acceptodds
Withdrawn from ICLR 2027

Do Plausible Structures Guarantee Coherent Dynamics? Contextual Anomaly Localization in Protein Trajectories

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Understanding how protein conformations evolve over time is essential for revealing the molecular mechanisms underlying biological function. Yet a trajectory can pass framewise structural checks and match a reference conformational distribution while containing temporally misplaced transitions, potentially misleading interpretations of conformational transition pathways. The challenge is to distinguish such contextual inconsistencies from legitimate fluctuations and conformational changes, which cannot be resolved by the magnitude of structural change alone. We introduce ProTAL, an offline framework that localizes trajectory anomalies by learning the compatibility between conformational transitions and their temporal context. ProTAL combines bidirectional latent predictions with residue-pair geometry across multiple temporal scales, integrating expectations of molecular evolution with direct evidence of structural change. Learning from unperturbed trajectories and controlled temporal corruptions, it localizes suspicious transitions without requiring paired reference trajectories at inference. On our ATLAS-derived protein-trajectory benchmark, ProTAL achieves substantially higher anomaly detection success rates than the conventional baselines evaluated. ProTAL adds transition-level diagnostics to structural and ensemble evaluation, helping researchers identify potentially misleading trajectory segments before drawing mechanistic conclusions. The code is available at the link: https://anonymous.4open.science/r/ProTAL-6EED.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for Science, Protein Conformational Dynamics, Anomaly Detection, Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_6n37OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proses ProTAL, a framework for detecting temporally inconsistent local transitions in protein dynamics trajectories. The authors argue that individual protein structures can be physically plausible but their ordering may be inconsistent. The model is trained on MD trajectories with artificial perturbations and they show experinemts on benchmarks like ATLAS.

Critical strengths

The authors take a different perspective which is interesting of identifying potential errors in protein dynamics trajectories, and i think developing better ways to evaluate these dynamics is perhaps an important and underexplored direction. But, I think the authors should shift their focus on local inconsistencies to more global trajectory behavior (lmetstable state occupancy, conform pathways, etc.)

Critical weaknesses

The main weakness is lack of biological motivation. MD trajectories do not have one canonical ordering (they are stochastic...), it is unclear why detecting local inconsistencies is useful when MD analysis is usually about overall metastable state occupancy and mode hopping transition pathways. The paper does not make a convincing case for why flagging individual transitions would provide useful biological insight or help assess these broader properties of the dynamics. Also the model is trained and evaluated on these artificial trajectory corruptions and there is not clear generalizability evidence for meaningful errors in generated protein trajectories.

Questions for authors

Major concerns with biological motivation. I am not convinced by the motivation of this paper. Protein dynamics is inherently stochastic, there is no single correct temporal ordering of conformations. What matters biologically is whether trajectories capture the correct energy landscape: metastable states, their populations and transition pathways / kinetics, rather than the precise plausibility of local fluctuation transitions. It is much less clear why these local inconsistencies in the frames would be useful and the paper does not motivate this need biologically. For example, a traj could have a few unusual transistions while still accurately reproducing the relevant conformation populations. Or conversely, the trj could appear locally smooth while misreprsenting the rates of transitions or occupancy of states. Moreover the authors do not convincingly establish that the anomalies they detect actually correspond to errors in biological meaningful quantities. With this, it is hard to motivate acceptance for the paper regardless of the method as this problem does not seem properly justified.

Limited generalization beyond synthetic anomalies. My second major concern is that the experimental setup is mainly focused on recognizing artificial perturbations. The authors train the modal by swapping/replacing/splicing frames and then evaluate on the same corruptions. This concern is reinforced by the results on unseen time warping where the recall drops significantly (~3%) compared to the main synthetic benchmark. It seems as if the model is just learning to predict the artifical perturbations. There needs to be much stronger experimental evidence along with better biological motivation.

Reviewer_1gZcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies local temporal inconsistencies in protein trajectories whose individual conformations remain structurally plausible. ProTAL is an offline detector combining bidirectional, multi-horizon latent prediction errors with temporal modeling of residue-pair geometric changes. It first learns from clean trajectories, then uses five families of synthetic temporal corruptions for supervised localization. Inference requires no paired reference trajectory. Evaluation covers an ATLAS-derived corruption benchmark, conventional and learned comparators, ablations, unseen time warping, MDGen trajectories, and selected conditional-MD studies. The results establish substantial improvements on the constructed corruption task; transfer beyond those corruption families remains less convincing.

Critical strengths
  1. On identical held-out windows, macro W3 recall is 55.00%, versus 12.62% for Transition innovation. At approximately 1% realized clean FPR, exact-edge W1 recall is 51.53%, versus 7.24%. The advantage therefore extends beyond tolerant localization. Protein-level analyses further support its breadth across the evaluated systems (Section 4.2, pp. 5–6; Table 1; Appendices C.14 and C.17, pp. 28 and 30). This is useful evidence that learned localization provides information beyond the evaluated conventional diagnostics.

  2. The evaluation includes branch ablations, unseen perturbations, and newly simulated conditional-MD endpoints. At matched 1% test FPR, full-model W3 recall is 56.20%, compared with 50.53% for Transitions-only. Among 20 selected cases receiving conditional-MD support, ProTAL detects 12, including five missed by the listed conventional comparators (Tables 3–4; Appendix C.11). The authors explicitly acknowledge selection effects and avoid treating endpoint evidence as a certificate of full-path validity. Together, these analyses support complementary information from the two branches and demonstrate some useful additional detections.

Critical weaknesses

The five corruption families define both supervised training and the principal high-recall evaluations. Their labels identify constructed temporal modifications, rather than independently established dynamical errors. The most direct unseen-corruption test yields 3.1843% time-warp W3 recall against a matched clean W3 background of 2.8984%. The difference is only +0.2859 percentage points, with the reported descriptive 95% interval [−0.4639,1.0156][-0.4639, 1.0156] (Table 8, p. 25). This provides no clear detection increment above the relevant background; ranking first among comparators alone does not establish useful transfer. The MDGen results do not yet close this gap. Of 4,092 candidate edges, 3,694 remain unresolved and must not be treated as negatives. Only 26 receive MD-reference support for deviation, of which ProTAL detects two. The headline 7.69% is the pooled analysis including historical seed 42; the designated primary analysis detects one of 18 supported deviations (Appendix C.10 and Table 9, pp. 25–26). Conditional-MD examples provide valuable evidence that useful detections exist, but selection depends on alarm strata and several cases share starting states. They consequently do not estimate population sensitivity or specificity for rare natural transitions. An aggregate clean FPR near 1% also cannot establish specificity within that particular subset. I appreciate the authors’ explicit caveats and do not require a physical-validity certificate; the remaining issue is whether the demonstrated transfer is sufficient to support the intended practical diagnostic contribution.

Questions for authors
  1. Given Table 8’s small increment over matched background and Table 9’s limited supported coverage, what practical scope do the authors consider established? Are there existing results, independent of model selection, demonstrating detection above paired background on other perturbations or generated trajectories, or false-alarm analyses specifically covering rare transitions in clean trajectories? Such evidence would materially affect my assessment.

  2. Table 5 selects ProTAL at epoch 132 using calibration W3, but uses fixed epoch 188 for Geometry-TCN/Transformer. Does the comparison persist when these controls use the same development selection criterion? Please clarify their supervised objectives and feature differences. Also, what do the error bars in Figures 13–16 represent, what is the sampling unit, and do they include retraining variability? These details matter because the closest branch ablations differ by roughly 5–6 percentage points, with gains concentrated in matched splicing (Table 4; Appendix D.7).

  3. Please specify how Test82 relates to the 200 training, 20 calibration, and 19 conventional-comparison protein entries, with reproducible membership and similarity information. Appendix C.1 states protein-entry-disjoint evaluation, but Test82’s construction and relationship to these sets remain insufficiently explicit. This is a clarification request, not a claim that leakage occurred.

Reviewer_98q3OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a method for checking whether generated protein dynamics follow transitions in the temporal order observed in MD trajectories. It combines a GVP-based trajectory encoder, transformer-based latent forecasting, and multiscale discrepancy scores. A separate component analyzes temporal changes in residue pairs. The scores are combined to flag frame swaps, block swaps, replica splices, and other temporal errors.

Critical strengths

The types of swaps and shuffles being considered here are hard to detect with other methods, or are simply undetectable because those methods do not even consider temporal dynamics or only look at ensembles.

Critical weaknesses

The manuscript is extremely difficult to follow and needs a complete rewrite. I suspect substantial AI use beyond what the authors acknowledge, although writing style alone cannot establish this. Regardless of how the text was produced, the lack of clear explanations prevents me from assessing the method and its claims.

Unclear wording appears throughout. For example, "Predictions are direct at each horizon, and targets are also detached." Direct in what sense? Detached from which computation? The auxiliary losses are described as follows: "The budgeted auxiliaries preserve representation diversity, distinguish correct-time targets, and anchor states and increments to geometry."This does not explain what the losses are or how they work.

Basic experimental terminology is introduced without explanation. The experimental setup suddenly states that "Test82 supports learned-detector comparisons, ablations, and the time-warp challenge." I had to examine the code to discover that 82 refers to the size of the test "shard"

The results are similarly opaque. The authors report that "ProTAL achieves 55.00% macro W3 recall, compared with 12.62% for Transition innovation" without making clear what "Transition innovation" means. Another claim states that "ProTAL leads at all three caps and in every family across the sampled positive budgets from 0.05% to 10%." The reader should not have to decipher the experimental setup from sentences like these. Likewise, "Candidate endpoints receive tail support when their descriptive rank is at most 0.05 and the native endpoint control is not also flagged" needs explanation.

The code documentation is even harder to follow than the manuscript. It is unnecessarily complicated and filled with vague terminology such as "compatibility contract". The README contains the sentence "Recovery is bound to the exact current source release and is not interchangeable across source releases" which appears to mean that results are only reproducible from the current version of the source code.

DATA_FORMAT.md states: "The materialized training windows additionally preserve entry, trajectory, replica, crop, shard, raw-frame, and source-index provenance. A six-field inference archive alone does not provide the complete training provenance or group/donor contracts." REPRODUCIBILITY.md states: “Scientific status fields remain meaningful after package renaming: research-only output is not a deployable candidate; a diagnostic pass is not a physical-stability result.” These statements do not give the reader clear, actionable information. This language is pervasive across the repository and makes the source code very difficult to understand or use.

Questions for authors
  • What does "Unless stated otherwise, thresholds are fixed by clean calibration q99" mean? Does q99 refer to the 99th percentile, and what exactly qualifies as "clean" calibration data?

  • The README states: "A selected research checkpoint can be used only when its parameter structure, configuration, and calibration policy pass validation; that compatibility does not authorize cross-release training recovery." What does "authorize cross-release training recovery" mean?

  • DATA_FORMAT.md states: "The model contract is 32 frames and 64 residue positions per window." Does the model process only 64 residues at a time? If so, how are these residues selected, and how are larger proteins handled?

  • What am I supposed to do with this information in REPRODUCIBILITY.md: "Existing selected research checkpoints may be computationally compatible when these checks succeed. This does not make an earlier source-bound recovery record compatible with this source release. Training recovery records include source/configuration identity and must remain tied to the release that created them." Are you providing checkpoints? Previous source releases? Is there a table tying previous checkpoints to older source releases?

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

Related papers

Discussion (0)

Sign in to comment.