acceptodds
Under review as a conference paper at ICLR 2027

A-CODE: Fully Atomic Protein Co-Design with Unified Multimodal Diffusion

11 views

Abstract

We present A-CODE, a fully atomic unified one-stage protein co-design model that simultaneously refines discrete atom types and continuous atom coordinates. Unlike predominant two-stage methods that cascade structure design with amino acid-level sequence design, our approach is fully atomic within a unified multimodal diffusion framework, in which residue identities are inferred solely from atom-level predictions. Built upon the powerful all-atom architecture, A-CODE achieves superior designability for unconditional protein generation, outperforming all existing one-stage and two-stage design models. For binder design, A-CODE rivals and even outperforms existing state-of-the-art two-stage design models and, compared with the existing one-stage co-design model, achieves a drastic tenfold improvement in success rate on hard tasks. The inherent flexibility of our atomic formulation enables, for the first time, seamless adaptation to non-canonical amino acid (ncAA) modeling. Our fully atomic framework establishes a new, versatile foundation for all-atom generative modeling that can be naturally extended to complex biomolecular systems.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein Design, Generative Model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_tCjZOpenReview ↗
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
  1. A-CODE demonstrates strong co-design performance, achieving 81.2% all-atom and 87.0% Cα-only codesignability.
  2. The atom-centric formulation enables a natural extension to non-canonical amino acids (ncAAs), without requiring architectural modifications.
  3. 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
  1. The binder design experiments do not compare against the recent state-of-the-art baseline such as BindCraft and Proteina-Complexa.
  2. 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).
  3. 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.
  4. 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
  1. 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?
  2. 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?
  3. 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
AuthorsOpenReview ↗

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.

TargetBHRF1H1IL17AIL7RAIRPDL1SC2RBDTNFaTrkAVEGFA
A-CODE (self)22.8755.711.244.0541.6728.7037.507.573.700.96
Proteina-Complexa56.4000.149.8625.4613.030016.557.69
BoltzGen (self)10.06006.6930.405.990.30011.973.02

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.

TargetBHRF1H1IL17AIL7RAIRPDL1SC2RBDTNFaTrkAVEGFA
BindCraft (100x more compute)12/410/8117/8146/7161/8149/8128/411/7149/7117/91

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.

AuthorsOpenReview ↗

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 η\eta serves as low-temperature structural sampling. Indeed, previous work like PXDesign has already shown that using a higher η\eta 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 tt) 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 η\eta, RDM constribution and carefully ablations, so I maintain my original score.

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

The paper proposes A-CODE, a co-design model for protein sequences and all-atom structures that naturally extends to non-canonical residue generation. A-CODE treats atom names and their presence as discrete generative variables, jointly denoises them with continuous coordinates, and infers residue identities through atom-name matching.

The paper evaluates unconditional protein generation, binder design against ten targets, and preliminary ncAA generation after fine-tuning. A-CODE achieves an all-atom codesignability of 81.2% on unconditional generation, substantially outperforming the other baselines.

Critical strengths
  1. A-CODE’s atomic composition representation has independent technical value. Including DMY as a predictable category allows side-chain atom presence, identity, and position to participate jointly in generation. The approximately 0.05% UNK rate for canonical residues indicates that this finer-grained prediction does not lead to a large number of unidentifiable outputs.

  2. The experiments report both co-design performance and designability after ProteinMPNN redesign, demonstrating clear benefits in capturing both the joint and marginal distributions.

  3. The proposed side-chain denoising lag investigates whether avoiding steric clashes early in generation can improve the quality of the final generated structures.

Critical weaknesses
  1. The gains on the unconditional protein generation benchmark may result from a narrower generated distribution. A-CODE obtains a diversity score of only 70, substantially below the strong La-Proteina baseline. This raises the possibility that A-CODE overfits to a limited set of high-quality modes.

  2. The ncAA analysis remains incomplete. Figure 4(b–c) appears to analyze structures generated by A-CODE, but lacks corresponding reference distributions from real proteins. There is also no benchmark assessing the overall quality of generated proteins containing ncAAs. The current analyses are insufficient to determine whether A-CODE exhibits particular ncAA preferences or how fine-tuning changes its behavior relative to the base model.

Questions for authors
  1. In Eq. 1, κt\kappa_t is defined as the probability of masking a real token with XPA, whereas αt\alpha_t in Sahoo et al.’s Simple and Effective Masked Diffusion Language Models denotes the probability of retaining the real token. This results in a negative CE coefficient in Eq. 1, implying that Ltype\mathcal{L}_{\mathrm{type}} encourages reducing the probability assigned to the correct class. Could the authors clarify whether this is an implementation error, a writing error, or a misunderstanding on my part?

  2. The scaling operation in Eq. 2 does not appear necessarily justified, since it scales the clean-coordinate component X0X_0 differently across atoms in XinputX_{\mathrm{input}}. While I understand that this operation can provide the network with inputs on a consistent numerical scale, could the authors provide an ablation with and without the scaling in Eq. 2? Is this operation necessary?

  3. Could the authors discuss in greater depth why the one-stage A-CODE (co-design) still underperforms two-stage methods on 7/10 conditional binder design tasks, and also underperforms A-CODE (PMPNN) on 7/10 tasks? Does this indicate a remaining design issue in A-CODE, or should two-stage methods be expected to substantially outperform one-stage methods on this benchmark?

3 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer xKeR, Part 1/2

We appreciate your high recognition of our work's novel atomic co-design framework, which yields negligible unknown generations while preserving atomic interactions. We further address your concerns as follows.

Q1 Gains and Tradeoff Between Diversity and Designability

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. This is likely due to the diversity-designability tradeoff, as shown by the drastic difference in diversity between the two La-Proteina variants. 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.

Q2 ncAA Evaluation

We understand your concerns about the ncAA evaluation pipeline. To the best of our knowledge, there is no widely agreed evaluation pipeline for ncAA-containing generations. Specifically, the standard binder evaluation pipeline will fail because AF2 refuses to take inputs containing ncAAs. We therefore resort to physical and geometrical evaluations such as interface hydrogen bonds and salt bridges that should contribute to stabilizing the complex, and RMSD compared to the CCD entry for geometric feasibility. In this way, the overall quality you suggested is less well-defined for ncAA evaluation.

Per your request, we additionally calculate the statistics of the interface ncAAs in our curated PDB complexes. 38.2% of the natural interface ncAAs form hydrogen bonds, and 14.0% of them form salt bridges. The median minimum distance to the target is 4.30 Å, which is higher than the median of our distribution. However, this does not mean our generations are better. We acknowledge that one existing limitation is that A-CODE tends to put ncAAs closer to the target, causing potential clashes and inflating interactions (e.g., $<$2 Å). Figure 4 serves more as a preliminary probe for A-CODE's generalizability to ncAAs.

We note that we have explicitly framed our exploration of ncAA modeling as preliminary and more of a proof-of-concept demonstration of 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.

Q3 Discrete Modality Objective

Thanks for pointing this out. This is a typo in our manuscript, and we will revise it. The standard CE loss as in MDLM is used, or equivalently, the weighted negative log-likelihood of the logit denoiser as −κ˙t1−κtlog⁡softmax(ϕθ(At,t))A0-\frac{\dot\kappa_t}{1-\kappa_t}\log \mathrm{softmax}(\phi_\theta(A_t,t))_{A_0}.

Q4 Side-Chain Lagging

The side-chain lagging technique is proposed to address early commitment of side-chain atoms, which hinders overall designability. For backbone atoms, there is no scaling at all during both training and sampling (see Algorithm 1). Scaling is only applied to side-chain atoms to defer their denoising. Therefore, the "no-scaling" variant as you requested is essentially the naive EDM denoising/training of the all-atom structures.

We already have a comprehensive ablation on its impact on both unconditional and binder design. For unconditional generation, side-chain lagging significantly boosts codesignability, leading to more consistent structure-sequence co-designs. For binder design, Figure 5 quantitatively evaluates the per-target impact, which is generally positive but varies across targets.

AuthorsOpenReview ↗

Rebuttal to Reviewer xKeR, Part 2/2

Q5 Discussion on One- vs Two-Stage Models

The gap between one- and two-stage models, even within the same model family, is a well-known issue across existing models. For example, BoltzGen explicitly suggests its inverse-folding sequences despite its co-design capability. Its self-generated sequences do lead to significantly worse designability, as demonstrated in our baseline results in Review NUSH, Q5. RFDiffusion3 also explicitly codes PMPNN redesigns in its official codebase. More recently, Proteina-Complexa, a partially latent flow all-atom flow matching model, also falls by as much as 5% on in silico metrics compared with PMPNN redesigns. Complexa's self-generated sequences are also compared as additional baselines in Review NUSH, Q5.

Nonetheless, we emphasize that in silico successes do not necessarily translate into wet-lab hits, as AF2 is also known to have prediction biases and a high false-positive rate. More direct and concrete evidence supporting all-atom one-stage modeling comes from Complexa. While Complexa’s self-generations are outperformed by PMPNN redesigns on in silico metrics, wet-lab phage display instead shows more successes for the self-generations. This inversion suggests that atomic-level co-design that captures finer-grained interactions may be favored by wet-lab experiments. Therefore, in a similar spirit and with a substantially improved one-stage alternative, we frame A-CODE as a strong and competitive generative co-design paradigm that may further push the limits toward a higher wet-lab success rate.

Thank you for the responses. My concerns about the loss formulation and side-chain scaling are resolved. The additional sequence-diversity analysis and natural ncAA reference statistics are helpful. However, higher sequence diversity does not fully address whether the designability gains partly reflect a narrower structural distribution. The ncAA evaluation also remains preliminary, with potential clashes and limited validation of overall protein quality. I continue to value the atomic co-design formulation and strong canonical co-design results, and therefore maintain my rating of 3.

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

This paper introduces a de novo protein generation framework that operates on an all-atom level. It combines discrete diffusion for predicting atom types with Euclidean diffusion to resolve 3D structure coordinates. Specifically, it represents each residue as a padded, fixed-size token slot of 14 atoms, using discrete diffusion to predict individual atom names, instead of predicting the residue identities directly. Residus are then decoded post-hoc by matching the predicted atom-name sets against canonical chemical templates. The authors compare their method against existing baselines and demonstrate its performance on both unconditional monomer generation and conditional binder design.

Critical strengths
  • The general development of one-stage co-design of proteins, operating on all-atom structures is an important and active area of research.
  • The paper’s methodology is sound and well-explained.
Critical weaknesses
  • The paper mentions that unidentified residues are resolved by a GLY mutation prior to folding with ESMFold. However, this post-processing step can introduce significant evaluation bias, affetcing designability and pLDDT metrics reported in Table 2. Can the authors clarify how many of these mutations occur in the results reported in Table 2 and report metrics without this Glycine mutation?
  • While the authors claim a low TV distance to the amino acid distribution frequencies in PDB in Section 4.1, Figure 3 reveals visible discrepancies in individual amino acid frequencies. Additionally, as unidentified residues are systematically mutated to Glycine, the reported Glycine frequency is naturally inflated by the decoding collapse rate. Can the authors clarify how they explain these divergences, and whether the TVD calculation was performed before or after the post-hoc GLY mutations?
  • In Section 4.1, the authors state that they use 1,000 Euler steps "for fair comparison" following PXDesign. However, this is signficantly different from what baselines reported in Table 2 use: La-Proteina uses N = 400 integration steps, for example. Protpardelle uses 500 steps, but in this benchmark it is evaluated at only 100 steps. Can the authors benchmark the performance aganist a compute-matched budget?
  • The evaluation is missing standard all-atom analyses, specifically atomistic motif scaffolding benchmarks (for example see section C.4 of Bolzgen or 4.2 of La Proteina for motiff scaffolding). Is the fact that residues are identified post-hoc the reason for which motif scaffolidng omitted? Can the authors clarify whether A-Code is capable of executing standard motif scaffolding tasks? Additionally rotamer quality metrics (such as rotamer/Ramachandran outlies, see Fig 14 of La-Proteina)
  • DISCO [1] is a directly relevant and comparable co-design model. Can the authors include benchmark against this baseline?

[1] Rector-Brooks, J., Lambert, T., Skreta, M., Roth, D., Long, Y., Li, Z.-Q., Zhang, X., Cretu, M., Li, F.-Z., Ganapathy, T., Jin, E., Bose, A. J., Yang, J., Neklyudov, K., Bengio, Y., Tong, A., Arnold, F. H., & Liu, C.-H. (2026). General Multimodal Protein Design Enables DNA-Encoding of Chemistry. https://arxiv.org/abs/2604.05181

Questions for authors

See above

2 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer NUsH, Part 1/2

We sincerely thank you for your acknowledgement of our paper's sound and well-explained methodology for addressing the protein co-design problem of practical significance. We will address your remaining concerns as follows.

Q1 Gly Default Impact

We understand your concerns regarding the glycine substitution, and we justify our choice as follows. First, we noted that unknown is extremely rare in our generations, about 0.05% of the total amino acids. Even if we mark every unknown-containing sequence as a failure (some of which still pass the in silico filter), A-CODE's codesignability remains 84.2%, with a large margin over existing co-design models. Other designability metrics remain unchanged, and A-CODE remains the best.

Additionally, such a substitution is a consistent pipeline, as we never hack toward any target or generative setup. Therefore, the same protocol can be readily applied to any binder design tasks. The PMPNN@1 and PMPNN@8 designability metrics use redesigned sequences and, by definition, do not use A-CODE's generated sequence. Therefore, these metrics mostly reflect structural fidelity and are not affected by UNK. Our model outperforms baselines on both.

Redesigning the sequence is a common step in protein design. In fact, all two-stage models can be considered to deliberately alter every amino acid. Co-design models like BoltzGen also explicitly prefer inverse folding results despite already supplying self-generated sequences. Therefore, substituting with glycine does not grant our model unfair advantages. We do agree that there are other better approaches, such as closest AA by graph matching or PMPNN partial redesign. We will leave these options as future explorations, as our model already outperforms existing baselines on unconditional benchmarks.

Q2 TVD to PDB AA Distribution

Figure 3 results convert unknowns into glycines. However, as we have explained in the previous section, the number of unknowns identified as glycine is extremely rare, about 0.05% in the unconditional generation, leading to negligible impact on the AA distribution, which reaches about 5% for G. Therefore, the TVD reported faithfully demonstrates that A-CODE's AA distribution resembles PDB/AFDB more.

Compared to La-Proteina, A-CODE significantly increases the occurrence of rare amino acids (H, M, W, and C), as post-hoc amino acid type inference suffers more from training data bias. A-CODE splits such a data bias into per-position objectives across the 10 learnable XPA tokens, therefore better capturing the long-tail distribution. Meanwhile, we do acknowledge that A-CODE still holds a preference for I, K, and N, probably because these AAs share the most frequent number of sidechain atoms (4 and 5).

Q3 Compute-Matched Comparison

We agree that computational efficiency is as important as final design results. We have already included a sampling time comparison for the unconditional generation setup in Table 3, where our proposed A-CODE, despite using more steps, achieves comparable sampling time to La-Proteina without triangular update and is significantly faster than the version with triangular update. These quantitative results demonstrate the efficiency of our model and further support the fair comparison with the compute you requested.

For Protpardelle-1c, we apologize for the typo in our manuscript. We used the official codebase, which defaults to 500 diffusion steps. Rerunning all generations, we get new results as:

TargetBHRF1H1IL17AIL7RAIRPDL1SC2RBDTNFaTrkAVEGFA
SR%3.660.27000.318.801.2202.290.82

which is close to our reported success rates in Table 4. We will correct this error.

Q4 Motif Scaffolding Benchmark

We acknowledge that motif scaffolding is another important design task. In principle, we can adapt A-CODE for motif scaffolding by retaining only the motif structure and AA types as conditions and co-designing the scaffold. However, this requires additional training, as the existing pipeline includes only monomers and multimers, without additional motif-scaffolding training as in BoltzGen and La-Proteina. Additionally, La-Proteina also trains a separate model specifically for motif scaffolding, rather than doing zero-shot inference or inpainting with the unconditional model. Therefore, due to time constraints, we defer this task to future work.

AuthorsOpenReview ↗

Rebuttal to Reviewer NUsH, Part 2/2

Q5 Additional Baselines

We thank you for pointing out related models and additional baselines. While DISCO adopts a similar co-design framework for simultaneously designing protein structures and sequence types, its current repo only supports small ligand, DNA, and RNA design, as it is specifically optimized for these tasks. It does not currently support protein binder design.

To provide more concrete evidence, we provide additional binder design baseline results of the more recent Proteina-Complexa model, a partially latent flow matching model that simultaneously denoises CA positions and per-residue latents on which the AA types are inferred post-hoc. BoltGen with self-generated sequences as a co-design model is also included for comparison. We use the same length-deterministic setup and the AF2-IG-Easy filter to report the results as follows:

TargetBHRF1H1IL17AIL7RAIRPDL1SC2RBDTNFaTrkAVEGFA
A-CODE (self)22.8755.711.244.0541.6728.7037.507.573.700.96
Proteina-Complexa56.4000.149.8625.4613.030016.557.69
BoltzGen (self)10.06006.6930.405.990.30011.973.02

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. Our A-CODE, by contrast, 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.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.