- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
A-CODE proposes a fully atomic, one-stage protein co-design model that jointly generates atom types and 3D coordinates within a unified multimodal diffusion process, rather than using the conventional two-stage pipeline of backbone generation followed by sequence design. By operating directly at the atom level, A-CODE can naturally support both canonical and non-canonical amino acids. The authors evaluate the model on unconditional protein generation and binder design, showing improved designability and co-designability over existing approaches.
Critical strengths
- A-CODE demonstrates strong co-design performance, achieving 81.2% all-atom and 87.0% Cα-only codesignability.
- The atom-centric formulation enables a natural extension to non-canonical amino acids (ncAAs), without requiring architectural modifications.
- A-CODE achieves strong performance on the binder design benchmark, with its one-stage co-design model consistently outperforming the listed one-stage baseline across all ten targets.
Critical weaknesses
- The binder design experiments do not compare against the recent state-of-the-art baseline such as BindCraft and Proteina-Complexa.
- The ncAA evaluation is relatively limited. Beyond RMSD, it would be important to assess chemical identity (molecular graph) and local geometry (bond lengths, angles, and dihedrals).
- The individual contributions of atom14 representation, dummy atoms, and the AF3-like architecture are not well disentangled. In particular, it remains unclear how much improvement comes from the unified multimodal diffusion itself, when controlling for architecture, parameter count, and training data.
- Although A-CODE achieves higher co-designability, its structural diversity is substantially lower than La-Proteina (70/65 vs. 206/129 structural clusters), raising the concern that its high co-designability may partly benefit from a narrower structural distribution.
Questions for authors
- How are ncAAs sampled during inference? Are their identities, positions, and copy numbers entirely emergent, or are they explicitly controlled? How is the foldability of the resulting ncAA-containing proteins evaluated? Could AF3 be used as an additional structure–sequence self-consistency check?
- La-Proteina shows a clear designability–diversity trade-off by varying sampling noise. Does A-CODE have a comparable control over sampling stochasticity, and how does it affect designability, diversity, and sequence distribution?
- A-CODE adopts RDM to remask atom types during sampling. How much does RDM contribute to the final performance compared with standard mask diffusion?
3 comments
Rebuttal to Reviewer tCjZ, Part 2/2
Q3 Gains from Unified Multimodal Diffusion
We do not claim atom14 as one of our contributions, as we already stated in the original manuscript that it was first used in AF2 and has since been adopted in recent design frameworks such as RFDiffusion3 and BoltzGen. In this way, we frame our major contributions as: a unified atomic framework in which both discrete (atom type) and continuous (atom coordinate) variables are denoised simultaneously as a "one-stage" co-design model, in contrast to most two-stage models. Specifically, the baselines PXDesign, RFDiffusion3, and BoltzGen are all built on an AF3-like architecture (middle in Figure 1). Therefore, adopting the same atom-transformer-based trunk, our architecture is comparable to baselines. This is further demonstrated by the sampling time comparison in Table 3, where A-CODE achieves comparable or better sampling efficiency.
In this way, we can better justify that most of the performance gain over similar AF3-like design models is due to our unified multimodal diffusion. The coordination techniques we proposed, including side-chain lagging that prioritizes backbone denoising over side chains and conservative discrete decoding that prioritizes the structure modality over discrete types, are also integral to our model design and help multimodal diffusion perform better on practical protein design tasks.
Q4 Diversity and Designability Tradeoff
While we agree that A-CODE generates less diverse structures, we note that it achieves the best sequence diversity. These results together demonstrate A-CODE tends to explore more plausible sequences even within a small set of structures. We have also quantatitively analyzed this effect in Table 9,where for each structure cluster, A-CODE generates 4.19 diverse sequences compared to La-Proteina's 1.57. The sequence identity within each structure cluster is also significantly lower (0.410 compared to La-Proteina's 0.872). In this way, we interpret the tradeoff as focusing more on structure diversity vs. sequence diversity.
We also thank you for pointing out the designability-diversity tradeoff, as also demonstrated in previous works such as La-Proteina. For AF3-like architectures, including our A-CODE, the noise scale serves as low-temperature structural sampling. Indeed, previous work like PXDesign has already shown that using a higher promotes designability but reduces diversity due to low-temperature sampling. As our A-CODE is built upon the same trunk as PXDesign, we expect a similar designability-diversity tradeoff.
Q5 RDM Constribution
Adopting RDM and conservative decoding primarily postpones the determination of discrete atom types unless the model is very confident. This strategy prioritizes accurate structure denoising to prevent early commitment to wrong atom types. In practice, about 15%-20% of the designable tokens are unmasked during the last step. In early model development, a uniform time schedule (i.e., each designable token is masked with probability ) tended to produce more unknown amino acids because of errors in predicted atom types due to strict atom-name matching.
Thank you for the additional experiments and results during the rebuttal. My concern about baseline is resolved however there's no more quantitative results regarding noise scale , RDM constribution and carefully ablations, so I maintain my original score.
Rebuttal to Reviewer tCjZ, Part 1/2
We thank you for your recognition of our strong co-design performance and atom-centric formulation that enables natural extension to ncAA modeling. We will further address your remaining concerns as follows.
Q1 Additional Baselines
Thank you for suggesting additional, more recent baseline models to better support our claims. To better compare against the one-stage co-design models, we emphasize diffusion baselines without additional redesign stages. We compare Proteina-Complexa (from co-denoised latents) and BoltzGen (from its structure-sequence mapping), which use self-generated sequences, using the identical length-deterministic sampling protocol and the same AF2-IG-Easy filter. The results are summarized as follows.
These two baselines offer stronger co-design performance. Nonetheless, A-CODE still outperforms it on 6/10 tasks. Remarkably, Complexa completely fails H1, SC2RBD, and TNFa, which are exactly the hardest targets. By contrast, our A-CODE leads by a large margin on these targets. As Complexa adopts a different latent-based co-design approach, these results show that A-CODE's atomic design and Complexa's offer complementary perspectives on binder design.
BindCraft uses hallucination to directly back-propagate gradient information into random logits with a carefully engineered four-stage optimization. Due to limited time and resources, and BindCraft's extremely long runtime, we report success rates evaluated on one sample per length instead of 8. This comparison is highly unfair to diffusion-based models because of the mismatched computational cost. As a reference, BindCraft requires 1836 seconds in average for a single generation, peaking 5904 seconds for the longest H1 target, whereas A-CODE and all other diffusion baselines range from 15 to 25 seconds. This leads to almost 100x more compute for BindCraft. Under a fixed compute budget, this would mean BindCraft achieves fewer total successes than most baselines, including our A-CODE. Despite the drastic computational cost, A-CODE still outperforms BindCraft on the hardest two targets, H1 and TrkA, as well as BHRF1.
Q2 ncAA Evaluation and Sampling Protocol
To allow flexible generations, we never enforced any constraints on the generations for the ncAA fine-tuning experiments. That is, we identify ncAA types using an exact graph isomorphism test, with element types identified by the discrete track after stripping dummy atoms. In this way, types purely emerge in a data-driven manner. The only additional post-processing stage is chirality identification of Ca to identify D-amino acids.
While we certainly agree that AF3-like cofolding models can handle ncAAs in the prediction, to the best of our knowledge, there has been no widely agreed evaluation protocol for ncAA-containing generations. We therefore leave such in-depth exploration for future work, as our ncAA results serve more as a proof-of-concept demonstration that no previous work has done.
We have already noted these limitations in our original manuscript in Appendix A.2, and we therefore treat our exploration of ncAA modeling as preliminary about how A-CODE can be readily adapted to ncAA modeling without changing its architecture. This flexible design differentiates A-CODE from existing co-design models such as BoltzGen or Protpardelle-1c and two-stage models that hard-code 20 standard AAs. We hope our approach can inspire future exploration in this domain with better generative frameworks.