gene set and the down-regulated gene set of the sensitive compound in the second sorted list corresponding to the compounds to be screened can be evaluated. Wherein, the up-regulated gene set and the down-regulated gene set of the sensitive drug are used as the reference gene set.

[0061] For each compound to be screened, the enrichment score of the compound to be screened is calculated according to formula (9) to formula (11):

[0062]

[0063] Among them, p is the number of genes in the up-regulated gene set corresponding to the sensitive drug, m p Indicates the mth gene in the up-regulated gene set corresponding to the sensitive drug p The order of genes in the first sorted list, q is the number of genes in the down-regulated gene set corresponding to the sensitive drug, m q Indicates the mth gene in the down-regulated gene set corresponding to the sensitive drug p The order of genes in the first sorted list, n c is the number of genes in the cell line (if the reference gene is linearly transformed, then n c Can be n u ), R(m p ) represents the mth gene in the up-regulated gene set corresponding to the sensitive drug p The ranking order of genes in the second ranking list, R(m q -1) The mth gene in the down-regulated gene set corresponding to the sensitive drug q -1 The sorting order of the gene in the second sorted list.

[0064] like Figure 3As shown, an embodiment of the present invention also provides a drug virtual screening system, which includes a data upload module 301, a model prediction module 302, a data storage module 303 and a result visualization module 304. The system supports users to select existing drugs and expression spectra before medication or upload expression spectra or drugs. After collecting the data uploaded by the user, the system will perform data preprocessing, and perform expression spectrum prediction after medication and virtual screening of drugs. The final system returns the prediction results and recommended drugs and subsequent data analysis, and visualizes the results. Among them, the data upload module 301 can upload the unperturbed expression spectrum and compound structure, or select the unperturbed expression spectrum and compound structure provided by the system. The model prediction module 302 respectively executes the aforementioned method of predicting the transcriptional response after perturbation and performing virtual screening of new drugs based on disease-sensitive drugs in the embodiment. The data storage module 303 is used to store the data uploaded by the user and the predicted results, and at the same time store the drug features of the expression spectrum after medication that have been predicted. The result visualization module 304 is used to analyze and count the predicted results and visualize the screening results. Through this system, users can upload undisturbed expression spectra and compound structures in the data upload module, or select undisturbed expression spectra and compound structures provided by the system. After receiving the data uploaded by the user, the system background preprocesses the data, and the model prediction module predicts the expression spectrum after medication and calculates the ranking of compounds in the screening compound library. The data storage module stores the data uploaded by the user and the predicted results, and also stores the drug features of the expression spectrum after medication that the system has predicted in advance. Finally, the result visualization module analyzes and statistics the predicted results and visualizes the screening results.

[0065] In another embodiment of the present invention, a computer-readable storage medium is also provided, on which a computer program or executable instructions are stored. When the computer program or executable instructions are executed, the aforementioned embodiment is implemented, and the implementation principle is similar, which will not be repeated here. In an embodiment of the present invention, a computer-readable storage medium can be any tangible medium that can store data and can be read by a computing device. Examples of computer-readable storage media include hard disk drives, network attached storage (NAS), read-only memory, random access memory, CD-ROM, CD-R, CD-RW, magnetic tape, and other optical or non-optical data storage devices. Computer-readable storage media may also include computer-readable media distributed on a network-coupled computer system so that computer programs or instructions can be stored and executed in a distributed manner.

[0066] In another embodiment of the present invention, there is provided an electronic device, comprising a processor and a memory, wherein the memory is used to store executable instructions executable by the processor, wherein the processor is configured to execute the executable instructions stored on the memory, and when the executable instructions are executed, the neural network accelerator design method described in the aforementioned embodiment is implemented, and the implementation principle is similar and will not be repeated here.

[0067] References in this specification to "various embodiments," "some embodiments," "one embodiment," or "an embodiment," etc. refer to a particular feature, structure, or property described in conjunction with the embodiment being included in at least one embodiment. Thus, the appearances of the phrases "in various embodiments," "in some embodiments," "in one embodiment," or "in an embodiment," etc., in various places throughout the specification do not necessarily refer to the same embodiment. Furthermore, particular features, structures, or properties may be combined in any suitable manner in one or more embodiments. Thus, particular features, structures, or properties shown or described in conjunction with one embodiment may be combined in whole or in part with features, structures, or properties of one or more other embodiments without restriction, as long as the combination is not illogical or inoperable.

[0068] The expressions of "including", "having" and similar meanings in this specification are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units inherent to these processes, methods, products or devices. "One" or "an" does not exclude multiple situations. In addition, the elements in the drawings of this application are for illustrative purposes only and are not drawn to scale.

[0069] Although the present invention has been described through preferred embodiments, the present invention is not limited to the embodiments described herein but includes various changes and modifications that may be made without departing from the scope of the present invention.

Claims

1. A method for predicting transcriptional responses to novel drug perturbations, characterized in that include: The chemical structure of the novel drug and its dosage information as well as the unperturbed transcriptional profile are input into the prediction model; The prediction model outputs a distribution of transcriptional responses perturbed by an input novel drug; Wherein, the prediction model includes: A perturbation adapter, which is used to encode the chemical structure of the input novel drug and its dosage information into a perturbation embedding; a perturbation encoder for generating a latent embedding based on the perturbed embedding and the unperturbed transcription profile; and A perturbation decoder is configured to predict a distribution of transcriptional responses to a novel drug perturbation of the input based on the latent embedding and the perturbation embedding.

2. The method according to claim 1, characterized in that After the prediction model outputs the distribution of transcriptional responses perturbed by the novel drug, the method further comprises: Sampling is performed from the distribution of transcriptional responses perturbed by the novel drug to generate an expression profile after perturbation by the novel drug.

3. The method according to claim 1 or 2, characterized in that: The disturbance adapter is also used to: The input SMILES format of the novel drug and the dosage information of the novel drug are used as a drug perturbation attribute set; Converting the SMILES format of each of the novel drugs in the drug disturbance attribute set into a corresponding first fingerprint embedding; Calculating a second fingerprint embedding of the drug disturbance attribute set according to the first fingerprint embedding of each of the new drugs in the drug disturbance attribute set and the dosage information thereof; The perturbation embedding is generated based on the second fingerprint embedding.

4. The method according to claim 1 or 2, characterized in that: The disturbance adapter, the disturbance encoder and the disturbance decoder are respectively multi-layer perceptrons with 2 or more layers.

5. The method according to claim 1 or 2, characterized in that: The training objective of the prediction model is to minimize the Gaussian negative log-likelihood loss.

6. A method for virtual screening of compounds to be screened based on disease-sensitive drugs, characterized in that: include: Using the prediction method as described in claim 2, respectively predicting the expression profiles of sensitive drugs and compounds to be screened after perturbing the cell lines; Based on the expression profiles before and after the cell line is disturbed by the sensitive drug, first up-regulation information or first down-regulation information of the gene level in the cell line is calculated, and the gene of the cell line disturbed by the sensitive drug is genetically marked according to the first up-regulation information or the first down-regulation information; Based on the expression profiles of the cell line before and after the cell line is disturbed by the compound to be screened, calculating the second up-regulation information or the second down-regulation information of the gene level in the cell line, and marking the gene of the cell line disturbed by the compound to be screened according to the second up-regulation information or the second down-regulation information; According to the gene markers corresponding to the sensitive drugs and the gene markers corresponding to the compounds to be screened, the enrichment scores of the compounds to be screened for the sensitive drugs of the cell line are calculated, and the compounds to be screened are screened according to the enrichment scores.

7. The method according to claim 6, characterized in that Based on the expression profiles before and after the cell line is disturbed by the sensitive drug, first up-regulation information or first down-regulation information of the gene level in the cell line is calculated, and the gene of the cell line disturbed by the sensitive drug is marked according to the first up-regulation information or the first down-regulation information, including: Comparing the expression profiles before and after the perturbation of the sensitive drug, and calculating the fold change value of the gene expression data of each gene in the cell line by the sensitive drug as the first fold change value; marking the gene perturbed by the sensitive drug based on the first fold change value; and / or, Based on the expression profiles of the cell line before and after the compound to be screened perturbs the cell line, calculating the second up-regulation information or the second down-regulation information of the gene level in the cell line, and marking the gene of the cell line perturbed by the compound to be screened according to the second up-regulation information or the second down-regulation information, including: For each compound to be screened: Comparing the expression profiles before and after the perturbation of the compound to be screened, and calculating the fold change value of the gene expression data of each gene in the cell line by the compound to be screened as the second fold change value; Genes perturbed by the compounds to be screened are marked based on the second fold change value.

8. The method according to claim 7, characterized in that According to the gene markers corresponding to the sensitive drugs and the gene markers corresponding to the compounds to be screened, the enrichment scores of the compounds to be screened for the sensitive drugs of the cell lines are calculated, including: sorting the genes of the cell line according to the first fold change value to obtain a first sorting list; According to the gene marker results of sensitive drug perturbation, the up-regulated gene set and the down-regulated gene set corresponding to the sensitive drug are determined; For each compound to be screened: sorting the genes of the cell line according to the second fold change value to obtain a second sorting list; According to the first ranked list, the second ranked list, the up-regulated gene set, the down-regulated gene set and the number of cell line genes, the enrichment score of the compound to be screened for sensitive drugs of the cell line is calculated.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are loaded and executed by a processor, the method according to any one of claims 1 to 8 is implemented.

10. A drug virtual screening system, characterized in that: include: The data upload module is used to upload the unperturbed expression spectrum and the compound structure of the drug, or to select the unperturbed expression spectrum and the compound structure of the drug provided by the system; The model prediction module is used to execute the method according to any one of claims 1 to 8; The data storage module is used to store the uploaded data and prediction results, and store the predicted expression spectrum drug features after medication; The result visualization module is used to analyze and count the prediction results and visualize the screening results.

Images