The first time a laboratory predicts a protein structure on a laptop, the response is usually wary optimism. A graduate student pastes an amino acid sequence into a web server, waits a few minutes to an hour, and receives a three-dimensional model with color-coded confidence scores. The model looks like a real structure: alpha helices, beta sheets, and loops that tangle into a plausible active site. It also carries a number that separates the parts to trust from the parts to ignore. That number, the predicted local distance difference test or pLDDT, has become one of the most read values in structural biology. Five years ago, most bench scientists had not heard of it.

Computational methods for predicting molecular structure are not new. Homology modeling, threading, fragment assembly, and molecular dynamics have been in use for decades. What changed is accuracy and access. Deep learning models that treat a fold as a pattern-recognition problem arrived with force at the CASP14 assessment in 2020, when AlphaFold2 produced predictions close to experimental accuracy for a large share of targets. In the months that followed, laboratories gained a credible structural hypothesis before purifying the protein, before crystallizing it, and sometimes before the gene synthesis order has even been signed off. That shift reverses the usual order of operations, and its full effect is still being sorted out.

From sequence to structure: how prediction models work

Proteins are built from chains of amino acids, and most fold into a preferred three-dimensional shape that determines how they bind other molecules, catalyze reactions, move ions, and respond to signaling partners. A model can't yet solve the physics of folding for every protein, so the best tools take a statistical shortcut. They scan sequences from thousands of related organisms and look for positions that tend to change together. This co-evolution signal indicates residues that sit close in three dimensions even when they are far apart in sequence. The multiple sequence alignment becomes the input, and the network predicts a set of coordinates.

This is why AlphaFold2 and its relatives improved so quickly: the growth of protein sequence databases gave them more evolutionary data. For proteins with deep sequence alignments, the prediction can be sharp. For orphan proteins or synthetic sequences with few relatives, the model has less evidence and the confidence scores drop. The output is therefore not a single answer. It is a hypothesis with a confidence label attached to every region.

The tools now showing up in lab protocols

AlphaFold2 remains the default for many protein-only tasks because its underlying method is described clearly and its predictions are easy to obtain. The AlphaFold2 method paper in Nature laid out the architecture, and the AlphaFold Protein Structure Database now hosts more than 200 million predicted structures from UniProt and other sources. RoseTTAFold, developed at the University of Washington, uses a three-track network that predicts coordinates with lower computational cost and is available through the Robetta server. ESMFold from Meta AI predicts structures quickly from single sequences rather than alignments, which suits metagenomic screening; its approach is described in Science. AlphaFold3 extends the task to protein-DNA, protein-RNA, protein-small-molecule complexes, and post-translationally modified protein forms.

No single tool is best for every problem. A lab studying a conserved enzyme may need only AlphaFold2. A metagenomics group handling millions of fragments may prefer ESMFold for speed. Bench scientists are beginning to choose models the way they choose expression systems: by the question and the budget.

What changes when a structure arrives before the experiment

The most immediate change is in construct design. Crystallographers and cryo-EM researchers have traditionally removed flexible loops or terminal regions by trial and error because unstructured regions prevent lattice contacts or complicate particle alignment. A predicted model with per-residue confidence can identify those regions before a single milligram of protein is produced. Membrane protein labs use predictions to choose truncations that may express better in detergent or nanodiscs. The time saved is not theoretical; it is measured in weeks of failed expression tests.

Predicted models also speed up experimental structure determination. In molecular replacement, a predicted model can serve as a search model when no experimental homolog is available. In cryo-EM, docking a prediction into a modest-resolution map helps trace the polypeptide chain more quickly than building from scratch. Journals and databases have adapted in parallel: Protein Data Bank depositors are now expected to state whether AlphaFold or similar models were used in model building. This is a practical acknowledgment that the two methods are increasingly intertwined.

Reading the confidence scores before trusting the model

The colored confidence plots are the first screen. pLDDT runs from 0 to 100; regions above 90 are generally reliable at the backbone level, while regions below 50 should be treated as unstructured or uncertain. For complexes, the predicted aligned error, or PAE, indicates whether the relative position of two domains is backed by evidence. A low PAE between two regions means the model is confident about their relative orientation; a high PAE means the shared position is a guess.

Researchers should resist the instinct to treat a global confidence score as permission to stop checking. A model can be right about a folded core and wrong about an active-site loop that matters for catalysis. The practical move is to read the per-residue scores first, then inspect the biology: does the predicted pocket line up with known functional residues? Do disulfide bonds form between cysteine residues that should be oxidized? Does the N-linked glycosylation site point outward?

• Check pLDDT and PAE before using any region in a functional claim.
• Compare the prediction with experimental data from small-angle X-ray scattering, circular dichroism, cryo-EM maps, or hydrogen-deuterium exchange when available.
• Confirm critical residues by mutagenesis or binding assays before drawing conclusions.
• Treat ligand docking positions from predictive models as hypotheses, not measured affinities or binding modes.

AlphaFold3 and the interaction problem

Proteins rarely act alone. The next practical test for molecular structure prediction is not another protein fold but a complex: a transcription factor on DNA, an antibody on an antigen, a kinase with a small-molecule inhibitor, or a viral spike protein with its receptor. AlphaFold3, described in Nature in 2024, extended prediction to these multi-component systems using a diffusion-based approach. It can produce plausible models of protein-ligand, protein-nucleic acid, protein-protein, and antibody-antigen assemblies without a template.

The catch is that plausible is not the same as physical. AlphaFold3 is not a thermodynamic model; it does not measure binding affinity, and it can hallucinate ligand positions in pockets that are only partially formed. The practical use inside a lab is to generate ideas, narrow a screen, or guide medicinal chemistry. The experimental confirmation must still happen, and it usually takes longer than the prediction.

The limits a bench scientist notices

Prediction tools are less reliable for intrinsically disordered regions, flexible loops, membrane protein topology in a lipid-bilayer environment, and proteins that require cofactors or post-translational modifications to fold. They don't model dynamics well: a static structure cannot show conformational changes, allostery, the effect of a mutation on stability, or binding kinetics. A predicted structure of an intrinsically disordered protein often appears as an extended chain with low pLDDT. That isn't a model failure. It is a signal that the protein does not have a stable fold on its own.

Single-point mutations are a related challenge. Most models do not consistently rank which variant will destabilize a protein, and they are not trained to predict temperature sensitivity or aggregation. When a lab asks which of 20 substitutions will break function, the model often answers with confidence intervals too wide to use. Experimental screening remains the backstop.

Validation remains laboratory work

Predictions are now fast enough that the bottleneck has moved to checking them. A structure prediction for a new protein can be generated in an afternoon, but establishing that the model is meaningful may take months: purifying the protein, measuring its oligomeric state, probing a binding site with mutants, or collecting a low-resolution cryo-EM map. The gap between time-to-prediction and time-to-validation is one reason that publications of purely predictive structures are treated with caution by reviewers.

In practice, a lab will combine methods. A prediction can justify a construct. A cryo-EM map can confirm the overall fold. A hydrogen-deuterium exchange experiment can test which regions are flexible. A mutagenesis panel can tie a residue to activity. The model enters this workflow as one piece of evidence, not a substitute for the others.

The question labs will face in year two is straightforward and awkward: did we stop doing the slower experiment because the model was good, or because the model was easy? The answer will vary by protein, by lab and by budget. Better tools can shorten the distance between a sequence and a structure. They can't remove the obligation to test the hypothesis.

Photo by Stephan HK on Unsplash