Absolute protein folding stability prediction that estimates ΔG by jointly modeling the folded and unfolded ensembles as residue-pair distograms.
No providers recorded yet. Browse all providers
Folding stability, the unfolding free energy ΔG, decides whether an engineered protein survives expression, purification, and use. It is also expensive to measure and awkward to predict. The most common computational workaround is to borrow a confidence metric from a structure predictor — pLDDT or pTM from AlphaFold 2 or ESMFold — and use it as a stability filter, even though those metrics were never trained on thermodynamics. A parallel line of work predicts ΔΔG, the stability change on mutation, but says nothing about absolute stability and is generally restricted to single point substitutions.
IFUM (In silico evaluation of unfolding Free energy with Unfolded state ensemble Modeling), from the Korea Institute of Science and Technology with collaborators at KAIST and Seoul National University, argues that the missing ingredient is the unfolded state. ΔG is by definition the free energy gap between folded and unfolded states, yet most deep learning approaches represent only the folded structure and treat the denatured state as a sequence-independent null. IFUM instead predicts both, jointly: a per-residue ΔG and an equilibrium ensemble expressed as a residue-pair Cα distance histogram that mixes a folded distogram with an unfolded one, weighted by the Boltzmann-derived folded population.
The unfolded distogram is not learned from scratch. It comes from the Flory random coil model, which gives root-mean-square Cα distances as a function of sequence separation alone. That single modeling choice is what the ablations credit: training IFUM without the ensemble objective costs as much accuracy as removing triangle multiplicative updates from the architecture.
IFUM is a transformer taking three inputs: a ProtT5 sequence embedding of shape [N, 1024], an ESM-IF1 structure embedding of shape [N, 512], and a one-hot binned Cα distogram of the folded structure (21 bins, 2–42 Å) predicted by ESMFold. Its main module is a modified Evoformer — triangle multiplicative updates retained, triangle attention dropped for memory, and all self-attention replaced by differential self-attention — stacked 11 blocks deep, with MLP heads for ΔG, the ensemble distogram, and an auxiliary sequence recovery task. Training used 648,650 Mega-scale proteins of 30–80 residues with proteolysis-derived ΔG labels plus 3,219 short DisProt sequences constrained toward ΔG below 0.5 kcal/mol, split by MMseqs2 clusters at 30% identity.
On the Mega-scale test set IFUM reached a Pearson correlation of 0.78 and RMSE of 1.16 kcal/mol; the ablated model without unfolded-state modeling fell to 0.70 and 1.39. On a common subset shared with ESMtherm, IFUM scored 0.91 versus 0.86 for ESMtherm and 0.44 for ESM-2 pseudo-likelihood. ΔΔG correlations were 0.81 for point mutants, 0.80 for indels, and 0.63 for doubles. Blind-tested engineering campaigns on IFN-λ, IL-10, and UGT76G1 gave melting-temperature correlations of 0.75, 0.62, and 0.87, where the AlphaFold 3 confidence metric correlated negatively on two of the three.
IFUM is aimed at protein engineers deciding which of many designed variants to express. Its two demonstrated uses are stabilization engineering — where loop truncations and insertions put the target outside the reach of point-mutation ΔΔG predictors — and de novo design triage, where filtering by predicted ΔG beat pLDDT filtering at identifying designs that express as soluble monomers across Rossmann, TIM barrel, helical repeat, beta barrel, and jelly roll folds. Code is MIT-licensed with a Google Colab notebook for users without local GPUs.
Published in Nature Communications, IFUM offers a concrete alternative to the widespread habit of filtering designs on structure-prediction confidence. Its limitations are stated plainly by the authors: accuracy tracks the quality of the input folded structure, absolute values become unreliable above roughly 200 residues, high loop content biases predictions low, and the model does not cover membrane proteins or obligate oligomers.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.