acceptodds
Under review as a conference paper at ICLR 2027

TwinMP: Adapting Inverse-Folding Models for Multi-Conformation Design and Mutation-Effect Prediction

Abstract

Protein inverse folding models learn strong priors from individual protein backbones, but typically overlook the conformational variability that is often essential to protein function. Existing multi-conformation adaptations either pool structural representations, potentially obscuring conformation-specific constraints, or process each conformation independently and combine their logits only during decoding, without information exchange between structural states. We introduce **Twin**-Stream with **M**essage **P**assing **(TwinMP)**, a model-agnostic post-training framework that adapts pretrained single-conformation inverse folding models to dual-conformation learning. TwinMP maintains two streams that separately preserve the structure-to-sequence mapping of each conformation while introducing bidirectional message passing between their hidden representations before sequence prediction. This design retains the single-state knowledge acquired during pretraining while enabling explicit exchange and reconciliation of shared and state-specific structural constraints. We instantiate TwinMP on ProteinReasoner, ESM-IF, and ProteinMPNN. Across all three model families, TwinMP improves mean amino-acid recovery over pooling and Logit mean, with backbone-dependent gains from message passing. Moreover, TwinMP with AF2 and MD-derived structural inputs achieves an official ProteinGym score of 0.486, compared with 0.473 for the single-conformation baseline. These results support TwinMP as a practical framework for adapting pretrained inverse-folding models to multi-conformation design and mutation-effect prediction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein Language Models, Conformational Ensembles, Protein Structure, Inverse Folding, Variant Effect 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_MsiXOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, the authors propose TwinMP, a post-training framework for adapting pretrained single-conformation inverse-folding models to dual-conformation design. The method keeps two parameter-tied streams for the two conformations, introduces residue-aligned cross-conformation message passing, and then fuses the hidden states for sequence prediction. The framework is evaluated on ProteinReasoner, ESM-IF, and ProteinMPNN using amino-acid recovery and structural consistency metrics, with an additional ProteinGym mutation-effect prediction experiment.

Critical strengths
  1. Reasonable method with broad evaluation. Adapting pretrained inverse-folding models to multi-conformation inputs is a sensible idea, and the paper evaluates it across three model families with both sequence- and structure-based metrics.
  2. Clear presentation. The paper is well organized, the figures are clear, and the method and experimental setup are easy to follow.
Critical weaknesses
  1. Limited novelty relative to DynamicMPNN. DynamicMPNN has already studied essentially the same multi-conformation inverse-folding problem. The paper does not clearly explain what important limitation of DynamicMPNN it addresses, especially relative to DynamicMPNN+DSS, which also fuses features during message passing.
  2. The key message-passing contribution is not clearly described in the Method section. And the ablations show only small gains from message passing.
Questions for authors
  1. Could the authors explain in substantially more detail how residue-aligned message passing is implemented for each backbone, especially ProteinMPNN?
  2. What is the precise methodological difference between TwinMP and DynamicMPNN+DSS? DynamicMPNN+DSS also allows intermediate representations from different conformations to interact. It would be helpful to explicitly compare the two update rules and explain what capability TwinMP provides that DSS does not.
Reviewer_EtwGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TwinMP adapts pretrained inverse-folding models to two residue-aligned conformations using shared model parameters, separate structural streams, and intermediate message exchange. It is evaluated with ProteinReasoner, ESM-IF, and ProteinMPNN on sequence recovery and target-template-conditioned structure prediction. A separate ProteinGym evaluation uses a fixed ProteinReasoner-150M checkpoint to compare duplicated AF2 inputs against AF2 paired with one MD equilibration endpoint. The reported hierarchical correlation improves from 0.479 to 0.486. This comparison measures a change in structural input under a fixed scoring procedure.

Critical strengths
  • A reusable adaptation framework with evaluation across three backbone families. The paper makes intermediate cross-conformation exchange implementable in substantially different inverse-folding architectures and evaluates held-out performance with backbone-specific decontamination. This breadth gives the approach practical value beyond a single bespoke model.
Critical weaknesses
  • The added value of intermediate interaction is insufficiently isolated. A central method claim concerns exchanging hidden states across conformations, but I need a post-trained Logit mean control with the same data, optimization, and comparable capacity to distinguish that mechanism from gains due to adaptation and fusion. Likewise, the fixed-checkpoint AF2+EQ improvement does not separate the value of a relaxed structure from the value of cross-conformation interaction. Single-structure EQ and matched no-interaction controls would materially affect whether I attribute this result to the proposed multi-conformation method.

  • The structural validation remains dependent on the target templates. The two-state evaluation predicts each generated sequence using the corresponding target conformation as a template. This provides evidence of compatibility under that conditioning, but leaves the central dual-conformation design claim insufficiently tested beyond the supplied targets. I would need validation that reduces their influence before giving this claim substantial weight.

Questions for authors
  1. Can you provide a Logit mean control post-trained on the same examples with matched optimization and comparable capacity? How does it compare with TwinMP on the same held-out pairs?
  2. For ProteinGym, can you compare AF2+EQ with single-structure EQ and a matched dual-input no-interaction model, keeping the WT sequence, MSA profile, scoring, and aggregation fixed?
  3. What evidence supports the generated sequences’ compatibility with both conformations when target-template influence is reduced? If such results are available, please report the protocol and paired outcomes. : )
Reviewer_bS7HOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors identify the suboptimal way single-state inverse folding models are used for multi-state protein design doing post-hoc logit averaging. The authors propose TwinMP, a multi-state inverse folding framework that better mixes the multi-conformational structure information of a protein before decoding it into a sequence. TwinMP is instantiated on multiple single-state inverse models and evaluate them on a multi-state backbone set. TwinMP is also benchmarked on the same multi-state protein design dataset against competing models. TwinMP multi-state logits are shown to increase zero-shot mutation effect prediction on the ProteinGym benchmark compared to single-state ones.

Critical strengths

The strongest TwinMP contribution is the cross-architecture applicability. The authors successfully show that TwinMP can be conveniently used finetune independently pretrained model families of single-state inverse folding models.

Critical weaknesses

TwinMP lacks architectural novelty. DynamicMPNN [1] implemented a cross-state encoder-level communication in its version DynamicMPNN-DSS (Deep Symmetric Set), which is highly similar to the TwinMP implementation. The authors incorrectly claim DynamicMPNN only performs pooling without cross-state communication, completely omitting the DSS variant of the model. The contributions need to be reframed in a more honest way.

The cross-architecture finetuning results in Table 2 are unreliable due to the highly varying sizes of the test set used for each architecture. The authors explain this was due to leakage removal between the multi-state test set and the different pretraining sets used for each single-state model. Retraining of the base models is needed for reliable metrics.

TwinMP lacks model adaptation novelty. While DynamicMPNN is centered around multi-chain inverse folding - where sequence(s) to be decoded are conditioned on the structures and unmasked sequences of protein binders – TwinMP only works with single chain inverse folding. This is a significant simplification given that alternative states of proteins are often conditioned on other binders.

The ProteinGym [2] is experiment is hard to follow and does not prove a relationship between protein conformations and mutation effects. For a given ProteinGym [2] mutation example, they use as a mutation score the difference between TwinMP logit value of the WT amino-acid at the mutation site and the logit value of the mutant amino-acid (after also adding in the MSA logits and applying softmax). They compare the single-state TwinMP logits correlation against the multi-state TwinMP logits one. However, the way the authors obtain the second conformation for multi-state TwinMP is via a short MD run, which risks to be highly correlated to the WT structure and perhaps just a relaxed version of the WT conformation which removes the experimental artifacts from it. A necessary ablation to support the multi-state logits being more expressive for mutation prediction is to run the ProteinGym [2] correlation of the MD single state TwinMP logits. For the multi-mutation examples, the authors do not account for epistasis, they rather only sum up the scores from each single mutation.

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

[2] Notin, P., et al. (2023). ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design. Advances in Neural Information Processing Systems, 36, 64331–64379.

Questions for authors

The DynamicMPNN metrics are significantly worse than the ones presented in [1], even though the dataset used in [1] is presented as a challenging selection of metamorphic and hinge proteins. Did you try to clarify if the difference comes from your new dataset splitting strategy? What is the reasoning behind choosing this splitting over the one in [1]? These questions directly affect the score of the submission.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.