| The current rapid development of single-cell multi-omics sequencing technology has generated a large amount of diverse omics data.By understanding gene expression information in multi-omics data,cellular signatures can be more fully explored and identified.It plays an important guiding role in the interaction between genes,the development of individual organs,the research and development of gene drugs,and clinical medical trials.At present,traditional research methods are affected by the noise and redundant information of various omics data,which makes it difficult for subsequent biomedical analysis.With the continuous development of deep learning,the mining of single-cell multi-omics data based on deep learning has become a research hotspot in bioinformatics.In this paper,based on deep learning,the data mining of single-cell multiomics data is carried out.The research contents are as follows:1.A method for denoising single-cell multi-omics data based on multi-head autoencoder network is proposed,called sc MAED.First,the model adds a classification decoder to the multi-head autoencoder network to remove data noise to the greatest extent in an unsupervised manner.Two encoders are used to independently learn the internal features of multi-omics data,and jointly decode the output low-dimensional features.Second,the classification decoder does not make any data distribution assumptions,and uses the predicted cell cluster labels to feed back data information to remove complex noise to the greatest extent.Finally,principal component analysis and t-SNE were used for visualization.In this experiment,four mature evaluation indicators were used to evaluate the performance of the model.The results show that sc MAED is superior to the comparison method in the experiment in terms of denoising effect,and can greatly improve the quality of single-cell multi-omics data.We have conducted multiple hyperparameter experiments,and the results show that sc MAED can effectively denoise data,make downstream analysis more accurate,and even provide insights for joint denoising in other omics.2.A dimensionality reduction method for single-cell multi-omics data based on a graph attention network is proposed,called sc MGAT.First,the sc MGAT model uses a graph neural network to jointly analyze and measure raw data,and adds a self-attention mechanism for cell weight optimization,allowing the model to focus on the expression information between different omics in the same cell.Secondly,the model simultaneously processes two sets of omics data with different feature information,uses an encoder to model each omics,and jointly processes latent feature information through a multi-head attention mechanism,while considering the data normalization problem,so that multiple Important information in omics is fused in low-dimensional hidden layers.Finally,the generated low-dimensional feature information is used to construct an intercellular structure map,and coordinated learning from multi-omics data in an unsupervised manner,and the optimal multi-omics low-dimensional data is obtained through multiple iterations.The performance of sc MGAT was evaluated based on eight publicly available sci CARseq and SNARE-seq datasets.In order to further explore the fault tolerance rate of the model,we conduct experiments under different batches,different embedding layer dimensions and different cell numbers.Experimental results demonstrate that sc MGAT can extract relevant features from different single-cell omics layers and construct weighted associations between omics data to obtain biologically meaningful low-level features.We experimentally demonstrate that sc MGAT outperforms current mature methods in terms of comprehensive performance. |
