acceptodds
Under review as a conference paper at ICLR 2027

CoMoFlow: Controllable Protein Conformational Ensemble Generation via Explicit Collective-Mode Modeling

Abstract

Protein conformational ensemble generation aims to efficiently characterize the structural variations accessible to a protein beyond a single static structure. Recent flow-based methods efficiently generate conformational ensembles, but they model coordinated structural variation implicitly, which can limit ensemble fidelity and hinder direct control. To address this limitation, we propose CoMoFlow, a controllable framework for generating protein backbone conformational ensembles by explicitly modeling collective variation. CoMoFlow constructs protein-specific collective modes from the equilibrium structure, learns the distribution of their coefficients across the molecular dynamics (MD) ensemble, and uses sampled collective variations to guide SE(3) flow matching through feature conditioning and prior recentering. Experiments on ATLAS and de novo proteins show that CoMoFlow improves reconstruction of MD conformational distributions and coordinated structural variation relative to strong flow-based baselines, while enabling direct control over targeted conformational changes.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Protein Conformational Ensemble Generation, Collective Conformational Modes, SE(3) Flow Matching, Controllable Generation, Protein Dynamics

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_y9A4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CoMoFlow, a flow-matching framework for protein backbone conformational ensemble generation that enables control through collective modes. A sampled coefficient vector conditions the flow for controlled generation, drawn from a learned distribution over the coefficients of protein-specific collective modes. The method is validated on the ATLAS and de novo protein benchmark, focusing on the reconstruction of MD conformational distribution, along with analyses of structural variation and controlled generation.

Critical strengths
  1. Generation steering through a low-dimensional coefficient space

Projecting collective variation onto a small set of coefficients allows simplified handling on conformational change; varying a single coefficient produces aligned responses along the target mode. Because the space is low-dimensional, it can also be searched directly with CEM (spelling out the abbreviation and a short explanation of CEM procedure somewhere would be good).

Critical weaknesses
  1. Control is shown only under fixed or unspecified noise, so the robustness of the c → conformation link is not established (Figure 4)

Fig. 4 shows a single generated conformation per coefficient offset, and its noise setting seems unspecified. Since controllability is a core contribution, it remains unclear whether a given c consistently yields the same conformational tendency. For each of the five offsets in Fig. 4, generating multiple samples with independent noise and reporting the variance of the realized change Δc^k\Delta \hat{c}_k would substantially strengthen the controllability claim. For example, a variation bar on Fig. 4(c) would suffice.

  1. Additional analyses only against BBFlow, omitting strong baseline in Table 1 (Figure 3, 5)

In Table 1, AlphaFlow-T and ESMFLow-T are competitive with CoMoFlow on several metrics. However, the analyses supporting the claim on variation recovery, targeted conformational exploration, and basin recovery are only done against BBFlow. As a result, it is unclear whether the advantages hold against some baselines. Including AlphaFlow-T and ESMFlow-T in Figures 3 and 5 would make the claims considerably more convincing.

Questions for authors
  1. Definition of long metric (Figure 3)

In Figure 3, r_long and MAE_long are shown, but I missed the definition of the long; is it simply a residue pair some how far away in the sequence?

  1. Analysis on transient-contact Jaccard JtrJ_{\mathrm{tr}}

JtrJ_{\mathrm{tr}} is notably lower than that of AlphaFlow-T and ESMFlow-T, while it is higher than other baselines. Do the authors have intuition on why CoMoFlow stands behind AlphaFlow-T and ESMFlow-T, while it outperforms them in other metrics?

  1. Physical validity of generated conformation

Physical validity is evaluated only on single-mode traversals in the Appendix, using residue-pair rates. Could the authors report per-conformation validity for the main ensembles, alongside some representative baselines?

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

The paper introduces a method called CoMoFlow to model a protein’s distribution of conformations (aka, ensemble generation). CoMoFlow is a flow matching based model that explicitly learns the collective variables (CVs) of proteins (I think? see weaknesses and questions about the author’s terminology) and uses these CVs to guide the ensemble generation process. They leverage molecular dynamics as a source of information of the ensemble distribution, which is often overlooked and important.

Critical strengths
  1. Modeling the collective variables first and then using the smaller number of degrees of freedom to guide conformation generation is quite interesting.
  2. Leveraging MD data (commonly unused) as a source of protein conformations is new (most people focus on the Protein Data Bank) and an important direction of work.
  3. The experiments are comprehensive and show that CoMoFlow improves upon existing baselines; I especially find the controllable generation section to be interesting and a novel capability enabled by modeling the protein's CVs.
Critical weaknesses
  1. The writing is difficult to read. Some examples include (1) the authors use terminology not commonly used within the protein or biomolecules field (see questions; I want to emphasize that much of their work rests on these unknown terms, which makes following the paper very hard), (2) repetitively describing existing works (e.g. flow matching in the introduction) multiple times, and (3) the methods section describes 4 different sub losses where each sub loss has lots of new notation; it’s not only hard to follow the notation and definitions, but even the motivation is not well established for some of these losses (e.g. Why is \mathcal{L}_{ens} even necessary).

  2. The “method” is chaining well established components into a pipeline, which in itself is not a negative, but even this pipeline is effectively a latent diffusion or flow matching model since Eq 8 and 9 is analogous to the VAE loss where the collective variables U_k is the latent space and the authors then use the latent space to conditionally generate the ensemble.

Questions for authors
  1. Throughout, the authors use jargon that I am not familiar with and is not common in the protein/biomolecule space (e.g. “collective variation”, “coordinated structural variation”, ). What do these terms mean?

    • Do you mean “collective variable” for “collective variation?”
  2. RE Equation 1, is Uk∈RNU_k \in \mathbb{R}^N?

  3. How is equation 7 minimized? Are there multiple c^* that minimize this equation?

  4. There are multiple components in the total objective (Line 304): (1) c^*, (2) the “collective variables” U_k, and (3) v_{\theta}. I assume you optimize these simultaneously? But is there any disadvantage to sequentially learning these components, aka first learning the collective variables with your \mathcal{L}{rec} and \mathcal{L}{coeff} and then plugging these terms into L_{flow} after?

Reviewer_A8V9OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CoMoFlow, a framework for generating protein backbone conformational ensembles from a single equilibrium structure. It predicts protein-specific collective modes and a joint distribution over their coefficients, then uses sampled coefficients to condition an SE(3) flow generator through feature conditioning and prior recentering. The authors evaluate on ATLAS and a fully held-out set of de novo proteins, comparing against flow-based and other ensemble-generation baselines. They report improvements on several ensemble reconstruction metrics, analyze the effects of collective-conditioning pathways and coefficient covariance, and evaluate coefficient-level control and targeted conformational exploration.

Critical strengths
  1. The method offers a coherent way to expose coordinated variation to a flow model. CoMoFlow learns protein-specific modes and a full-covariance coefficient distribution, then uses these both to condition the generator and recenter its translational prior. The ablations on the conditioning pathways and covariance parameterization provide evidence that these components contribute to the reported ensemble metrics.

  2. The evaluation examines multiple aspects of ensemble reconstruction. The paper reports results on ATLAS and de novo proteins, uses five random seeds, and evaluates flexibility, coordinated motion, conformational spread, transient contacts, and basin recovery. The control experiments also examine responses to coefficient interventions rather than relying only on aggregate reconstruction scores.

Critical weaknesses
  1. The data split used to select the number of modes and mixture components is unclear. Appendix B.1 states that the main model uses K=8K=8 based on Appendix D.1 and C=2C=2 based on Appendix D.2. Those appendices motivate these choices using results on ATLAS and de novo proteins, but do not identify the split used for those results. The selected K=8K=8 and C=2C=2 rows match the CoMoFlow results reported in both Table 1 and Table 2. The paper describes the de novo benchmark as being used exclusively for OOD evaluation, which makes it important to establish whether those results informed model selection. If ATLAS test proteins or the de novo benchmark were used to choose KK or CC, the final evaluation would not be independent of that selection.

  2. The control results establish latent-coordinate steering more directly than physically or functionally meaningful control. The main control metrics measure whether perturbing a coefficient produces the expected change when the generated structure is projected back onto the same learned mode basis. This is useful evidence of coefficient-level controllability, but by itself does not establish that the modes correspond to independently specified or biologically meaningful motions. The targeted opening/closing experiment provides a useful additional check, but its geometric plausibility analysis is limited to Cα clashes and Cα–Cα bond-length deviation (Table 10). Thus, the evidence for broader claims about physically plausible controlled conformational changes remains limited.

Questions for authors

Q1: Which split was used to produce Appendix D.1 and D.2? Were the ATLAS test set or de novo benchmark used to choose K=8K=8 or C=2C=2? If so, can the authors provide results for configurations selected using validation data only, evaluated on an untouched test set?

Q2: Are the modes intended to represent specific interpretable motions, beyond directions in the learned coefficient space? Table 10 checks Cα clashes and Cα–Cα bond-length deviation; please clarify whether the physical-plausibility claim is limited to these checks or supported by additional backbone-geometry validation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.