acceptodds
Under review as a conference paper at ICLR 2027

EnvGeoBench: A Benchmark and MM Adapter for QM-Region Geometry Generation

Abstract

In multiscale molecular simulations of complex chemical systems, the geometry of a chemically active quantum-mechanical (QM) region is shaped by its surrounding environment, typically represented using molecular mechanics (MM). Existing geometry-generation methods largely treat the QM system in isolation, without a systematic approach to incorporating information from surrounding MM regions. This motivates a benchmark beyond isolated-molecule geometry generation that evaluates both structural accuracy and the benefit of environmental information. In this work, we introduce EnvGeoBench, a benchmark for environment-conditioned QM geometry generation from QM/MM simulation records. It defines four tasks for generating QM conformations with decreasing amounts of available QM-side information: reaction endpoint conformations and a requested path stage, reaction endpoint conformations alone, a reaction coordinate value alone, or no per-record QM-side input. For each task, we specify model inputs, hold out groups of related simulation records, and evaluate generated geometries against their reference structures. For systematic incorporation of MM information and performance evaluation, we introduce a reusable MM adapter that injects a compact representation of the local environment into the QM geometry model, with implementations for a conditional flow model and direct coordinate predictors. We compare four MM conditions to assess the benefit of including the spatial organization and record correspondence of the MM environment: the correct corresponding environment (matched-MM), no environment information (no-MM), randomized atomic directions (shuffled-MM), and an environment from another record (mismatched-MM). Across the four tasks, matched-MM lowers the reference models' reconstruction error relative to no-MM, with a median relative reduction of 14.3%. In the Window-Coordinate and Environment-Only tasks, it also outperforms shuffled-MM and mismatched-MM. We further evaluate the adapter with flow and diffusion generators to assess its transferability across generative-model families; matched-MM again yields the lowest mean reconstruction error among the four conditions for both families. EnvGeoBench enables systematic quantification of how molecular environments improve QM geometry generation, and its reusable adapter provides a general mechanism for integrating MM context into existing geometry generators.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
benchmark, conditional generative models, geometric deep learning, 3D molecular generation, AI for science, QM/MM

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

The paper introduces a benchmark for generating QM-region coordinates conditioned on the surrounding MM environment and varying amounts of reaction information. It also proposes a common MM adapter and evaluates several model architectures. The central concerns are the limited chemical diversity of the benchmark, insufficient description of the source data, and unclear practical motivation for reconstructing QM geometries from contemporaneous MM configurations.

Critical strengths

The benchmark provides explicit task definitions and grouped splits intended to reduce leakage between related simulation records. Comparisons with absent, shuffled, and mismatched MM environments provide useful controls. The finding that matched MM context improves reconstruction is a reasonable, although expected, sanity check.

Critical weaknesses

The manuscript is difficult to follow. Although the appendix provides dataset counts and provenance, the molecular systems and simulation protocols are insufficiently described, requiring readers to consult the source publications.

The datasets comprise relatively small, system-specific collections. Held-out replicas, pathways, and windows evaluate generalization within these collections, but do not establish transfer to unseen molecular systems.

More fundamentally, practical utility is not demonstrated. The MM inputs come from QM/MM calculations that already produced the target QM coordinates. It is unclear how these models would be used when the target geometries are genuinely unavailable, or how the proposed workflow would avoid the cost of generating them.

Many gains in the endpoint- and coordinate-conditioned tasks are only hundredths of an ångström. Larger improvements occur in other settings, but their downstream scientific significance remains unclear. The adapter’s advantage over alternative MM-conditioning approaches is also not established.

Figure 1 and its caption use inconsistent labels: the caption references T1, T3, and T4 without identifying them in the figure, and refers to a reaction coordinate that the figure labels a window coordinate.

Questions for authors

What concrete application permits obtaining suitable MM inputs without already computing the corresponding QM geometries, and what computational savings would result?

Can the benchmark evaluate transfer to unseen molecular systems or chemical conditions?

Can the paper provide a self-contained description of each source system, QM-region composition, simulation protocol, and sampling coverage?

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

This paper introduces EnvGeoBench, a benchmark for environment-conditioned QM-region geometry generation with four tasks that vary the available reaction information. It also provides MM-conditioning adapters for direct predictors, flow models, and diffusion models. Comparisons using matched, absent, shuffled, and mismatched environments show that matched MM context improves mean reconstruction accuracy across eleven task subsets. The main contribution is a structured evaluation of environmental information, supported by task-specific baselines and controlled experiments.

Critical strengths
  1. The four tasks and explicit input contracts provide a useful framework for evaluating how environmental information complements reaction-specific priors.

  2. The complementary MM controls, independently trained models, grouped splits, and group-level confidence intervals strengthen the empirical analysis.

  3. Improvements across multiple predictors and generator families, together with supplementary reaction-coordinate results, support the usefulness of matched MM context.

Critical weaknesses
  1. T4-HC7 selects its MM neighborhood using the target QM centroid, making the input preprocessing target dependent. The effect of this choice remains unquantified, despite being acknowledged in Appendix C.

  2. Evaluation remains within known systems with fixed QM atom schemas. The short temporal blocks used for HC7 and MHETase also lack a clear decorrelation justification, leaving effective train–test independence uncertain.

  3. The dominant D-MAE metric measures internal QM geometry and cannot establish correct placement or physically reasonable contacts within the MM environment. Aggregate environmental and chemical validity checks are limited.

  4. The ablations establish the value of MM information but do not isolate the proposed adapter’s contribution relative to simpler or alternative conditioning mechanisms.

Questions for authors
  1. How can the HC7 neighborhood be constructed without the target QM centroid, and are results using a fixed MM-only anchor available?

  2. What evidence supports the chosen temporal block lengths for HC7 and MHETase as sufficient to prevent strong correlations across partitions?

  3. Beyond internal-distance and reaction-coordinate errors, is there evidence that the predictions preserve reasonable QM–MM contacts and chemical validity?

  4. Which aspect of the adapter improves upon simply providing MM coordinates, and are comparisons with simpler conditioning methods available?

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

This paper introduces EnvGeoBench, a benchmark for environment-conditioned quantum-mechanical (QM) geometry generation in QM/MM simulation settings. The benchmark consists of four tasks that differ in the amount of reaction information available at inference time: Anchored-Stage, Endpoint-Bridge, Window-Coordinate, and Environment-Only. To evaluate the contribution of the surrounding molecular-mechanical (MM) environment, the authors introduce four conditioning settings: matched-MM, no-MM, shuffled-MM, and mismatched-MM. The paper also proposes a reusable MM adapter that encodes local environmental information for integration into different QM geometry-generation models.

Critical strengths
  1. The benchmark problem is well-motivated. The paper addresses an important limitation of existing molecular geometry-generation benchmarks, which typically focus on isolated molecular systems without explicitly evaluating the effects of their surrounding environments.

  2. The design of environmental conditioning controls is good, such as the comparison between matched-MM, no-MM, shuffled-MM, and mismatched-MM conditions. These controlled experiments distinguish the benefit of environmental information from the contributions of its spatial organization and correspondence to the target configuration.

  3. Good reproducibility. The authors explicitly define task-specific information contracts and grouped data splits to reduce leakage between correlated simulation records. The provision of evaluation protocols, baseline implementations, and reproducibility artifacts is also valuable for future research.

Critical weaknesses
  1. Generalization to unseen molecular systems. My primary concern is that the benchmark evaluates reconstruction on held-out records from the same underlying molecular systems. It does not necessarily demonstrate generalization to unseen reactions, enzymes, molecular compositions, or chemical environments.

  2. Non-Physical metrics. The benchmark primarily relies on pair-distance MAE and aligned Cartesian MAE. While these metrics provide useful measures of geometric reconstruction accuracy, they do not necessarily reflect whether a generated QM structure is chemically meaningful or physically plausible. I would encourage the authors to include more chemically relevant evaluation metrics, such as bond-length and bond-angle errors, local geometric validity, and environment-contact accuracy.

  3. Heterogeneous tasks. The proposed tasks differ not only in their reaction information but also in the baseline architectures and evaluation metrics. I would appreciate a clearer discussion of the relationship between the four tasks and guidance on how researchers should use EnvGeoBench to compare future methods.

Questions for authors
  1. Generalization: Is it possible to construct evaluation splits involving unseen molecular systems, enzymes, or reaction types? How would the authors expect performance to change under such settings?

  2. Chemical validity: Do improvements in pair-distance or Cartesian MAE translate into improvements in chemically meaningful properties, such as bond geometry, QM/MM interaction energies, or reaction-coordinate accuracy beyond the reported T1 experiments?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.