Cryo-EM Troubleshooting: Is the Sample or the Processing Your Bottleneck?
Weeks get lost reprocessing data that no algorithm can rescue, and samples get remade when one classification job would have solved it. The diagnostics that tell you which side of the microscope your problem lives on.
The Short Answer
One question separates the two failure modes faster than anything else: do any of your 2D class averages show secondary structure? If even a handful of classes show crisp alpha-helical rods or clear domain boundaries, your sample is delivering high-resolution information to the detector, and whatever is holding the map back lives downstream, in collection strategy or processing choices, where more compute and better decisions can reach it. If no class shows internal features after several hundred thousand particles from decent ice, the information is not in the images, and no software will put it there. That version of the problem is solved at the bench, not in cryoSPARC. Everything below is elaboration of that rule, plus the two measurements that make it quantitative: the orientation distribution and how resolution responds to particle number.
Read the Micrographs Before Anything Else
The cheapest diagnostics never require a reconstruction. Look at raw micrographs from a screening session and ask four things. Can you see individual particles at all, with the size and shape you expect from the construct? Is the ice thin enough that particles show contrast, but thick enough that they are not excluded from holes? Are particles evenly dispersed, or clumped into aggregates and clustered at hole edges? And is the concentration in a workable range, since both nearly empty holes and lawns of overlapping particles are collection problems you fix before spending microscope time? A surprising amount of wasted processing traces back to skipping this step: classification cannot separate what vitrification already destroyed, and motion correction cannot recover particles you cannot find.
The 2D Class Test, Properly Applied
Two-dimensional class averages are the single most information-dense diagnostic in the pipeline, because they show you the best your data can currently do with the least processing risk. Apply the test fairly: use enough particles per class, let classification converge, and look across all classes rather than cherry-picking. Secondary structure in some views but not others usually means orientation or flexibility, both workable. Sharp edges but uniformly smooth interiors at high particle counts is the signature of a sample problem, most often damage at the air-water interface or genuine conformational disorder. One honest caveat: a bad 2D result can also be produced by bad upstream processing, careless picking that buries real particles under carbon edge and ice contamination, or classification run with too few iterations. So before you indict the sample, clean the picks once, rerun, and only then read the verdict.
Orientation Bias: The Problem That Lives on Both Sides
Preferred orientation is the diagnosis people miss because the nominal resolution number can look fine while the map is unusable, stretched and streaky along one axis, with helices smeared into sheets. The measurement that catches it is the angular distribution plot plus the 3DFSC sphericity score, both standard outputs now. The defining published case is influenza hemagglutinin: Tan and colleagues (Nature Methods 14, 793, 2017) showed that HA trimer adsorbs to the grid almost exclusively in top views, that untilted data produced reconstructions with severe directional anisotropy despite respectable-looking global FSC values, and that collecting with the stage tilted to 40 degrees restored the missing views and yielded an interpretable 4.2 angstrom map. The instructive part is where the fix lived: not in the software and not in a new sample, but in collection strategy. Orientation problems sit exactly on the boundary this guide is about, and they are why the angular distribution belongs in your standard post-collection checklist.
When It Really Is the Processing
Good 2D classes plus a map stuck in the 3.5 to 4.5 angstrom range is the classic profile of processing headroom. The usual suspects, roughly in order of payoff: junk particles still in the stack inflating the noise floor; heterogeneity that needs 3D classification or a heterogeneity-aware method rather than one consensus refinement; symmetry imposed that the particle does not actually have, or symmetry available but not exploited; masks that cut through real density or include disordered regions; and per-particle motion and CTF refinement left unrun. Flexibility deserves special mention because it masquerades as a resolution problem: a two-domain protein refined as one rigid body will hold both domains hostage, and local refinement with focused masks routinely buys half an angstrom in the region that matters. None of these require new sample. They require someone who has seen the pattern before, which is why a stalled project is often one afternoon of experienced eyes away from moving again.
The Particle-Number Diagnostic
The quantitative version of the bottleneck question is how resolution responds to particle count. The Rosenthal-Henderson framework makes the expectation precise: plotting the inverse squared resolution against the logarithm of particle number gives a line whose slope is set by the B-factor of your dataset, and a well-behaved dataset improves predictably as you add particles. Run the experiment cheaply by refining with random subsets, a quarter, a half, and all of your stack, and watching the trend. If resolution is still climbing at your full particle count, collect more and keep processing. If it has plateaued well short of where the 2D classes say the information should reach, adding data will not help, and the limit is orientation coverage, unmodeled heterogeneity, or damage. That plateau is the strongest quantitative signal for redirecting effort from the cluster back to the bench, or from more collection to smarter classification.
The Honest Counterweight: What Processing Can Never Fix
The air-water interface is the reason the bench usually wins when 2D classes are featureless. Tomographic surveys of vitrified grids by Noble and colleagues (eLife 2018) showed that in typical preparations the large majority of particles sit at the air-water interface rather than suspended in the ice, and proteins that adsorb there can partially denature or adopt a single orientation in the milliseconds before vitrification. A particle that unfolded before the grid froze contributes nothing recoverable at any particle count with any algorithm. The fixes are physical: additives and surfactants, different grid supports including graphene oxide or gold grids, reconstitution into nanodiscs for membrane proteins, or faster spot-to-plunge times with dispensing instruments built for exactly this failure mode. Budget expectations accordingly: seasoned practitioners assume sample optimization consumes most of a difficult project's calendar, and treating grid conditions as a first-class experimental variable is what separates projects that converge from projects that circle.
The Cost Math, and When to Hand It Over
The bottleneck question is ultimately a resource-allocation question, so put numbers on it. A week of postdoc time costs more than a screening session at most academic facilities, and months of iterating alone costs more than either. A useful discipline: if you have collected on three grids of the same construct and hit the same wall each time, the wall is the construct or the grid chemistry, and the next dollar belongs at the bench. Conversely, if your 2D classes show secondary structure and the map does not deliver it, the next dollar belongs in processing, and it is the cheaper dollar, because reprocessing needs no new sample, no travel, and no instrument time. Be honest about a third possibility as well: some projects stall not because either side is broken but because nobody involved has processed this particular pathology before. Orientation bias, pseudosymmetry, and compositional heterogeneity all have standard playbooks that are obvious in retrospect and invisible the first time you meet them.
Frequently Asked Questions
How many particles do I need before concluding my 2D classes are genuinely bad?
There is no universal number, but a practical threshold: if several hundred thousand picked particles from reasonable ice, cleaned of obvious junk, produce no class with any internal feature, the burden of proof has shifted to the sample. Small particles under about 150 kDa show features later and more subtly, so apply the test more cautiously there.
My map is streaky and stretched along one axis. Sample or processing?
That is the signature of preferred orientation, and it sits on the boundary. Check the angular distribution plot and the 3DFSC sphericity first. Fixes run from collection strategy (stage tilt, as in the published hemagglutinin case) to grid chemistry (additives, supports that change how particles adsorb). Pure reprocessing rarely solves it, but it diagnoses it in an afternoon.
I am stuck at 3.8 angstroms and my 2D classes look sharp. What is the likely headroom?
That profile usually means processing headroom: residual junk in the stack, unaddressed heterogeneity, missing per-particle CTF and motion refinement, or a flexible region holding back a consensus refinement. Focused classification and local refinement are the usual first moves, and half an angstrom is a realistic prize in the region you care about.
Does more collection time ever substitute for fixing the sample?
Within limits. The Rosenthal-Henderson logic says resolution improves with the logarithm of particle number, so brute force pays off only while the trend is still climbing. If a subset experiment shows resolution has plateaued, more movies mostly buy you a larger hard drive bill. Run the subset test before requesting another session.
When should I bring in outside processing help instead of iterating myself?
When the same wall has survived three of your own serious attempts, or when the calendar cost of learning a new pathology exceeds the cost of someone who has processed it before. Stalled-at-3.8, orientation bias, and heterogeneous complexes are patterns an experienced processor recognizes quickly, and a reprocessing pass needs no new sample or microscope time.
Related Resources
Need Help Finding a Facility?
Browse our directory or get a fast-track quote from a private CRO.