Fig. 4

Machine learning and MR analysis of hub MitoDEGs nominate three causal MitoDEGs for COPD. a The mean decrease in accuracy (MeanDecreaseAccuracy) (left) and mean decrease in Gini (MeanDecreaseGini) (right) were used to rank the relative importance of the 18 selected feature MitoDEGs for COPD which were selected by RF analysis following the filter criteria as both MeanDecreaseAccuracy and MeanDecreaseGini > 5. b Assessing the performance of the RF model using the receiver operating characteristic curve (ROC), the area under the curve (AUC) values were 0.916 in the training set, 0.902 in the test set, 0.881 in the internal validation dataset (GSE57148), and 0.828 in the external validation dataset (GSE76925), reflecting the overall stability and reliability of the RF model in (a). c LASSO regression showed log(λ) = 5.726 when the error of the model is minimized (top), and 13 MitoDEGs from 18 feature MitoDEGs identified by RF analysis in (a) were selected (bottom). d Forest plot showing the MR estimates of 6 hub MitoDEGs (MRPL27, MRPL2, NDUFS2, CAT, SCP2, HSD17B4) and COPD risk using the inverse variance weighting (IVW) method together with the results of sensitivity analyses. Pval represents p-value for MR analysis, P.pleio represents p-value for horizontal pleiotropy analysis, P.heter.ivw represents p-value for heterozygosity with IVW method, P.heter.mr represents p-value for heterozygosity with MR-Egger method. e Scatter plots showing the causal effect of MRPL27, MRPL2, NDUFS2, CAT, and HSD17B4 on COPD. f Leave-one-out plot demonstrating the causal effect of MRPL27, MRPL2, NDUFS2, CAT, and HSD17B4 on COPD when leaving one SNP out. g Venn diagram showing the intersection of 18 feature genes identified by RF analysis (green), 13 signature genes selected by LASSO regression analysis (blue), and 5 causal genes nominated by MR analysis (red). Only NDUFS2, MRPL2, and CAT appeared simultaneously in the results of three analyses