A little more information would be useful. How many particles? Is it a membrane protein? Is it symmetric? Does it have any other unusual characteristics (e.g.: long, needle shape)? What do micrographs look like? What do the 2D classes look like? What does the initial model look like? What does the 3D volume that corresponds to that FSC curve look like?
My initial impression, lacking any information beyond that FSC curve, is that whatever dataset you’re processing is still heterogeneous, and/or it is aligning poorly (if membrane protein) and converging too quickly. What might or might not help will depend on what the sample is, and whether or not 2D looks good enough that high resolution 3D might be achieved.
Thank you for your feedback. This is not a membrane protein but a soluble protein, which might have high flexibility. It should have C2 symmetry. However, whether I use C1 or C2 symmetry, the FSC curve still shows a sharp drop.Here are its 2D classification images. I have used approximately 990,000 particles and performed two rounds of heterogeneous refinement.
Your 2D looks like the sample is quite high concentration. You could try another round of 2D with a large number of classes, slow convergence (Initial classification uncertainty factor to 8 or 10) and increased number of iterations (say, 70-100 subsets, 3 full). Because those good looking classes do look good! Then try to get a good initial model from a very clean particle set, before heterogeneous refinement into [lots of classes] with the larger stack. 
Looks like you need to clean up your data. Still a lot of bad particles in this 2D set.
I’ll give it a try, thank you ！
Hi @xinyi, we see this all the time with flexible proteins, even when the set is clean enough. Once you’ve cleaned it as stated above, it will help if you can solve the flexibility.