AlphaFold vs. Cryo-EM: When You Still Need the Microscope
Structure prediction answers the first question about a new protein but not the ones that follow. Where prediction stops, where cryo-EM starts, and how to use them together, with a worked example from a published structure.
The Short Answer
Structure prediction changed the first question you ask about a new protein. It did not remove the microscope from the workflow. The distinction that matters in practice: a prediction is a hypothesis about one shape, and cryo-EM is a measurement of which shapes your protein actually occupies under the conditions you care about. Three situations still require experimental structure determination. First, when you need absolute confidence in the fold rather than a confident guess. Second, when you need the dynamic landscape, meaning the set of conformations a protein can adopt, not one representative pose. Third, when you need to observe your protein in a specific context, because folds change in the presence of substrates, inside cells, and in different buffers. If none of those three apply to your question, a prediction may genuinely be all you need, and paying for microscope time will not tell you anything new.
What Prediction Actually Gives You
AlphaFold2, AlphaFold3, and open implementations such as ColabFold are extremely good at the problem they were trained on: recovering the fold of a domain that resembles something already in the Protein Data Bank. For a single well-behaved globular domain with close structural relatives, predictions are frequently accurate to within a couple of ångströms at the backbone level, which is enough to design constructs, place domain boundaries, choose crystallization or grid targets, and reason about which surfaces are likely to be interfaces. The confidence metrics are useful if you read them for what they measure. pLDDT is a per-residue estimate of local geometric reliability, and PAE tells you how much to trust the relative placement of one region against another. Neither is a measurement of your sample. A high pLDDT means the model is highly consistent with the patterns the network learned; it does not mean your protein is in that conformation in your tube. This is also why the training data matters so directly. A protein whose domains have no close relative in the PDB, or a novel or designed fold, will often predict poorly, and the confidence scores are less reliable exactly where you most need them to be.
Reason 1: You Need Certainty, Not a Confident Guess
There is a category of project where being probably right is not acceptable. If a structure is going into a drug program, a mechanistic claim in a paper, or a patent, the difference between a measurement and an inference is the whole point. Prediction also does not tell you several things a map does. It does not tell you what is actually bound: cofactors, nucleotide state, lipids, metals, and small molecules are either absent from the prediction or placed with much lower reliability than the protein backbone. It does not reliably tell you stoichiometry or the oligomeric state your sample adopts. It does not tell you post-translational modifications or proteolytic processing, which are often the biologically interesting part. And it cannot tell you that the complex you assumed exists actually assembles. A map either shows density for a subunit or it does not, and that is information no prediction can supply.
Reason 2: You Need the Landscape, Not One Pose
Prediction methods return a dominant conformation. Even when you generate many models, what you get is a set of plausible variants rather than a population with measured occupancies. For a rigid enzyme that is fine. For a machine, it is the wrong kind of answer. Transporters cycle between inward-open and outward-open states. Receptors have active and inactive conformations that differ by a few ångströms in the places that matter for drug design. Motors and unfoldases progress through nucleotide-dependent steps. Cryo-EM is unusually good here, because heterogeneity that ruins a crystal is data in a single-particle experiment. From one dataset, 3D classification and heterogeneity analysis can separate coexisting states and estimate how populated each one is, which is a direct readout of the energy landscape. Our work on the Cdc48 AAA+ ATPase used exactly this property to resolve a series of nucleotide states along the substrate unfolding pathway, published in Nature Communications 15, 7505 (2024). No prediction of Cdc48 would have produced that sequence of states, because the question was not what shape the protein has but which shapes it moves through and in what order.
Reason 3: You Need the Protein in Its Actual Context
Folds are not fixed properties of sequences. They respond to their environment. A protein can order a loop only when substrate is engaged, change domain arrangement with buffer ionic strength or pH, adopt a different register when membrane-embedded rather than in detergent, and behave differently as part of a larger assembly than in isolation. A prediction has no access to any of that, because the input is a sequence and the output is the network's single best guess at a canonical form. If your question is specifically about the conformation under condition X, condition X has to be in the experiment. This is the most common reason a good prediction still does not close a project: the model is not wrong so much as it is unconditioned.
A Worked Example: Prediction as a Search Tool, Density as the Arbiter
The most productive use of prediction in a cryo-EM project is often not modeling a known protein but identifying an unknown one. In our Cdc48 study we had unassigned density in the consensus map above the central pore, adjacent to the D1 large subdomain: an extended β-strand plus roughly ten residues of α-helix. That is far too little to identify by sequence from the map alone. Instead we used prediction as a hypothesis generator. We ran ColabFold predictions over roughly 100-residue sections of five candidate interactors, Glc7, Sds22, Ypi1, Doa1 and Shp1, against the D1 large subdomain and asked which scored best. Shp1 did. We then made a refined prediction using Shp1 residues 125 to 139 with the Cdc48 D1 large subdomain, residues 219 to 379, with template mode set to pdb70 and 24 recycles. That prediction fit the density as a rigid body and became the starting model for refinement in Phenix. Two things are worth taking from this. Prediction alone would not have identified Shp1; the experimental density is what selected among the candidates and what the final model was refined against. And in the same structure, Ypi1 was present in the sample but could not be assigned in the modeled density. Prediction narrows the field, the map decides, and sometimes the map declines to answer. Note also that the finished model was a mosaic: predicted coordinates for the Shp1 segment, prior crystal structures rigid-body fit for PP1-SDS22 (PDB 6OBN) and the UBX domain (PDB 6OPC), and density-driven refinement over all of it.
Where Cryo-ET Goes Beyond Both
Cryo-electron tomography answers a question neither prediction nor single-particle cryo-EM can, and it charges for it in time. In tomography the protein is imaged inside a cell or a lamella milled from one, so you keep the physiological context that X-ray crystallography and single-particle cryo-EM both discard by design. You see where the complex sits relative to membranes and cytoskeleton, what it neighbours, its local concentration, and how it is oriented in a real environment. Purification removes all of that before you ever get to the microscope. The trade is honest: cryo-ET is by far the slowest route to a structure. Sample preparation often requires FIB milling, resolution is typically lower than single-particle work unless you can collect and average enough copies by subtomogram averaging, and the processing burden is substantially higher. Choose it when in-context information is the actual result you need, not as a general-purpose substitute for a single-particle experiment.
How to Combine Them in Practice
The efficient modern workflow treats prediction as an input to the experiment rather than a replacement for it. Before collecting data, use predictions to design constructs and pick domain boundaries, decide what is likely to be flexible and therefore worth masking or removing, and estimate whether particle size and shape are workable. After you have a map, use predictions to dock and identify subunits, assign sequence register in weak regions, and generate starting models for refinement. Then validate against the density rather than against the prediction confidence. Check real-space correlation per residue, confirm that bulky side chains have density where the register says they should, and look at whether loop paths follow the map or float above it. The failure mode to watch for is a predicted model that has been docked into a 3.5 to 4 Å map and looks convincing at a glance while being locally wrong, because the eye rewards a complete-looking model. If a region of the prediction has no supporting density, the correct action is to remove it, not to leave it in because it was confidently predicted. And if the answer you need is which state, in what context, at what occupancy, then the microscope is not optional and prediction is not the shortcut it appears to be.
Frequently Asked Questions
If AlphaFold gives me high pLDDT, can I skip cryo-EM?
It depends on the question, not the score. High pLDDT means the prediction is internally consistent with known structural patterns, and for construct design or a general sense of architecture that is often enough. It is not enough when you need the specific conformation present in your sample, what is bound, the oligomeric state, the population of coexisting states, or a measurement you can defend in a paper or drug program.
Can I use a predicted model as a starting model for refinement?
Yes, and it is now standard practice. Dock it as a rigid body, then refine against the map in Phenix, Refmac or ISOLDE. The discipline is to let the density govern: trim regions without supporting density, correct the register where side-chain density disagrees with the prediction, and report per-residue map correlation rather than the prediction confidence.
Does AlphaFold predict conformational changes?
Not reliably. Prediction methods converge on a dominant conformation, and sampling tricks such as subsampling the MSA or varying seeds can surface alternatives but do not give you populations or occupancies. If the biology is about a cycle of states, cryo-EM 3D classification and heterogeneity analysis measure that directly from a single dataset.
How do I tell whether a predicted model is wrong in my map?
Look for local disagreement rather than global fit. Check per-residue real-space correlation, verify that large aromatic and charged side chains sit in density, and confirm that loops follow the map instead of hovering in empty space. Systematic register errors show up as side chains consistently one or two positions off from their density.
What about ligands and drug design?
Ligand placement is the weakest part of prediction and the most consequential for medicinal chemistry, because a pocket that is roughly right at the backbone level can still be wrong in the details that determine binding. For structure-based design, get experimental density for the bound complex. Predicted apo models are useful for triage and for building expectations, not for committing chemistry.
When is cryo-ET the right choice over single-particle cryo-EM?
When the location and surroundings of your complex are part of the result, for example a membrane contact, a cytoskeletal association, or native spatial organization. It preserves physiological context that purification destroys, at the cost of being the slowest and most processing-intensive route, usually at lower resolution unless you can average many copies.
Related Resources
Need Help Finding a Facility?
Browse our directory or get a fast-track quote from a private CRO.