Every protein engineer has experienced it: you pick the mutation that looks most promising on the structure, clone it, express it, run the DSC, and the Tm drops instead of rising. The computational filter you were hoping for sent you in the wrong direction. Why does this happen, and what does it reveal about the limits of current ΔΔG prediction methods?
The formalism and why it understates the difficulty
The standard formalism is straightforward. A protein's thermostability is characterized by its folding free energy ΔGfold, the free energy difference between the folded and unfolded states. When a single residue is mutated from the wild type (WT) to a variant (MUT), the effect is captured as:
ΔΔG = ΔGfold(MUT) − ΔGfold(WT)
A negative ΔΔG means the mutation stabilizes the fold; a positive ΔΔG means destabilization. The sign and magnitude are what protein engineers want to predict before committing to synthesis and expression.
The mathematical definition is clean. The physical reality underneath it is not. ΔGfold is a thermodynamic average over an entire conformational ensemble — the protein in solution is not sitting in the single conformation captured in a crystal structure. When you introduce a point mutation, you perturb not just the single crystal-state geometry, but the entire ensemble of conformations the protein can access. The X-ray or cryo-EM structure you feed into a scoring function represents one snapshot of that ensemble. How well that snapshot approximates the relevant thermodynamic average varies by protein, by position, and by the chemical nature of the substitution.
The four dominant contributors — and why each is hard to model accurately
van der Waals packing. Replacing a small side chain (alanine, glycine) with a large one (phenylalanine, tryptophan) in a buried hydrophobic core almost always costs packing energy — steric clashes in the crystal structure are a reliable signal. The difficulty is in the other direction: replacing a large residue with a smaller one creates a cavity. Cavity formation in hydrophobic cores is stabilizing when the cavity is filled by water or accommodated by slight backbone shifts, and destabilizing when neither happens. Physics-based scoring functions compute cavity energy from the static structure. Whether a cavity actually forms — versus the backbone relaxing to fill the void — requires either explicit energy minimization or MD sampling, neither of which is fast enough for scanning 6,650 substitutions per protein.
Electrostatics and buried charged residues. Charge burial penalties are among the largest ΔΔG values in experimental datasets — mutations that bury a charged residue in a hydrophobic environment typically produce ΔΔG values in the +3 to +6 kcal/mol range. These are the easy cases: the sign is almost always positive and the magnitude is large. The hard cases are surface salt bridges. A salt bridge between Asp and Arg on the protein surface can be stabilizing (ΔΔG −0.5 to −1.5 kcal/mol) or nearly neutral, depending on the local solvation geometry and competing interactions. Continuum electrostatics models (Poisson-Boltzmann, generalized Born) capture average contributions but miss the geometry-specific cases that account for most of the variance in experimental data.
Backbone strain and torsion entropy. Mutations that alter backbone torsion angles — particularly glycine substitutions, which remove the side chain and increase backbone flexibility, or proline substitutions, which rigidify the backbone — have effects that extend well beyond the mutated residue. A glycine at a position that would normally be in α-helix conformation reduces helix propensity; the perturbation propagates along the helix. These long-range effects are difficult to capture from local energy calculations on a single structure.
Solvent exposure and hydrophobic burial. The transfer free energy of amino acids from water to a hydrophobic environment is the most tractable contribution to ΔΔG, and it's also the one that's best covered by classical scoring functions. But even here, the magnitude depends on how deeply buried the residue actually is in the mutant — which requires modeling structural reorganization that the static structure doesn't show.
Why ML models improved things — and where they still struggle
The emergence of learned ΔΔG models (most famously Rosetta's cartesian_ddg, and more recently transformer-based models trained on ProThermDB and ProtaBank) improved prediction Pearson correlations from the low-0.6 range typical of pure physics methods to 0.75–0.82 on held-out test sets. What did the learned component add?
First, it implicitly learned systematic corrections for the cases where physics models are consistently wrong — for example, the tendency of Rosetta-based methods to overestimate destabilization for mutations in loop regions, or to underestimate the stabilizing effect of hydrophobic core mutations involving methionine-to-leucine substitutions.
Second, transformer architectures trained on protein sequences learned evolutionary context: positions that are highly conserved across homologs are likely functionally or structurally critical, and mutations there are more likely to be destabilizing regardless of what the local geometry suggests. This evolutionary signal acts as a prior that prevents physics scoring from suggesting mutations that evolutionary pressure has already selected against.
The remaining failures cluster in two areas. Proteins with few homologs in sequence databases — fewer than ~50 effective sequences in a multiple sequence alignment — lose the evolutionary conservation signal almost entirely. And conformationally flexible regions (intrinsically disordered segments, long surface loops) violate the assumption underlying both physics and structure-based learned methods: that the crystal structure meaningfully represents the ensemble relevant to thermostability.
What this means in practice
For protein engineers running stability campaigns, the practical implications are these:
- For well-folded, well-conserved globular domains with an experimental structure or a high-confidence AlphaFold model (pLDDT > 85), ΔΔG predictions from ensemble methods achieve 0.75–0.80 Pearson correlation with experiment. That's good enough to collapse a 6,000-candidate substitution scan to a shortlist of 5–10 where the true stabilizing variant is in the top candidates ~80% of the time.
- For flexible loops, N- and C-terminal tails, and proteins with pLDDT below 70 in the region of interest, ΔΔG predictions are unreliable. Treat them as uninformative and apply experimental judgment.
- Absolute ΔΔG magnitudes are less reliable than rankings. A prediction of −1.8 kcal/mol vs −1.5 kcal/mol for two variants does not reliably predict which will show the larger Tm shift in DSC. The ranking is informative; the precise value is not.
- Glycine and proline substitutions, and mutations that alter net charge at buried positions, are the highest-variance predictions. Flag them explicitly rather than trusting the numerical ΔΔG alone.
ProtSynq reports per-position confidence intervals that widen for the cases described above. When you see a wide confidence band on a top-ranked variant, that's a flag to either validate the structure quality in that region or to treat the prediction as lower-confidence than the number suggests.
The honest current state
ΔΔG prediction has improved substantially over the past decade, driven by larger experimental datasets, better structure prediction (which expands what structures are available to input), and learned models that correct for systematic physics-model errors. But no current method achieves the prediction accuracy that would let you run a single computational scan and trust the result without any experimental validation. The right framing is: computational prediction as a filter to prioritize which five variants to test first, not as a replacement for testing.
The gap between a Pearson r of 0.78 on a held-out test set and the kind of predictive power that would make wet-lab validation optional is still real. Being explicit about that gap is more useful to protein engineers than pretending the problem is solved.