I had very similar classes with a small transmembrane protein. For me playing around with follwing parameters helped me to get some reasonable classes.
Number of final iterations: 5
online em iterations: 40
Batchsize: 400
Non negativity enforced
Also it helped, in my case, to use a white noise model in the first iterations to get junk and noise classes out. Later I was changing it again to the standard noise model to get higher Resolution 2D averages.
In the end however, I went back to sample/grid preperation as I couldn’t get much out of such datasets.
A few recommended threads in this regard:
Additionally, you could try running your 2D classifcation in Relion - smth I always do with poorly aligning particles. Sometimes I get totally different results there, especially when using the EM algorithm.