One reason could be due to highly correlated features specially in selected feature in second case. You can check this, by using correlation matrix analysis and see if new selected features have high correlation with previous one or not. If this is the case, it means that you actually use less than 50 feature in second case and it not surprising if accuracy reduced.