An Adaptive Unsupervised Feature Selection Algorithm Based on MDS for Tumor Gene Data Classification

Six-gene prognostic signature for non-alcoholic fatty liver disease susceptibility using machine learning

BACKGROUND

nonalcoholic fatty liver disease (NAFLD) is a common liver disease affecting the global population and its impact on human health will continue to increase. Genetic susceptibility is an important factor influencing its onset and progression, and there is a lack of reliable methods to predict the susceptibility of normal populations to NAFLD using appropriate genes.

METHODS

RNA sequencing data relating to nonalcoholic fatty liver disease was analyzed using the "limma" package within the R software. Differentially expressed genes were obtained through preliminary intersection screening. Core genes were analyzed and obtained by establishing and comparing 4 machine learning models, then a prediction model for NAFLD was constructed. The effectiveness of the model was then evaluated, and its applicability and reliability verified. Finally, we conducted further gene correlation analysis, analysis of biological function and analysis of immune infiltration.

RESULTS

By comparing 4 machine learning algorithms, we identified SVM as the optimal model, with the first 6 genes (CD247, S100A9, CSF3R, DIP2C, OXCT 2 and PRAMEF16) as predictive genes. The nomogram was found to have good reliability and effectiveness. Six genes' receiver operating characteristic curves (ROC) suggest an essential role in NAFLD pathogenesis, and they exhibit a high predictive value. Further analysis of immunology demonstrated that these 6 genes were closely connected to various immune cells and pathways.

CONCLUSION

This study has successfully constructed an advanced and reliable prediction model based on 6 diagnostic gene markers to predict the susceptibility of normal populations to NAFLD, while also providing insights for potential targeted therapies.

An adaptive feature selection algorithm based on MDS with uncorrelated constraints for tumor gene data classification

The developing of DNA microarray technology has made it possible to study the cancer in view of the genes. Since the correlation between the genes is unconsidered, current unsupervised feature selection models may select lots of the redundant genes during the feature selecting due to the over focusing on genes with similar attribute. which may deteriorate the clustering performance of the model. To tackle this problem, we propose an adaptive feature selection model here in which reconstructed coefficient matrix with additional constraint is introduced to transform original data of high dimensional space into a low-dimensional space meanwhile to prevent over focusing on genes with similar attribute. Moreover, Alternative Optimization (AO) is also proposed to handle the nonconvex optimization induced by solving the proposed model. The experimental results on four different cancer datasets show that the proposed model is superior to existing models in the aspects such as clustering accuracy and sparsity of selected genes.