acceptodds
Under review as a conference paper at ICLR 2027

RoCO-ET: Structured Conditional Observation Learning for Enzyme Temperature Prediction

Abstract

Predicting temperature properties from enzyme sequences is important for selecting enzymes suited to target temperature conditions and guiding the design of experimental conditions. However, temperature observations are affected by multiple experimental conditions, and the relevant context is incompletely documented. Observations with large deviations can therefore exert excessive influence on model fitting. Predicting active temperature ranges additionally requires learning both range location and span. To address these challenges, we propose RoCO-ET (Robust Conditional Observation for Enzyme Temperature Prediction), which unifies the modeling of scalar temperatures and ordered active temperature ranges in a sequence-conditioned location–scale framework. The model uses conditional scales to measure observation deviations, modulates the influence of large standardized residuals on fitting through a Student-t objective, and treats the two range endpoints jointly as a complete observation during training. A readout parameterized by a center and a positive half-width produces ordered endpoints by construction, allowing record-level modulation to act on both range location and span. On three public tasks, compared with PatchET under the same protocol, RoCO-ET reduces root mean square error (RMSE) by 10.88% for optimum temperature prediction and by 6.46% for operational stability temperature prediction, and increases mean intersection over union (IoU) for active temperature ranges from 0.516 to 0.581. Matched ablations and training interventions further show that sequence-dependent scales and joint modulation of complete records together improve temperature prediction and range fitting.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
enzyme temperature prediction, protein sequence modeling, robust regression, conditional scale modeling, active temperature range prediction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Wz71OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates enzyme temperature prediction from amino acid sequences and proposes RoCO-ET (Robust Conditional Observation for Enzyme Temperature Prediction), a framework for robustly learning scalar temperature properties and active temperature ranges under heterogeneous experimental observations.

The main contributions of this work include:

  1. A sequence-conditioned location-scale framework that incorporates heteroscedastic residual scales and a Student-t observation objective to reduce the influence of large standardized deviations.

  2. A structured center–positive-half-width parameterization that guarantees ordered range endpoints, combined with a joint multivariate Student-t objective that modulates both endpoints at the observation-record level.

  3. An analytical characterization of the resulting gradient modulation, together with empirical evaluations, matched ablations, and controlled perturbation experiments.

Experimental results on three enzyme-temperature prediction tasks demonstrate improvements over PatchET. Specifically, RoCO-ET reduces RMSE by 10.88% and 6.46% for optimum temperature and operational stability temperature prediction, respectively, and improves active temperature range IoU from 0.516 to 0.581.

Critical strengths
  1. Practical Motivation: The paper addresses an important problem in enzyme temperature prediction, where incomplete experimental conditions and heterogeneous observations can affect model reliability. The motivation for incorporating robust observation modeling is reasonable and practically relevant.

  2. Structured Prediction Framework: The combination of sequence-dependent residual scales, a joint Student-t objective, and an ordered range parameterization provides a coherent framework for handling scalar and interval-valued temperature observations.

  3. Experimental Improvements: The proposed method consistently improves the main prediction metrics across three tasks. The comparisons with Gaussian, Huber, and independent Student-t objectives provide useful evidence for the effectiveness of the proposed design.

  4. Ablation and Robustness Analysis: The paper includes component-level ablations, controlled observation perturbations, sequence-group holdout evaluation, and an alternative protein representation interface. These experiments help clarify the benefits and limitations of record-level modulation.

Critical weaknesses
  1. Limited Methodological Novelty: The main components, including heteroscedastic Student-t regression, robust likelihood optimization, and center–half-width interval parameterization, are closely related to established techniques. The joint multivariate Student-t formulation follows standard probabilistic modeling principles. Although their integration is useful, the paper does not sufficiently establish a fundamentally new learning mechanism beyond existing robust regression frameworks. The authors should more clearly distinguish the methodological contributions from prior work on heteroscedastic regression and multivariate robust estimation.

  2. Insufficient Justification of the Observation Model: The proposed framework assumes that sequence-conditioned Student-t distributions provide an appropriate model for heterogeneous temperature observations. However, the paper provides limited empirical evidence that the actual observation noise follows the assumed heavy-tailed structure or that both range endpoints exhibit the shared deviation patterns implied by the joint model. Furthermore, the theoretical analysis mainly characterizes gradients with fixed scales and does not establish robustness guarantees for the full jointly optimized network. Additional validation of the learned scales and observation assumptions would strengthen the methodological foundation.

  3. Limited Experimental Generalization and Robustness Validation: Although the paper includes sequence-group holdout experiments and controlled perturbations, the evaluation remains concentrated on three related enzyme-temperature tasks. The active-range test set contains only 182 records, and results are reported over three random seeds. More importantly, synthetic perturbations may not accurately represent real assay variability. The joint objective also performs substantially worse than Independent-t under strong single-endpoint perturbations, suggesting that its robustness depends on the structure of the observation errors. Additional external validation and experiments under more realistic noise conditions are needed.

  4. Incomplete Comparisons and Reproducibility Concerns: The experimental comparisons cover several relevant objectives but do not fully investigate alternative heteroscedastic regression approaches, such as beta-NLL or Faithful Heteroscedastic Regression, both of which are discussed in the related work. Comparisons with more flexible multivariate observation models would also help establish whether the improvements specifically arise from shared record-level modulation. In addition, several experimental figures contain the annotation "Illustrative workbook data," making the provenance of the plotted results unclear. The authors should clarify this annotation and ensure that all reported figures are directly traceable to experimental measurements.

Questions for authors
  1. What are the essential methodological differences between RoCO-ET and existing heteroscedastic Student-t regression frameworks? In particular, which aspects of the joint observation objective constitute new technical contributions rather than direct applications of standard multivariate Student-t modeling?

  2. What empirical evidence supports the assumption that enzyme temperature observations exhibit sequence-dependent heavy-tailed deviations and shared endpoint perturbations? Could the authors evaluate the calibration of the learned scales and analyze whether the assumed observation distribution matches the empirical residual distributions?

  3. How does RoCO-ET perform under realistic assay variability, different sequence similarity thresholds, and external enzyme datasets? Given its weaker performance under single-endpoint corruption, could the authors further characterize the conditions under which joint modulation becomes less effective than independent endpoint modeling?

  4. Could the authors provide additional comparisons with beta-NLL, Faithful Heteroscedastic Regression, or alternative multivariate robust objectives? Additionally, could they clarify the meaning of "Illustrative workbook data" in the experimental figures and explain how the plotted results were generated?

Reviewer_c4CROpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work presents RoCO‑ET, a sequence‑based probabilistic deep‑learning framework for enzyme temperature prediction, handling both scalar temperature targets (optimum temperature, operational stability temperature) and ordered active‑temperature interval outputs. Real‑world enzyme temperature datasets suffer from heterogeneous, noisy observations caused by incompletely‑recorded experimental assay metadata; outlier measurements can corrupt model fitting, and range prediction must jointly learn both the location and span of active windows.

Critical strengths
  1. Prior enzyme‑temperature prediction research predominantly focuses on improving protein sequence encodings, while largely ignoring heterogeneous noisy labels arising from missing experimental assay metadata. RoCO‑ET introduces a multivariate Student‑t record‑level robust regression setup with sequence‑dependent conditional scales and a mathematically guaranteed ordered‑range readout.
  2. Beyond standard benchmark comparisons against prior state‑of‑the‑art PatchET, the paper uses carefully matched ablations.
Critical weaknesses
  1. No prospective experimental validation on real new enzyme measurements; evaluation remains confined to curated static benchmark splits. The experiments can be evaluated on new released datasets. The enzyme-related tasks also have these problem, like pH, etc.

  2. Degrees‑of‑freedom ν is fixed to 4 without principled data‑driven selection. The Student‑t degrees‑of‑freedom hyperparameter ν, which directly controls heavy‑tailed outlier‑downweighting behaviour, is manually set to ν=4.

Questions for authors
  1. For the Student‑t degrees‑of‑freedom ν: could you clarify why you chose to fix ν=4 rather than learning ν or optimising it over the validation set?
Reviewer_55zgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RoCO-ET to predict enzyme optimum temperature, operational stability temperature, and active temperature ranges from sequences. It combines frozen ESM2 representations with sequence-dependent residual scales and a Student-t objective. For range prediction, a center-positive-half-width parameterization guarantees ordered endpoints. A joint likelihood couples their influence during training. Across three public benchmarks, the paper reports RMSE reductions of 10.88% and 6.46% on the scalar tasks and an increase in range IoU from 0.516 to 0.581 relative to locally evaluated PatchET. Validation includes three-seed comparisons, component ablations, controlled label perturbations, and additional split and representation settings.

Critical strengths
  1. The tightly controlled ablation and perturbation experiments isolate record-level joint modulation from conditional scaling and readout design. The predicted advantage under record-wide corruption is consistent with a mechanistic claim.
  2. The structured center–half-width readout and interval-level evaluation, including the widening control, show that the IoU improvement reflects better range fitting rather than merely wider predictions.
Critical weaknesses
  1. The methodological contribution is modest. The individual components (heteroscedastic Student-t regression, the centre+positive half-width parameterisation for ordered intervals and the shared precision representation of the multivariate Student-t) are all well established techniques. The principal conceptual contribution of the paper is to treat the two endpoints as a single observation, both sharing a residual dependent modulation factor and then a brief analysis of the resulting coupling between the range centre and the span. This is perfectly reasonable and does deliver some tangible benefits, but the advance over the individual techniques is fairly modest. Its importance therefore hinges very much on whether the record level coupling is in fact a general feature of interval valued observations or simply an advantage in this particular case and under these particular training conditions.
  2. The experiments do not establish the proposed mechanism on real data. The intervention study clearly shows that joint modulation helps when corruption affects an entire range, while independent endpoint fitting is preferable when only one endpoint is corrupted. However, the paper does not show which type of deviation predominates in the enzyme-temperature datasets. Consequently, the experiments demonstrate behavior under constructed perturbations but do not establish that realistic assay heterogeneity causes the observed benchmark gains. Without this connection, it remains unclear when the proposed joint observation model should transfer to other interval-prediction problems.
Questions for authors
  1. Is there empirical evidence (e.g., from repeated measurements or assay metadata) showing that real-world dataset noise manifests primarily as whole-record variations rather than isolated single-endpoint errors? How does this justify the advantage of joint modulation over Independent-tt?
  2. What specific clustering algorithm, thresholds, and duplicate sequence mapping rules were used for the sequence-group holdout split? What is the maximum cross-partition sequence identity remaining?
  3. Given the test sample size (N=182N = 182), can statistical tests (e.g., paired tt-test) or confidence intervals be provided to quantify test sample variability alongside training variability?
  4. What exact mathematical functions, reference formulas, and sampling distributions generated the synthetic shift, span, and single-endpoint perturbations?
  5. What does "Illustrative workbook data" mean in Figures 3, 4, and A1–A4?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.