I'm just another grad student doing something between comp bio and structure so take this all with a grain of salt, but I'd read about cryo-ET, FIB milling, structural proteomics, and spatial proteomics (Matthias Mann's group's work comes to mind as an example of some of the wildest stuff people can do). Plenty of groups are working on what you describe but the details are tough. To give a non-technical description of the problem, most structure techniques rely on either explicit or implicit averaging to get good resolution. Small crystals or particle numbers, conformational heterogeneity, and beam-induced radiation damage are big resolution killers. People have figured out all kinds of tricks that amount to purifying a ton of protein, making sure it crystallizes well/is relatively homogenous with a good distribution of ways it sticks to a grid, and minimizing sample beam exposure during diffraction/SPA, which require relatively little, anyway. With cryo-ET, though, you aren't taking a ton of micrographs of a ton of particles or quickly zapping (well, you're rotating it and exposing it for some time under a little jet of super cold air, but hey, it works) a crystal that diffracted precisely because you got a large, relatively structurally homogenous lattice, you're taking a series of micrographs of the same cell, which is structurally probably way more different from other cells than one protein conformation is from another (and that’s if you culture/sort your cells ahead of time instead of working with tissue). Meanwhile, you're frying your sample from the prolonged exposure time a tilt series requires and not even getting a full tilt series because you can't rotate the darn sample holder 360 degrees (though people train CNNs to get around the missing wedge problem now). Once you've reconstructed a structure from your tomographs, how are you going to assign the blobs and squiggles you think you see in your gray static? You either need to know what you're looking ahead of time which requires pre-existing structures and good guesses on location (presumably reasonable to do with something like actin or microtubules, pretty tough with John Doe globular, cytosolic protein that looks like everything else at bad resolution), some kind of labeling (so now you probably need a cryo-CLEM or whatever too or to add something pretty big/distinct which comes with other problems), or you add mass-spec somehow. Oh, and I took the FIB milling for granted when describing this, btw (if you want to look at an entire cell, trying to FIB mill the entire thing is like trying to replace all your walls with glass, so we're back to sampling the right parts/averaging). If you add mass-spec, you’re basically trying to make the smallest laser micro dissection slices ever on a sample that you might have to almost destroy if you want to see more than figments of your imagination. So, hundreds of stupid little problems but people are working on it, and there's work for people with strong computer vision/ML backgrounds.

RNA is a totally different problem since it's often described as pretty "floppy" compared to protein and we don't have that many RNA structures, so presumably it's harder to assign things unless you're looking at well-studied complexes. Aren’t transcriptomics people working on this, though? Maybe MERFISH does kind of what you want just without pipelines for depth assignment yet (or maybe these exist).