Analyzing the structure of proteins in terms of their structural polymorphism has been recently performed using their two-dimensional electron-density maps obtained through cryo-electron microscopy (cryo-EM) experiments. In performing such analyses for a protein, the initial step is to classify its maps in terms of its structural polymorphism. Although current programs used for analyzing cryo-EM experimental data implement classification algorithms, they require the number of classes as input prior to conducting the classification. However, the number of classes is generally unknown, and choosing the wrong number of classes leads to difficulties in performing the structural analyses. Manifold learning is a candidate to resolve this issue because it has been successfully used for the classification of two-dimensional electron-density maps. However, a low signal-to-noise ratio of the maps would lead to the failure of the classification, especially for small proteins. Here, we investigated the effects of a low-pass filter, which can reduce noise, on the classification of two-dimensional electron-density maps using manifold learning. We performed a simulation for a cryo-EM experiment of a small protein that predominantly adopts two states. We found that while the classification failed for the raw two-dimensional electron-density maps, it was successful for the maps where a low-pass filter was applied. We also investigated the dependence of the filter’s parameters on the classification and found a relation between the values of the parameters and the degree of success of the classification.