acceptodds
Under review as a conference paper at ICLR 2027

NATIVE UPDATE ORIENTATION IN FROZEN PROTEIN STRUCTURE PREDICTORS

Abstract

We study whether adapter parameterization changes the effect of output-channel rotation under a fixed training budget. Our Factor head composes residue-level increments through a frozen predictor’s native pair-generating operators. The generic head G+ explicitly receives the interface anchors and has a free output layer, making its function class closed under output rotation. Our four-cell design trains Native and Rotated versions of each head separately from the same zero- residual function under paired settings. At identical inputs, interface anchors and parameters, orthogonal rotations preserve local writer spectra. The interaction Ψ is the difference between the two heads’ signed Native–Rotated score differences, rather than their absolute responses. On 192 prospectively locked new targets, fixed Protenix/ESMC models establish a Cα pair-lDDT interaction of +0.02515 [+0.01832, +0.03213]; OpenFold’s Fresh96 primary interaction remains unestab- lished. In an observed-panel follow-up, the Protenix interaction remains positive after a prespecified doubling of the update budget, while Factor’s own Native– Rotated score difference decreases. An amended OpenFold schedule study on observed targets and one fixed rotation finds the largest interaction at the lowest absolute quality. We then test whether a shared trainable orthogonal output map gives rotated Factor a larger compensation gain than Native Factor. Rotated Factor improves in both executions, but this gain difference was not re-established in the same-seed repeat. A prespecified reconstruction diagnostic did not establish the predicted ranking of post-training Native–Rotated score differences across eight tested rotations. Together, these findings motivate reporting absolute performance, signed within-head score differences and their interaction under explicit training budgets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein structure prediction, protein language models, parameter-efficient adaptation, factorized adapters, native update alignment

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hSwNOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates whether adapter parameterization affects sensitivity to output-channel orientation when adapting frozen protein structure predictors under a fixed training budget. The authors introduce Factor, which predicts residue-level factor increments and constructs pairwise residual updates through the predictor’s frozen native operators. They compare it with G+, an anchor-conditioned pairwise MLP with a free output layer. Both adapters are independently trained with native and fixed orthogonally rotated outputs, starting from zero residual updates. The central quantity is the difference between their signed native–rotated performance gaps. A positive interaction is established for a fixed Protenix/ESMC configuration on new evaluation targets and remains positive after doubling the training budget on the same panel; the corresponding new-target OpenFold test is inconclusive. Additional experiments examine injection schedules, output compensation, and reconstruction-based diagnostics, but leave the underlying mechanism unresolved.

Critical strengths
  1. A carefully constructed test of the interaction between parameterization and output orientation. The four-cell design independently trains native and rotated versions of both adapters from the same zero-residual function, with matched feature access and training settings. G+ explicitly receives the native anchors and has a free output layer that can absorb rotations at the function-class level. This provides an informative comparator for Factor and tests whether adaptation can compensate for the intervention within the training budget. The resulting interaction is independent of the low-scoring Query baseline, giving the central empirical claim a meaningful basis.

  2. New-target validation of a nontrivial empirical effect. For the fixed Protenix/ESMC configuration, the authors evaluate previously trained adapters on 192 targets selected before prediction and establish a positive interaction of 0.02515, with a conditional 95% confidence interval of [0.01832, 0.03213]. The interaction also remains positive after a prespecified budget extension on the same panel. This provides substantive evidence that the observed effect transfers to new evaluation targets and persists beyond the original stopping point, supporting the paper’s value as a focused empirical study.

Critical weaknesses
  1. The conceptual significance of the observed interaction remains insufficiently established. The four-cell experiment demonstrates a parameterization-dependent response to rotation, but does not identify what produces this difference. Factor and G+ differ in their function classes, residue-versus-pair organization, nonlinearities, and anchor entry points. The reconstruction diagnostic does not predict the post-training direction gaps, and the preferential benefit of output compensation is not re-established in the repeat. Consequently, the study does not yet provide a validated mechanism, predictive criterion, or actionable design principle. Although a descriptive finding can be valuable, the manuscript does not sufficiently explain how this particular interaction changes our understanding or design of adapters.

  2. The evidence supporting a broader adaptation phenomenon is limited. The primary interaction is established for one fixed Protenix/ESMC configuration, whereas the new-target OpenFold test remains inconclusive and AtlasFold shows no established adaptation benefit. The budget extension reuses an observed evaluation panel, and the reported confidence intervals condition on the fitted models rather than capturing variability across new training runs. These limitations are transparently disclosed, but leave substantial uncertainty about the finding’s stability across training configurations and architectures. Since the interaction also does not rank absolute adaptation quality, its broader relevance to the adapter community remains unclear, limiting the significance of the empirical contribution.

Questions for authors
  1. Contribution clarification: What specific insight for adapter design is established beyond the observed parameterization-by-rotation interaction?

  2. MSA baseline: How does the adapted model compare with standard MSA-enabled Protenix in lDDT and TM-score? Does the adapter provide further improvement when MSA is available?

  3. Additional evaluation suggestion: Could you compare against strong baselines on a representative subset of CASP15/16 monomer targets?

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

This paper investigates whether fixed orthogonal rotations of adapter outputs affect adaptation differently across two parameterizations in frozen protein structure predictors. It compares Factor, which constructs residual updates through native pair-generating operators, with G+. The main interaction is established for Protenix/ESMC but not confirmed on new OpenFold targets, and follow-up experiments leave its mechanism unresolved.

Critical strengths
  1. The 2×2 comparison of adapter type and output rotation provides a useful controlled comparison of adapter parameterization and output rotation, with matched information access and zero-residual initialization.
  2. The paper transparently reports negative findings and distinguishes absolute prediction quality, within-head rotation effects, and their interaction.
Critical weaknesses
  1. The primary interaction is established for Protenix/ESMC but not on the new-target OpenFold evaluation; AtlasFold shows no established adaptation benefit. The evidence therefore supports a configuration-specific observation rather than a general principle.
  2. The compensation gain difference is not re-established in the repeat, and the reconstruction diagnostic does not support the predicted relationship. Consequently, the study does not clearly explain why the observed interaction occurs.
  3. Although Factor is an interesting construction, its advantages are not consistent across backbones. With neither a broadly supported explanatory insight nor a consistently superior adaptation method, the overall contribution remains limited.
Questions for authors
  • What actionable guidance for adapter design follows from the observed interaction, given that it is not consistently established across backbones and a larger interaction does not imply better prediction quality?
Reviewer_gg9HOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors ask a interesting question which is whether an adapter benefits from expressing updates in the native output-channel orientation of a frozen protein structure predictor. It compares two similarly parameterized adapter families:

  • Factor, which predicts residue-level increments and composes them through a frozen predictor's native pair operator.
  • G+, which predicts pair updates using an explicit-anchor pair MLP and whose representable output function class is closed under an orthogonal rotation of its output channels.

For each family, the authors train Native and Rotated variants and define a difference-in-differences interaction

Psi = (Factor_Native - Factor_Rotated) - (G+_Native - G+_Rotated).

The main positive result is on a newly selected Fresh192 panel with fixed Protenix/ESMC adapters. Pair-lDDT means are 0.8281, 0.8042, 0.7280, and 0.7293 for Native Factor, Rotated Factor, Native G+, and Rotated G+, respectively. The reported interaction is +0.02515 [0.01832, 0.03213], with 140/192 target-level interactions positive and all nine seed-by-rotation means positive. A separate Fresh96 OpenFold/ESM2 evaluation gives an interaction of +0.01014 [-0.00318, 0.02347], which is not established. AtlasFold shows essentially no adaptation benefit or interaction.

The paper supplements these results with observed-panel studies on training budget, recycle supervision schedule, signed permutations, feature substitution, training-set size, learned-update norm/source interventions, reconstruction, a trainable compensating channel map, anchor dependence, and AtlasFold propagation. The prospective target evaluation, complete four-cell reporting, preservation of null findings, and precise separation between primary and observed-panel analyses are meaningful strengths.

My main reservation is that the strongest supported conclusion is narrower than the title and motivating narrative suggest: under one selected Protenix/ESMC training recipe, at one finite update budget, and for three fixed rotations and three fixed training seeds, Factor has a larger Native-versus-Rotated score gap than G+. It does not by itself isolate a privileged “native orientation,” establish a causal geometric mechanism, or show that the result generalizes over random rotations, optimization runs, inference noise, architectures, or target regimes.

Critical strengths
  1. The signed difference-in-differences is defined explicitly, and the paper specifies that a positive interaction neither implies that every Native-minus-Rotated contrast is positive nor that Factor is uniformly better in absolute score. This avoids common interpretive errors.

  2. Fresh192 was locked before prediction and evaluates fixed adapters without evaluation-label selection. The primary Protenix effect is not a marginal artifact: 140/192 target interactions and every seed-by-rotation marginal are positive. The OpenFold Fresh96 null is retained rather than hidden.

  3. The manuscript distinguishes new-target tests from reused observed panels, records the Fresh96 screening-rule revision, labels the single-rotation schedule amendment, and does not promote post hoc or follow-up analyses to confirmatory status.

  4. The paper reports that the reconstruction association fails (Spearman rho = 0.02381, exact one-sided p = 0.48839), that the compensator's preferential benefit does not reappear in a same-seed repeat (T_C = 0.00065 [-0.01223, 0.01338]), and that the anchor intervention does not establish the proposed dependence. This is scientifically valuable.

  5. Protenix, OpenFold, and AtlasFold give materially different outcomes. Although this weakens a universal claim, it makes the empirical record more informative and reduces concern that contradictory systems were omitted.

Critical weaknesses
  1. The central scientific variable is the output rotation, yet the primary intervals use targets as the resampling units while averaging three fixed rotations and three fixed training seeds inside each target. Thus the precise Fresh192 interval quantifies target variation conditional on 24 already fitted adapters. It does not estimate uncertainty for a new random rotation or a new training run. The paper also uses one inference sample per target. For predictors with stochastic structure generation, this leaves inference-sampling variation outside the uncertainty analysis. A fixed or paired inference seed is useful for variance reduction, but it does not establish that an effect of roughly 0.025 is stable across sampling trajectories. If the intended estimand is strictly conditional on these rotations, seeds, and inference samples, the abstract and conclusions should say so more prominently. If the intended claim concerns random channel rotations more generally, rotations and independent training runs need to be experimental units. I would like to see substantially more rotations, independent training seeds, and multiple paired inference samples, analyzed with a hierarchical or randomization-based procedure.

  2. The proof that G+'s representable function class is closed under output rotation is correct and useful, but function-class closure is not optimizer invariance. The paper itself notes that AdamW's coordinate-wise moments are not invariant to a dense rotation. More importantly, Factor and G+ differ beyond rotation handling: one predicts residue-level factors passed through a frozen native composition, while the other forms pairwise updates through an explicit-anchor MLP. They differ in computation graph, parameter sharing, nonlinearities, anchor access, and optimization geometry. A strong experiment would use a much closer matched head in which the represented function, parameterization, initialization, optimizer state mapping, and per-step computation can be placed in a one-to-one rotated correspondence. Alternatives could include mapped initialization and optimizer states, an optimizer with an appropriate invariance, or a compensating map tested prospectively across enough rotations. The current compensator intervention is conceptually relevant but its preferential effect fails to reproduce.

  3. The main comparisons use 1,536 updates. Continuing the same Protenix/ESMC models to 3,072 updates improves all arms but changes them unequally: G+ gains substantially more in absolute score, and the interaction decreases from 0.02515 to 0.01820. The interaction remains positive, but the manuscript explicitly states that neither convergence nor budget invariance is established. This matters because Factor and G+ organize their computation differently. Equal update counts do not demonstrate equal compute, equal optimization maturity, or equally tuned regularization and learning rates. G+ may simply have a different learning curve under the shared recipe. The authors intentionally avoid head-specific tuning, which is defensible for one tightly defined fixed-recipe estimand, but it limits conclusions about intrinsic orientation sensitivity or head quality.

  4. While Fresh192 is new with respect to prediction, the Protenix/Train384/ESMC recipe was chosen after examining extensive observed-panel evidence across backbones, training-set sizes, feature families, and adapter variants. The new panel therefore validates the selected recipe on new targets; it does not independently validate the broad hypothesis over the recipe space.

Questions for authors
  1. What population is the primary confidence interval intended to cover: only the fixed 192 targets under the fitted adapters, or new rotations and retrainings as well? Why are rotations not treated as sampling/randomization units?
  2. How stable is the Fresh192 interaction across inference seeds or multiple structure samples per target?
  3. Can the authors provide compute, wall-clock, and memory-matched comparisons, as well as learning curves showing whether the interaction stabilizes?
  4. Which experimental results were visible when Protenix/Train384/ESMC was selected for Fresh192? Could the full recipe-selection chronology be summarized in one table?
  5. Why is G+ the most informative counterfactual rather than a more tightly matched parameterization with a one-to-one rotation and optimizer-state correspondence?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.