AlphaFold2 changed structural biology more rapidly than anything since the introduction of NMR spectroscopy. Within two years of its public release, the AlphaFold DB contained predicted structures for over 200 million proteins — more structures than the RCSB PDB had accumulated in 50 years of experimental crystallography and cryo-EM. Protein engineers gained access to structural models for targets that had no crystallographic data and no reasonable path to getting it. That's genuinely valuable. But a significant confusion propagated alongside it: if AlphaFold can predict structure this accurately, can it also predict the effect of mutations on stability and binding? The answer is no — and the reasons why matter for anyone designing a protein engineering campaign.
What AlphaFold was trained to do
AlphaFold2 is trained on a specific task: predict the three-dimensional structure of a protein given its amino acid sequence. More precisely, it is trained to predict the most probable conformation — the structure closest to the ground-state fold. The training signal is the experimentally determined structure in the PDB, and the objective function rewards predicting the correct atomic coordinates.
This training objective has an important implication for mutation analysis. AlphaFold learned to predict "what fold does this sequence adopt?" — a single-point prediction for each sequence. It did not learn to predict "how does the thermodynamic stability of the fold change when this position changes from alanine to valine?" — a differential prediction across two closely related sequences. These are different tasks.
An analogy: a model trained to predict what town you'll end up in if you follow a GPS route is not the same as a model trained to predict how long the route will take if one road is closed. Structural prediction and thermodynamic perturbation prediction are related problems, but they require different training signals and different model architectures.
The specific gaps for mutation scanning
AlphaFold predictions are not thermodynamic calculations. A ΔΔG value requires knowing the free energy of both the folded and unfolded states for the wild type, and again for the mutant — a total of four thermodynamic estimates whose difference yields the ΔΔG. AlphaFold does not compute free energies. It predicts the most likely structure. When you feed AlphaFold a point mutant sequence, you get a predicted structure for that mutant — but no information about whether that structure is more or less thermodynamically stable than the wild type. The pLDDT confidence score reflects the model's confidence in its structural prediction, not the thermodynamic stability of the fold.
Structural similarity ≠ thermodynamic similarity. For most single-residue substitutions, AlphaFold will predict a structure for the mutant that is nearly identical to the structure for the wild type — because most single-residue substitutions don't substantially alter the fold. This means the AlphaFold outputs for a wild-type protein and a point mutant look almost the same, even when the experimentally measured ΔΔG is −2 kcal/mol (substantially stabilizing). The signal you're looking for is largely invisible in the structure comparison.
The pLDDT score does not report on mutation effects. A common misapplication of AlphaFold for mutation analysis involves comparing pLDDT scores between wild-type and mutant structures: "the pLDDT dropped from 87 to 82 for this mutant — it must be destabilizing." This is not a valid inference. pLDDT reflects structural prediction confidence, which is determined primarily by evolutionary depth in the multiple sequence alignment and structural plausibility. A single amino acid substitution rarely changes pLDDT significantly, and the change when it does occur reflects uncertainty in structure prediction, not a thermodynamic stability estimate.
ColabFold / ESMFold have the same limitation. Both ColabFold (which uses AlphaFold2 with accelerated MSA generation) and ESMFold (ESM protein language model adapted for structure prediction) share the fundamental limitation: they are structure predictors, not ΔΔG predictors. Running a library of mutant sequences through ColabFold and ranking by pLDDT difference or predicted structure quality metrics does not produce a reliable stability prediction.
What AlphaFold is good for in a mutation scanning workflow
None of this means AlphaFold is not useful for protein engineering. Quite the contrary — it belongs in the pipeline, just not in the mutation scoring step.
Structural input generation. For proteins without experimental structures (the majority of enzyme engineering targets), AlphaFold2 provides a high-quality structural model that can be used as input to ΔΔG scoring methods. This is the primary role: AlphaFold as structure provider, not as stability predictor. ProtSynq accepts AlphaFold model outputs (.pdb from AF2/AF3) directly as structural inputs. The AF2 model gets you the geometry; ProtSynq's scoring ensemble handles the thermodynamic perturbation prediction.
Structure quality assessment. pLDDT scores are a useful guide to which regions of the AlphaFold model are reliable for downstream mutation analysis. Positions with pLDDT > 85 are typically well-predicted and suitable as structural inputs for ΔΔG calculations. Positions with pLDDT below 70 — disordered regions, flexible loops, N/C-terminal tails — should be treated as low-confidence for both structural and thermodynamic predictions. This is not a limitation unique to AlphaFold-based workflows; it applies to any structure-based prediction in disordered regions.
Multi-chain complex modeling. AlphaFold-Multimer can predict the structure of protein–protein complexes, including antibody–antigen and enzyme–substrate complexes, where no experimental complex structure is available. These predicted complex structures can serve as inputs for binding ΔΔG prediction — though with lower confidence than experimentally determined complex structures, and with the proviso that large predicted interface errors in AlphaFold-Multimer models (which occur more frequently than in single-chain predictions) will propagate into unreliable binding ΔΔG scores.
Methods that do predict mutation effects on stability
The methods actually suited to ΔΔG prediction share a common feature: they are trained or parameterized on experimentally measured thermodynamic data, not on structural accuracy.
Physics-based methods (Rosetta ddg_monomer, FoldX, CUPSAT) estimate ΔΔG from changes in molecular mechanics energy terms — van der Waals, electrostatics, solvation — computed from the structural coordinates. These methods have been available for two decades and form the basis of most current hybrid approaches.
Learned ΔΔG models (ProteinMPNN with energy heads, sequence-based DMS-trained models, and ensemble methods like ProtSynq's scoring engine) are fine-tuned on databases of experimentally measured ΔΔG values from mutagenesis studies. ProThermDB and ProtaBank are the primary sources. The learned component captures the systematic correction between what physics models predict and what experiment measures — the calibration that converts a physics energy estimate into a biologically meaningful ΔΔG prediction.
The distinction matters: AlphaFold's training data is structural (PDB coordinates); ΔΔG prediction requires thermodynamic training data (calorimetric and spectroscopic stability measurements). These are different datasets with different experimental origins.
The practical workflow
For protein engineering campaigns, the correct integration of AlphaFold and ΔΔG methods looks like this:
- Use AlphaFold2 or ColabFold to generate a structure for your protein of interest, if no experimental structure is available.
- Evaluate pLDDT scores to identify well-predicted regions (> 85) and exclude disordered regions (< 70) from the active scanning space.
- Submit the AlphaFold model (.pdb) to a ΔΔG scoring engine (ProtSynq, FoldX, Rosetta ddg_monomer, or equivalent). This step is where the mutation effect prediction actually happens.
- Use the ΔΔG scores — not the AlphaFold pLDDT, not structural similarity between wild-type and mutant AF2 outputs — to rank the substitution landscape and generate the experimental shortlist.
AlphaFold got you the structure you couldn't otherwise access. The ΔΔG scoring method does the mutation analysis that AlphaFold wasn't designed for.