REVIEW
Predicting protein structures at any folding stage from sequence
Biomolecules · Submitted 3 Nov 2026 · 168 references
Pith's verdict: Real novelty in Q-conditioned ensemble generation, but chain-level accuracy across Q for held-out sequences is not yet demonstrated.
T0 · AI review
Why Pith says this
Highlight any text to comment on it.
Referee report
The paper
The paper claims
FoldScape generates protein conformations at any requested folding stage directly from amino acid sequence, without per-protein simulation.
The reading
What carries the argument
The mechanism that carries the argument is the ordered triple of conditioning descriptors $(Q, \tilde R_g, f_{ss})$ — fraction of native contacts, normalized radius of gyration, and secondary-structure fraction — injected at every stage of a four-stage cascade. The design principle is problem decomposition: instead of predicting $3N$ coordinates at once, Round 1 regresses one-dimensional $C_\alpha$–$C_\alpha$ distances at the seven shortest sequence separations inside 15-residue windows; Round 2 learns only the residual error of Round 1, so its gains are attributable by construction; Stage 1 lifts window-local features to the chain index by cross-window attention pooling and produces the single representation that conditions the decoder; and a rectified-flow decoder samples the final ensemble from a learned conditional velocity field. Chains beyond the 200-residue single-pass ceiling are handled by stride-one crop-and-superpose, in which overlapping crops are predicted independently and merged by Kabsch superposition, so length extrapolation is carried by parameter-free assembly rather than by any learned component.
Extended reading notes
Core claim
The paper's central claim is that the entire folding coordinate is learnable from sequence as a conditional generation problem, and that accuracy holds across the coordinate rather than only at the native end. FoldScape reduces three-dimensional reconstruction to a hierarchy of one-dimensional regressions: for every 15-residue sliding window it predicts seven $C_\alpha$–$C_\alpha$ distances at sequence separations 1 through 7 plus a windowed trace; a residual round refines those predictions; a chain-level stage pools the window features into a single representation; and a rectified-flow decoder turns that representation into an ensemble by Euler integration from Gaussian noise. Conditioning on the triple $(Q, \tilde R_g, f_{ss})$ puts the user, not the simulation, in control of the folding stage. The paper reports that reconstruction accuracy stays approximately flat along $Q$ — $C_\alpha$ RMSD between 0.67 and 1.76 Å across the partially folded and compact states of a 706-residue chain, with the fully extended frame's 125.8 Å value explained as an artifact of global RMSD on a nearly one-dimensional object — while the ensemble genuinely spans the coordinate, and it argues that holding those two properties together is what separates a folding generator from a native-state predictor.
Load-bearing premise
The model's unfolded and intermediate conformations are learned from coarse-grained folding runs at one fixed reduced temperature and from constant-velocity mechanical pulling in implicit solvent, so the claim that it generates real folding stages collapses if those simulated trajectories are not representative of the actual conformational continuum of folding.
Editorial extensions
If this is right
- A user can request a folding stage by number and receive diverse, geometrically plausible conformations across the whole coordinate, from high-entropy unfolded chains (median $Q \approx 0.24$, with 29.3% of conformations below $Q = 0.1$) to compact native states.
- Sequences several times longer than the 200-residue single-pass limit become tractable through crop-and-superpose assembly, demonstrated on chains of 706 and 857 residues with reconstruction accuracy stable along $Q$.
- The residue-level contact survival at the force peak of the training trajectories correlates with experimental mechanical $\phi$-values across three proteins and four pulling speeds, and correlates negatively with chemical $\phi$-values, showing the corpus encodes the mechanical unfolding pathway rather than the reverse of chemical folding.
- Extension-conditioned observables such as $Q$ and $R_g$ are reproducible across two decades of pulling velocity, which licenses pooling trajectories from different speeds into one training set as long as targets stay continuous and reaction-coordinate-conditioned.
- Unfolded segments of partially stretched chains follow a linear extension law ($\alpha = 0.947$ globally, residual MAE of 3.9 Å across 557,550 segments), which justifies fragment-level supervision as a match to the physical modularity of intermediates.
Reading between the lines
- The largest open step the paper leaves implicit is thermodynamic calibration: the pipeline emits conformations at a prescribed $Q$ with no claim about their equilibrium weights, so the natural next experiment is to attach populations — from experiment, Markov-state models, or simulations rerun at each protein's own transition temperature — and check whether the weighted ensemble reproduces measure
- Because the corpus was generated by mechanical, not chemical, denaturation, a direct test of the method's logic would be to train the identical architecture on chemical-unfolding or equilibrium trajectories for the same proteins and compare the two $Q$-conditioned ensembles; the paper's own ubiquitin control shows the two routes give genuinely different residue-level intermediates, so agreement is
- The fragment view of unfolding intermediates predicts sharp folded-to-unfolded boundaries and length-determined segment extensions that could be tested experimentally with single-molecule force spectroscopy or hydrogen-exchange data, independently of any prediction from the model.
- The secondary-structure coverage axis is the most exposed to back-mapping bias because the reconstruction tools carry a folded-structure prior; a cheap sensitivity test is to evaluate $f_{SS}$ coverage directly on $C_\alpha$-only geometry instead of DSSP on back-mapped heavy atoms.
Original abstract
Protein structure prediction has largely focused on native conformations, leaving the continuum of unfolded states, folding intermediates, and native structures difficult to access. We present FoldScape, a sequence-conditioned deep learning framework for generating protein conformational ensembles at specified folding fractions $Q$. Trained on approximately one million conformations from 10,000 protein sequences, FoldScape combines coarse-grained folding and all-atom mechanical-unfolding trajectories. It reconstructs three-dimensional $C_\alpha$ chains through hierarchical sliding-window regressions of inter-residue distances, chain-level assembly, and a rectified-flow decoder. By conditioning on folding progress, FoldScape generates conformations from unfolded states through intermediates to native-state structures without protein-specific molecular dynamics at inference. In matched-sample comparisons, FoldScape achieves median conformational-space coverage of 0.989, 0.722, and 0.877 along radius of gyration, fraction of native contacts $Q$, and secondary-structure fraction, respectively, versus 0.05, 0.05, and 0.10 for seed-varied AlphaFold 3 ensembles. Coverage advantages are consistent across the evaluated proteins and robust to binning choices. Reconstruction accuracy remains approximately stable across the evaluated stages of a folding trajectory, and stride-one crop-and-superpose inference extends predictions to chains several times longer than the single-pass limit. Analyses of mechanical-unfolding trajectories show residue-resolved pathways, reproducible extension-conditioned observables across pulling velocities, and geometrically coherent segments in partially unfolded chains. FoldScape extends sequence-based structure prediction beyond a single folded endpoint to controllable conformational-ensemble generation across protein folding.
The compiled record
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.
The abstract, sentence by sentence
Protein structure predictors reliably identify native conformations, but do not describe the continuum connecting unfolded ensembles to folded states. Here we present FoldScape, a sequence-conditioned deep learning framework that generates C$\alpha$ conformational ensembles at a specified folding fraction $Q$. FoldScape reconstructs three-dimensional chains through hierarchical, sliding-window regressions of one-dimensional inter-residue distances, followed by chain assembly and sampling with a rectified-flow decoder. It is trained on multi-scale coarse-grained folding trajectories and all-atom mechanical-unfolding trajectories. Across 20 human proteins (36-857 residues), at matched sample size, FoldScape covers about an order of magnitude more of the folding coordinate than seed-varied AlphaFold 3 ensembles, robustly across proteins, order parameters and binning choices. AlphaFold 3 better anchors the immediate native vicinity, which FoldScape does not reach, making the approaches complementary. FoldScape reconstruction accuracy remains approximately constant along $Q$, from collapsed coil to near-native globule, rather than deteriorating away from the native state. A stride-one crop-and-superpose scheme extends inference to chains several times longer than a single chain-level forward pass. We also show that training trajectories resolve mechanical unfolding at residue resolution, that elongation-conditioned ensembles exhibit reproducible continuous observables across pulling velocities, and that partially unfolded chains form geometrically coherent segments. FoldScape thus combines broad coverage of folding stages with stable reconstruction accuracy, without per-protein simulation.
background contribution method result interpretation main claim
| # | sentence | role | the authors' own work | states the main conclusion | strength |
|---|---|---|---|---|---|
| s0 | Protein structure predictors reliably identify native conformations, but do not describe the continuum connecting unfolded ensembles to folded states. | gap 0.87 | 0.06 | 0.04 | asserted 1.00 |
| s1 | Here we present FoldScape, a sequence-conditioned deep learning framework that generates C$\alpha$ conformational ensembles at a specified folding fraction $Q$. | contribution 0.99 | 0.94 | 0.14 | asserted 1.00 |
| s2 | FoldScape reconstructs three-dimensional chains through hierarchical, sliding-window regressions of one-dimensional inter-residue distances, followed by chain assembly and sampling with a rectified-flow decoder. | method 1.00 | 0.88 | 0.06 | asserted 1.00 |
| s3 | It is trained on multi-scale coarse-grained folding trajectories and all-atom mechanical-unfolding trajectories. | method 1.00 | 0.79 | 0.03 | asserted 1.00 |
| s4 | Across 20 human proteins (36-857 residues), at matched sample size, FoldScape covers about an order of magnitude more of the folding coordinate than seed-varied AlphaFold 3 ensembles, robustly across proteins, order parameters and binning choices. | result 1.00 | 0.95 | 0.27 | asserted 0.99 |
| s5 | AlphaFold 3 better anchors the immediate native vicinity, which FoldScape does not reach, making the approaches complementary. | limitation 0.48 | 0.84 | 0.11 | asserted 1.00 |
| s6 | FoldScape reconstruction accuracy remains approximately constant along $Q$, from collapsed coil to near-native globule, rather than deteriorating away from the native state. | result 1.00 | 0.94 | 0.27 | asserted 1.00 |
| s7 | A stride-one crop-and-superpose scheme extends inference to chains several times longer than a single chain-level forward pass. | method 0.97 | 0.90 | 0.04 | asserted 1.00 |
| s8 | We also show that training trajectories resolve mechanical unfolding at residue resolution, that elongation-conditioned ensembles exhibit reproducible continuous observables across pulling velocities, and that partially unfolded chains form geometrically coherent segments. | result 0.98 | 0.95 | 0.07 | asserted 1.00 |
| s9 | FoldScape thus combines broad coverage of folding stages with stable reconstruction accuracy, without per-protein simulation. | interpretation 0.86 | 0.92 | 0.83 | asserted 1.00 |
What was examined, and what was not
- arXiv listing abstract v1
- full text
- figures and tables
- references and citing sentences
- relations to other papers
- P abstract
- S sections
- T body sentences
- C citations
- R cross-paper relations
Figures
Figures reproduced from arXiv:2611.04287 by the authors.
References
Discussion
Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.