acceptodds
Under review as a conference paper at ICLR 2027

BLOCS: Scalable Block-Conditional Generative Proposals with Exact Local Metropolis–Hastings Correction

Abstract

Generative molecular samplers propose large conformational changes, but full-system statistical correction becomes inefficient as proposal mismatch accumulates across molecular degrees of freedom. We introduce BLOCS, a transferable block-conditional proposal model for accelerating local equilibrium transitions. Each learned torsional block receives an exact full-system Metropolis–Hastings correction using the molecular energy, coordinate measure, and forward/reverse proposal densities. A state-dependent Hastings-score diagnostic connects correction difficulty to stationary proposal-pair overlap. In controlled scaling, global-proposal acceptance falls from at to at , whereas fixed-scope block correction remains usable; complete-pool scans show that moved dimension has a substantially larger effect on acceptance than total protein size. The residual autoregressive proposal, trained on proteins with , transfers zero-shot to proteins with up to 250 residues. Under the canonical , fixed setting, end-to-end ESS/GPU-s improves by – over Rotamer MH across six proteins, with a median gain of . On the five-system common-pool thermodynamic benchmark, gains over a matched-data Robin-style autoregressive control are –. In a paired protein-scale exact-composition benchmark with Metropolized dynamics, the gains remain – over Rotamer MH and – over the matched learned control. Acceptance calibration on frozen primary and prospective pools reaches median relative error.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular sampling, generative models, Markov chain Monte Carlo, Metropolis-Hastings, block-conditional proposals, protein conformational dynamics, zero-shot transfer

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_ya2JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper speeds up sampling of protein side-chain conformations. A learned autoregressive model, conditioned on the surrounding 3D structure, proposes changes to a block of side-chain angles (12 in the headline setting) drawn from a fixed, state-independent pool. Each move is accepted or rejected with a Metropolis–Hastings test on the full-system energy. These block moves are interleaved with ordinary Langevin MD. The model is trained on small proteins and applied unchanged to larger ones. Experiments on five proteins compare free-energy accuracy and sampling efficiency (ESS per GPU-second) against a rotamer-library sampler and a retrained learned sampler.

Critical strengths

As described (Table 23), the block moves are a valid Metropolis–Hastings step, so they preserve the right equilibrium distribution. The ordinary MD steps between block moves are not Metropolized in the headline runs (Table 24), but the authors measure that bias (Table 44) and repeat the comparison with Metropolized dynamics. The learned-baseline comparison is fair-minded. The authors retrain a Robin-style autoregressive proposal (their own reimplementation) on the same data and training budget, and still beat it by about 1.6–1.9× in ESS/GPU-s, including with Metropolized dynamics on three systems (Table 45).

Critical weaknesses
  1. The global-versus-local acceptance comparison behind the scaling claim can't be checked, and the paper's figures disagree on it. The abstract and Sec. 5.1 rest it on Fig. 2a. There, acceptance of a "global proposal" falls from 2×10⁻³ (N=20) to 3×10⁻⁹ (N=76), while BLOCS stays flat near 0.42. I could not find a description of that global proposal, or of which systems make up the ladder, beyond a one-line mention of a "global-reweighting stress test" (App. D.5). The paper's own Fig. 7e, over the same range of N, shows BLOCS dropping to about 0.16–0.17 at N=56–76. That is consistent with the real-protein acceptances in Table 36 (about 0.16 for Protein G and ubiquitin). Fig. 2a's flat 0.418 matches no protein-scale table. The direction (local beats global as size grows) is the expected one, and the complete-pool scan (Table 51) supports it. The problem is the specific headline curve, and no trained global sampler is compared on the scaling axis.
  2. Several efficiency numbers are hard to trace, and the classical baseline isn't tuned symmetrically. The BLOCS/rotamer speed-up is 3.0× in Table 16 (13-observable common pool), but 2.0× in Tables 12/30, where the pool is not stated. Rotamer MH is run at BLOCS's chosen block size and MD interval (m=12, r=100), tuned only in temperature (Table 11), and checked at one other setting (Protein G, m=4, Table 50). It is never swept over m or r, while BLOCS's settings were. BLOCS's free-energy errors (0.033–0.055 k_BT) are also comparable to the reference uncertainty (±0.034–0.057, Table 37), so the ranking against baselines is clearer than the absolute accuracy.

Overall: If the global comparator is defined, Figs. 2a and 7e are reconciled, and rotamer MH is run at its own best settings, I would raise to a weak accept.

Questions for authors
  1. What exactly is the global proposal in Fig. 2a, and which systems make up that ladder? Why do Figs. 2a and 7e disagree at the same N?
  2. How does rotamer MH do at its own best block size and MD interval, beyond temperature and the single Protein G m=4 point in Table 50?
  3. Table 17 names the pools for Tables 1/16/31/37/51, but not for Tables 12/30 or Figs. 2a, 3b, 3c, 4c, 7e. Please add those.
  4. Fig. 3a colours Trp-cage and Villin as training-range systems, while Table 14 lists them as held-out. Which is correct?
5 comments
AuthorsOpenReview ↗

Response to Reviewer ya2J: Condition 1, global comparator and Fig. 2a vs. Fig. 7e

Thank you for the precise conditions. We uploaded a revised PDF (changes in blue) and address each condition below with full results.

What the submitted Fig. 2a showed. The submitted "BLOCS" curve came from a controlled embedded-block ladder (CTRL-EMBED), not from real protein pools. It used Trp-cage, Villin HP35, Protein G and ubiquitin (N=20,36,56,76N=20,36,56,76; the tick "35" was a labelling error). In each system BLOCS moved one fixed two-torsion block, the alanine-dipeptide block of App. F.4, grafted as a solvent-exposed pendant more than 9 Å from every protein atom. Its context, and hence its acceptance, is independent of NN by construction, and 0.418 is the isolated-block value (Table 28). The global comparator was a per-system neural-spline flow (8 coupling layers, 4.1M parameters) over all eligible side-chain torsions, fitted by forward KL on 50 ns of that system's MD, with one full-system MH step per draw. Its Rao–Blackwellised acceptance α^=n−1∑imin⁡(1,eSi)\hat\alpha=n^{-1}\sum_i\min(1,e^{S_i}) falls from 2.0×10−32.0\times10^{-3} at N=20N=20 to 3.1×10−93.1\times10^{-9} at N=76N=76. The panel now appears in App. D.5 under the CTRL-EMBED label.

New real-protein ladder (main-text Fig. 2a,b). We ran six held-out proteins (FIXED6-M12) with three proposals, each with one full-system MH step and r=100r=100. (a) In the all-block joint proposal, the frozen BLOCS network proposes all eligible side-chain torsions at once, so proposal scope is the only variable that differs from (c). (b) The trained global proposal is a transferable autoregressive model over all torsions, retrained on our corpus with the same 39.4 GPU-h budget. (c) BLOCS uses m=12m=12. For (a) and (b), acceptance is the RB estimate from 2×1052\times10^5 attempts per seed (5 seeds), with realised accepts from seed 1 in the last column:

SystemNN(a) Joint(b) Global(c) BLOCSESS/GPU-s (a) / (b) / (c)Accepts (a) / (b)
Chignolin103.6×10−23.6\times10^{-2}6.2×10−26.2\times10^{-2}0.2960.384 / 0.517 / 2.1987214 / 12388
Trp-cage202.3×10−32.3\times10^{-3}5.4×10−35.4\times10^{-3}0.2700.071 / 0.112 / 1.891461 / 1080
Villin HP35363.9×10−53.9\times10^{-5}1.8×10−41.8\times10^{-4}0.2380.0093 / 0.0128 / 1.6738 / 36
Protein G565.8×10−75.8\times10^{-7}4.1×10−64.1\times10^{-6}0.1540.0041 / 0.0043 / 1.4660 / 1
Ubiquitin764.4×10−94.4\times10^{-9}6.9×10−86.9\times10^{-8}0.1460.0033 / 0.0034 / 1.2040 / 0
Lysozyme1291.7×10−141.7\times10^{-14}2.2×10−122.2\times10^{-12}0.1180.0022 / 0.0022 / 0.9440 / 0

Changing only the proposal scope moves acceptance at N=129N=129 from 1.7×10−141.7\times10^{-14} to 0.118. The trained global proposal improves on the joint proposal by about 10× and collapses in the same way, and at N≥56N\ge56 both are at or below MD alone on Protein G (0.0044 ESS/GPU-s). BLOCS acceptance falls by 23% from Protein G to lysozyme, consistent with the 17.4% size penalty at fixed mm in Table 51 and with the ≈0.16 values the reviewer cited.

Reconciling Fig. 7e. Fig. 7e comes from a separate legacy run whose training set contained Chignolin, Trp-cage and Villin HP35, and it uses a different proposal from Fig. 2a. Its local curve is single-χ1\chi_1 acceptance: 0.50–0.54 on those three training proteins and 0.191 / 0.157 on held-out Protein G / ubiquitin. Its global curve is a weight-overlap estimate for a product-form global proposal, which is zero at N=56,76N=56,76. The drop between N=36N=36 and 56 coincides with that run's train/held-out boundary, so Fig. 7e does not isolate size dependence; with every protein held out, size dependence is measured by the new Fig. 2a and Table 51. Fig. 7 and App. D.5 now label the legacy run, its training set and its estimator.

AuthorsOpenReview ↗

Response to Reviewer ya2J : Condition 2, Rotamer MH at its own best settings; 3.0× vs. 2.0×

Symmetric tuning. We swept Rotamer MH over m∈{1,2,4,8,12,16}×r∈{10,25,50,100,200}m\in\{1,2,4,8,12,16\}\times r\in\{10,25,50,100,200\} (30 cells) on Protein G, ubiquitin and lysozyme, re-tuning temperature at every cell (5 seeds per cell), and swept BLOCS over m∈{2,4,8,12,16}×r∈{25,50,100,200}m\in\{2,4,8,12,16\}\times r\in\{25,50,100,200\}. Each mm uses the THERMO-COMMON construction rule at that mm, so the (12,100)(12,100) cell is the Table 1 run. Rotamer MH prefers small blocks and shorter MD intervals, and BLOCS prefers (8,100)(8,100). Best cells (ESS/GPU-s):

SystemRotamer best (m,r)(m,r)RotamerBLOCS best (m,r)(m,r)BLOCSBest-vs-bestCanonical (12,100)
Protein G(4, 50)0.641(8, 100)1.5382.40×3.01×
Ubiquitin(4, 50)0.552(8, 100)1.2712.30×2.86×
Lysozyme(2, 50)0.468(8, 100)0.9892.11×2.69×
Protein G, complete eligible pool(2, 50)0.551(4, 100)1.1292.05×2.16×
Protein G, Metropolized dynamics(4, 50)0.303(8, 100)0.6722.22×2.79×

Symmetric tuning lowers the gain (e.g. Protein G from 3.01× to 2.40×) and leaves it above 2× in every setting, including exact Metropolized composition. Tables 59 and 60 give the full grids, and the abstract and Sec. 5.4 now report both the canonical and the best-vs-best numbers.

3.0× (Table 16) vs. 2.0× (Tables 12/30). The two numbers come from different block populations, now named in every caption. THERMO-COMMON-M12 (Tables 1, 16) uses one common pool per system, with ESS as the median over its 13 locked observables. EXT-UNION-M12 (Tables 12, 20, 29, 30) uses, per system, the union of all frozen observable-centred candidate lists, run in the external-wrapper path required by the public checkpoints. With eEe_E the ESS per 10510^5 target evaluations and vEv_E the evaluations per GPU-s, the decomposition ESS/GPU-s=eE vE/105\mathrm{ESS/GPU\text{-}s}=e_E\,v_E/10^5 (Eq. 8) gives:

PopulationRotamer eEe_E / vEv_EBLOCS eEe_E / vEv_EStatistical × executionRatio
THERMO-COMMON-M12281 / 167.31216 / 116.04.33 × 0.6931.410/0.470 = 3.00×
EXT-UNION-M12312 / 167.3912 / 114.62.92 × 0.6851.045/0.522 = 2.00×

The execution factor is the same in both populations (0.693 vs 0.685), and the difference lies in the statistical factor. Every block in an observable-centred list contains a slow residue, so Rotamer library jumps reach the slow χ2\chi_2 rotamers directly (eEe_E 281 → 312). The same lists over-represent buried, tightly packed blocks, where BLOCS acceptance is lower (eEe_E 1216 → 912). On the common population, Rotamer MH, the matched AR and BLOCS reach 0.470, 0.805 and 1.410 ESS/GPU-s, now also given in the Table 12 caption.

AuthorsOpenReview ↗

Response to Reviewer ya2J : free-energy precision, pools, Q5

Errors vs. reference uncertainty. We ran two experiments. First, we extended the references by doubling the umbrella windows and running each window 4× longer:

SystemObservableΔF⋆\Delta F^\star (orig. ± σ)σ extendedShiftBLOCS ϵF/σext\epsilon_F/\sigma_{\rm ext}
Trp-cageARG16 χ2\chi_20.87 ± 0.0340.0160.0031.9
Villin HP35PHE11 χ2\chi_2−1.26 ± 0.0410.0190.0041.9
Protein GTYR33 χ2\chi_2−2.08 ± 0.0470.0220.0021.9
UbiquitinILE44 χ2\chi_20.73 ± 0.0510.0240.0041.8
LysozymeTRP62 χ2\chi_2−1.51 ± 0.0570.0270.0041.9

σ is halved and every reference value moves by less than 0.005 kBT0.005\,k_BT, so all errors in Tables 1 and 16 are unchanged at the reported precision; over all 13 observables, re-scoring changes no system median by more than 0.003 kBT0.003\,k_BT and changes no ranking. Second, we measured error against compute on Protein G with Metropolized dynamics (mean ∣ΔF−ΔF⋆∣/kBT|\Delta F-\Delta F^\star|/k_BT over 3 observables × 5 seeds):

GPU-h124.06.7712.324
Rotamer MH0.610.440.310.2360.2250.0980.071
Matched AR0.290.210.1600.1000.0940.0740.058
BLOCS0.210.1420.0990.0560.0520.0410.031

BLOCS has the lowest error at every budget; its gap to Rotamer MH is 0.40 kBT0.40\,k_BT at 1 GPU-h and 0.04 kBT0.04\,k_BT at 24 GPU-h. BLOCS falls below 2σext2\sigma_{\rm ext} (0.044 kBT0.044\,k_BT for TYR33) after about 10 GPU-h, so its remaining error is resolved above reference noise. The 4.0, 6.7 and 12.3 GPU-h columns are the median times to 0.10 kBT0.10\,k_BT in Table 45, and the 7 GPU-h column reproduces its error column; median first-passage time and checkpoint mean are different statistics.

Pools. Table 17 now defines all populations, and every caption names its own. Tables 12 and 30 use EXT-UNION-M12. Fig. 2a uses FIXED6-M12 (the submitted panel is CTRL-EMBED, App. D.5). Fig. 3b uses CORE-MULTIDOF (buried core residues, relative SASA < 10%, m∈{1,2,4}m\in\{1,2,4\}). Fig. 3c uses THERMO-COMMON-M$m$ with each method at its best (m,r)(m,r). Fig. 4c is now an illustrative event model, labelled as such. Fig. 7e uses LEGACY-PRODUCT-GLOBAL.

Q5 (Fig. 3a vs. Table 14). Table 14 is correct: Trp-cage and Villin HP35 are held out from the model behind Fig. 3a. The submitted bars come from the G1 checkpoint: the Trp-cage, Protein G, ubiquitin and lysozyme bars are the OBS-SPEC-CONTROL acceptances of Tables 36/38 (0.270, 0.159, 0.159, 0.141), and the BBA bar is the reference row of Table 43 (0.110). The "train" legend and shading were wrong for these values; they followed the training set of a legacy run (Chignolin, Trp-cage, Villin HP35) that produced only the submitted Fig. 7b and 7e. An earlier pilot produced the submitted Fig. 7c,d, which marked Protein G and ubiquitin as training proteins; those two panels are removed. No main-text result uses either run. In the revision, Fig. 3a uses the common FIXED6-M12 population with every protein marked held out and only the training size range shaded (Chignolin 0.296, Trp-cage 0.270, Villin 0.238, Protein G 0.154, ubiquitin 0.146, lysozyme 0.118), and the submitted held-out-structure bars (BBA, WW, Protein G, Protein L, ubiquitin, lysozyme) are kept, correctly labelled, as Fig. 7d.

Reviewer_Jx2iOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces BLOCKS, that applies MH transition kernels at local torsional blocks but corrects it with global density values. The method is very interesting and insightful, it demonstrates applying such transition yields desirable computational gains over regular MD and other AI-based global samplers at similar compute budgets.

Critical strengths

The paper provides useful insights in utilizing local transitions for exploring global energy surfaces of a biological system. The method yields impressive computational gains over baseline. The method also demonstrates one-shot transfer to larger protein by just training on smaller systems, which is very useful takeaway for simulating for real world biomolecular modelling tasks.

Critical weaknesses

The main weakness of this paper lies in its presentation. The writing is dense and reads as though it was largely AI-generated, which makes it difficult to follow. The methods section introduces numerous symbols and notation without defining or explaining them. The experiments and results section consists largely of short, disconnected statements that are heavy on technical detail but lack continuity, explanation, and reasoning. Furthermore, the authors do not discuss the key takeaways in their results sections.

Questions for authors

To improve the paper’s readability, I encourage the authors to clarify the following points:

  1. There are several undefined symbols in the equations in the methods section. For ex. mm, rr, CB(ϵ)C_B(\epsilon), please provide adequate definitions/context for them

  2. The training protocol is unclear. What was the loss function used and how does this relate to the forward/reverse KL? Do you train to predict the torsions at the next time step?

  3. The inference procedure needs further explanation. At what interval do you alternate between standard Langevin steps and Metropolis–Hastings proposals? How does this procedure compare to those of the Boltzmann Generator baselines? How is ESS per GPU computed?

  4. Please state the key takeaway of each subsection in the experiments and results section. Expanding on these in clear, plain language, would help convey the paper’s main messages.

  5. The paper also does not have a reproducibility section, do the authors plan to release the code?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Jx2i : notation, training objective, inference

Thank you for the encouraging assessment of the method. We rewrote Sections 3–5 for readability; the changes are in blue in the revised PDF.

Notation. Every symbol, including those the reviewer listed, is now defined at first use, and App. A.0 adds a notation table. BB is the block of torsions and m=∣B∣m=|B| the number of torsions moved jointly. cB(ξ)c_B(\xi) is the context: all non-block heavy atoms, backbone included, within rc=9r_c=9 Å of the block anchors. The move touches no context atom, so cB(ξ′)=cB(ξ)c_B(\xi')=c_B(\xi) and the same context enters both densities. ξ=(ξB,ξBˉ)\xi=(\xi_B,\xi_{\bar B}) splits the internal coordinates into moved and frozen parts, and J(ξ)J(\xi) is the coordinate Jacobian (J′/J=1J'/J=1 for torsion-only moves). ρdk,μdk,κdk\rho_{dk},\mu_{dk},\kappa_{dk} are the weight, mean and concentration of the von-Mises mixture components. SθS_\theta is the log Hastings score, with α=min⁡(1,eSθ)\alpha=\min(1,e^{S_\theta}) and Dtr=−E[Sθ]D_{\rm tr}=-\mathbb{E}[S_\theta]. rr is the number of MD steps between block moves.

Training protocol. BLOCS is trained by maximum likelihood with teacher forcing on MD transition pairs: L(θ)=−E(xt,xt+τ) EB∑d=1mlog⁡qθ,d(Δχd∣Δχ<d,χB(t),cB(xt)),Δχ=wrap(χB(t+τ)−χB(t)).\mathcal L(\theta)=-\mathbb E_{(x_t,x_{t+\tau})}\,\mathbb E_B\sum_{d=1}^{m}\log q_{\theta,d}\big(\Delta\chi_d\mid\Delta\chi_{<d},\chi_B(t),c_B(x_t)\big),\quad \Delta\chi=\mathrm{wrap}(\chi_B(t+\tau)-\chi_B(t)). This is the forward KL from the empirical transition distribution to the model, with no reverse-KL or energy term. The target is the torsion residual after lag τ=50\tau=50 ps, not the next integrator step. Forward-KL training gives MH acceptance 0.352, against 0.073 for reverse-KL training (Fig. 6c).

Inference. One cycle (Algorithm 1 in the revision):

repeat:
  r = 100 Langevin-middle steps (2 fs each, i.e. 0.2 ps)
  B ~ Uniform(frozen pool)                       # state-independent
  sample Δχ_1..Δχ_m autoregressively from the von-Mises mixtures given (χ_B, c_B)
  χ' = wrap(χ + Δχ); rotate distal side-chain subtrees
  S = -β[U(x') - U(x)] + log J'/J + log q(-Δχ | χ', c_B) - log q(Δχ | χ, c_B)
  accept x' with probability min(1, e^S)         # full-system energy

MD and MH alternate every r=100r=100 steps, and Tables 59–60 sweep r∈{10,…,200}r\in\{10,\dots,200\}. Boltzmann Generator baselines propose all coordinates at once and correct the whole sample, by importance weights or by an independence MH or SMC step. BLOCS proposes a small block conditioned on the current state, corrects only that move, and interleaves dynamics. This is why global acceptance collapses with size while BLOCS stays usable: on lysozyme (N=129N=129), proposing all torsions jointly gives acceptance 1.7×10−141.7\times10^{-14}, while 12-torsion blocks give 0.118.

ESS per GPU-second. ESS =N/τint=N/\tau_{\rm int} per observable, with τint\tau_{\rm int} from FFT autocorrelation and initial-positive-sequence truncation, taking the median over a system's observables. It is divided by end-to-end GPU-seconds, which include proposal inference, every energy evaluation (rejections included) and all MD (Sec. 4.2).

AuthorsOpenReview ↗

Response to Reviewer Jx2i: takeaways and reproducibility

Key takeaways. Each subsection of Sec. 5 now opens with a bold takeaway, followed by its evidence.

Sec. 5.1: Limiting the scope of each correction keeps MH usable as proteins grow. On six held-out proteins, proposing all torsions jointly drops acceptance to 1.7×10−141.7\times10^{-14} at N=129N=129, while 12-torsion blocks keep 0.118. Moving from m=2m=2 to 12 costs 71–73% of acceptance, whereas growing the protein from N=56N=56 to 129 costs 10–17%.

Sec. 5.2: One model trained on systems of at most 36 residues transfers to proteins of up to 250 residues. Transfer holds without crops of large proteins, under increasing environment novelty, and with a 4.5 Å context.

Sec. 5.3: Faster local barrier crossing becomes lower free-energy error at fixed compute. BLOCS has the lowest error at every budget, and accepted moves persist under MD.

Sec. 5.4: The gain persists against a data-matched learned proposal, a best-tuned classical sampler and exactly composed dynamics. BLOCS is 1.67–1.91× faster than the matched AR, 2.11–2.40× faster than best-tuned Rotamer MH, and 2.79–2.87× faster than Rotamer MH with Metropolized MD.

Reproducibility. The revision adds a Reproducibility Statement, and the code will be released upon acceptance. The paper specifies one sampling cycle (Algorithm 1), the corpus construction and training (Algorithm 2), all symbols (App. A.0), the corpus, split and homology control (App. S), and the architecture, partitions, block populations, transition, physical target and estimators (Tables 10, 14, 17, 23, 24 and 35), together with the frozen artifact fields behind the headline tables (Table 22).

I thank the authors for responding to my comments

Dear Reviewer Jx2i, Thank you for reading our revision and for raising your score. Your comments on notation, the training objective, the inference procedure, and section takeaways substantially improved the clarity of the paper. We will continue to polish the presentation in the final version. Best regards, The Authors

Reviewer_R6NzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces BLOCS, a transferable proposal for Metropolis–Hastings sampling of protein conformations. The authors train an autoregressive model that takes a small block of torsion angles along with the local 3D environment as an input and outputs updates to the input angles. These updates are based on a Metropolis-Hastings step using the full system. The authors claim that this construction helps avoid the collapse of global proposals while learning transferable updates on larger systems.

Critical strengths
  1. The separation of local proposals from global corrections is a sound framework to avoid collapse of global proposals.
  2. The compute time is very carefully accounted for across all experiments.
  3. The authors also demonstrate multiple dynamics validation metrics like S2, which further bolsters their claim.
Critical weaknesses

Dataset poorly specified

The authors mention that their training set comprised of proteins with at most 36 residues. However, in the later sections, they also specify that their training set comes from ATLAS and MdCATH. The smaller protein in MdCath has amino acid sequence length of 50. This makes the data set curation quite dubious.

Torsion update physicality is underspecified

After applying 100 steps of MD, it is not clear if the force field cancels out / "undoes" the sidechain updates. Or if the sidechain updates applied to the model are physical from a stereochemical perspective.

Lack of landscape coverage metrics

For smaller proteins like Chingolin, TRP-cage, etc. the authors do not show any landscape coverage metrics like TICA/PCA projections against a reference, RMSD to native distributions, or fold/unfolded populations. This is important especially if the paper claims to sample from rare distributions

Miscellaneous

The framework is quite reliant on running MD after a sidechain update. Also, there is no specification of the difference between MD and long MD. How different are the final structures?

Questions for authors
  1. Can the authors clarify the structures from mdCath that were included in the training set and the size of the amino acid sequence for these proteins?
  2. How physical are the sidechain updates? Do these sidechain updates persist after multiple calls to the autoregressive model? Also, how physical are these torsion updates? is there a way they avoid clashes?
  3. How is the 3D environment encoded into the model? Which atoms are included, is the 9A cutoff a hard or soft threshold?
  4. Does the zero-shot transfer property apply when we use a smaller environment radius like 4-5A? Because with 36 residues training data and 9A environment, almost the entire protein will usually be included as an input.
  5. Can the model capture non local interaction / allosteric effects? The way the model is constructed, seems like this is not possible.
5 comments
AuthorsOpenReview ↗

Response to Reviewer R6Nz : dataset, MD vs. long MD

Thank you for the careful questions. We uploaded a revised PDF (changes in blue) and give every new result in full below.

Q1. Which mdCATH structures, and how large. The limit N≤36N\le36 applies to every simulated training system. We take 1,217 mdCATH domains (50–500 residues) and 403 ATLAS chains and cut them into crops of 18–36 residues, and add 384 complete small PDB proteins:

SourceSystemsNNBlock examples
Complete small PDB proteins38410–361.02M
Crops of 1,217 mdCATH domains2,43118–365.95M
Crops of 403 ATLAS chains51518–361.34M
Total3,330≤36\le368.31M

A crop is a seed residue plus its nearest residues by Cβ distance, up to 36 in total, capped (ACE/NME) at chain breaks, with crop-boundary backbone heavy atoms restrained. The mdCATH and ATLAS trajectories are not used: these databases supply structures only, and every system is re-simulated in ff14SB/GBn2 at 300 K (3 × 20 ns; 199.8 µs; 1,262 GPU-h), matching the production force field. Every MMseqs2 30%-identity cluster or CATH superfamily that contains an evaluation protein is removed before cropping, and nearest-training sequence identity is 9.7–20.0% for all evaluation proteins. The full list of training systems will be released with the code upon acceptance (App. S, Table 57).

"MD" vs. "long MD". "MD" is unbiased Langevin dynamics given the same GPU budget as the method under comparison. "Long MD" is the independent reference ensemble, e.g. ubiquitin 12 × 1 µs (about 410 GPU-h), Chignolin 8 × 5 µs and Trp-cage 8 × 10 µs. Both definitions are now in Sec. 4.2.

How different are the final structures. Final BLOCS ensembles match long MD closely: χ1/χ2\chi_1/\chi_2 JS divergence is 0.012 / 0.019, and mean backbone RMSD to native is 1.21 Å against 1.18 Å for long MD (KS 0.034).

Reliance on MD. The interleaved dynamics samples the backbone and collective motions, and BLOCS accelerates side chains, which is the stated scope. The two legs work together: block moves alone give backbone φ/ψ JS 0.214, and composing them with MD lowers it to 0.021 while side-chain χ\chi JS stays at 0.019 (Table 26). The dynamics leg can also be made exact: with Metropolized dynamics (GHMC), BLOCS keeps 2.79–2.87× over Rotamer MH (Table 2).

AuthorsOpenReview ↗

Response to Reviewer R6Nz : physicality, persistence, 3D encoding

Q2a. Are the side-chain updates physical, and do they avoid clashes? By construction, moves rotate distal subtrees about rotatable bonds, so bond lengths, bond angles and chirality are exactly preserved (Jacobian residual ≤ 2.3×10−92.3\times10^{-9}), and ring and closure torsions, including proline, are ineligible. We also audited 2.4×1052.4\times10^5 attempts × 5 seeds on Protein G, counting a clash as any heavy-atom overlap above 0.4 Å:

ProposalαClash (proposed)Clash (accepted)Median βΔU\beta\Delta U (proposed)KS(accepted UU, equilibrium UU)
Uniform torsion jump0.02661.8%0.00%+38.20.009
Rotamer library + MH0.05824.6%0.00%+9.70.010
BLOCS0.1548.3%0.00%+2.10.008

BLOCS proposes clashing states 7.4× less often than uniform jumps, which is where its higher acceptance comes from. The full-energy MH test removes every clash, accepted energies match the equilibrium distribution, and rotamer outliers occur in 0.7% of accepted states against 0.6% of long-MD frames.

Q2b. Does MD undo the updates, and do they persist? We measured the probability that the new rotamer survives a given MD time after an accepted BLOCS move, and compared it with survival from equilibrium MD frames in the same rotamer (Protein G, all moved χ\chi):

MD time after move0.2 ps (=r=r)2 ps20 ps200 ps
After accepted BLOCS move0.9910.9570.8020.412
Equilibrium MD (reference)0.9930.9610.8110.425

MD relaxes accepted states no faster than equilibrium states, so the force field does not undo the moves. Because both kernels preserve the Boltzmann distribution, a later return to the old rotamer is ordinary equilibrium dynamics. The net effect is faster mixing: τint\tau_{\rm int} falls from 81.6 (MD) to 9.2 (BLOCS), and cross-mode transitions at matched GPU time rise from 31 to 486 (Tables 72–73).

Q3. How the 3D environment is encoded. The context contains all non-block heavy atoms, backbone included, within 9 Å of the block anchor atoms, each with its atom type, residue type and position in the anchor frame. Membership uses a hard 9 Å cutoff, and each pairwise distance is expanded in 32 radial basis functions multiplied by a cosine envelope that vanishes at 9 Å, so atoms enter and leave the context smoothly. A 4-layer SE(3)-invariant message-passing network aggregates the context, which has on average 238 atoms per m=12m=12 block in training, 291 on Protein G and 318 on lysozyme. The move touches no context atom, so cB(ξ′)=cB(ξ)c_B(\xi')=c_B(\xi) and the same context enters the forward and reverse densities (Sec. 3.3, Fig. 8a).

AuthorsOpenReview ↗

Response to Reviewer R6Nz : smaller radius, non-local effects, landscape coverage

Q4. Zero-shot transfer with a 4–5 Å environment. We measured context coverage and retrained BLOCS at 4.5 and 6 Å with the same budget. Coverage is the fraction of system heavy atoms inside the context of an m=12m=12 block:

Train (complete)Train (crops)Trp-cageVillinProtein GUbiquitinLysozymeTIM (N=250N=250)
9 Å0.710.660.740.620.410.330.210.12
4.5 Å0.290.260.310.240.140.110.070.04

Efficiency after retraining (ESS/GPU-s; retention = lysozyme/Trp-cage):

rcr_cTrp-cageVillinProtein GUbiquitinLysozymeRetention
9 Å1.8911.6731.4661.2040.9440.499
6 Å1.0400.9200.8030.6260.4810.463
4.5 Å0.5110.4430.3810.3010.2270.444

The reviewer is right that a 9 Å context covers most of a training system (0.71). At test time it covers only 0.21 of lysozyme and 0.12 of a 250-residue protein, so the model already extrapolates from partial context. Zero-shot transfer holds at every radius: retention changes only from 0.499 to 0.444 while training coverage drops from about 0.7 to about 0.3. What shrinks at a small radius is absolute efficiency; at 4.5 Å BLOCS falls below best-tuned Rotamer MH (Protein G 0.381 vs 0.641), so local packing information carries the gain (Table 68, Fig. 3d,e).

Q5. Non-local and allosteric effects. Correctness does not depend on locality. The MH test uses the full-system energy, so every non-local term enters each acceptance exactly, and locality affects only proposal efficiency. Decomposing ΔU=ΔUin+ΔUout\Delta U=\Delta U_{\rm in}+\Delta U_{\rm out} over attempted moves on Protein G, ubiquitin and lysozyme:

BlocksMedian ∥βΔUout∥\|\beta\Delta U_{\rm out}\|95th pct.99th pct.MH decisions changed if droppedcorr(out-term, accept)
All0.0310.190.440.38%0.04
Containing Lys/Arg/Asp/Glu0.0720.410.931.2%0.07

Charged residues carry larger out-of-cutoff terms through long-range GB electrostatics, and their moves are still corrected exactly. BLOCS proposals do not model allosteric coupling and do not accelerate large-scale cooperative transitions; such motions are sampled by the interleaved dynamics, as now stated in the Limitations.

Landscape coverage (Chignolin, Trp-cage). TICA (lag 5 ns, Cα pair distances) is fitted on long reference MD. MD and BLOCS+MD each receive 24 GPU-h over 5 seeds; "folded" means RMSD to native below 2 Å, and W1W_1 is the Wasserstein-1 distance between RMSD-to-native distributions.

SystemMethodTICA JSRMSD W1W_1 (Å)Folded populationFold/unfold events
ChignolinLong MD––0.583 ± 0.021–
MD0.0640.310.551 ± 0.04711
BLOCS+MD0.0390.180.574 ± 0.03114
Trp-cageLong MD––0.694 ± 0.018–
MD0.0770.360.732 ± 0.0617
BLOCS+MD0.0460.210.701 ± 0.0399

Coverage improves about 1.6-fold and folded populations move closer to the reference, because faster side-chain relaxation lets the backbone relax faster. Folding events increase only modestly, since block moves leave the backbone fixed, consistent with our scope of side-chain equilibrium sampling (Table 65).

Reviewer_bU7YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BLOCS learns a transferable proposal for blocks of protein side-chain torsions. A residual autoregressive model uses the current torsions and nearby three-dimensional context to propose an action, which is accepted using the full molecular energy and the forward/reverse proposal densities. The paper also studies the relationship between acceptance and proposal-pair overlap, and tests transfer, block size, context, thermodynamic accuracy, and computational efficiency.

Critical strengths
  1. The detailed-balance argument seems sound, and I like that the authors carefully distinguish the exactness of the learned actions from potential bias introduced through the stochastic dynamics from Langevin. I also appreciate that the authors are clear about the scope of their work, focusing on local side-chain sampling rather than making broader claims about protein folding.

  2. I found the experimental evaluation relatively convincing overall. The authors go beyond simply reporting acceptance rates and include several thoughtful comparisons and ablations. I appreciate the fair comparison with an autoregressive baseline, the attention to actual computational cost, and the experiments showing why spatial context and state conditioning matter. The reported efficiency gains of around upwards of 1.5x are encouraging, especially alongside improvements in sampling accuracy.

Critical weaknesses
  1. My main concern is that many important details about training and evaluation remain quite unclear, making it difficult to fully assess the claimed generalization. In particular, I was confused by the use of mdCATH given the stated training protein sizes and force-field setup. I would also like to better understand how the training pairs are constructed and how the data is split for improved clarity (and reproducibility). Similarly, the adaptations of Robin and Prose, as well as the Metropolized dynamics control, need more explanation to judge whether the comparisons are fair and reproducible. These details matter because the paper's main contribution depends not just on theoretical correctness, but also on demonstrating that the learned proposal generalizes effectively.

  2. I also have some reservations about the broader efficiency and sampling-quality claims. While I find the matched autoregressive comparison somewhat encouraging, I am less convinced by some of the comparisons with other methods, where the evaluation procedures and adaptations are not sufficiently clear. I would especially like to see a more complete accounting of training costs and when those costs can realistically be amortized. The proposed diagnostic is interesting, but I am not yet convinced that it offers a clear advantage over simpler alternatives.

Questions for authors
  1. Small-to-large generalization: Could the authors repeat the BLOCS versus matched-AR comparison using only training proteins with at most 36 residues, excluding fragments from larger proteins? I think this would help clarify whether the method truly generalizes from small proteins to larger ones, or mainly benefits from local environments already seen during training. It would be useful to compare sampling accuracy and efficiency under the same training budget.

  2. Baseline comparisons: Could the authors provide a more controlled comparison between BLOCS, matched AR, and the adapted learned baselines on a small system with a reliable reference distribution? I would be particularly interested in whether these methods recover the same joint torsion distributions and how their efficiency compares at similar computational costs.

  3. Protein-level quality and computational cost: I would like to see how BLOCS compares with recent methods such as PLaTITO, MarS-FM, and/or STAR-MD on a shared protein benchmark. Comparing sampling quality and total computational cost, including training when appropriate, would help establish where BLOCS is most useful. Even if its advantages are mainly limited to side-chain sampling, I think that would still be a valuable contribution, but the practical scope should be clearer.

  4. Value of the proposed diagnostic: Could the authors compare the Gaussian-overlap diagnostic with a simpler approach that uses direct acceptance rates and the same utility model? I think this would help determine whether the Gaussian approximation itself provides a meaningful benefit. It would also be useful to compare how reliably each approach selects good sampling settings and how much computational effort that selection requires.

5 comments
AuthorsOpenReview ↗

Response to Reviewer bU7Y : training data, split, and baselines

Thank you for the detailed review. We uploaded a revised PDF (changes in blue) and give every new result in full below.

mdCATH vs. "≤36 residues". The limit N≤36N\le36 applies to every simulated training system, not to the protein it is taken from. The corpus has 3,330 systems:

SourceSystemsNNBlock examples
Complete small PDB proteins38410–361.02M
Crops of 1,217 mdCATH domains (50–500 res.)2,43118–365.95M
Crops of 403 ATLAS chains51518–361.34M
Total3,330≤36\le368.31M

A crop is a seed residue plus its nearest residues by C$\beta$ distance, up to 36 in total, capped with ACE/NME at chain breaks, with crop-boundary backbone heavy atoms restrained (1.0 kcal mol$^{-1}A˚Å^{-2}$). mdCATH and ATLAS supply structures and starting conformations only. Every system is re-simulated under the production target, ff14SB/GBn2 at 300 K (3 replicas × 20 ns; 199.8 µs and 1,262 A100 GPU-h in total), so the training force field matches the target and no original mdCATH/ATLAS trajectory is used.

Training pairs. Each example is (χB(t), cB(xt), wrap(χB(t+τ)−χB(t)))(\chi_B(t),\,c_B(x_t),\,\mathrm{wrap}(\chi_B(t+\tau)-\chi_B(t))) at lag τ=50\tau=50 ps, with frames saved every 10 ps and blocks drawn by the benchmark pool rule. The model is trained by maximum likelihood (forward KL) with teacher forcing.

Split. Sequences are clustered with MMseqs2 at 30% identity. Every cluster or CATH superfamily that contains a validation, benchmark, prospective or extended-transfer protein is removed before cropping, and model selection uses an 18-protein validation partition. Nearest-training sequence identity / TM-score is 17.4% / 0.41 for Trp-cage, 16.1% / 0.39 for Villin, 14.6% / 0.35 for Protein G, 12.9% / 0.31 for ubiquitin and 9.7% / 0.275 for lysozyme (App. S, Table 57; Algorithm 2).

Robin-local+MH and PROSE-local+MH (frozen public checkpoints). The block is drawn from the same pool. The checkpoint proposes the block torsions, and its density over those torsions, with all other coordinates fixed, enters the same full-target MH ratio. Both models see a capped window of at most 8 residues centred on the block, matching their 2–8-residue training range; longer blocks are split into consecutive windows proposed in causal order. For Robin, the block torsions are placed last in the autoregressive order and the non-block window coordinates are teacher-forced, which gives the exact conditional. For PROSE, the coupling order transforms the block coordinates last, conditioned on the clamped non-block coordinates, so the log-determinant is tractable. In both cases the non-block window coordinates are identical before and after the move, so any normaliser over them cancels in the MH ratio, and atoms outside the window enter through the full-system energy.

Matched-data Robin-AR. This control (24.0M parameters) is retrained from scratch on the same proteins, trajectories, split, block examples and 39.4 GPU-h budget as BLOCS, with a 16-configuration hyperparameter search and the same validation criterion for each method.

Metropolized dynamics control. This is GHMC with Langevin-middle (BAOAB) splitting. The Ornstein–Uhlenbeck step is exact, and the deterministic B–A–A–B part of each step is one Metropolis proposal accepted with min⁡(1,e−βΔH)\min(1,e^{-\beta\Delta H}); on rejection, positions are restored and momenta flipped. We use SHAKE/RATTLE (tolerance 10−810^{-8}), Δt=2\Delta t=2 fs and γ=1\gamma=1 ps$^{-1}$. The dynamics leg is accepted 0.962 (Trp-cage), 0.951 (Protein G) and 0.938 (lysozyme) of the time. Under this exact composition, BLOCS keeps 2.792.79–2.87×2.87\times over Rotamer MH and 1.571.57–1.68×1.68\times over the matched AR (Table 2). App. E.2 and Table 46 give these specifications.

AuthorsOpenReview ↗

Response to Reviewer bU7Y Q1 small-to-large generalization, Q2 controlled joint-torsion comparison

Q1. Training only on complete N≤36N\le36 proteins. We retrained BLOCS and the matched AR on the 384 complete small proteins with every crop removed (1.02M examples, 12.3% of the corpus). Both use the same 39.4 GPU-h budget and validation set and are evaluated zero-shot (THERMO-COMMON-M12, 5 seeds). Entries are acceptance / ESS/GPU-s / median ϵF\epsilon_F (kBTk_BT):

Training dataModelProtein GUbiquitinLysozyme
Full corpusMatched AR0.107 / 0.805 / 0.1290.094 / 0.676 / 0.1410.079 / 0.565 / 0.155
Full corpusBLOCS0.154 / 1.466 / 0.0410.146 / 1.204 / 0.0440.118 / 0.944 / 0.055
Complete N≤36N\le36 onlyMatched AR0.088 / 0.662 / 0.1620.075 / 0.538 / 0.1810.061 / 0.431 / 0.204
Complete N≤36N\le36 onlyBLOCS0.137 / 1.271 / 0.0490.127 / 1.019 / 0.0540.099 / 0.772 / 0.069
BLOCS/AR, full → small-only1.82× → 1.92×1.78× → 1.89×1.67× → 1.79×

Removing the crops costs BLOCS 13–18% of its ESS/GPU-s and the matched AR 18–24%, so the BLOCS advantage grows when training is restricted to complete small proteins. Small-only BLOCS still exceeds best-tuned Rotamer MH by 1.98×, 1.85× and 1.65×.

Environment novelty. For each held-out block, novelty is the distance of its context embedding to the nearest training environment, normalised by the 95th percentile of validation-to-train distances; values above 1 are more novel than 95% of validation environments. Pooled over Protein G, ubiquitin and lysozyme:

QuintileQ1 (familiar)Q2Q3Q4Q5 (novel)
Median novelty0.420.610.780.971.31
BLOCS α, full corpus0.1600.1510.1400.1290.116
BLOCS α, small-only0.1490.1380.1240.1100.093

Acceptance declines smoothly and does not collapse in the most novel quintile, which contains blocks more novel than 95% of validation environments (27.6% of all held-out blocks have novelty above 1).

Q2. Controlled joint-torsion comparison. Chignolin and Trp-cage have long references (8 × 5 µs and 8 × 10 µs). Every method received the same 2.0 GPU-h over 5 seeds; distribution metrics are averaged over both systems and all residues:

Methodχ1\chi_1 JSχ2\chi_2 JSKL$(\chi_1,\chi_2)$|ΔMI$_{\chi\chi}$| (nats)Rot.-pop. RMSEESS/GPU-s Chig. / Trp
MD0.0780.0880.2120.0310.0940.0078 / 0.0061
Rotamer MH0.0510.0790.1680.0390.0620.658 / 0.582
PROSE-local+MH0.0420.0630.1310.0240.0510.502 / 0.442
Robin-local+MH0.0360.0550.1120.0210.0450.652 / 0.576
Matched AR0.0270.0410.0810.0140.0341.139 / 0.990
BLOCS0.0140.0210.0380.0060.0182.198 / 1.891

All corrected samplers approach the same distribution. At equal cost, BLOCS is closest to the reference on the marginals, on the joint (χ1,χ2)(\chi_1,\chi_2) structure and on inter-residue coupling. Rotamer MH has the largest coupling error because its library proposes each residue independently (App. X, Table 69).

AuthorsOpenReview ↗

Response to Reviewer bU7Y : Q3 PLaTITO / MarS-FM / STAR-MD, Q4 diagnostic, training cost

Q3. Recent generative models of protein dynamics. From the published papers:

PLaTITOMarS-FMSTAR-MDBLOCS
RepresentationCα onlyresidue frames + 7 torsions per residueresidue frames + side-chain torsionsall side-chain torsions, full-system energy
Correction to force-field Boltzmannnone; the authors note no formal detailed-balance guaranteenone describednone describedblock MH; exact with Metropolized MD
Training datamdCATH, 4,471 domains (≤200 res.)mdCATH, 5,398 domains (≤500 res.), 320–450 KATLAS, 100 nsff14SB/GBn2 systems, N≤36N\le36

These models emulate long-lag transitions quickly, while BLOCS gives exactly corrected all-atom side-chain sampling whose error keeps falling with compute. The two are complementary: a backbone-level generator can supply large moves while BLOCS corrects side chains. On a shared benchmark, we added MarS-FM to the ubiquitin NMR evaluation (10⁴ frames from MSM-state transitions, side chains as generated):

MDRotamer MHBLOCSMarS-FMLong MD
Methyl S2S^2 RMSD ↓0.0910.0640.0380.0610.035
Methyl S2S^2 Pearson ↑0.780.860.940.880.95
χ1\chi_1 rot.-pop. RMSD ↓0.1020.0730.0410.0660.039
GPU-h to within 10% of ref. S2S^296.012.84.2not reached–

MarS-FM generates its ensemble in 0.6 GPU-h and plateaus at S2S^2 RMSD 0.061, whereas BLOCS reaches the long-MD level within 4.2 GPU-h (Tables 41, 71).

Q4. Gaussian overlap vs. direct acceptance with the same utility. Direct acceptance is computed on the same attempted pairs at no extra cost. On four prospective systems with m∈{2,4,8,12}m\in\{2,4,8,12\} and a 0.65 GPU-h pilot (entries: selected mm / efficiency relative to the best cell):

StrategyBPTIBarnaseRNase ACalmodulinMean regretSpearman ρ
Direct acceptance ranking2 / .7362 / .7122 / .7672 / .68427.5%−0.18
Short-run ESS8 / 1.0004 / .9458 / 1.0004 / .9303.1%0.88
Direct acceptance + utility4 / .9678 / 1.0004 / .9004 / .9305.1%0.86
Gaussian overlap + utility4 / .9678 / 1.0004 / .9008 / 1.0003.3%0.91

At this budget the utility-aware strategies are within about 2% of each other, and direct acceptance predicts acceptance slightly better (4.1% vs 4.8% error). The difference appears at smaller pilots (mean efficiency relative to the best cell):

Pilot GPU-h0.080.160.330.65
Gaussian + utility0.9310.9520.9620.967
Direct + utility0.8420.8970.9310.949
Short-run ESS0.6110.7440.8730.969

The revision therefore presents the Gaussian diagnostic as a low-cost, interpretable analysis tool, most useful when the pilot is 1/72 of a full 5.80 GPU-h sweep, that explains each selection through one quantity, DtrD_{\rm tr} (Tables 54–55).

Training cost and amortisation. Data generation costs 1,262 GPU-h and BLOCS training 39.4 GPU-h; the matched AR uses the same data and budget, and Rotamer MH has no training cost. Against Rotamer MH, the time to 0.10 kBT0.10\,k_BT drops by 5.3, 8.3 and 11.1 GPU-h on Trp-cage, Protein G and lysozyme (Table 2). Break-even, in number of systems:

Single 0.10 kBTk_BT target10 GPU-h per system50100500
Training only4.82.20.430.220.04
Training + data15871.514.37.21.4

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.