It is hard to say about the 2D classes that are damaged, e.g. Figure 1 and others in https://elifesciences.org/articles/42747
Bad 2D classes could be from bad picking, incorrect centring in extraction/2D, sub-optimal box size, too much junk, etc - a lot of reasons. In this case the tell tale sign is seeing “high resolution” secondary structure in 2D classification, but also some damaged part in the same 2D class. I have had a lot of cases where the protein looks great in 2D, but the dynamic area just does not align and sort of a white cloud. Then doing all of the usually focus/local refine with masks, more 3D classification, 3D-Flex/cryoDRGN/DynaMight/etc all can help.
Regardless, it is really hard to say from just 2D classification and even with 3D classification you still can find a junk class. To say for sure is this comes from particles at the AWI, you would need to tilt the stage. see https://discuss.cryosparc.com/t/how-to-tell-if-an-ab-initio-model-is-reliable-correct-or-not/5525/4?u=mark-a-nakasone
From what you say, it sounds like a complex is falling apart. Also if it is much smaller than expected this could bias some picking approaches. Depends if you can reconstruct part of it, but your initial 2D classes did not seem promising. Could not hurt to try different picking methods.