Single cell transcription reaction prediction algorithm based on cross-domain feature cross migration

Abstract

The invention provides a single cell transcription reaction prediction algorithm based on cross-domain feature cross migration. Drug screening is carried out by using a deep learning technology. The algorithm combines feature decoupling, cross-domain transfer learning and linear modeling methods, and specifically comprises encoder feature decoupling, feature cross transfer, disturbance state modeling and transcriptional spectrum prediction. According to the method, cross-domain transfer learning is applied to the field of drug or gene disturbance-induced cellular response prediction for the first time, and the model is constructed by using gene transcriptional profile data. In the embodiment, the whole genome determination coefficient between the predicted drug perturbation transcriptional profile and the real transcriptional profile reaches 0.81, the differential genome determination coefficient reaches 0.62, efficient prediction of the unicellular external perturbation reaction is successfully achieved, a new thought and method are provided for the field of tumor treatment, and beneficial reference is provided for future drug research and development.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, which is an application of deep learning technology in drug research and development, and in particular to a single-cell transcriptional response prediction algorithm based on cross-domain feature cross-migration. Background Art

[0002] The interaction between drug molecules and targets is affected by the complex cellular environment, which makes it difficult for drugs designed to precisely bind to targets to induce specific cellular phenotypic changes in an expected manner, limiting the success rate of targeted drug discovery. Therefore, researchers have turned to phenotypic drug discovery. Transcriptomics analysis is a commonly used form of molecular phenotypic data used to reveal the response of cells to various interferences. Single-cell RNA sequencing can detect subtle changes in transcription levels within individual cells, identify cell subpopulations that are resistant to specific drugs, and analyze phenotypic effects at single-cell resolution.

[0003] Despite the power of these high-throughput screening techniques, their capacity is still limited compared to the huge space of all possible cell types and perturbation combinations. Therefore, it is very necessary to develop computational models that can use existing experimental data to predict phenotypic changes caused by different perturbations. Various methods have been proposed to predict cellular responses to drug treatments, but most machine learning models perform poorly when processing high-dimensional single-cell RNA sequencing (scRNA-seq) data, resulting in poor performance in high-dimensional response prediction.

[0004] To overcome these limitations, deep learning techniques have been widely used in the analysis of scRNA-seq data in recent years. For example, some deep encoder-decoder frameworks can generate interpretable drug and cell type embeddings, support drug similarity analysis and prediction of unseen doses and drug combinations, or predict single-cell transcriptional responses of unseen drugs by introducing drug molecule encoders. In addition, some models use knowledge graphs of gene-gene relationships and graph neural networks to model the effects of multi-gene interference on transcriptional responses, or use deep generative models to learn continuous mappings between interference space and cell state space.

[0005] However, these previous studies have not been committed to learning transferable drug representations, and therefore perform poorly when applied to new drugs. To make up for this shortcoming, this paper proposes a single-cell transcriptional response prediction algorithm based on cross-domain feature cross-migration, aiming to build a more robust and transferable model through deep learning technology to achieve efficient prediction of single-cell drug perturbation responses. Summary of the invention

[0006] The present invention aims to solve the defects in the prior art and proposes a new single-cell drug perturbation prediction algorithm by combining feature decoupling, cross-domain transfer learning and linear modeling. The algorithm performs well in single-cell drug perturbation prediction, provides new ideas and methods for related research in the field of tumor treatment, and also provides a useful reference for future drug development and treatment.

[0007] The idea of ​​the present invention is to construct a deep learning model through cross-domain transfer learning technology to realize drug perturbation prediction at the single-cell level. 1) Screen single-cell transcription profile data with drug dosages of 10μM and 0μM and effective gene expression numbers greater than 500, and perform differential gene analysis on the transcription profile data to screen out differential genes that have significant effects on biological properties. 2) Standardize the data and divide them into training sets, validation sets, and test sets, and pair them according to the principle of different drugs for the same cells. 3) Input paired transcription profile data into the model, decouple the drug perturbation features from the perturbation feature encoder, decouple the ground state feature encoder from the ground state feature, and impose orthogonal constraints and similarity constraints. 4) Linear modeling of feature vectors is performed so that the model focuses on learning transferable perturbation representations. 5) Adopt a cross-domain migration method, reconstruct and cross-domain cross-migration through the transcription profile generator, and impose reconstruction loss constraints to train the model. 6) Use the trained model to predict the drug perturbation response of single cells.

[0008] The specific technical solution of the present invention is: a single cell transcriptional response prediction algorithm based on cross-domain feature cross-migration, comprising the following steps:

[0009] The data required for the model is single-cell gene transcription profile data, which is obtained from public data sources.

[0010] The data were screened according to drug dose and effective gene expression number, and single-cell transcription profile data with doses of 10μM and 0μM and effective gene expression number greater than 500 were retained. Highly variable gene screening was performed, and the top 2000 differential genes were retained. The data were then standardized and divided into training sets, validation sets, and test sets according to the drug-level data isolation method.

[0011] Taking the training set as an example, the data is first divided into multiple subsets according to cell types, and then the types of drugs in each subset are counted and divided into two groups: Group 1 and Group 2. It is ensured that the drug information of the two groups of data is completely different. Each group is sorted according to the drug name, and then Group 1 and Group 2 are cyclically matched to obtain multiple pairs of transcriptional profile data.

[0012] Paired transcription profile data are input into the model, and the perturbation feature encoder decouples the corresponding drug perturbation features, and the cell ground state feature encoder decouples the corresponding ground state features. In order to achieve complete feature decoupling, orthogonal constraints are imposed on drug perturbation features and ground state features to ensure their independence, and similarity constraints are imposed on ground state features.

[0013] Linear modeling of feature vectors mapped to the same latent space allows linear combination of two perturbation representations with the cell ground state, making the model more focused on learning transferable perturbation representations. The perturbation feature encoder and the ground state feature encoder work together to achieve decoupled representation learning. With the help of the linear model, the combination from the homologous representation can not only restore the original information, but also restore the corresponding information after cross-transfer.

[0014] The cross-domain migration method is used to cross-migrate the two drug perturbation representations and the cell basal state characteristics, and two sets of reconstructed transcription profiles and two sets of cell gene transcription profiles that achieve drug migration are obtained through the transcription profile generator. The reconstruction loss and cross-domain cross loss are then used to train the model as a whole.

[0015] The trained perturbation feature encoder, cell ground state feature extractor and transcription profile generator are used to extract perturbation features and cell ground state features, migrate perturbation features and generate gene transcription profiles after migration, so as to predict the drug perturbation response of single cells.

[0016] The model network includes a perturbation feature encoder for extracting perturbation features, a ground state feature encoder for extracting ground state features, and a transcription spectrum generator for generating transcription spectrum.

[0017] The present invention has the following beneficial effects:

[0018] 1. This single-cell transcriptional response prediction algorithm based on cross-domain feature cross-migration is the first to apply cross-domain transfer learning technology to model cellular responses caused by drug perturbations. Through cross-domain migration constraints, our model can effectively learn transferable drug perturbation representations, thereby enhancing the model's generalization ability to unknown drugs and new cell environments. This approach breaks through the limitations of traditional models in dealing with new drugs and new cell lines, and significantly improves the adaptability and accuracy of the model.

[0019] 2. The single-cell transcriptional response prediction algorithm based on cross-domain feature cross-migration was comprehensively evaluated on different types of cell response datasets. The experimental results show that our model outperforms existing methods in all indicators and exhibits higher prediction accuracy and robustness. Especially when dealing with complex drug-cell interactions, our model demonstrates excellent performance, verifying its potential and advantages in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is the model algorithm framework diagram;

[0021] Figure 2 The bar chart shows the coefficient of determination of the other four algorithms on the same data set. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0023] Reference Figure 1 , a single-cell transcriptional response prediction algorithm based on cross-domain feature cross-migration, including the following steps:

[0024] S1: Obtain single-cell gene transcription profile data from public data sources, and screen single-cell transcription profile data with dosage values ​​of 10μM and 0μM and effective gene expression numbers greater than 500 based on drug dosage and effective gene expression numbers, and perform differential gene analysis on the gene transcription profile data to screen out differential genes that can better affect biological properties.

[0025] S2: Standardize the gene transcription profile data and divide the data into training set, validation set and test set according to the drug-level data isolation strategy. Then divide each data set into multiple subsets according to the cell line. Then count the types of drugs in each subset and divide them into two groups: Group 1 and Group 2. Ensure that the drug information of the two groups of data is completely different. Each group is sorted by drug name, and then group 1 and group 2 are matched cyclically to obtain multiple pairs of transcription profile data with the same cells but different drugs.

[0026] S3: Paired transcriptional profile data are input into the model, denoted as X(a) and X(b), i.e., the transcriptional profiles of cell X under perturbations a and b. The goal of the ground state feature encoder Es is to extract the cellular ground state from X(a) and X(b) and map it to the latent space, denoted as S(a) and S(b). Ideally, if perturbations a and b act on the same cell or cell line, the ground states S(a) and S(b) extracted from their induced gene transcription profiles should be the same. Therefore, we try to minimize the distance between them during model training. The similarity loss is defined using the Kullback-Leibler divergence

[0027]

[0028] At the same time, in order to achieve high-performance decoupling while avoiding information redundancy and making the extracted perturbation features and cell ground state features independent of each other, we strengthen the orthogonality of the expression profile decoupling between the base features and the perturbation features. Defined as:

[0029]

[0030] S4: We expect that the ground state features S(a) and S(b) extracted by the ground state feature encoder can reconstruct the actual unperturbed transcription profile X. Therefore, the transcription profile generator D aims to recover the unperturbed transcription profile from the ground state in the latent space. To this end, we define the mean square error between the actual unperturbed transcription profile and the inferred transcription profile as the ground state reconstruction loss and minimize it during model training. The ground state reconstruction loss is defined as follows:

[0031]

[0032] In addition, we assume that in the latent space, the effect of perturbation on the cell state follows a linear superposition rule, so the perturbed transcriptional profile is restored by the same transcriptional profile generator. Specifically, suppose the transcriptional profile X(a) is decoupled into the base state S(a) and the perturbation feature P(a), then the sum of S(a) and P(a) in the latent space represents the perturbation imposed on the base state. Then, the transcriptional profile generator reconstructs the perturbed transcriptional profile X(a) by inputting (S(a)+P(a)). The same operation applies to X(b). The perturbation state reconstruction loss is defined as follows:

[0033]

[0034] S5: To learn transferable perturbations, we exchange the two representations P(a) and P(b) extracted from the transcriptional profiles X(a) and X(b). Under the assumption of the linear superposition rule, the cross-domain transferred perturbations are added to the base states S(a) and S(b), respectively, to simulate the effect of the perturbation on the cell state, i.e., (S(a)+P(b)) and (S(b)+P(a)). Next, the corresponding perturbed transcriptional profile is restored using the shared transcriptional profile generator D. The cross-perturbation loss function is defined as follows:

[0035]

[0036] S6: The two encoders Es and Ep collaborate to achieve feature disentanglement, while the cross-domain constraints force them to learn transferable perturbation representations. The total loss function is defined as follows:

[0037]

[0038] The encoder / decoder network is implemented using a fully connected feed-forward network with ReLU activation function.

[0039] In the network model:

[0040] Cell ground state feature encoder Es: Extracting cell ground state features from perturbed gene transcription profiles.

[0041] Perturbation Feature Encoder Ep: Extracting drug perturbation features from drug-induced gene transcription profiles.

[0042] Transcription profile generator D: maps the ground state and perturbation features in the latent space back to the corresponding transcription profiles.

[0043] Cross-domain transfer module: Exchange perturbation features in latent space to simulate the effects of different drugs on cell states.

[0044] In this example, single-cell transcription profiles from public data sources were used as experimental data. This data set contains three human cancer cell lines, and high-throughput screening technology is used to provide the measured values ​​of more than 7,000 drug-sensitive genes in hundreds of thousands of cells on these three cell lines. Transcription profile data with a dose of 10 μM and 0 μM and an effective gene expression number greater than 500 were used, and differential gene analysis was performed. The transcription profile data were paired according to the principle of different drugs for the same cell and used as model input, and the determination coefficient (R) between the predicted perturbed whole gene and differential gene transcription profile data and the real perturbated whole gene and differential gene transcription profile data was calculated. 2 ) to represent the model performance and compared with four baseline models.

[0045] Experiments show that the model is more accurate than the four benchmark models of Baseline, chemCPA, chemCPA+pretraining, and cycleCDR on the whole genome data. 2 The results were improved by 60%, 58%, 19% and 4.4% respectively. 2 The increases were 113%, 93%, 14.8% and 12.7% respectively.

[0046] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A single-cell transcriptional response prediction algorithm based on cross-domain feature cross-migration, characterized by: Includes steps: 1) Data screening and preprocessing: Screen single-cell transcriptional profile data with drug doses of 10 μM and 0 μM and effective gene expression greater than 500, and perform differential gene analysis on the transcriptional profile data to screen out differential genes that can better affect biological properties. 2) The data is standardized and divided into training set, validation set and test set according to the type of drug perturbation to which the cells are subjected. The transcriptional profile data in the training set, validation set and test set are then paired according to the principle of the same cells with different drugs. 3) Paired transcriptional profile data are input into the model, and the perturbation feature encoder decouples the corresponding perturbation features, and the cell ground state feature encoder decouples the corresponding ground state features. In order to achieve complete feature decoupling, orthogonal constraints are imposed on drug perturbation features and ground state features to ensure independence between the two, and similarity constraints are imposed on ground state features to ensure the accuracy of the cell ground state feature extractor. 4) Linear modeling of feature vectors mapped to the same latent space allows linear combination of the two perturbation representations with the cell ground state, making the model more focused on learning transferable perturbation representations. 5) The cross-domain migration method is used to cross-migrate the two drug perturbation representations and the cell basal state characteristics, and two sets of reconstructed transcription profiles and two sets of cell gene transcription profiles that have achieved drug migration are obtained through the transcription profile generator. The reconstruction loss and cross-domain cross loss are then used to train the model as a whole. 6) Use the trained perturbation feature encoder, cell ground state feature extractor and transcription profile generator to extract perturbation features and cell ground state features, migrate perturbation features and generate gene transcription profiles after migration, so as to predict the drug perturbation response of single cells. In the step 1), single-cell gene transcription profile data is obtained from a public data source. In the step 2), the data is paired in the following manner: taking the training set as an example, the data is first divided into multiple subsets according to the cell line, and then the types of drugs in each subset are counted, and the subsets are equally divided into two groups: group 1 and group 2, to ensure that the drug information of the two groups of data is completely different, each group is sorted according to the drug name, and then group 1 and group 2 are matched cyclically, so that the pairing rule of different drugs for the same cells is implemented.

2. According to claim 1, a single cell transcriptional response prediction algorithm based on cross-domain feature cross-migration is characterized in that: The transcriptional profile data were screened for highly variable genes, the top 2000 genes were retained, and then the dataset was divided strictly according to the drug-level data isolation strategy.

3. The single-cell transcriptional response prediction algorithm based on cross-domain feature cross-migration according to claim 1, characterized in that: In step 3), the perturbation feature encoder and the ground state feature encoder are fully connected neural networks responsible for decoupling and separating the perturbation features and the ground state features from the transcription spectrum data and mapping them to a low-dimensional latent space.

4. The single-cell transcriptional response prediction algorithm based on cross-domain feature cross-migration according to claim 1, characterized in that: In step 4), the perturbation feature encoder and the ground state feature encoder work together to realize decoupled representation learning. With the support of the linear model, the combination from the homologous representation can not only restore the original information, but also restore the corresponding information after cross-transfer.

5. The single-cell transcriptional response prediction algorithm based on cross-domain feature cross-migration according to claim 1, characterized in that: In step 5), a low-dimensional perturbation representation in the latent space is constructed based on the linear model, and the transcription profile generator is responsible for reconstructing the low-dimensional perturbation representation into a fully connected neural network corresponding to the cell transcription profile.

6. The single-cell transcriptional response prediction algorithm based on cross-domain feature cross-migration according to claim 1, characterized in that: The transcription profile generator processes the perturbation representation in the latent space and can accurately predict and output the differential gene transcription profile data of single cells after being perturbed by specific drugs.

Images