REVIEW 5 major objections 5 minor 18 references
Predicting protein folding dynamics using sequence information
T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read This paper presents a method that converts a multiple sequence alignment of a protein family into quantitative predictions of how any member sequence folds and unfolds.
desk verdict A clean, useful protocol that generalizes the authors' earlier Potts-plus-Ising work, but the 'sequence alone' claim is not validated by new experiments in this paper. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is a finite-chain Ising model of foldons. Each foldon $j$ carries a two-state spin (folded or unfolded), an internal folding free energy $\epsilon_j^i$, and surface interactions $\epsilon_{jk}^s$ with other folded foldons, plus an entropic penalty $s_j = L_j s$ with $s = 5\,\mathrm{cal\,mol^{-1}\,K^{-1}\,res^{-1}}$; all energetic terms are computed by summing Potts couplings and fields inside and between foldons for the target sequence and scaling by $k_B T_{\mathrm{sel}}$. The Potts model, learned from the MSA by DCA-type inference, provides the evolutionary energy field, and $T_{\mathrm{sel}}$ converts evolutionary energy differences into physical free-energy differences. Metropolis Monte Carlo then samples the chain across temperatures, from which thermal unfolding curves, free-energy profiles, folding temperatures, subdomain boundaries, and cooperativity scores are read off.
What would settle it
Apply the protocol to a protein family that already has experimental thermal denaturation curves and a deep mutational scan, then compare predicted per-foldon folding temperatures, the overall unfolding curve, and mutation-induced shifts in stability and cooperativity with the measurements. A disagreement beyond calibration noise in the rank order of stabilities across homologs, or in the location of predicted subdomain boundaries, would show that the Potts-to-free-energy mapping is not generally valid.
Extended reading notes
Core claim
The central claim is that folding dynamics, not just the native structure, are accessible from sequence information alone. For any sequence belonging to a family with a sufficiently deep and clean alignment, the equilibrium folding curve can be computed: the protein is divided into foldons, each treated as a folded/unfolded spin, with internal folding free energies $\epsilon_j^i$ and inter-foldon surface energies $\epsilon_{jk}^s$ derived from the Potts evolutionary energy of that sequence and rescaled by $k_B T_{\mathrm{sel}}$. The selection temperature is obtained from experimental $\Delta\Delta G$ data when available, or from the ratio of standard deviations of evolutionary energy changes for single mutations relative to a reference family. Subdomains are identified when neighboring foldons have folding temperatures closer than 5 K, cooperativity is measured by the fraction of intermediate folded-count states $Q$ that are never free-energy minima, and single-mutation effects are extracted from the wild-type simulation without rerunning it. The paper thus offers an end-to-end pipeline from alignment to folding dynamics.
Load-bearing premise
The load-bearing premise is that one family-level selection temperature $T_{\mathrm{sel}}$, together with a fixed additive per-residue entropy, converts evolutionary Potts energies into real folding free energies for every sequence of the family; if those calibrations are family-dependent, or if the user-chosen foldon partition does not match the true cooperative units, the predicted curves, subdomains, and mutation effects would be artifacts.
Editorial extensions
If this is right
- For any family with a deep MSA, natural sequences can be ranked by predicted folding temperature and cooperativity, which may help select thermostable variants for biotechnology.
- Mutation effects on stability and cooperativity can be predicted for all single-point mutants from one wild-type simulation, providing a sequence-only proxy for deep mutational scanning.
- Sequences with more favorable total evolutionary energy are predicted to be more stable by construction, so family-level energy ranking is itself a stability predictor.
- Topology is predicted to matter: elongated $\alpha$-helical families can show sequence-dependent folding mechanisms, while compact $\beta$-containing families are more conserved in mechanism.
- Synthetic sequences sampled from the Potts field can be placed in the heterogeneity–interaction phase space to estimate cooperativity before running any folding simulation.
Reading between the lines
- If the relative-scale estimate of $T_{\mathrm{sel}}$ transfers across families, the pipeline could be applied to metagenomic or poorly characterized families with no experimental stability data at all.
- Because the foldon partition is a user input, averaging predictions over neutral-model partitions (as the paper suggests) could serve as a robustness check for whether predicted subdomains are real or an artifact of the chosen boundaries.
- Positions conserved for reasons other than folding stability—such as binding or catalysis—may show systematic local errors, and frustration-based analysis could flag those positions ahead of time.
- Substituting protein language model logits for Potts couplings (mentioned as an alternative in the paper) would make the same Ising machinery applicable to families too shallow for DCA, at the cost of trusting the language model's energy interpretation.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript presents a protocol for predicting equilibrium protein folding behavior from sequence information. The method learns a Potts model from a multiple sequence alignment, maps the resulting evolutionary energy parameters to a coarse-grained Ising chain of foldons using a selection temperature and an empirical per-residue entropy, and then runs Metropolis Monte Carlo simulations to compute thermal unfolding curves, per-foldon folding temperatures, free energy profiles, a cooperativity score, and apparent subdomain emergence. It also describes how to estimate changes in folding temperature and cooperativity upon single-point mutations, and illustrates the protocol on a DHFR sequence, a ubiquitin sequence, and an analysis of 7490 sequences from 15 families.
Significance. If the quantitative mapping from Potts energies to folding free energies could be shown to transfer across families without per-family experimental calibration, this would be a valuable method for converting sequence alignments into testable predictions about folding stability, cooperativity, and subdomain organization. The paper is clearly written as a step-by-step protocol, ships a reproducible Colab implementation, and builds on well-established DCA methodology; the 7490-sequence phase-space analysis is also a useful descriptive resource. However, the central claim of sequence-only prediction is not yet demonstrated, because the quantitative content rests on fitted constants and a user-specified foldon partition, and no independent experimental benchmark is presented in this manuscript.
major comments (5)
- [Methods, Step 2 (and Abstract/Box 1)] The abstract and Box 1 claim that folding curves and mutation effects are predicted from sequence information alone, but Methods Step 2 requires experimental delta-delta-G data for at least one family member to fix T_sel, or transfers a PDZ-derived T_sel under the assumption that the standard deviation of delta-delta-G is family-independent. This fitted constant multiplicatively scales all energy differences and therefore directly sets every predicted T_f and delta-delta-G; without a sensitivity analysis across families or an independent benchmark, the 'sequence alone' claim is not supported.
- [Methods, Step 3] The per-residue entropy s = 5 cal/mol/K/res is taken from ref [29], an empirical fit for repeat proteins, and is assumed additive and sequence-independent for all foldons. Because the Boltzmann weights in the Hamiltonian make the unfolding curves exponentially sensitive to this entropy, the transfer of this fit to globular proteins with different topologies (e.g., DHFR and ubiquitin in Figures 2 and 4) needs justification; at minimum, a sensitivity analysis over a plausible range of s should be provided.
- [Materials, Section 3, and Methods, Step 5] The foldon partition is user-specified, and the apparent-domain definition uses a fixed threshold |T_fj - T_fk| < 5 K. The inferred subdomains and cooperativity score therefore depend on the chosen partition and threshold. The manuscript mentions a neutral-model averaging option but does not apply it, so there is no evidence that domain emergence is robust rather than an artifact of the specific foldon assignment chosen in Figures 2 and 4.
- [Figures 2, 3, and 4] The central claim that the method 'predicts' folding dynamics is not tested against new experimental data in this manuscript. Figures 2 and 4 are illustrative applications, and Figure 3 is a descriptive scatter of the cooperativity score across 7490 sequences with no comparison to measured folding quantities. As a protocol paper, the manuscript should either provide an experimental validation for at least one family or explicitly frame the curves as untested predictions; the current abstract overstates what is demonstrated.
- [Title and Methods, Step 4] The title and abstract promise 'folding dynamics,' but the Monte Carlo simulation is an equilibrium Metropolis scheme that computes thermal unfolding curves and free energy profiles, not kinetic rates or pathways. The word 'dynamics' should either be replaced by 'thermodynamics' or the model must be extended to non-equilibrium simulation; as written, the terminology overstates the scope of the method.
minor comments (5)
- [Methods, Step 4] The critical-points factor and the autocorrelation analysis are described qualitatively; please give the default values and the convergence criteria used in the reported simulations so that results are reproducible.
- [Figure 3] The axis labels 'energetic heterogeneity' and 'average interaction strength' are not defined quantitatively; please provide the exact formulas used to compute these quantities.
- [Methods, Step 6] The statement that 'using a linear fit, useful predictions can be made' should cite the specific figure or Table that supports this claim and should report the fit quality (e.g., R^2 or mean absolute error).
- [Note 3 and Abstract] Note 3's scope restriction, that compact beta proteins do not require a sequence-sensitive Potts model, should be reflected in the abstract, since it substantially narrows the class of proteins for which the protocol is claimed necessary.
- [Methods, Steps 1-3] The equations in the Methods section as provided do not render in the manuscript text, making it impossible to verify the mapping from Potts couplings to foldon energies; please ensure all equations are fully displayed.
Circularity Check
Mutation 'predictions' and absolute folding curves inherit a ΔΔG-fitted T_sel and an empirically fitted entropy, so the sequence-only claim reduces to a calibrated pipeline.
-
fitted input called prediction
[Methods, Step 2 (Estimate the Selection Temperature)]
"If experimental folding free energy changes upon mutations 𝛥𝛥 G are available for at least a protein of the studied family, k_B T_sel can be explicitly calculated as the proportionality constant between 𝛥𝛥 G and the corresponding changes scores as the energy difference of the wild-type and mutant according to Potts field."
T_sel is obtained by regressing experimental ΔΔG against Potts energy differences, i.e., it is a fitted slope. In Step 3 the Potts energies are converted into the coarse-grained folding Hamiltonian by dividing by k_B T_sel, so every predicted stability change caused by a mutation is the same linear transformation that was used to calibrate T_sel. The deep-mutational-scan 'predictions' (Figure 4 and 'the impact of all possible single-site mutations can be estimated directly from the wildtype simulation', citing [21]) therefore restate the calibration line rather than deriving mutation effects from sequence information.
-
other
[Abstract and Methods, Step 2]
"Here we present a method to infer protein folding dynamics based only on sequence information. ... Alternatively, assuming that the standard deviation of 𝛥𝛥 G is nearly constant irrespectively of protein families, T_sel can be estimated in a relative scale to a known reference, using the ratio of the standard deviation of evolutionary energy changes by single mutations [35]. The reference should be a T_sel obtained from 𝛥𝛥 G experimental data, for instance for PDZ family..."
The headline claim 'based only on sequence information' is not realized: the absolute temperature scale of every predicted folding curve is fixed by T_sel, which is either fit to experimental ΔΔG for the family or transferred from a PDZ reference T_sel that was itself obtained from experimental ΔΔG. Thus the quantitative content of the folding curves is inherited from experimental stability data through a fitted constant, contradicting the sequence-only framing.
1 more flagged steps
-
fitted input called prediction
[Methods, Step 3 (Assign the coarse-grained folding hamiltonian)]
"To compute the entropic terms ( s_j ), we recommend to approximate to be independent of amino acid identity and strictly additive and to use an average entropy per residue of s = 5 cal mol −1 K −1 res −1 that was empirically fitted [29]. ... Alternatively, if experimental Temperature denaturation scans of the protein of interest (or a protein of the family) and mutants is known, the entropic term can be fitted by minimizing the difference between the simulated and the experimental data [29]."
The predicted unfolding curves and cooperativity scores depend directly on the entropy term. The default entropy is an empirical constant imported from the authors' prior repeat-protein work [29], so the curves' shape is not a first-principles consequence of the sequence. The alternative mode explicitly fits the entropy to experimental denaturation scans of the protein whose folding curve is then reported, making agreement between simulation and experiment a by-product of the fit rather than an independent sequence-based prediction.
full rationale
The derivation chain has an independent core: the Potts evolutionary field is learned from an MSA and is a legitimate sequence-based input. However, the bridge from that field to physical folding is not first-principles. Step 2 defines T_sel as the proportionality constant between experimental ΔΔG and Potts energy differences; Step 3 uses that same T_sel to build the Ising Hamiltonian. Consequently, the reported mutation stability and ΔT_f 'predictions' are the calibration line itself, not a new sequence-only result. The absolute scale of the folding curve additionally depends on an entropy per residue (s = 5 cal/mol/K/res) that is an empirical fit from the authors' previous work, and the paper explicitly allows fitting this entropy to the same experimental denaturation data to which the output is compared. Note 1 and Note 3 further concede that positions under other selection forces and topology-only models narrow the scope of the sequence-information claim. These are partial rather than total circularities: relative ordering of stabilities within a family and the phase-space map in Figure 3 retain some independent content, and the method is self-consistent. But the central quantitative 'predictions' for mutations and the 'based only on sequence information' framing reduce by construction to fitted inputs, so a moderate-to-high circularity score is warranted.
Assumptions & free parameters
free parameters (4)
- Selection Temperature T_sel =
calibrated per family from experimental delta-delta-G values (e.g., PDZ reference)
- Per-residue foldon entropy s =
5 cal mol^-1 K^-1 res^-1
- Foldon partition =
user-defined (secondary structure, exons, repeats, or neutral geometric sampling)
- Subdomain emergence threshold =
|T_f,j - T_f,k| < 5 K
assumptions (4)
- domain assumption Protein stability has been the main local evolutionary constraint on sequences
- domain assumption The evolutionary ensemble of homologous sequences is Boltzmann-distributed at an effective temperature T_sel
- domain assumption Protein folding can be coarse-grained to a one-dimensional Ising chain of independent foldons with a single internal free energy and pairwise surface interactions
- standard math The mean evolutionary energy of random sequences is zero (zero-sum gauge)
Cite this review
Pith. "Pith review of Predicting protein folding dynamics using sequence information." pith.science (2026). https://pith.science/paper/Y42VIOEJ
@misc{pith2026250517237,
author = {Pith},
title = {Pith review of: Predicting protein folding dynamics using sequence information},
year = {2026},
howpublished = {\url{https://pith.science/paper/Y42VIOEJ}},
note = {Machine review of arXiv:2505.17237}
}
read the original abstract
Natural protein sequences somehow encode the structural forms that these molecules adopt. Recent developments in structure-prediction are agnostic to the mechanisms by which proteins fold and represent them as static objects. However, the amino acid sequences also encode information about how the folding process can happen, and how variations in the sequences impact on the populations of the distinct structural forms that proteins acquire. Here we present a method to infer protein folding dynamics based only on sequence information. For this, we will rely first on the obtention of a precise 'evolutionary field' from the observed variations in the sequences of homologous proteins. We then show how to map the energetics to a coarse-grained folding model where the protein is treated as a string of foldons that interact. We then describe how, for any given protein sequence of a family, the equilibrium folding curve can be computed and how the emergence of protein folding sub-domains can be identified. We finally present protocols to analyze how mutations perturb both the folding stability and the cooperativity, that represent predictions for a deep-mutational scan of a protein of interest.
Reference graph
Works this paper leans on
-
[1]
A target protein sequence
-
[2]
A Multiple sequence alignment of proteins homologous to the target sequence
-
[3]
A partition of the sequence into contiguous folding elements, often referred to as foldons . We propose using a sequence-based energy function that evaluates the probability of observing a protein sequence of a given length within an equilibrated ensemble. These "evolutionary" energy parameters can be computationally derived from a Multiple Sequence Align...
-
[4]
Target Protein Sequence A target amino acid sequence is required to initiate the folding simulation. This sequence can represent the full length of a protein, or a fragment (e.g., a domain) that is assumed to fold independently of the rest of the chain. Variations in folding dynamics are expected across protein families, particularly for protein topologie...
-
[5]
Multiple Sequence Alignment (MSA) To derive a sequence-based evolutionary energy, a set of homologous and aligned sequences is necessary. A Multiple Sequence Alignment (MSA) can be obtained from the 4 Pfam database [22] (now hosted by InterPro [23]) by searching for the protein family containing the target sequence. Alternatively, an MSA can be constructe...
work page 2000
-
[6]
Folding Elements (Foldons) The protein sequence must be divided into distinct, non-overlapping folding elements, or foldons, which are groups of amino acids that fold and unfold cooperatively. While a common foldon assignment for each protein family (or MSA) is recommended, the method can accommodate sequence-specific partitions. We describe here some heu...
-
[7]
Learn the Potts model parameters from a MSA
-
[8]
Estimate the Selection Temperature
Show all 18 references
-
[9]
Assign the coarse-grained folding hamiltonian
-
[10]
Set up the Monte Carlo parameters and run simulations
-
[11]
Visualize and analyze results for a single protein
-
[12]
7 Model definition and methods pipeline are summarized in figure 1
Extra: Exploit family results for protein design. 7 Model definition and methods pipeline are summarized in figure 1. Figure 1. Model Definition. We learn the Evolutionary Energy field parameters from a Multiple Sequence Alignment (MSA) using a Direct Coupling Analysis (DCA). ...
-
[13]
The evolutionary energy parameters can be inferred from a MSA through a Direct Couple Analysis (DCA)
Learn the Potts model parameters from a MSA. The evolutionary energy parameters can be inferred from a MSA through a Direct Couple Analysis (DCA). A Potts model is defined, containing local fields and pairwise couplings, to account for the occurrence of a given amino acid in a...
-
[14]
Estimate the Selection Temperature. The Selection Temperature T sel is the apparent temperature at which sequences of a particular family were selected by nature and quantifies how strong the folding constraints have been during evolution [34]. If experimental folding free ene...
-
[15]
The protein is modeled as an array of interacting folding elements (foldons) that can be either folded (F) or unfolded (U), as 2-state spin variables
Assign the coarse-grained folding hamiltonian. The protein is modeled as an array of interacting folding elements (foldons) that can be either folded (F) or unfolded (U), as 2-state spin variables. The system is represented as a finite-size Ising chain of N elements, where the...
-
[16]
Set up the Monte Carlo parameters and run simulations Once the Hamiltonian has been set for a sequence, the Metropolis Monte Carlo simulation of the Ising Model must be set up. To describe the folding dynamics, the simulation must be repeated for several temperatures T , such ...
-
[17]
As the temperature increases, the folding elements are expected to transition from the folded to the unfolded state
Visualize and analyze results for a single protein. As the temperature increases, the folding elements are expected to transition from the folded to the unfolded state. For each element, the algorithm outputs the simulated thermal unfolding curves and single-element folding te...
-
[18]
THEORY OF PROTEIN FOLDING: The Energy Landscape Perspective,
Extra: Exploit family results for protein design. Once the Ising model has been set up for computing the folding dynamic of a given sequence, it is straightforward to also apply it for many other sequences of the same protein family. The evolutionary Potts model, the foldon as...
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.