Optimising Multi-Layer Neural Networks for Single-Cell Data Modelling Using Gaussian Processes

In this study, single-cell gene expression data were modelled using a Gaussian process optimised multilayer neural network to assess the effects of different compounds on gene expression in multiple cell types. The dataset used for the study was derived from the LINCS connectivity graph dataset of the Kaggle platform and included differential gene expression values for T cells, B cells, NK cells and myeloid cells. Through initial pre-processing and analysis, we evaluated the gene expression counts in individual cells using the Limma method to derive the extent of the effect of compounds on gene expression. During the model construction process, we tried various activation functions, and finally selected Leaky Relu as the activation function through Gaussian process optimisation, combined with 5-fold cross-validation and mean square error as the objective function for optimisation. In the process of constantly updating historical data and exploring new parameters, Gaussian process optimisation effectively improves the performance of the model. Eventually, we obtained the optimal parameter configuration, and the model performed well on both public and private test sets, with MRRMSE scores of 0.645 and 0.839, respectively. The results of the study showed that the larger absolute value of the gene expression difference value indicated its higher sensitivity to compounds, which provided an important basis for drug discovery and development. In addition, methods such as Gini index and K-Means clustering play an important role in gene screening and analysis. This study not only demonstrates the potential application of Gaussian process optimisation in multilayer neural network modelling, but also provides new ideas and methods for the application of single-cell technology in drug development, which has a wide range of applications. In the future, we will continue to optimise the model structure and explore more hyperparameter combinations, with a view to further improving the prediction accuracy and practical application value of the model.