acceptodds
Withdrawn from ICLR 2027

PLDesign: Unifies All-Atom Complex Representation and Geometry-Guided Protein Design

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Modeling protein complexes requires coordinating atomic detail with cross-chain geometry, placing substantial demands on generative models of interacting systems. Autoencoder-based structural compression offers a way to separate atomic reconstruction from distribution learning, opening the possibility of modeling these coupled relationships in a compact space. We introduce PLDesign to investigate this direction through joint encoding of complete all-atom complexes. An Atom14 autoencoder learns a shared representation that preserves intrachain structure and interchain organization, while latent perturbations reveal a correspondence with decoded geometric variation. A downstream latent flow model uses this representation for unconditional backbone generation and target-conditioned binder design. Geometry-guided trajectory search further makes the latent information actionable during sampling, ranking intermediate states before coordinate decoding or external structure prediction. This search improves unconditional designability and enriches candidates in a matched computational binder screen, whose retained complexes contain locally packed interfaces with potential value for interface seed development. PLDesign connects joint complex encoding with latent-guided exploration, opening a route to studying and designing interacting molecular systems through compressed structural representations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Binder design, AutoEncoder, Latent-based generation model

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ozmjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PLDesign proposes a model that projects the all-atom protein complex structure into a latent space. The latent space can be used for unconditional protein generation as well as target-conditioned binder design. A search algorithm applied in the latent space helps to better steer the generative process during inference time.

Critical strengths
  1. Training an autoencoder to compress all-atom complex structure into latent space and further training flow-based generative model over the latent space is a novel and interesting exploration in the field of protein generative modeling.

  2. Broad evaluation task coverage, spanning from reconstruction (to check latent quality) and de novo monomer/binder generation. Strong reconstruction performance.

  3. Novel idea of search within the latent space.

Critical weaknesses
  1. Metrics used for each task are not comprehensive, missing some important metrics. See details in Questions.

  2. Relatively weak performance compared with baseline in generative tasks.

Questions for authors

Questions regarding model evaluation:

  1. The table 1 shows strong dimer reconstruction performance of the PLDesign, which is very good. Besides the geometrical metrics such as RMSD, lDDT, and DockQ, can the authors also evaluate the side-chain quality? Things like bond length, dihedral angle distribution, or a combined metric such as MolProbity would make the reconstruction evaluation more comprehensive for side-chain quality.

  2. In table 2, the authors benchmark PLDesign with other baselines on unconditional monomer design. For monomer generation with ProteinMPNN+ESMFold eval pipeline, related works usually report "designability" along with "diversity" measured using Foldseek. Why did the authors choose to report DPT instead of regular diversity? Also, given the PLDesign is an all-atom generative model, can the authors also evaluate the co-designability of the model to see its native sequence generation capability?

  3. The in-silico success rate of binder design filtered by AF3 scores seems to be very low for PLDesign, compared with other baselines. Proteina-Complexa is not even included in the baselines in Table 23, let alone other hallucination-based methods such as BindCraft. Even though the AF3 scores do not correlate perfectly with wet-lab success, they are the best surrogate tools we can use to filter for actual binders. I can hardly say PLDesign, at the current stage, is a good model for binder design. Maybe the authors can try to further improve it by scaling up the training data or model size?

  4. How does the PLDesign model's performance (reconstruction, de novo generation) scale with the size of the complex?

Questions about modeling:

  1. From the architecture design, it seems to me that multimers with more than 2 chains can be encoded into latent space. However, the authors only trained on complexes with no more than 2 chains. Is there any specific reason that the authors did not try training and evaluating on samples with more than 2 chains? They are important test cases for multimer reconstruction and multi-chain target binder design tasks.

  2. The search score lat_self_lddt functionally trims diffusion trajectories with "outlier" lddt value in latent space, which makes it de-facto similar to low-temperature sampling. How does it compare to simply using lower temperature during the sampling process?

Reviewer_WyULOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces PLDesign, a protein binder generative model which operates in a fully latent space. Previous all-atom generative models can be categorized as fully explicit, or partially latent. Models like RFDiffusion3, PXDesign, or BoltzGen are fully explicit and represent all atoms, whereas models like Protein-Complexa only model the backbone Ca explicitly, and use an autoencoder latent to represent the side chains and residue types. Motivated by the computational savings of latent approaches, PLDesign instead fully operates in the latent space of an all-atom complex autoencoder. The model supports target template and hotspot conditioning, and they also introduce scores on the latents to enable test-time search in the latent space without the need to decode to an all-atom structure.

Critical strengths

None

Critical weaknesses

The main motivation for this approach is quite weak. The computational savings between a fully-latent approach and a partially latent approach are small, since the number of tokens that are handled in each case is the same. There are also several downsides to the latent approach, since the latents are less interpretable, and therefore more difficult to steer, and there is extensive literature about the difficulties of learning a "diffusable" latent space. The partially latent approach, in contrast, retains the explicit meaning of the backbone geometry with most of the computational savings from collapsing the multiple sidechain atoms into a single latent. Moreover, the empirical results are poor. Monomer generation designability is only comparable to Proteina-Complexa with the use of search (74% vs. 71%, and down to 53% without search), while in the more important binder design evaluations PLDesign only has a 3.15% success rate, against 32.29%, 60.95% and 41.76% for BoltzGen, PXDesign and RFDiffusion3 respectively. There is no comparison to Proteina-Complexa, one of the state-of-the-art methods and the main partially-latent baseline. Finally, the paper is very poorly written. Many paragraphs are vague and confusing, lack detail, refer to terms that have not been defined, and defer any description of the actual method to the appendix.

Questions for authors

The authors are welcome to rebut the main criticisms outline in critical weaknesses. In addition, I have the following questions:

  • Did the authors compare the performance of their method to Proteina-Complexa for binder design, or why not? There is a comparison for unconditional generation.
  • Given that you evaluate PLDesign with search, why not also compare to Proteina-Complexa with its own search methods under a fixed compute budget?
  • Where did the latent scores come from in Table 17? It looks as if you treat the latents z like coordinates, computing quantities like angles and distances between them. So are your latents just encoding the backbone coordinates?
Reviewer_54NzOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

PLDesign proposes a continuous autoencoder that jointly encodes all-atom protein dimers into per-residue latent representations, aiming to preserve both intrachain structure and interchain geometry. A downstream flow-matching model generates structures in this latent space for unconditional backbone generation and target-conditioned binder design. The method also uses latent-space geometric scores to guide trajectory search without decoding coordinates or running external structure predictors during ranking. Selected outputs are decoded into backbone candidates, followed by ProteinMPNN sequence design and AF3 evaluation. The paper reports strong complex reconstruction accuracy and improved computational designability with latent search, although binder screening success remains substantially below the compared baselines.

Critical strengths
  • The motivation is valid: generating protein complexes entirely in a continuous latent space could potentially improve both structural accuracy and computational efficiency. The paper makes a meaningful effort to develop an all-atom complex representation, with reconstruction experiments and several representation-level ablations.
Critical weaknesses

Overall, the central idea is to represent all-atom protein complexes using a continuous autoencoder and then generate binders in that latent space. However, I find both the framing and the completeness of the work insufficient. Most of the convincing results concern representation learning, while the protein design component is not developed enough to support the broader claims. I currently lean toward rejection.

  • The title suggests a broader scope than the experiments support. The work mainly studies protein dimers and target-conditioned binder generation, rather than general protein complex design. There is also a mismatch between the emphasis on “all-atom” modeling and the actual generation pipeline: generated structures are used as backbone seeds, followed by ProteinMPNN sequence design and AF3 structure prediction. It is therefore unclear what practical benefit the all-atom representation provides for the demonstrated design task.

  • The dimer reconstruction experiments rely on CASP14 and CASP15. A more recent benchmark, such as CASP16 or a carefully filtered set of PDB structures released in 2026, would provide a stronger test of generalization. Sequence-based filtering alone does not rule out substantial structural similarity between training and test examples. I would like to see explicit structural-overlap analysis. In addition, PLDesign and ProteinAE are trained on different data, particularly with respect to complex structures. This makes it difficult to determine how much of the reconstruction improvement comes from the proposed representation, rather than differences in training data and coverage.

  • Search improves scRMSD-based designability, but it also reduces diversity. This raises the concern that the model or search procedure favors a narrower set of easily refolded structures, potentially dominated by simple alpha-helical/loop-rich scaffolds. The current evidence does not establish whether this is the case, but the possibility deserves direct analysis. Higher self-consistency alone is not enough to demonstrate useful binder design, especially given the low conditional screening yield.

The work feels incomplete as a design paper. The representation component is substantially more developed than the design component. More evidence is needed to connect accurate reconstruction to useful generation and to demonstrate why this representation is preferable for downstream design. In its current form, the paper is more convincing as a study of complex representations than as a protein design method.

Questions for authors

The underlinying strength of a all-atom design should be addressed to the sidechain and according interfaces. why didn't you design such experiments for illustration?

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

Related papers

Discussion (0)

Sign in to comment.