acceptodds
Under review as a conference paper at ICLR 2027

EASER: Energy-Aware Symbolic Ensemble Readout for Conformer-based Molecular Prediction

Abstract

Flexible molecules are naturally represented by ensembles of three-dimensional conformers, yet many conformer-based predictors rely on neural readouts whose aggregation mechanisms are difficult to inspect and do not explicitly account for relative conformer energies. We introduce EASER, an Energy-Aware Symbolic Ensemble Readout for molecular conformer ensembles. EASER trains and freezes a SchNet encoder to map each conformer to a fixed-dimensional representation. A cross-fitted descriptor model captures readily predictable two-dimensional signal, while energy-conditioned genetic-programming experts construct compact symbolic features to model the remaining residual from conformer embeddings. Their outputs are linearly combined into conformer-level residual scores and aggregated using a learned effective energy scale. Across multiple MoleculeNet benchmarks under both random and scaffold splits, EASER consistently improves over a matched descriptor-only baseline while remaining competitive with established molecular prediction methods. Ablation studies support complementary contributions from descriptor residualization, symbolic feature construction, and energy-conditioned ensemble modeling, while conformer-budget experiments show that most predictive gains can be obtained with relatively small ensembles. At inference, evolutionary search is no longer required: EASER uses only the frozen encoder, descriptor baseline, fixed algebraic expressions, linear coefficients, and explicit energy-aware aggregation. EASER thus provides a compact and auditable readout over learned 3D molecular representations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular property prediction, molecular conformer Eensembles, genetic programming, 3D molecular representation learning, symbolic regression

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

This work proposes EASER. It uses a frozen pretrained 3D GNN as the encoder, and, instead of using a standard neural readout, it uses am explicit symbolic pipeline for conformer ensemble property prediction of molecules. It uses cross fitted residualization to filter cheap descriptor signals, and go on to use genetic programming to explore further nonlinear features over the latent space. The paper separates 2 roles of conformer energy, one is the deterministic routing to energy-bin-specific symbolic experts, while the other is the learned Boltzmann-form aggregation of conformer scores. The method is evaluated on 6 tasks from MoleculeNet, against the vanilla neural net and other baseline models.

Critical strengths
  1. The experiments are thorough and meticulous, covering major baselines and diverse tasks on MoleculeNet, with ablation studies and clean budget control.

  2. The methods is physically well motivated, providing an interpretable readout that can be potentially useful for chemists to audit and justify a prediction.

Critical weaknesses
  1. None of the discovered symbolic expressions are reported in the manuscript, and only aggregate structural statistics are provided. It is unclear what the actual scientific benefit would be of those symbolic expressions to a chemist.

  2. There are no case studies or benchmarking on harder realistic molecular tasks or OOD settings, and everything is carried out on MoleculeNet. Also, "audit" would mean that by looking at the symbolic output, we can identify failure modes of a model, but this is also absent from the current version of the paper.

  3. The recurrence metric on ClinTox (0.52) and BBBP (0.57) seems to suggest that the readout learned function varies a lot across the different runs, thus making a stable audit not plausible.

Questions for authors
  1. There is a mismatch between two places reporting random-split FreeSolv: 0.89 in Table 1, and 0.96 in Sec 4.2. Which one is correct?

  2. Can you provide a few examples of the symbolic expressions discovered by the protocol, and try to do a audit by interpreting those symbolic expressions?

  3. Given the functional stability is only 0.52 on ClinTox, do you suggest a chemist to use this for reliably auditing the predictions? Can you also provide the prediction-level agreement across the GP seeds alongside the feature-level recurrence?

1 comment

Thank you for your careful response, which has addressed some of my concerns. I am still not fully convinced that the reconstructed expressions would provide further insight into the underlying chemical principle guiding the system, or that they could shed further insights that can be verified by the chemists. I encourage the authors to involve chemists in the evaluation of the qualities of those expressions, and to provide a practical case study for how those expressions would enlighten chemists in a principled way.

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

This paper proposes EASER, an energy-aware symbolic ensemble readout for molecular property prediction. The method first uses 2D molecular descriptors to establish a baseline prediction, then employs genetic programming (GP) to construct symbolic experts over frozen SchNet 3D conformer representations to predict the residual. Relative conformer energies are incorporated through two mechanisms: energy-conditioned routing to specialized symbolic experts and energy-aware aggregation of conformer-level predictions. The proposed method is evaluated on six molecular property prediction benchmarks covering regression and classification tasks under both random and scaffold splits. Experimental results show that EASER consistently improves upon the descriptor-only baseline and achieves competitive performance against existing methods and matched neural readouts. Ablation studies further demonstrate the positive contributions of individual components, including residual learning, symbolic feature construction, energy-conditioned routing, and energy-aware aggregation.

Critical strengths
  • Comprehensive experimental evaluation. The paper evaluates EASER on six molecular property prediction benchmarks under both random and scaffold splits, with comparisons against various baseline methods. The ablation studies systematically examine the contributions of individual components, particularly the two distinct uses of conformer energy. These experiments provide empirical evidence that the proposed components contribute positively to predictive performance.
  • Compact and transparent readout with competitive performance. EASER achieves competitive performance compared with a matched neural readout while using substantially fewer trainable parameters in the readout module. Furthermore, the symbolic expressions and explicit energy-aware aggregation make the prediction process more transparent and easier to inspect, although this does not necessarily translate into chemical interpretability.
Critical weaknesses
  • Limited methodological novelty and insufficient motivation for the proposed components. The proposed framework largely combines established techniques, including residual learning, genetic programming, mixture-of-experts, and energy-aware aggregation, without introducing a substantially new methodological insight. More importantly, the motivations for several key design choices are not sufficiently justified. For example, it remains unclear why symbolic regression is particularly suitable for modeling residuals from SchNet representations, or why conformers from different energy regimes should require different symbolic experts. The paper lacks a clear and unified rationale connecting these components to a specific limitation of existing methods. This substantially weakens the methodological contribution and limits the insights that future research can draw from the proposed framework.
  • Limited performance gains and insufficient evidence of the practical value of the proposed complexity. Although the ablation studies demonstrate positive contributions from individual components, the improvements are generally modest and do not convincingly establish the necessity of the proposed design choices. Furthermore, EASER achieves performance comparable to the matched MLP-based readout, but the paper does not sufficiently demonstrate what unique practical advantages symbolic regression provides beyond reducing the number of trainable parameters in the readout. Consequently, the experimental results do not fully justify the complexity of the overall framework or establish a sufficiently compelling advantage over simpler alternatives.
Questions for authors
  • What specific limitations of existing neural readouts motivate the use of genetic programming over frozen SchNet representations? Given that the latent dimensions do not have explicit chemical interpretations, what unique interpretability or practical advantages does the symbolic readout provide over a conventional MLP, beyond having fewer trainable parameters?
  • What is the underlying rationale for assuming that conformers from different energy regimes require different prediction functions?
  • Although individual components consistently improve average predictive performance, the gains are generally modest. Can the authors provide stronger statistical evidence that these improvements are reliable?
Reviewer_bKs9OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper propose EASER, a method using energy-conditioned experts and learned exponential pooling for molecular property prediction using multiple conformers. The task is important, and main contribution is a compact, explicit way to learn from conformers ensemble .

Critical strengths

The method is clearly formulated, and the matched component ablations help distinguish the contributions of residualization, symbolic features, routing, and aggregation. The authors also conduct many experiments against various baselines and datasets.

Critical weaknesses

The energy partition is insufficiently explained. Splitting conformers into three energy box needs further clarification. The single-expert ablation supports specialization, but simple experiments, like sensitivity to the number of bins, or comparison with continuous energy conditioning, are helpful.

The empirical evidence does not sufficiently distinguish the method from simpler uses of multiple conformers. Gains over the matched MLP are modest, while published Uni-Mol results are not controlled comparisons. Matched conformer-augmentation and prediction-averaging baselines, together with an explicitly ensemble-dependent MARCEL task such as Kraken, would strengthen the evidence. The discussion should also engage with Hamakawa and Miyao’s Understanding Conformation Importance in Data-Driven Property Prediction Models (2025), which directly examines augmentation versus aggregation. Thus, case study on some molecules will be helpful, for example, to check the learned "temperature", see if the learned distribution is meaningful. I think is heavily realied on the training targets. If so, the method will downgrade to another method for data augmentation.

Questions for authors

See weakness part.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.