A method for analyzing ligand-receptor interaction-mediated cell communication

Abstract

The application discloses a kind of analysis ligand-receptor interaction mediated cell communication method, first collect multiple ligand-receptor interaction data sets, secondly by a kind of combined heterogeneous newton Boosting model and deep neural network's heterogeneous deep integration model to screen and identify potential ligand-receptor interaction, third based on single cell transcriptome data to filter known and predicted ligand-receptor interaction in organization, fourth based on filtered ligand-receptor interaction and expression threshold method, expression product method and joint score strategy prediction cell communication, finally cell communication result is visualized, this framework is further applied to analysis human melanoma tissue and head and neck squamous cell carcinoma tissue in cell communication, the application helps the prediction of cell communication in tumor microenvironment, and further help anticancer drug design and tumor targeted therapy.

Description

Technical Field

[0001] This invention belongs to the field of systems bioinformatics, specifically a method for analyzing cell communication mediated by ligand-receptor interactions. Background Technology

[0002] In multicellular organisms, cell communication induces multiple cells to coordinate with each other, forming tissues, organs, and systems, and further enabling various life activities. Many cancers rely heavily on communication between cancer cells and between cancer cells and normal cells. Cell communication is crucial for understanding tumorigenesis and development, tumor immunity, and treatment resistance. Its predictions contribute to understanding the molecular mechanisms of tumor progression and metastasis, and further provide guidance for anticancer drug design and targeted cancer therapy.

[0003] Despite significant advancements in biomedical experimental techniques, our understanding of cell communication remains incomplete. With the continuous maturation of single-cell sequencing technology, identifying cell communications has become a research hotspot. Cell communication is typically mediated by ligand-receptor interactions (LRIs), where ligands either bind to receptors in a soluble form or in a membrane-bound form, requiring the two communicating cell types to be physically similar. When the significance of ligand-receptor interactions for patient prognosis is fully elucidated, treatments targeting cell signaling pathways will be a powerful strategy in clinical practice. To date, several computational methods have been proposed to predict cell communication based on single-cell transcriptome data and known ligand-receptor interactions.

[0004] These methods fall into four main categories: differential expression combinatorial methods, network-based methods, expression perturbation-based methods, and tensor-based methods. However, most differential expression combinatorial methods rely solely on the sum of ligand and receptor expression in known ligand-receptor interaction data to predict cell communication specifically, without comprehensively considering ligand-receptor interactions across all cell types. Network-based methods fail to account for gene regulatory mechanisms. Tensor-based methods struggle to interpret cell communication strength from a tensor decomposition perspective. More importantly, all four methods require addressing the issues of false positives and false negatives.

[0005] Patent CN112466403B discloses a cell communication analysis method and system. This method includes cell communication prediction and ligand-target gene regulation prediction. Cell communication prediction includes ligand-receptor pair expression abundance analysis, ligand-receptor pair number analysis, significantly enriched ligand-receptor pair number analysis, and construction of a cell interaction network diagram. Ligand-target gene regulation prediction includes ligand activity analysis and ligand-target gene regulatory potential analysis. This invention's cell communication analysis method mainly describes the relationships between cells. It uses CellphoneDB software to construct cell communication networks based on single-cell gene expression matrices and NicheNet software to analyze ligand activity and the regulatory potential of ligands on target genes based on ligand-target gene expression relationships. This patent needs to address the issues of false positives and false negatives. Summary of the Invention

[0006] To address the problems mentioned in the background section, this invention provides a method for analyzing cell communication mediated by ligand-receptor interactions, thereby resolving the issues of the prior art.

[0007] The technical solution adopted in this invention is:

[0008] This invention provides a method for analyzing ligand-receptor interaction-mediated cell communication, comprising the following steps:

[0009] S1: Ligand-receptor interaction data arrangement, collecting four ligand-receptor interaction datasets;

[0010] S2: Ligand-receptor interaction prediction, using a heterogeneous deep ensemble model that combines a heterogeneous Newton Boosting model and a deep neural network to screen and identify potential ligand-receptor interactions;

[0011] S3: Ligand-receptor interaction filtering, which filters ligand-receptor interactions by combining single-cell transcriptome data, known ligand-receptor interactions, and identified ligand-receptor interactions;

[0012] S4: Cell communication prediction;

[0013] S5: Visualization of cell communication.

[0014] Furthermore, the heterogeneous deep ensemble model screening and identification of potential ligand-receptor interactions in step S2 includes the following steps:

[0015] S2.1: Feature extraction;

[0016] S2.2: Dimensionality Reduction;

[0017] S2.3: Ligand-Receptor Interaction Classification. Assume D = (X, Y) represents a ligand-receptor interaction dataset with n samples (ligand-receptor pairs), where (X, Y) represents a training sample, x ∈ X is a d-dimensional feature vector, and y ∈ Y represents its label. For the i-th (i = 1, 2, ..., n) sample x... i If the ligand-receptor pair interacts, then y i =1, otherwise y i =0.

[0018] S2.3.1 The interaction probability of each ligand-receptor pair is calculated using a heterogeneous Newtonian Boosting machine;

[0019] To predict x i The labels are defined by the objective function as follows:

[0020]

[0021] Where: y i and f(x) i ) represent x respectively i The true and predicted labels. Loss function l(y i ,f(x i )) is twice differentiable, corresponding to f(x) i ), l′(y i ,f(x i )) and l″(y i ,f(x i )) represent the first and second derivatives, respectively.

[0022] Assume that each boost iteration can be obtained from one of K distinct subclasses, and H (k) This represents the Kth (K = 1, 2, ..., K) subclass defined by the following formula:

[0023]

[0024] in Describes a finite class function b(x) i ):R→R satisfies

[0025]

[0026] For the domain defined by the formula above, in each boosting iteration, a subclass is randomly selected to contain several binary decision trees, with a maximum depth of D. min and D max The elements are randomly and uniformly distributed. Therefore, K = N. D (N D =Dmax -D min +1) The only choice for the subclass. The corresponding probability mass function Φ can be defined by the following formula:

[0027]

[0028] Let u m (u m =1,2,...,K) is the sampling subclass index at the m-th boosting iteration. The basic assumption for insertion at the m-th boosting iteration is determined by the following formula:

[0029]

[0030] Where g i =l′(y i ,f m-1 (x i )), h i =l″(y i ,f m-1 (x i ))

[0031] To ensure global convergence, the model ∈>0 is updated by applying the learning rate:

[0032] f m (x i )=f m-1 (x i )+∈b m (x i )

[0033] Finally, the optimized model can be solved using Algorithm 1. Based on Algorithm 1, the target ligand-receptor pair x can be calculated. i Interaction probability

[0034] Algorithm 1: Heterogeneous Newton Boosting Machine

[0035] 1: Initialize f 0 (x i ) = 0

[0036] 2: Entering the for loop, when m = 1,...,M, the following statement is executed:

[0037] 3: Calculate g i =l′(y i ,f m-1 (x i ))

[0038] 4: Calculate h i =l″(y i ,fm-1 (x i ))

[0039] 5: Sample subclass index μ based on Φ m

[0040] 6: To ensure compliance with the basic assumptions,

[0041] 7: Update the model: f m (x)=f m-1 (x i )+∈b m (x i )

[0042] 8: End the for loop

[0043] 9: Output:

[0044] S2.3.2 The interaction probability of each ligand-receptor pair is calculated using a deep neural network;

[0045] like Figure 9 As shown, the DNN model consists of two Long Short-Term Memory (LSTM) network layers, three fully connected layers, and one output layer. The first two layers are LSTM network layers with d neurons, used to output the transformed d-dimensional feature vector b. LSTM networks are an important deep learning model based on backpropagation, which can better convey information. An LSTM block consists of an input gate L, a forget gate, and a forget gate. The output gate ω consists of a single unit. The weights of the autorecursive model can be dynamically changed by adding gates. In LSTM networks, the input and the output in the forget gate can be based on the input and previous unit states. Assume... x represents i At time t, the j-th input (feature) represents one of c storage units. CL , and w cω These represent the peephole weights from unit c to the three doors. This represents the state of unit c at time t. Let f represent the output of the h-th neuron at time t. f, g, and h represent the activation functions in the three gates, respectively.

[0046] The input and output of an input gate can be represented by the following formulas:

[0047]

[0048] The input and output of the forget gate can be described by the following formula:

[0049]

[0050] The input and state of unit c at time t can be represented by the following formula:

[0051]

[0052] The output of unit c at time t can be described by the following formula:

[0053]

[0054] The input and output of an output gate can be represented by the following formula.

[0055]

[0056]

[0057] Based on the above formula, the final output b can be obtained, which is a d-dimensional feature vector transformed by two LSTM network layers.

[0058] The next three layers are fully connected layers consisting of 256, 128, and 64 neurons, respectively. In the first layer, the input is a transformed d-dimensional feature vector b, representing a ligand-receptor pair. ReLU is used as the activation function for all three layers to mitigate the vanishing gradient phenomenon and further improve generalization performance. Its definition is expressed below:

[0059]

[0060] Where a represents a parameter, a = 1.

[0061] The output layer consists of a single neuron and outputs each target ligand-receptor pair x based on a sigmoid function defined by a formula. i The interaction probability. The formula is as follows:

[0062]

[0063] In the output layer, the binary cross-entropy function is minimized to clarify that the predicted ligand-receptor interaction is expressed by the following formula:

[0064]

[0065] Where: y i Indicates target ligand-receptor pair x i The true label, x represents i The predicted interaction probabilities were calculated. The experiment was trained for 100 epochs with a minimum batch size of 128.

[0066] S2.4: Ensemble learning integrates a heterogeneous Newton Boosting (HNBM) model and a deep neural network to obtain a heterogeneous deep ensemble model for the final ligand-receptor interaction. During the learning process, HNBM and DNN serve as basic classifiers, and the outputs of both models are integrated to obtain the final classification result. For a ligand-receptor pair x... i Assuming and Let HNBM and DNN represent the interaction probabilities calculated by HNBM and DNN, respectively, and the final ligand-receptor interaction probability. This can be obtained from the following expression:

[0067]

[0068] Wherein: parameter α represents the relative importance of HNBM and DNN for the classification performance of ligand-receptor interactions, and needs to be learned through cross-validation.

[0069] Furthermore, in step S2.4, α is set in the range of [0.5, 0.8], preferably 0.6. When α is set to 0.6 on most datasets, the proposed LRI-HDEnHD model can achieve higher AUC and AUPR on the four ligand-receptor interaction datasets.

[0070] Furthermore, the feature extraction in step S2.1 includes the following steps:

[0071] S2.1.1: Obtain ligand-receptor sequence information from the UniProt database;

[0072] S2.1.2: Biological characteristics of proteins are extracted using methods including kmer, combined autocovariance and crosscovariance (ACC), distance-based Top-n graph (DT), distance-pair-based pseudo-amino acid composition (PseAAC-DP), parallel correlated pseudo-amino acid composition (PC-PseAAC), and serial correlated pseudo-amino acid composition (SC-PseAAC). Let p... o This represents a protein (ligand or receptor) sequence containing L amino acid residues:

[0073] p o ={R1R2…R i …R L}

[0074] Where R i p o The amino acid residue at position i.

[0075] Let t represent the distance between each residue pair. Extract the biological characteristics of the protein, including Kmer, ACC, DT, PseAAC-DP, PC-PseAAC, and SC-PseAAC, to represent the ligand / receptor.

[0076] Kmer: Based on the Kmer method proposed by Liu et al., proteins can be represented based on the frequency of occurrence of k adjacent amino acids. Therefore, biological characteristics of each protein based on Kmer are extracted, and the ligand / receptor is described as a 400-dimensional vector.

[0077] ACC: ACC is a combination of autocovariance and crosscovariance. Autocovariance measures the association between the same properties of each pair of residuals at a distance t.

[0078]

[0079] Where P U (R i ) represents R i The physicochemical property values, This represents the average value of physical and chemical properties.

[0080]

[0081] Cross-covariance evaluates the association between two distinct attributes of a residual pair with a distance t using the following formula:

[0082]

[0083] U1 and U2 represent two different physicochemical indices.

[0084] Assuming that the N1 and N2 features are calculated using the autocovariance and crosscovariance methods, the ACC features of the ligand / receptor can be represented as a (N1+2) dimensional feature vector.

[0085] DT: DT is used to measure the distance between the top n gram pairs. It obtains the top-n-gram feature vector based on the relative position information of each top-n-gram pair in the sequence composed of top-n-grams. This invention constructs a 1220-dimensional distance-based top-1-gram feature vector to describe ligand / receptor.

[0086] PseAAC-DP describes a protein description based on a reduced alphabet method and the distance between each pair of residues. Using the amino acid cluster spectrum cp provided by Liu et al. as the reduced alphabet method, a feature vector is defined, representing the ligand / receptor pair using the following formula:

[0087]

[0088]

[0089] Where f t (u)=f(R i ,R j |t) represents a residue pair separated by a distance t (R i and R j The frequency of occurrence of ).

[0090] The PseAAC method can extract the main features of 20 amino acid components, but it loses other features besides these 20 components. PC-PseAAC integrates adjacent local and global sequence order information based on the protein sequence.

[0091] Sequence order correlation factor:

[0092]

[0093]

[0094] Where h 1 (R i ), h 2 (R i ) and M(R i ) represent R respectively i The values ​​of hydrophobicity, hydrophilicity, and side chain mass are derived from the following formula:

[0095]

[0096] in, and M0(R) i ) represent R respectively i The original values ​​of hydrophobicity, hydrophilicity, and mass.

[0097] Finally, the PC-PseAAC characteristics of the ligand / receptor are expressed by the following formula:

[0098]

[0099]

[0100] Where f i This represents the normalized frequency of the i-th amino acid.

[0101] SC-PSEAC is a variant of PC-PseAAC. Considering that different types of proteins have different hydrophobic and hydrophilic characteristics, Chou et al. defined sequence order characteristics using the following formula:

[0102]

[0103] in j and Representing two residues R respectively i and R j The correlation between hydrophobicity and hydrophilicity is expressed by the following formula:

[0104]

[0105]

[0106] An augmented discrete form was achieved by incorporating the 2λ amphiphilic correlation factor into the classical amino acid composition, as shown in the following formula:

[0107]

[0108]

[0109] S2.1.3: Obtain the number of features for each type, fuse these features to describe the ligand-receptor pair, and a ligand-receptor pair can be represented as a 4576-dimensional vector using cascade operations.

[0110] Furthermore, step S2.2 dimensionality reduction includes the following steps: after feature extraction, a ligand-receptor pair is characterized as a vector with 4576 dimensions. Dimensionality reduction based on principal component analysis retains 99% of the information. Finally, a d-dimensional feature vector is obtained to describe each ligand-receptor pair.

[0111] Further, step S3, ligand-receptor interaction filtering, includes the following steps:

[0112] S3.1: Download single-cell transcriptome data of the target tissue from the GEO database;

[0113] S3.2: Update ligand-receptor interaction data by merging predicted ligand-receptor interactions with known ligand-receptor interactions;

[0114] S3.3: Ligand-receptor interactions are removed when the ligand or receptor from the ligand-receptor interaction is not expressed in the cells of the corresponding tissue's single-cell transcriptome data.

[0115] Furthermore, the joint scoring strategy in step S4 includes the following steps:

[0116] S4.1: Given p filtered LRIs {(l1,r1), (l2,r2), ..., (l... p ,r p )}, where l i / ri Represents single-cell transcriptome data and m-cell types {C1, C2, ..., C} k ,…,C m The expression vectors of ligand i / receptor j in the cell type are calculated first. k1 Arithmetic mean expression value of ligand i and cell type C k2 Arithmetic mean of receptor j Then, a joint scoring strategy combining the expression threshold method and the expression product method is used to describe C. k1 and C k2 The communication potential between them.

[0117] First, C is described based on the expression threshold method. k1 and C k2 The communication potential between them;

[0118] For two cell types C mediated by ligand-receptor interactions k1 and C k2 When the gene expression levels of ligands and receptors from ligand-receptor interactions exceed a given threshold, a potential signaling pathway may exist. Therefore, the expression threshold method can be used to test C. k1 and C k2 The "communication" between them. First, the arithmetic mean expression value of all cells was calculated based on the expression levels of ligand i / receptor j. and standard deviation σ i / σ j Secondly, when Greater than At that time, the ligand was thought to be in C k1 Highly expressed. Similarly, when Greater than At that time, the receptor is thought to be in C k2 Highly expressed in C10 cells. For ligand-receptor interactions, when ligand i and receptor j from the ligand-receptor interaction are in cell type C10... k1 and C k2 When expressed at medium to high levels, they are respectively regarded as C k1 and C k2 Potential ligand-receptor interactions between The detailed definition is as follows:

[0119]

[0120] The recognized ligand-receptor interactions also indicate the direction of the communication path (sender → receiver). Finally, C k1 and C k2 The communication score g1(k1,k2) between them is obtained by applying the formula to all LRIs. 1The calculation is done by summation, expressed as follows:

[0121]

[0122] Then, C is described based on the expression product method. k1 and C k2 The communication potential between them;

[0123] Based on single-cell transcriptome data and filtered ligand-receptor interactions, the expression product method was used to score cell communication strength. First, cell type C... k1 and C k2 One ligand-receptor pair (l i ,r i The expression for the product is defined by the following formula:

[0124]

[0125] C k1 and C k2 The communication score between them, g2(k1,k2), is then calculated using the following formula for all LRIs. 2 To calculate by summation:

[0126]

[0127] S4.2: Normalize the two expressions above using the min-max method, and the expression is as follows:

[0128]

[0129] Wherein represents represents g1(k3,k4)(k3,k4=1,2,…,m);

[0130]

[0131] Wherein represents represents g2(k3,k4)(k3,k4=1,2,…,m);

[0132] S4.3: Finally, C k1 and C k2 The communication scores between them can be added together. and We obtain the following formula:

[0133]

[0134] The cell communication score g1(k1,k2) of the joint scoring strategy was obtained.

[0135] Furthermore, step S5, analyzing and visualizing cell communication in human melanoma tissue and head and neck squamous cell carcinoma tissue, includes the following steps: illustrating the top 3 ligand-receptor interactions of each cell communication intensity in human melanoma tissue and head and neck squamous cell carcinoma tissue, and visualizing cell communication using heatmaps and network views.

[0136] Furthermore, the analysis and visualization of cell communications in human melanoma tissue revealed interactions between CAF cells, macrophages, and endothelial cells and melanoma cancer cells.

[0137] Furthermore, the analysis and visualization of cell communication in head and neck squamous cell carcinoma (HNSCC) tissue concluded that endothelial cells, macrophages, and fibroblasts interact with head and neck squamous cell carcinoma.

[0138] Compared with the prior art, the beneficial effects of the present invention are:

[0139] The heterogeneous deep ensemble model in this invention achieves optimal performance, demonstrating its powerful ligand-receptor interaction classification ability and improving ligand-receptor interaction recognition performance. The joint scoring strategy in this invention can identify gene expression heterogeneity and further aid in differential gene expression analysis, resolving the false positive and false negative problems caused by expression threshold methods. Simultaneously, this invention also addresses the issue of expression product methods where, when there are significant differences in the transcriptional level between ligands and receptors from ligand-receptor interactions, one ligand or receptor dominates the interaction signal, thereby contributing to cell communication prediction in the tumor microenvironment and further assisting in anticancer drug design and tumor-targeted therapy. Attached Figure Description

[0140] Figure 1 Flowchart of a method for analyzing cell communication mediated by ligand-receptor interactions;

[0141] Figure 2 A method for analyzing ligand-receptor interaction-mediated cell communication: Step S2, Ligand-receptor interaction recognition method flowchart;

[0142] Figure 3 LSTM network;

[0143] Figure 4 AUC of the five methods on dataset 1;

[0144] Figure 5 AUC of the five methods on dataset 2;

[0145] Figure 6 AUC of the five methods on dataset 3;

[0146] Figure 7 AUC of the five methods on dataset 4;

[0147] Figure 8 AUPRs for the five methods on dataset 1;

[0148] Figure 9 AUPRs for the five methods on dataset 2;

[0149] Figure 10 AUPRs for the five methods on dataset 3;

[0150] Figure 11 AUPRs for the five methods on dataset 4;

[0151] Figure 12 Heatmap of cell communication in human melanoma tissue based on filtered ligand-receptor interaction and expression threshold methods;

[0152] Figure 13 A network view of cell communication in human melanoma tissue based on filtered ligand-receptor interactions and expression thresholding methods;

[0153] Figure 14 Top 3 ligand-receptor interactions in cell communication in human melanoma tissues based on filtered ligand-receptor interactions and expression product methods;

[0154] Figure 15 Heatmap of cell communication in human melanoma tissue based on filtered ligand-receptor interaction and expression product method;

[0155] Figure 16 A network view of cell communication in human melanoma tissue based on a filtered ligand-receptor interaction and expression product approach;

[0156] Figure 17 The top 3 ligand-receptor interactions in human melanoma tissue cell communication analysis based on filtered ligand-receptor interactions and a joint scoring strategy;

[0157] Figure 18 Heatmap of cell communication analysis in human melanoma tissues based on filtered ligand-receptor interactions and a joint scoring strategy;

[0158] Figure 19 A network view of cell communication analysis in human melanoma tissues based on filtered ligand-receptor interactions and a joint scoring strategy;

[0159] Figure 20Heatmap of HNSCC tissue cell communication analysis based on filtered ligand-receptor interactions and expression threshold method;

[0160] Figure 21 Network view of HNSCC tissue cell communication analysis based on filtered ligand-receptor interactions and expression thresholding;

[0161] Figure 22 The top 3 ligand-receptor interactions in HNSCC tissue cell communication analysis based on filtered ligand-receptor interactions and expression product method;

[0162] Figure 23 Heatmap of HNSCC tissue cell communication analysis based on filtered ligand-receptor interaction and expression product method;

[0163] Figure 24 Network view of HNSCC tissue cell communication analysis based on filtered ligand-receptor interaction and expression product method;

[0164] Figure 25 The top 3 ligand-receptor interactions in HNSCC tissues based on filtered ligand-receptor interactions and a joint scoring strategy;

[0165] Figure 26 Heatmap of cell communication analysis in HNSCC tissues based on filtered ligand-receptor interactions and a joint scoring strategy;

[0166] Figure 27 A network view of cell communication analysis in HNSCC tissues based on filtered ligand-receptor interactions and a joint scoring strategy;

[0167] Figure 28 Comparison of cell communication analysis results of LRI-HDEnHD-joint with CellPhoneDB, LIANA, and CellChat in human melanoma tissue;

[0168] Figure 29 Comparison of cell communication analysis results of LRI-HDEnHD-joint with CellPhoneDB, LIANA, and CellChat in human HNSCC tissues. Detailed Implementation

[0169] To clearly illustrate the technical features of this solution, the present invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application; however, this application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of this application is not limited to the specific embodiments disclosed below. Unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are for the purpose of describing specific embodiments only and are not intended to limit the scope of protection of the present invention.

[0170] Example 1

[0171] One embodiment of the present invention provides a method for analyzing ligand-receptor interaction-mediated cell communication, comprising the following steps:

[0172] S1: Ligand-receptor interaction data processing. Four ligand-receptor interaction datasets were collected, as shown in Table 1.

[0173] Datasets 1 and 2 were obtained by Shao et al. through manual searching of protein-protein interactions in the STRING database using text mining techniques. These two datasets contain 3398 human ligand-receptor interactions and 2033 mouse ligand-receptor interactions, respectively. For both datasets, this invention removed duplicate ligand-receptor interactions and those lacking sequence information for either the ligand or receptor in the Uniprot database. After preprocessing, Dataset 1 contains 3390 ligand-receptor interactions between 812 human ligands and 780 human receptors. Dataset 2 contains 2031 ligand-receptor interactions between 650 mouse ligands and 588 mouse receptors.

[0174] Dataset 3 was constructed by Skelly et al. using the human ligand-receptor interaction dataset provided by Ramilowski et al. This invention preprocesses Dataset 3 to obtain 2006 ligand-receptor interactions containing 574 human ligands and 559 receptors.

[0175] Dataset 4 was constructed by Ximerakis et al., who extracted 6638 ligand-receptor interactions from iRefIndex, Pathway Commons, and BioGRID. After similar preprocessing as datasets 1 and 2, this invention yielded 6585 ligand-receptor interactions between 1129 ligands and 1335 receptors in mice.

[0176] Furthermore, there were 351 overlapping ligands, 335 overlapping receptors, and 352 overlapping ligand-receptor interactions on the two human datasets, while there were 423 overlapping ligands, 402 overlapping receptors, and 914 overlapping ligand-receptor interactions on the two mouse datasets.

[0177]

[0178] S2: Ligand-receptor interaction prediction uses a heterogeneous deep ensemble model (LRI-HDEnHD) that combines a heterogeneous Newtonian Boosting model and a deep neural network to screen and identify potential ligand-receptor interactions. LRI-HDEnHD includes the following four steps:

[0179] S2.1: Feature extraction, including the following steps:

[0180] S2.1.1: Obtain ligand-receptor sequence information from the UniProt database.

[0181] S2.1.2: Use kmer, combined autocovariance and crosscovariance (ACC), distance-based Top-n plot (DT), distance-pair-based pseudo-amino acid composition (PseAAC-DP), parallel correlated pseudo-amino acid composition (PC-PseAAC), and serial correlated pseudo-amino acid composition (SC-PseAAC) methods to extract biological characteristics of proteins.

[0182] S2.1.3: Obtain the number of features for each type, as shown in Table 2. Integrate these features to describe each ligand-receptor pair. A ligand-receptor pair can be represented as a 4576-dimensional vector using cascade operations.

[0183]

[0184] S2.2: Dimensionality Reduction. After feature extraction, a ligand-receptor pair is represented as a 4576-dimensional vector. To reduce computational costs, this invention achieves dimensionality reduction based on principal component analysis while retaining 99% of the information. Finally, a d-dimensional feature vector is obtained to describe each ligand-receptor pair.

[0185] S2.3: Classification of ligand-receptor interactions, using the heterogeneous Newtonian Boosting model. and deep neural networks Calculate the interaction probability for each ligand-receptor pair. and

[0186] S2.4: Ensemble learning, fusing heterogeneous Newton Boosting models and deep neural networks to obtain heterogeneous deep ensemble models. And calculate the probability of the final ligand-receptor interaction. To obtain the final ligand-receptor interaction, where α represents the relative importance of the heterogeneous Newton Boosting model and the deep neural network model for the classification performance of ligand-receptor interactions, which needs to be learned through cross-validation.

[0187] S3: Ligand-receptor interaction filtering, including the following steps:

[0188] S3.1: Download single-cell transcriptome data of the target tissue from the GEO database;

[0189] S3.2: Update ligand-receptor interaction data by merging predicted ligand-receptor interactions with known ligand-receptor interactions;

[0190] S3.3: When the ligand or receptor from the ligand-receptor interaction is not expressed in the cells of the corresponding tissue's single-cell transcriptome data, the ligand-receptor interaction is removed;

[0191] S4: Cell communication prediction based on filtered ligand-receptor interactions and expression thresholding, expression product and combined scoring strategies;

[0192] S5: Visualization of cell communication.

[0193] Performance comparison of the LRI-HDEnHD model with four state-of-the-art protein interaction prediction methods (PIPR, XGBoost, DNNXGB, and OR-RCNN):

[0194] To evaluate the performance of the proposed LRI-HDEnHD model, it was compared with four state-of-the-art protein interaction prediction methods (PIPR, XGBoost, DNNXGB, and OR-RCNN). This invention performed 20 five-fold cross-validations, with 80% of the ligand-receptor pairs randomly selected as training data for each experiment, and the remainder used as test data. On the four ligand-receptor interaction datasets, the dimension d of each ligand-receptor interaction feature vector after dimensionality reduction was set to 327, 295, 280, and 346, respectively.

[0195] Parameter settings for each model

[0196]

[0197]

[0198] This invention uses AUC and AUPR to evaluate the performance of ligand-receptor interaction prediction methods. AUC and AUPR have been widely used to evaluate various classification models. AUC represents the area under the ROC curve. It can be calculated by the relationship between the true positive rate and the false positive rate. AUPR represents the area under the precision-recall curve. It can be calculated by plotting the ratio of observed ligand-receptor interactions in all predicted ligand-receptor interaction data for each given recall rate. Higher AUC and AUPR indicate better performance. A grid search was performed to set the optimal value of α, and it was found that the proposed LRI-HDEnHD model achieved the highest AUC and AUPR on four ligand-receptor interaction datasets when α = 0.6.

[0199] Performance comparison of five ligand-receptor interaction prediction models

[0200] according to Figure 4-7 It can be observed that LRI-HDEnHD achieves the highest AUC on datasets 1 and 4. While its calculated AUC is slightly lower than DNNXGB on datasets 2 and 3, the difference is very small. For example, on datasets 2 and 3, the AUC values ​​obtained from LRI-HDEnHD are only 0.12% and 0.38% lower than DNNXGB, respectively. More importantly, according to... Figure 8-11 It can be observed that LRI-HDEnHD achieved the highest AUPR on the four datasets, with values ​​of 0.8537, 0.8112, 0.8052, and 0.8399, respectively, outperforming the second-best method by 5.29%, 3.55%, 3.60%, and 2.66%. LRI-HDEnHD achieved the best performance in most cases, demonstrating its strong ability to classify ligand-receptor interactions. Therefore, this invention utilizes predicted ligand-receptor interactions to improve cell communication analysis.

[0201] Performance comparison between single models and ensemble models

[0202] This invention compares the performance of LRI-HDEnHD with HNBM and DNN in classifying ligand-receptor interaction data. According to the results in Table 4, it can be found that LRI-HDEnHD achieves the best AUC and AUPR on the four datasets, which is significantly better than HNBM and DNN. The results show that LRI-HDEnHD is a good heterogeneous deep ensemble model that can improve the recognition performance of ligand-receptor interactions.

[0203]

[0204] Sensitivity selection of parameter α

[0205] To investigate the impact of parameter α on ligand-receptor interaction classification performance, this invention considers the prediction accuracy when α is set in the range [0.5, 0.8]. As shown in Table 5, in most datasets, the LRI-HDEnHD model achieves better AUC and AUPR on the four ligand-receptor interaction datasets when α is set to 0.6. Therefore, this invention selects α as 0.6 to balance the influence of HNBM and DNN on ligand-receptor interaction classification performance.

[0206]

[0207]

[0208] Performance Comparison Conclusion: LRI-HDEnHD achieves strong ligand-receptor interaction classification accuracy due to the following features: (i) it extracts multiple biological features of ligands and receptors. (ii) HNBM has strong generalization ability and better global convergence without affecting classification performance. (iii) DNN can handle datasets of different sizes and improve classification accuracy. Furthermore, LSTM networks can dynamically add gates to adjust the weights of the autorecursive model. (iv) Heterogeneous deep ensemble models consider the diversity of complex heterogeneous fusion structures and demonstrate good flexibility and strong generalization performance.

[0209] Validation of ligand-receptor interaction data

[0210] The LRI-HDEnHD model can screen for novel ligand-receptor interaction data to improve cell communication analysis. This invention validates predicted ligand-receptor interactions in three ways: three representative cell communication analysis tools, existing PPI and PubMed databases, and molecular docking.

[0211] Validation of predicted ligand-receptor interaction data using three representative cell communication analysis tools

[0212] First, we analyzed the overlapping ligand-receptor interaction data predicted by LRI-HDEnHD, CellTalkDB, CytoTalk, and CellPhoneDB. CellTalkDB is a comprehensive, manually curated database containing 3398 human ligand-receptor interactions and 2033 mouse ligand-receptor interactions. CytoTalk uses single-cell transcriptome data to construct cell-specific networks, effectively identifying potential information pathways. CellPhoneDB is a novel ligand-receptor interaction database that considers the substructures of ligands and receptors, enabling prediction of cell communication from single-cell transcriptome data. These three methods were used to screen for predicted ligand-receptor interaction data. Table 6 shows the number of overlapping ligand-receptor interactions predicted by LRI-HDEnHD, CellTalkDB, CytoTalk, and CellPhoneDB across the four datasets.

[0213]

[0214]

[0215] In dataset 1, LRI-HDEnHD and CellTalkDB predicted 202 overlapping ligand-receptor interactions, while LRI-HDEnHD and CytoTalk predicted 13 overlapping ligand-receptor interactions. After removing duplicate ligand-receptor interactions (DEFB4A-CCR6), LRI-HDEnHD and the three cell communication analysis tools predicted 214 overlapping ligand-receptor interactions. Similarly, the other three datasets showed 384, 322, and 9 overlapping ligand-receptor interactions, respectively.

[0216] Validation of ligand-receptor interactions using five existing PPI databases and the PubMed database.

[0217] This invention validates predicted ligand-receptor interactions by manually searching five PPI databases (BioGrid, STRING, Pickle, Intact, and Pathway Commons) and the PubMed database. The results show that the 313, 57, 30, and 827 ligand-receptor interactions identified by LRI-HDEnHD on the four ligand-receptor interaction datasets can be validated using one or more of the aforementioned six databases.

[0218] Verification of ligand-receptor interactions via molecular docking

[0219] Thirty predicted ligand-receptor interaction data were randomly selected from each dataset for molecular docking experiments to validate the results. The structures of the ligands and receptors were downloaded using the online tool ZDOCK, and molecular docking was then performed using the molecular docking tool PDBePISA with default parameters. Table 7 shows the ligand, receptor, binding energy (kcal / mol) (BE), hydrogen bond (HB), and interfacial area (IA) of the ligand-receptor interaction with the highest binding energy among the 30 randomly selected ligand-receptor interactions from each dataset via molecular docking.

[0220]

[0221]

[0222] Example 2

[0223] One embodiment of the present invention provides: an application of a method for analyzing cell communication mediated by ligand-receptor interactions to the analysis of cell communication in human melanoma tissue.

[0224] Studies have shown that all cells possess the ability to communicate with each other. One approach to assessing the strength of communication between two cell types is to separately count the number of ligand-receptor interactions linked to each cell type and consider the expression levels of ligands and receptors from these interactions. This invention uses data from seven cell types in melanoma-derived single-cell suspensions provided by Zhou et al. to analyze cell communication in melanoma: melanoma cancer cells, cancer-associated fibroblasts (CAF), macrophages, endothelial cells, T cells, B cells, and natural killer (NK) cells. The specific steps include:

[0225] Based on steps S1 and S2 in Example 1.

[0226] S3: Ligand-receptor interaction filtering, which filters ligand-receptor interactions by combining single-cell transcriptome data, known ligand-receptor interactions, and identified ligand-receptor interactions, including the following steps:

[0227] S3.1: Download single-cell transcriptome data related to melanoma from the GEO database.

[0228] S3.2: Ligand-receptor interactions were filtered by integrating known ligand-receptor interactions, predicted ligand-receptor interactions, and single-cell transcriptome data from melanoma tissue. In Dataset 1, 67,595 ligand-receptor interactions were identified in addition to 3,390 known interactions.

[0229] S3.3: When a ligand-receptor interaction is not expressed in a cell of a single-cell transcriptome data set from melanoma tissue, that interaction is removed. Therefore, 24,881 ligand-receptor interactions associated with melanoma tissue were obtained in dataset 1.

[0230] S4: Cell communication prediction. Cell communication is predicted based on filtered ligand-receptor interactions and a combined scoring strategy of expression threshold and expression product. The communication scores between the seven cell types in melanoma tissue are calculated using the expression threshold method, expression product method and combined scoring strategy.

[0231] S4.1 Cell Communication Analysis Based on Filtered Ligand-Receptor Interactions and Expression Thresholds

[0232] Based on known ligand-receptor interactions and screened ligand-receptor interactions, expression thresholding was used to calculate cell communication between melanoma cancer cells and six other cell types (i.e., CAF, macrophages, endothelial cells, T cells, B cells, and NK cells). The results are shown in Table 8, which displays the communication specificity ranking between melanoma cancer cells and the other six cell types, the number of ligand-receptor interactions mediating communication between melanoma cancer cells and the other six cell types, and the number of ligand-receptor interactions “talked back” between other cell types and melanoma tumor cells.

[0233] Conclusion: CAFs showed the strongest communication with human melanoma cancer cells, followed by endothelial cells and macrophages. Notably, among all six cell types, T cells exhibited relatively weaker communication with melanoma cancer cells.

[0234] S5.1 Cell communication visualization: Heatmaps and network views visualize cell communication.

[0235] Heatmap, such as Figure 12 As shown, rows and columns represent cells expressing the ligand and receptor, respectively. Communication scores between the two cell types are represented by different colors. Darker colors indicate stronger communication.

[0236] Network view, such as Figure 13 As shown, each edge represents communication between two cell types, starting at a cell type expressing one ligand and ending at a cell type expressing another receptor. The thickness of the edge is proportional to the cell communication score.

[0237] Conclusion: Ren et al. found that the communication strength between human melanoma cancer cells and T cells is very low. This invention, based on known and screened ligand-receptor interactions and expression threshold methods, obtained the same results as Ren et al., and constructed a cell-cell communication network in melanoma tissue using heatmaps and network diagrams.

[0238]

[0239] S4.2 Cell communication analysis based on filtered ligand-receptor interaction and expression product method

[0240] The expression product method was used to measure the strength of cell communication in human melanoma tissue based on filtered ligand-receptor interactions. Table 9 shows the top three pairs of ligand-receptor interactions mediating communication from cancer cells to the other six cell types in melanoma tissue, as well as the top three pairs of ligand-receptor interactions mediating communication from the other six cell types to cancer cells.

[0241] S5.2 Cell Communication Visualization Figure 14 The illustration shows the top three ligand-receptor interactions in each cell communication in human melanoma tissue. The darker-colored ligand-receptor interactions are more likely to mediate the corresponding cell communication. Figure 15 heatmap and Figure 16 The network view visualizes cell communication.

[0242] Conclusion: CAFs exhibited the strongest communication with human melanoma cancer cells, followed by macrophages and endothelial cells. Furthermore, melanoma cancer cells communicated with CAFs via B2MRPSA, HLA-A-RPSA, and HLA-C-RPSA, while CAFs communicated with melanoma cells via B2M-RPSA, HLA-CRPSA, and HLA-A-RPSA.

[0243]

[0244] S4.3 Cell communication analysis based on filtered ligand-receptor interactions and co-scoring strategy

[0245] A joint scoring study was conducted on cell communication in human melanoma tissue. Table 9 shows that the joint scoring strategy achieved the same cell communication ranking as the expression product method. Tables 8-9 indicate that CAFs, macrophages, and endothelial cells may communicate with melanoma cancer cells.

[0246] S5.3 Cell Communication Visualization Figure 17 The top 3 pairs of ligand-receptor interactions mediating cell communication in human melanoma tissues Figure 18 heatmap and Figure 19Network views were used to evaluate cell communication between human melanoma tissue and eight other cell types. Multiple studies have confirmed that cell-associated flow factors (CAFs) are closely related to tumor growth, metastasis, and treatment resistance. Melanoma cells communicate with tumor-associated macrophages by releasing soluble factors that enhance or inhibit tumor growth. Metastatic B16 melanoma cells can produce macrophage cytotoxic elements, enabling tumor cells to evade host immune surveillance and prevent metastasis. Notch signaling can mediate communication between melanoma cells and endothelial cells, and surface-diffuse melanoma cells can transform into endothelial cells at metastatic sites.

[0247] Comparison of LRI-HDEnHD with CellPhoneDB, LIANA, and CellChat in cell communication analysis in melanoma tissue

[0248] To further analyze the performance of the proposed cell communication prediction framework (LRI-HDEnHD+joint), this invention compared it with three cell communication analysis tools: CellPhoneDB, LIANA, and CellChat. CellPhoneDB is a novel ligand-receptor interaction database used to predict communication between different cell types from single-cell transcriptome data. It analyzed how to assess communication between 10,000 cells and 19 cell types. LIANA is an open-source ligand-receptor analysis framework. It systematically compared 16 cell communication analysis resources and 7 methods, evaluating the consistency of cell communication prediction methods through spatial colocalization, cytokine activity, and receptor protein abundance. CellChat quantitatively identifies and analyzes cell communication networks from single-cell RNA sequencing data based on network analysis and pattern recognition methods. CellPhoneDB, LIANA, and CellChat demonstrated good cell communication prediction performance and are state-of-the-art cell communication analysis tools. Table 10 shows the cell communication analysis results in human melanoma tissue obtained from the above four cell communication analysis tools.

[0249]

[0250]

[0251] Table 10 shows that the combined LRI-HDEnHD analysis predicts that CAFs are more likely to communicate with melanoma cancer cells, which is consistent with the results obtained by LIANA and CellChat. Furthermore, it predicts that macrophages may have signaling pathways with melanoma cancer cells, while CellPhoneDB predicts that macrophages are most likely to communicate with melanoma cancer cells. Figure 28A heatmap of cell communications predicted by LRI-HDEnHD-joint, CellPhoneDB, LIANA, and CellChat is shown in human melanoma tissue. Figure 28 The results from LRI-HDEnHD-joint, CellPhoneDB, LIANA, and CellChat showed that CAF was more likely to communicate with melanoma cancer cells across all seven cell types.

[0252] Comparison of LRI-HD, EnHD, and NicheNet

[0253] Further comparisons were made of overlapping ligand-receptor interaction data predicted by LRI-HDEnHD and NicheNet in melanoma tissue. NicheNet accurately predicts ligand-receptor interactions mediating cell communication based on prior knowledge of expression information and gene regulation and signaling networks. Unlike most current cell communication prediction methods, NicheNet studies cell communication from single-cell expression data, taking into account ligand gene regulation information and integrating intracellular signaling and transcriptional regulation. NicheNet is primarily used for ligand-receptor interaction identification and is a very useful tool for predicting ligand-receptor interactions. Therefore, this invention analyzed overlapping ligand-receptor interaction data predicted by LRI-HDEnHD and screened by NicheNet that were not observed in human melanoma tissue. In melanoma tissue, LRI-HDEnHD ultimately obtained 23,768 ligand-receptor interactions from ligand-receptor pairs not observed in dataset 1. NicheNet obtained 10,555 ligand-receptor interactions in melanoma tissue. Finally, in human melanoma tissue, there were 1981 overlapping ligand-receptor interactions between LRI-HDEnHD and NicheNet-predicted ligand-receptor interactions.

[0254] in conclusion:

[0255] In this study, an expression thresholding method was used to analyze the specificity of cell communication. Cell types in human melanoma tissues were determined by setting thresholds for the expression levels of the ligand and receptor from each ligand-receptor pair. A ligand-receptor pair was considered "active" when the expression levels of both genes were above the threshold; otherwise, it was considered "inactive." That is, ligand-receptor pairs were assigned 1 and 0, respectively. All ligand-receptor pairs were classified as "active" or "inactive" based on different thresholds.

[0256] However, the expression threshold method assumes that ligand-receptor interactions require higher gene expression and necessitates the selection of a gene expression threshold, which can lead to false positives and false negatives. More importantly, many proteins exhibit different biological activities depending on their concentration, and the association between mRNA and protein levels varies from gene to gene.

[0257] The expression product method calculates continuous values ​​based on the expression product of ligand and receptor, thus successfully identifying differences in a ligand-receptor pair between cells. However, the expression product method can be problematic when the ligand and receptor in a ligand-receptor interaction differ significantly at the transcriptional level, or when one ligand or receptor dominates the interaction signal.

[0258] Therefore, this invention designs a method for analyzing ligand-receptor interaction-mediated cell communication based on a joint scoring strategy using a heterogeneous deep integration model and single-cell transcriptome data, in order to identify the heterogeneity of gene expression across cells and further aid in differential expression analysis.

[0259] Example 3

[0260] One embodiment of the present invention provides: an application of a method for analyzing cell communication mediated by ligand-receptor interactions to the analysis of cell communication in human HNSCC tissue.

[0261] Human hepatic noncellular carcinoma (HNSCC) originates from mucosal epithelial cells and is one of the most aggressive and recurrent cancers. Puram et al. systematically studied the ligand-receptor expression patterns between HNSCC cancer cells and eight other cell types (i.e., fibroblasts, B cells, myocytes, macrophages, endothelial cells, T cells, dendritic cells, and mast cells) in HNSCC-derived single-cell suspensions to measure the strength of cell communication in human HNSCC tissues, including the following steps:

[0262] Based on steps S1 and S2 in Example 1.

[0263] S3: Ligand-receptor interaction filtering, which filters ligand-receptor interactions by combining single-cell transcriptome data, known ligand-receptor interactions, and identified ligand-receptor interactions, including the following steps:

[0264] S3.1: Download single-cell transcriptome data related to HNSCC from the GEO database (GSE103322).

[0265] S3.2: By integrating known ligand-receptor interactions, predicted ligand-receptor interactions, and single-cell transcriptome data in HNSCC tissues, 68,400 ligand-receptor interactions associated with HNSCC tissues were obtained.

[0266] S4: Cell communication prediction. Cell communication is predicted based on a combined scoring strategy of expression threshold and expression product after filtering ligand-receptor interaction. The communication scores between the above nine cell types in HNSCC tissue are calculated by expression threshold method, expression product method and combined scoring strategy.

[0267] S4.1 Cell Communication Analysis Based on Filtered Ligand-Receptor Interactions and Expression Thresholds

[0268] The expression thresholding method was used to calculate the cell communication strength between HNSCC cancer cells and eight other cell types. Table 11 shows the communication information between HNSCC cancer cells and the other eight cell types.

[0269] Conclusion: Dendritic cells showed the strongest communication with HNSCC cancer cells, followed by endothelial cells and mast cells.

[0270] S5.1 Cell Communication Visualization Figure 20 heatmap and Figure 21 Network views were used to assess the communication strength between HNSCC cancer cells and eight other cell types.

[0271]

[0272] S4.2 Cell communication analysis based on filtered ligand-receptor interaction and expression product method

[0273] This invention further uses the expression product method and filtered ligand-receptor interactions to identify cell communication specificity in HNSCC tissues. Table 12 shows the communication information between HNSCC cancer cells and eight other cell types.

[0274]

[0275] Conclusion: The results indicate that macrophages exhibit the highest specificity in communication with HNSCC cancer cells, followed by endothelial cells and fibroblasts. Furthermore, HNSCC cancer cells communicate with macrophages via B2M-CD74, MIF-CD74, and CALM2-CD74, while macrophages communicate with HNSCC cancer cells via B2M-CD9, B2MIFITM1, and B2M-RPSA.

[0276] S5.2 Cell Communication Visualization Figure 22 The diagram illustrates the top three ligand-receptor interactions that influence communication in each cell. Figure 23 heatmap and Figure 24 The network view characterizes the communication between HNSCC cancer cells and eight other cell types.

[0277] S4.3 Cell communication analysis based on filtered ligand-receptor interactions and co-scoring strategy

[0278] A joint scoring strategy was used to score cell communication between HNSCC cancer cells and eight other cell types. Table 13 shows the communication specificity ranking between HNSCC cancer cells and the other eight cell types, the top three ligand-receptor interactions mediating communication between HNSCC cancer cells and the other eight cell types, and the top three ligand-receptor interactions mediating communication between other cell types and HNSCC cancer cells.

[0279] S5.2 Cell Communication Visualization Figure 25 The diagram illustrates the top three ligand-receptor interactions in each cell communication pair. Figure 26 heatmap and Figure 27 Network views were used to characterize cell communication between HNSCC cancer cells and eight other cell types.

[0280] Conclusions: The results showed that, unlike the expression product method, the combined scoring strategy predicted the strongest communication ability between endothelial cells and human HNSCC cancer cells, followed by macrophages and fibroblasts. Furthermore, the proposed framework, based on an expression threshold method, predicted the potential communication between dendritic cells and HNSCC cancer cells. Hypoxia is one of the most prominent features of HNSCC. It induces tumor progression and is closely associated with tumor recurrence, chemotherapy resistance, and poor survival. Upregulating the secretion of pro-inflammatory cytokines by dendritic cells can improve the plasticity of immune cells.

[0281]

[0282] Comparison of LRI-HDEnHD-joint with CellPhoneDB, LIANA, and CellChat in HNSCC tissues

[0283] To further analyze the performance of the proposed LRI-HDEnHD-joint framework, this invention compared it with three cell communication analysis tools in HNSCC tissue: CellPhoneDB, LIANA, and CellChat. Table 14 shows the cell communication analysis results obtained by the four cell communication analysis tools in human HNSCC tissue.

[0284]

[0285]

[0286] The results in Table 14 indicate that LRI-HDenHD-join predicted the strongest communication specificity between endothelial cells and HNSCC cancer cells, followed by macrophages and fibroblasts. Furthermore, CellPhoneDB, LIANA, and CellChat validated strong communication between endothelial cells and HNSCC cancer cells. Figure 29 A heatmap of cell communications identified by LRI-HDEnHD-joint, CellPhoneDB, LIANA, and CellChat in human HNSCC tissue is displayed.

[0287] from Figure 29 In the study, it was found that among all eight cell types that could potentially communicate with HNSCC cancer cells, LRI-HDenHD-joint predicted the strongest communication between endothelial cells and HNSCC cancer cells. CellPhoneDB, LIANA, and CellChat also reached this conclusion. A comparison of LRI-HDenHD and NicheNet is also provided.

[0288] This invention further compares the ligand-receptor interaction prediction results of LRI-HDEnHD and NicheNet in human HNSCC tissues. In human HNSCC tissues, LRI-HDEnHD-joint obtained 65,145 ligand-receptor interactions from ligand-receptor pairs not observed in dataset 1, while NicheNet obtained 29,060 ligand-receptor interactions. Ultimately, there are 4,969 overlapping interaction data between the ligand-receptor interactions predicted by LRI-HDEnHD and NicheNet in HNSCC.

[0289] Computation time and storage analysis

[0290] Finally, this invention analyzes the computation time and memory usage of four PPI prediction models—XGBoost, DNNXGB, OR-RCNN, and PIPR—for LRI-HDEnHD. The experimental server configuration was as follows: AMD EPYC 7302 CPU, GeForce RTX2080 Ti, 256GB of memory, and Ubuntu 20.04.4 LTS operating system. Table 15 shows the computation time (m) and memory usage (MB) required for a single 5-fold cross-validation experiment on four ligand-receptor interaction datasets for the five ligand-receptor interaction prediction models.

[0291]

[0292] The results show that DNNXGB has the shortest runtime on the four ligand-receptor interaction datasets, while XGBoost requires the least storage space. Among all five ligand-receptor interaction prediction methods, although LRI-HDEnHD requires more runtime and space than XGBoost and DNNXGB, it requires significantly less runtime and space than OR-RCNN and PIPR. Specifically, LRI-HDEnHD's runtime on the four datasets is 134.94m, 91.66m, 86.17m, and 258.89m, respectively, which is 250.40%, 258.12%, 306.97%, and 210.55% less than OR-RCNN, and 52.62%, 43.44%, 52.77%, and 16.42% less than PIPR. It requires 2880.02MB, 2225.61MB, 2303.09MB, and 3957.66MB of space on four datasets, which is 690.66%, 981.24%, 946.12%, and 456.55% less than OR-RCNN, and 992.99%, 1938.04%, 1922.03%, and 647.40% less than PIPR, respectively. With the rapid development of computer hardware technology, the computation time required by LRI-HDEnHD is relatively less important compared to memory and the predictive performance it achieves. Therefore, LRI-HDEnHD is more suitable for screening novel ligand-receptor interactions than the other four state-of-the-art ligand-receptor interaction prediction models.

[0293] in conclusion:

[0294] In this study, an expression thresholding method was used to analyze the specificity of cell communication. Cell types in human melanoma tissues were determined by setting thresholds for the expression levels of the ligand and receptor from each ligand-receptor pair. A ligand-receptor pair was considered "active" when the expression levels of both genes were above the threshold; otherwise, it was considered "inactive." That is, ligand-receptor pairs were assigned 1 and 0, respectively. All ligand-receptor pairs were classified as "active" or "inactive" based on different thresholds.

[0295] However, the expression threshold method assumes that ligand-receptor interactions require higher gene expression and necessitates the selection of a gene expression threshold, which can lead to false positives and false negatives. More importantly, many proteins exhibit different biological activities depending on their concentration, and the association between mRNA and protein levels varies from gene to gene.

[0296] The expression product method calculates continuous values ​​based on the expression product of ligand and receptor, thus successfully identifying differences in a ligand-receptor pair between cells. However, the expression product method can be problematic when the ligand and receptor in a ligand-receptor interaction differ significantly at the transcriptional level, or when one ligand or receptor dominates the interaction signal.

[0297] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for analyzing cell communication mediated by ligand-receptor interactions, characterized in that, Includes the following steps: S1: Organizing ligand-receptor interaction data, collecting multiple ligand-receptor interaction datasets; S2: Ligand-receptor interaction prediction, using a heterogeneous deep ensemble model that combines a heterogeneous Newton Boosting model and a deep neural network to identify potential ligand-receptor interactions; S2.1: Feature extraction; S2.2: Dimensionality Reduction; S2.3: Classification of ligand-receptor interactions, using the heterogeneous Newtonian Boosting model. and deep neural networks Calculate the interaction probability of each ligand-receptor pair. and ; S2.3.1 The interaction probability of each ligand-receptor pair is calculated using a heterogeneous Newtonian Boosting machine; Among them, the heterogeneous Newton Boosting model ,in Indicates the first One sample, Representation model pair The calculation of the first The sub-class probability is updated to... , ; S2.3.2 The interaction probability of each ligand-receptor pair is calculated using a deep neural network; the deep neural network ,in , For parameter 1, express 3D feature vector; S2.4: Classification results integration: A heterogeneous deep ensemble model is obtained by fusing the heterogeneous Newton Boosting model and the deep neural network. And calculate the probability of the final ligand-receptor interaction. To achieve the final ligand-receptor interaction; where This indicates the relative importance of heterogeneous Newtonian Boosting models and deep neural network models for ligand-receptor interaction classification performance. The range is [0.5, 0.8]; S3: Ligand-receptor interaction filtering, which filters ligand-receptor interactions by combining single-cell transcriptome data, known ligand-receptor interactions, and identified ligand-receptor interactions; S4: Cell communication prediction, based on filtered ligand-receptor interactions and a joint scoring strategy to predict cell communication; S4.1: Calculate the cell communication scores using the expression threshold method and the expression product method, respectively. and Its expression is: in , They all said and Potential ligand-receptor interactions between them and This represents two cell types mediated by ligand-receptor interactions. Indicates ligand, Represents receptor, and Indicates ligand and receptor Mean expression value in all cells and Indicates ligand and receptor In cell type and The average expression value in and Indicates the corresponding deviation; S4.2: Normalize the two expressions above using the min-max scaling method, and the expression is as follows: in express ; = ,in express ; S4.3: Obtain the joint cell communication score Its expression is ; S5: Visualization of cell communication.

2. The method for analyzing ligand-receptor interaction-mediated cell communication according to claim 1, characterized in that, In step S2.4 It is 0.

6.

3. The method for analyzing ligand-receptor interaction-mediated cell communication according to claim 1, characterized in that, The feature extraction in step S2.1 includes the following steps: S2.1.1: Obtain the sequence information of the ligand and receptor from the UniProt database; S2.1.2: Using kmer, combinations of autocovariance and crosscovariance, distance-based Top-n plots, distance-pair-based pseudo-amino acid composition, parallel correlation pseudo-amino acid composition, and serial correlation pseudo-amino acid composition methods to extract biological characteristics of proteins. S2.1.3: These features are fused to describe each ligand or receptor, so that a ligand-receptor pair can be represented as a 4576-dimensional vector using a concatenation operation.

4. The method for analyzing ligand-receptor interaction-mediated cell communication according to claim 1, characterized in that, Step S2.2 dimensionality reduction includes the following steps: After feature extraction, a ligand-receptor pair is described as a sample. Principal component analysis is used to reduce the dimensionality of its features while retaining 99% of the information. Finally, a sample is obtained. Each ligand-receptor pair is described by a dimensional eigenvector.

5. The method for analyzing ligand-receptor interaction-mediated cell communication according to claim 1, characterized in that, Step S3, ligand-receptor interaction filtering, includes the following steps: S3.1: Download single-cell transcriptome data of the target tissue from the GEO database; S3.2: Integrate predicted ligand-receptor interaction data with known ligand-receptor interaction data to update ligand-receptor interaction data; S3.3: When a ligand or receptor of a ligand-receptor interaction is not expressed in the cells of the target tissue single-cell transcriptome data, the ligand-receptor interaction is considered not to mediate cell communication in that tissue and is therefore removed.

6. The method for analyzing ligand-receptor interaction-mediated cell communication according to claim 1, characterized in that, Step S5 is to visualize cell communication in human melanoma tissue and head and neck squamous cell carcinoma tissue, including the following steps: illustrating the top 3 ligand-receptor interactions of each cell communication intensity in human melanoma tissue and head and neck squamous cell carcinoma tissue, and visualizing cell communication with heatmaps and network views.

7. A method for analyzing ligand-receptor interaction-mediated cell communication according to claim 6, characterized in that, The analysis and visualization of cell communication in human melanoma tissue revealed strong communication between CAF cells, macrophages, endothelial cells, and melanoma cancer cells.

8. A method for analyzing ligand-receptor interaction-mediated cell communication according to claim 7, characterized in that, The analysis and visualization of cell communication in head and neck squamous cell carcinoma tissue revealed strong communication between endothelial cells, macrophages, fibroblasts, and head and neck squamous cell carcinoma.

Images