acceptodds
Under review as a conference paper at ICLR 2027

EVORESOLVE: Preserving Evolutionary Heterogeneity in Protein Conformation Prediction

Abstract

Evolution preserves a protein family's capacity for multiple functional conformations, leaving partially distinct structural constraints across its homologs. However, AlphaFold3 (AF3)-style folding models merge these signals by averaging MSA-row outer products, leaving the folding model with single evolutionary summary. Therefore, we introduce EVORESOLVE, which learns multiple evolutionary summaries before MSA aggregation. Motivated by the richer row-wise variation observed in multi-conformation proteins, EVORESOLVE organizes homolog evidence into compact, query-conditioned branches using sliced wasserstein transport. Multi-conformation training thus supervises the new aggregation process, encouraging branches to emphasize different evidence in support of alternative states. The resulting representations drive a shared folding backbone, producing multiple predictions from one searched MSA without target-specific alignment editing. Across various conformational benchmarks, EVORESOLVE achieves the lowest reported RMSD among the compared methods, indicating that it can retain state-relevant differences while preserving the evolutionary constraints across homologs.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein conformation prediction, multiple sequence alignment, coevolutionary information reservation, protein understanding

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_2bjbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors introduce the multi-state conformer generation strategy for AF3-style co-folding models. Using SWE approach on AF3's OPM operation, multiple pair representation can be obtained. The authors evaluated their method on multi-conformation systems, including multi-state states including membrane, domain motion and fold-switching, and ligand-induced systems including GPCR, kinase, and cryptic pocket. Moreover, the ablation studies are provided to evaluate the individual contributions of SWE and QBOP.

Critical strengths

The motivation of this work is clear. While AF3-style co-folding models use stochastic diffusion methods, they often fails to recover multiple conformational states and generate a single state. Therefore, these methods rely on stochasticity of MSA subsampling and require many inference seeds to recover diverse states. Instead, this method suggests how to preserve multiple views from a single MSA information on single trunk run.

Critical weaknesses
  1. It is unclear how this methodology can be combined with the common recycling strategy involving MSA subsampling employed in AF3-style co-folding models. In particular, much of the structural diversity observed in AF3 is driven by MSA subsampling, yet the paper does not analyze the synergy between the proposed method and conventional subsampling.

  2. It is difficult to argue that the ability to reconstruct the holo state without ligand information supports the method’s capacity to preserve evolutionary heterogeneity. Unlike in monomer-focused methods such as AFCluster or BioEmu, evolutionary signals guide co-folding models to generate either the apo or holo state depending on the presence or absence of a binding ligand. Consequently, it is unclear whether the improved recovery performance comes at the expense of the co-folding model’s ability to distinguish between conformational states based on binding context.

Questions for authors
  1. The generation settings used for evaluation are unclear. Author should report the inference settings including inference seeds, number of diffusion samples, and recycling iterations.
  2. How does the proposed method behave under the different binding conditions? For example, AF3Cluster, which incorporates MSA clustering of AFCluster to AlphaFold3, investigates how to find both apo and holo states with binding context.

[1] Cheng, Xiwei, et al. "Generalized Multi-State Protein Design with AlphaFold3." ICLR 2026 Workshop on Generative and Experimental Perspectives for Biomolecular Design.

Reviewer_R522OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2(−1)
Soundness
1
Read review
Paper summary

Despite breakthroughs in protein structure prediction (e.g. AlphaFold2 and AlphaFold3), protein conformation prediction for multi-state, flexible proteins that can adopt multiple distinct conformations has been lagging behind. Crucially, such different conformations often reveal complex protein function and mechanisms. The authors identify as one of the main causes of poor AF3 conformation prediction the compression of the row-wise outer product as part of the MSA information encoding. To mitigate this information loss, EvoResolve introduces an end-to-end multi-branch MSA aggregation mechanism into Protenix, replacing the default OPM with learned branch-spacific Sliced Wasserstein Embeddings and query-based aggregation. Multi-conformation supervision encourages the branches to encode alternative structural states.

Critical strengths

The main strength of the paper is the formulation of a clear gap in present protein folding models and the clear presentation of a new DL-based framework to solve it. This work flags the poor performance of protein structure prediction models in predicting alternative conformations/states of proteins, which is indeed one of the main bottlenecks in protein design. They correctly identify the suboptimal way protein folding models process the MSA signal: they average out the complex evolutionary signals (MSA subfamilies) across homologues, including evolutionary modes that could reveal distinct well-defined structural states/conformations. The solution the authors propose is theoretically neat and can be integrated in present folding models in an end-to-end workflow, which is a great property. They design parallel branches that are meant to decompose the heterogeneous signal into evolutionary modes for different conformations. On each branch, they use a branch-specific set of learnable anchors to compress the original MSA P homologue rows into a much smaller number S of pseudo-MSA rows. These are then processed by a query-biased outer product, which adaptively weights their contribution based on their relevance to the original input sequence. A weighted average of the pseudo-MSAs gives the final branch MSA, that is then fed into the Pairformer for conformation prediction. Conformation-specific distogram losses and adaptive diversity constraints provide gradients that optimize each branch to recover distinct structural states. I find the conformation-specific branches, MSA compression into mode-specific pseudo-MSAs and pseudo-MSA-input sequence compatibility weights to be intuitive ideas showing a deep understanding of the proposed problem.

Critical weaknesses

Improvements on reported benchmarks are not convincing. For example, Protenix-v1 achieves competitive performance in Table 1 compared to EvoResolve using the default MSA encoding. This agrees with findings in [1] and [2] that show limited use of MSA-subsampling or MSA-clustering approaches compared to default AF2 performance on multi-state protein structure state prediction. Moreover, AF-Cluster [3] and especially MSA-subsampling already explored exactly this idea of finding protein MSA subfamilies using a much simpler, lighter approach, so arguing for an end-to-end DL model trying to achieve multi-conformational structure prediction should show stronger prediction improvements in Table 1. Crucially, STDs and error bars are not reported in any of the tables or plots respectively.

Table 2 experiments on disordered proteins dilute the message of the paper, which is otherwise focused on conformational changes in structured proteins, which show strong (and clustered in the case of multi-state proteins) evolutionary signals in their MSAs. Disordered proteins, however, are known to function based on a much looser “sequence grammar” which breaks the assumption of positional homology that MSAs for structured proteins are built on [4]. Moreover, the metrics in Table 2 are weak ensemble-average proxies for what the model aims to do, i.e. full conformational prediction.

[1] Schafer JW, Lee M, Chakravarty D, Thole JF, Chen EA, Porter LL. 2025. Sequence clustering confounds AlphaFold2. Nature 638: E8-E12

[2] Chakravarty D, Schafer JW, Chen EA, Thole JF, Ronish LA, et al. 2024. AlphaFold predictions of fold-switched conformations are driven by structure memorization. Nature Communications 15: 7296

[3] Wayment-Steele, H. K., Ojoawo, A., Otten, R., et al. (2024). Predicting multiple conformations via sequence clustering and AlphaFold2. Nature, 625, 832–839.

[4] Lindorff-Larsen, Kresten, and Birthe B. Kragelund. "On the Potential of Machine Learning to Examine the Relationship Between Sequence, Structure, Dynamics and Function of Intrinsically Disordered Proteins." Journal of Molecular Biology, vol. 433, no. 20, 2021, p. 167196.

Questions for authors

Why did you not discuss more the multi-conformational data scarcity as a fundamental bottleneck for this task? Besides CoDNas, there have been larger multi-state datasets curated from PDB based on maximising conformational divergence for lower sequence similarity thresholds, like in [5].

Alternative states are usually reinforced by binders (other proteins, small molecules, metal ions, etc). Have you considered running benchmarks on such binder-conditioned states?

[5] Abrudan, A., et al. "Multi-state Protein Sequence Design with DynamicMPNN." International Conference on Learning Representations, 2026.

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

This paper aims to preserve conformation-related evolutionary information by integrating branch-specific sequence-weighted embeddings, a decoding mechanism based on transport theory, and an outer-product aggregation approach incorporating query bias. In contrast, such evolutionary information is typically discarded in AF3 due to the averaging process applied during optimal alignment matching. The proposed method is built upon Protenix-v1 and fine-tuned on relevant datasets, yielding promising results. Experimental evaluations across various tasks—including sensitivity analysis of optimal alignment matching, six conformational benchmarks, and the study of intrinsically disordered proteins—demonstrate that the proposed algorithm achieves competitive performance.

Critical strengths
  1. In the MSA aggregation stage of this paper, multi-conformation supervision is introduced, and a diversity constraint based on branch matching and reference state adaptation is combined. This provides a goal-oriented framework for learning multiple conformation-related representations from a single MSA.
  2. This paper provides certain support for the validity of the model through a series of experiments.
Critical weaknesses
  1. The current experimental results are not sufficient to clearly demonstrate whether the model can accurately predict the two conformations. It is recommended to present representative predicted structures and list the results separately according to the target and state. Additionally, the same sampling and filtering budget should be used to compare different methods, and the uncertainty of the results should be reported.
  2. The existence of only one reference structure does not prove that a protein has only a single conformation. Such targets should not be automatically classified as "single conformation" control groups. To determine whether the observed differences are indeed related to conformational characteristics, variables such as protein length, category, and the information content of multiple sequence alignment must be controlled.
  3. The current experiments mainly target proteins with a large number of homologous sequences; it is unclear whether this method is also effective for "shallow" multiple sequence alignments (i.e., limited homologous sequence data). It is recommended to report the number and characteristics of excluded targets, conduct additional experiments under different MSA depths, or clearly define the application scope of the method's conclusion.
  4. The current experiments have not sufficiently distinguished the performance improvement due to architectural improvement from the improvement brought by the addition of new multi-conformation training data. To accurately evaluate the effectiveness of architectural optimization, the same training data, supervision targets, and computing budget should be used to compare the original architecture and the new architecture.
Questions for authors

refer to weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.