acceptodds
Under review as a conference paper at ICLR 2027

AtomBloom: All-Atom Branching Flows for 3D Binder Design

Abstract

Designing target-specific small molecules, peptides, and antibodies requires generating atomistic 3D structures whose cardinality, atom types, and geometry are jointly adapted to the target protein environment. Existing atom-level diffusion and flow models typically fix atom count throughout generation, while hierarchical latent models determine residue or fragment types before all-atom decoding. We introduce AtomBloom, a pocket-specific generative framework that directly generates ligand atoms and allows them to split or disappear along the generative trajectory. This formulation jointly generates atom counts, types, and 3D geometry within a unified framework for small molecules, peptides, and antibodies. For practical generation, AtomBloom combines exact interval event probabilities with dynamic rescaling that adapts the predictor's coordinate scale along the trajectory. We further introduce model-rolled self-conditioning during training and propagate previous predictions along branching lineages during inference. Across small-molecule, peptide, and antibody benchmarks, AtomBloom achieves the best overall ranking on small molecules. For peptides and antibodies, AtomBloom-R adaptively generates each residue’s atoms at fixed residue count and improves sequence recovery, structural accuracy, and binding scores. Further adapting residue count yields additional binding gains, with favorable peptide designs occurring at both longer and shorter lengths than the reference. These results support treating binder cardinality as a design variable alongside atom types and geometry.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Structure-based binder design, Branching flows, Variable-size generation, Geometric deep learning, Molecular generative modeling

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_34zNOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

Built on Branching Flows, AtomBloom is a pocket conditional framework that allows for the generation of binders such as small molecules, peptides and antibodies with the possibility of making atoms to split and disappear during the generation trajectory. To make the atoms appear, split and disappear they introduced: a strategy to delete or split atoms (see Eq.15, Eq.16, Eq.18), a dynamic coordinate rescaling (Eq.19) to compress coordinates initially and restore them gradually and a self conditioning strategy to make the model see the lineage of the previous atom state (see Eq.20). The authors benchmark on LNR peptides, RAbD antibodies, and CrossDocked small molecules.

Critical strengths

Finding the right molecule and cardinality of a molecule is difficult and relevant problem to solve. The paper provide an interesting solution treating the cardinality as a jointly generated with the molecular structure (see Summary). The formulation is rigorous, even though inconsistent in some instances and difficult to understand (see Questions and Weaknesses for details). I find the benchmarks relevant even if some relevant metrics seems to be missing (see Questions and Weaknesses for details).

Critical weaknesses
  • The method doesn't have explicit control on the final atom count and atoms emerge during the generative process. This can be a drawback in certain use cases where the size need to be constrained a priori.

  • The related work section is too synthetic and doesn't provide depth and breadth on how the method is positioning. You can see my questions on this point below.

  • The method part is quite hard to read, with some inconsistencies. See my questions related to the method part below.

  • Several relevant metrics in the experiments are missing. This makes it really hard to understand the advantages of the method against competitors. I am listing here some relevant ones missing. For small molecules, chemical validity (RDKit chemically valid), fraction of fragmented molecules, uniqueness (fraction of distinct SMILES among valid outputs), novelty (fraction of not in training set). For peptides, Ramachandran plot, C$\alpha$ chirality correctness.

  • Recent methods for molecular generation are not referenced in the paper ie. [1]. I leave the authors to find others.

  • The paper doesn't provide inference time comparisons against the baselines.

[1]. Cremer et al. (2025) FLOWR: Flow Matching for Structure-Aware De Novo, Interaction-and Fragment-Based Ligand Generation. https://arxiv.org/abs/2504.10564

Questions for authors
  • There are several issues in the reference section, ie inconsistent citations format, missing url. I give here two examples, and the authors can fix the rest.

    Example 1

    Xiangzhe Kong, Yinjun Jia, Wenbing Huang, and Yang Liu. Full-atom peptide design with geometric latent diffusion. arXiv preprint arXiv:2402.13555, 2024

    They cite this paper as arXiv, despite being a NeurIPS 2024 (https://neurips.cc/virtual/2024/poster/95784). Please, always reference to the full pubblication if the paper was accepted.

    Example 2

    Xiangzhe Kong, Zishen Zhang, Ziting Zhang, Rui Jiao, Jianzhu Ma, Wenbing Huang, Kai Liu, and Yang Liu. Unimomo: Unified generative modeling of 3d molecules for de novo binder design. 2025. URL https://arxiv.org/abs/2503.19300.

    Here, another example of a paper accepted at ICML 2024 (https://icml.cc/virtual/2025/poster/45639). Despite both listed references should be linked to where they were pubblished, the point here is different. The two ArXiv you cited follow for some reasons different formats, which should not be the case.

  • Questions method.

    • I think that there is a notation mismatch between Eq. 46 and Eq. 18.

    • Why isn't θ\theta referenced in the proposition of Eq.18?

    • I would avoid to use in Eq.13 "BCEWithLogits" to refer to the operation.

    • Can you provide more details on how this term Γ is sampled ? Here an illustration referenced in the appendix could really help to understand how is sampled.

    • In several occasion things are explained in the Appendix. For example, Eq.19 uses αt\alpha_t, βt\beta_t and γt\gamma_t defined in Appendix F.

  • Related works questions.

    • The lack of a priori control over output size is a limitation for applications requiring fixed-size binders. Could you expand it in the related works to discuss this?

    • Can you expand the discussion of the different existing methods in the related works? Currently, most of the related works are in the appendix and very little discussion of the paper positioning is provided in the main paper.

    • Can you better highlight the novelty and differences from [1]?

    • Can you discuss how the method position against autoregressive 3D molecule generation models or string based generators + docking like REINVENT4 [2] or PepINVENT [3]?

  • Can you add the metrics I mentioned in the weaknesses?

  • Can you provide inference time comparisons against the baselines?

[1]. Billera et al. Branching flows: Discrete, continuous, and manifold flow matching with splits and deletions.

[2]. Loeffler et al. Reinvent 4: Modern AI–driven generative molecule design.

[3]. Geylan et al. PepINVENT: generative peptide design beyond natural amino acids.

11 comments
AuthorsOpenReview ↗

Response to Reviewer 34zN (1 of 5)

We thank the reviewer for the valuable comments, and we answer the reviewer's questions as follows.

W1: The method doesn't have explicit control on the final atom count and atoms emerge during the generative process. This can be a drawback in certain use cases where the size need to be constrained a priori. The lack of a priori control over output size is a limitation for applications requiring fixed-size binders. Could you expand it in the related works to discuss this?

Thank you for raising this use case. AtomBloom supports both prescribed and adaptive binder sizes. AtomBloom-R already supports a prescribed residue count for peptides and antibodies. In this revision, we additionally provide a constrained sampler for applications requiring an exact atom count.

Prescribed residue counts for peptides and antibodies. AtomBloom-R takes the peptide or CDR length as input and generates residue identities and all-atom structures at that length. This supports applications with specified length requirements and allows direct comparison with existing baselines under the same residue-count constraint. Under this setting, AtomBloom-R achieves strong structural recovery and binding-score improvement, as shown in Tables 1 and 2.

Adaptive size for small molecules. When atom count is not specified in advance, molecular size is itself a design variable that depends on the target pocket. The baselines in Section 4.3 also produce molecules of different sizes. Some select atom count from a size prior before generating the structure, while autoregressive methods determine size through successive additions and a stopping decision. AtomBloom incorporates this choice into the atom-level generation process, allowing atom count to evolve together with atom types and coordinates. Under the shared benchmark protocol, AtomBloom achieves the best overall ranking among the 15 methods in Table 3.

Exact atom-count control added in this revision. To support applications requiring an exact atom count, we additionally provide a constrained sampler that enforces the user-specified count at inference time without retraining. It conditions the sampling of split and deletion events on the requested final count, using the model’s existing predictions without additional network evaluations. We have added the algorithm and its evaluation to Appendix T.

We evaluated the constrained sampler on all 100 CrossDocked2020 test pockets with five attempts per pocket, setting the target heavy-atom count N to that of the reference ligand. The free and constrained runs use the same checkpoint, settings, and random seeds. We also tested user-specified counts of 20, 30, and 40 heavy atoms, with four attempts on each of eight pockets for each count. Validity and single-fragment rates are calculated over all attempts.

SamplerAtom count = NMean |n − N|ValidSingle fragment
Free45 / 5004.2899.0%98.8%
Constrained, N = reference size500 / 5000.0099.2%98.8%
Constrained, N ∈ {20, 30, 40}96 / 960.00100.0%100.0%

The constrained sampler meets the requested atom count in every tested attempt, while maintaining validity and single-fragment rates comparable to free sampling. We have also expanded Section 2 to clarify the distinction between selecting size before generation and adapting it during generation.

AuthorsOpenReview ↗

Response to Reviewer 34zN (2 of 5)

W2: The related work section is too synthetic and doesn't provide depth and breadth on how the method is positioning. You can see my questions on this point below. Can you expand the discussion of the different existing methods in the related works? Currently, most of the related works are in the appendix and very little discussion of the paper positioning is provided in the main paper. Can you better highlight the novelty and differences from [A]? Can you discuss how the method position against autoregressive 3D molecule generation models or string based generators + docking like REINVENT4 [B] or PepINVENT [C]?

Thank you for these questions. We have rewritten Section 2 to explain the central comparisons directly in the main paper. It now covers atom-level and hierarchical binder generation, variable-dimensional generation, and molecular string-based design and optimization. Appendix A expands the mechanisms and modeling trade-offs.

Differences from Branching Flows. AtomBloom inherits the splitting and deletion dynamics and conditional generator-matching objective of Branching Flows [A]. We develop this framework into a pocket-conditioned, all-atom binder-design method that supports adaptive atom and residue counts. The table below distinguishes the inherited framework from the task requirements and our additions.

What Branching Flows providesWhat binder design requiresWhat AtomBloom adds
Splitting, deletion, and element-state evolutionGenerating atom types and coordinates conditioned on a target pocketPredicts atom states, splitting, and deletion from the current atom set and target pocket for small-molecule, peptide, and antibody generation (Sec. 3.3)
Continuous-time event dynamicsAccurate event probabilities under finite-step samplingComputes event probabilities over each sampling interval and analyzes the atom-count bias introduced by the original approximation (Sec. 3.4)
Continuous-state evolutionHandling different coordinate scales across stages of generationAdjusts the coordinate scale seen by the predictor, with stronger compression early in generation and the original scale at the endpoint (Sec. 3.5)
Ancestor–descendant correspondenceReusing earlier predictions when the atom set changesPasses earlier predictions from ancestors to descendants and trains the model on states influenced by its own predictions (Sec. 3.6)
Variable-cardinality generationDetermining residue count and types from an evolving atom setPredicts how atoms are grouped into residues and each residue’s type, then adjusts atom counts and assigns atom identities and connectivity (Sec. 3.3, Appendices E and G)

Figure 2 and Tables 6–9 provide the sampling, rescaling, and self-conditioning analyses.

Positioning against autoregressive 3D generators. Autoregressive methods such as 3D-SBDD, Pocket2Mol, and GraphBP also support variable-size molecular generation. The distinction is how size and geometry evolve during generation. Autoregressive methods construct molecules through successive decisions conditioned on the preceding structure. AtomBloom jointly updates the evolving atom set while adding or deleting atoms, allowing existing coordinates and molecular size to adapt together to the pocket. These autoregressive baselines are included in Table 3, where AtomBloom achieves the best overall ranking among 15 methods.

Positioning against string-based generation with docking. REINVENT 4 and PepINVENT provide flexible chemical generation and property optimization through string representations. REINVENT 4 supports configurable scoring objectives, including docking, while PepINVENT extends peptide design to non-natural amino acids. In these string-generation-plus-docking workflows, the pocket structure is not explicitly provided as input to the generator, and the final 3D geometry is constructed by a separate component. AtomBloom instead generates atom types and 3D coordinates directly in the target pocket while adapting atom count. These approaches emphasize different capabilities. String-based frameworks facilitate optimization against configurable molecular objectives, whereas AtomBloom focuses on jointly learning composition, size, and geometry from structural data.

[A] Billera et al. Branching Flows: Discrete, Continuous, and Manifold Flow Matching with Splits and Deletions. NeurIPS, 2026.

[B] Loeffler et al. Reinvent 4: Modern AI-driven generative molecule design. Journal of Cheminformatics, 16, 20, 2024.

[C] Geylan et al. PepINVENT: generative peptide design beyond natural amino acids. Chemical Science, 16, 8682–8696, 2025.

AuthorsOpenReview ↗

Response to Reviewer 34zN (3 of 5)

W3: The method part is quite hard to read, with some inconsistencies. See my questions related to the method part below.

Thank you for the careful reading. We have revised the method section to make the notation consistent and introduce definitions and assumptions where they are first used.

I think that there is a notation mismatch between Eq. 46 and Eq. 18.

The two equations refer to the same quantities. Above Eq. 17, we define

\qquad d_i=d_{\theta,i}(X_s,c,s).$$ Eq. 18 previously used the parameterized symbols, while Eqs. 45–46 used this shorthand. We now use the shorthand consistently throughout Section 3.4 and Appendix B. > **Why isn't $\theta$ referenced in the proposition of Eq. 18?** **The dependence on $\theta$ is already included in the shorthand definitions $\lambda_i=\lambda_{\theta,i}(X_s,c,s)$ and $d_i=d_{\theta,i}(X_s,c,s)$ introduced above Eq. 17.** Proposition 1 uses these abbreviations, so $\theta$ is not written explicitly. > **I would avoid using “BCEWithLogits” in Eq. 13.** **We have replaced the implementation-specific name with the explicit binary cross-entropy**, $$\mathcal L_{\mathrm{del}}=\mathbb E\left[-d_i(\Gamma,t)\log d_{\theta,i}-\bigl(1-d_i(\Gamma,t)\bigr)\log\bigl(1-d_{\theta,i}\bigr)\right],$$ where $d_{\theta,i}=\sigma(\hat d_{t,i})$, as defined in Eq. 8. The loss itself is unchanged. > **Can you provide more details on how Γ is sampled? An illustration could help.** $\Gamma$ specifies a sampled branching history that connects an initial set of particles to the observed ligand atoms. It contains the branching forest, event times, and base-process endpoints needed to construct a training state and its supervision. It is used to generate training examples and is neither provided to the network nor required during inference. **We have added a step-by-step description and an illustration in Appendix C.1.** Given a training ligand, the construction proceeds as follows. 1. *Add auxiliary leaves.* The observed ligand atoms form the terminal leaves. Auxiliary leaves are added near existing atoms and marked for deletion, providing supervision for the deletion process. 2. *Construct the branching forest backward.* Adjacent nodes are repeatedly merged into common ancestors. Full AtomBloom uses one root. AtomBloom-R retains one root per residue, with merging restricted to each residue. A leaf’s anchor contains its coordinate and atom-type encoding. An internal node’s anchor is the average of its descendant leaf anchors, including auxiliary leaves. 3. *Sample split times.* For a forest with $m$ internal nodes, draw $m$ times from the split-time distribution and sort them. 4. *Assign split times to internal nodes.* Assign the sorted times in a sampled topological order, ensuring that each internal node splits after its parent. 5. *Sample deletion times.* Assign deletion times to auxiliary leaves according to the deletion-time schedule. A cut through this history at time $t$ determines both the active particles and their training targets. OU bridges provide the coordinates and continuous atom-type states of the active particles, forming $X_t$. Each active node $v_i(t)$ has three supervision targets, $$b(v_i(t)),\qquad \lambda_i(\Gamma,t)=|D(v_i(t))|-1,\qquad d_i(\Gamma,t)=\mathbf 1\\{v_i(t)\text{ is an auxiliary leaf}\\}.$$ These specify its endpoint anchor, remaining split count, and deletion indicator, where $D(v)$ contains the terminal and auxiliary leaves below $v$. A worked example with four terminal atoms and one auxiliary leaf is given in Appendix C.1 and illustrated in Figure 4. > **Several quantities are defined only in the appendix, including the coefficients in Eq. 19.** **We have moved the definitions of $\alpha_t,\beta_t,\gamma_t$ directly above Eq. 19 in Section 3.5.** > **W4: Several relevant metrics in the experiments are missing. This makes it really hard to understand the advantages of the method against competitors. I am listing here some relevant ones missing. For small molecules, chemical validity (RDKit chemically valid), fraction of fragmented molecules, uniqueness (fraction of distinct SMILES among valid outputs), novelty (fraction of not in training set). For peptides, Ramachandran plot, C chirality correctness.** > **Can you add the metrics I mentioned in the weaknesses?** Thank you for the suggestion. **We have added chemical validity, connectivity, uniqueness, and novelty metrics for small molecules, together with Ramachandran plots and Cα chirality analysis for peptides, to Appendix P.** These metrics provide a more complete assessment of the generated candidates.
Reviewer_94wQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper extends Branching Flows for ligand and binder design, with pocket conditioning and numerical changes to sampling such as using exact interval event probabilities, dynamic rescaling, and self conditioning. The main contributions seem to be these numerical changes which authors ablate for, and application to pocket conditioned small molecule, peptide and antibody design.

Critical strengths
  • Pocket conditioned generation and small molecule, peptide and ab are important application spaces for variable length design. The paper targets an important area of application.
  • For some of the introduced numerical changes to the sampling, e.g. dynamic scaling, there are ablations to show they may improve sampling success. The work also spans a wide range of modalities.
Critical weaknesses
  • Although there is some modest success in small molecule task, in peptide task the diversity is significantly low and this would limit the usefulness of the model. Authors acknowledge that diversity suffers mostly due to self conditioning but does not seem to see it as severe as this reviewer. Since the model novelty is already limited, a strong application result would have made a difference.
  • Expected to see more modern peptide or antibody comparisons, the baselines seem to be outdated.
  • Expected to see more references related to other variable length design models, and their use in binder design, or discussion about theoretical reasons why one model may be better than other.
Questions for authors
  • would the authors consider adding more baselines for the peptide and antibody tasks, eg bindcraft or boltzgen or similar generation?
  • diversity is an important metric for biologics. The authors train on the very diversity-limited SabDab. What strategies the authors would suggest for increased diversity without sacrificing other metrics?

EDIT post rebuttal: increasing score based on added baselines and how the model stands positively compared to those.

7 comments
AuthorsOpenReview ↗

Response to Reviewer 94wQ (1 of 3)

We thank the reviewer for the valuable comments, and we answer the reviewer's questions as follows.

W1: Although there is some modest success in small molecule task, in peptide task the diversity is significantly low and this would limit the usefulness of the model. Authors acknowledge that diversity suffers mostly due to self conditioning but does not seem to see it as severe as this reviewer. Since the model novelty is already limited, a strong application result would have made a difference.

Thank you for emphasizing this. Lower overall diversity raises the concern that users may obtain fewer distinct candidates suitable for further testing. To examine this directly, we counted distinct designs meeting the same binding-score and geometry criteria under a fixed sampling budget. We also tested whether adjusting self-conditioning at inference could recover diversity while retaining favorable binding scores. These analyses are included in Appendix Q.

Distinct qualifying designs per pocket. Overall diversity measures variation across all generated candidates, but does not distinguish whether those candidates meet quality requirements. We therefore additionally measured diversity among clash-free designs with ΔG at least as favorable as the native binder. For each pocket, we generated 100 candidates and counted sequence and structural clusters among qualifying designs, averaging over all 93 pockets. UniMoMo was reproduced using its official code on the same pockets.

MethodOverall diversity, seq / structClash-free with ΔG < 0Qualifying designsQualifying clusters per pocket, seq / struct
AtomBloom-R0.189 / 0.10394.6%30.7%4.92 / 3.04
AtomBloom0.346 / 0.15895.3%27.9%8.99 / 3.89
UniMoMo0.632 / 0.62176.4%6.6%4.39 / 3.24

Lower overall diversity does not translate into fewer qualifying sequence alternatives. AtomBloom produces 8.99 qualifying sequence clusters per pocket, compared with 4.39 for UniMoMo. AtomBloom-R and AtomBloom find at least one qualifying candidate for 65 and 73 of the 93 pockets, respectively, compared with 36 for UniMoMo.

Increasing diversity without retraining. The reviewer’s concern also motivates improving diversity itself. Since self-conditioning contributes to the reduction, we tested delaying its activation until t≥0.5t \geq 0.5, alongside the default and fully disabled settings, using the same trained checkpoints.

ModelSelf-conditioning at inferenceAAR %C / L-RMSD ÅΔGIMP %Clash in / out %Seq / struct diversity
AtomBloom-RDefault42.881.92 / 0.76−43.9470.970.079 / 0.4350.189 / 0.103
Only for t ≥ 0.545.011.38 / 0.70−43.2373.120.156 / 0.7410.381 / 0.380
Off44.251.41 / 0.62−41.1260.220.375 / 0.9330.440 / 0.399
AtomBloomDefault40.551.18 / 0.69−45.3879.570.127 / 0.3840.346 / 0.158
Only for t ≥ 0.538.421.21 / 0.65−44.4577.420.409 / 0.6600.520 / 0.375
Off38.041.41 / 0.79−42.1266.670.63 / 0.950.593 / 0.404

Delaying self-conditioning substantially increases diversity with a smaller change in binding scores than disabling it entirely. For AtomBloom-R, sequence diversity doubles and structural diversity increases 3.7-fold, while ΔG changes from −43.94 to −43.23 and IMP increases from 70.97% to 73.12%. For AtomBloom, delayed self-conditioning raises sequence and structural diversity to 0.520 and 0.375 while retaining ΔG = −44.45 and IMP = 77.42%.

Thus, the default configuration provides more qualifying sequence alternatives under a fixed budget, and inference-time adjustment offers additional diversity without retraining.

W2: Expected to see more modern peptide or antibody comparisons, the baselines seem to be outdated. Q1: would the authors consider adding more baselines for the peptide and antibody tasks, eg bindcraft or boltzgen or similar generation?

Thank you for the suggestion. We have added BoltzGen comparisons for peptides and antibodies, and BindCraft comparisons for peptides. All results use the same evaluator as our original experiments and are included in Appendix O. Note that BoltzGen reports using all PDB structures up to June 1, 2023, subject to its preprocessing filters [A]. All 93 LNR complexes (released between 1995 and 2019) and all 60 RAbD complexes (1994–2017) fall within this training-data time window, whereas our training splits exclude these test entries.

AuthorsOpenReview ↗

Response to Reviewer 94wQ (2 of 3)

BoltzGen on peptides. We evaluated the official peptide pipeline on all 93 LNR targets, using the pocket as the binding constraint and the native peptide length. We followed its default SolubleMPNN and Boltz-2 refolding settings and evaluated 100 designs per target without filtering or retries. AtomBloom-R and BoltzGen receive the native residue count; AtomBloom generates the length.

MethodAAR %C / L-RMSD ÅΔGIMP %Seq / struct diversity
AtomBloom-R42.881.92 / 0.76−43.9470.970.189 / 0.103
AtomBloom40.551.18† / 0.69†−45.3879.570.346 / 0.158
BoltzGen45.701.51 / 0.49−43.8278.500.027 / 0.366

AtomBloom-R achieves similar interface energy to BoltzGen, while AtomBloom achieves a more favorable ΔG without a prespecified length. BoltzGen achieves stronger sequence and fixed-length structural recovery.

To assess how many distinct candidates are available for downstream testing, we also counted clash-free designs with ΔG at least as favorable as the native binder. Cluster counts are averaged over all 93 pockets under the same budget of 100 attempts per pocket.

MethodQualifying designsQualifying sequence clusters per pocketQualifying structural clusters per pocket
AtomBloom-R30.7%4.923.04
AtomBloom27.9%8.993.89
BoltzGen20.5%1.465.69

Both AtomBloom variants produce higher qualifying rates and more qualifying sequence clusters, whereas BoltzGen provides more qualifying structural clusters.

BoltzGen on antibodies. We evaluated the official antibody pipeline on all 60 RAbD targets, providing the full antigen chain and antibody variable domain. Each target has 100 designs, split equally between the diverse and adherence checkpoints, followed by SolubleMPNN and Boltz-2 refolding. Sample@100 results are below; Sample@1 and Sample@10 results are included in Appendix O.2.

MethodAAR %RMSD ÅIMP %Clash in / out %JSD bb / sc
AtomBloom-R54.290.8371.670.10 / 0.030.182 / 0.231
AtomBloom48.150.92†78.330.13 / 0.040.180 / 0.235
BoltzGen46.210.9448.330.05 / 0.210.169 / 0.176

Both AtomBloom variants achieve higher AAR and IMP than BoltzGen. AtomBloom-R also achieves lower RMSD. BoltzGen has lower internal clash rates and dihedral JSDs, while both AtomBloom variants have lower external clash rates.

BindCraft on peptides. We evaluated the official code with its peptide settings, native peptide lengths, and pocket residues as hotspots. All methods are evaluated at 10 designs per target, with the first 10 generated designs used for both AtomBloom variants. We report BindCraft’s AlphaFold 2 predictions and the same designs after FastRelax.

The unfiltered comparison uses the same 92 LNR targets for all methods, excluding one target whose 1,289-residue receptor exceeded GPU memory.

MethodAAR %C / L-RMSD ÅΔGIMP %Seq / struct diversity
AtomBloom-R30.732.21 / 1.17−39.5559.780.512 / 0.325
AtomBloom27.411.74† / 1.06†−39.5563.040.811 / 0.477
BindCraft, AlphaFold 2 prediction26.5013.46 / 1.92−14.2117.390.493 / 0.967
BindCraft, relaxed26.5013.61 / 2.08−15.6515.220.493 / 0.971

We additionally evaluated BindCraft after applying its official peptide filters. This comparison includes the 17 LNR targets with at least 10 filter-passing designs, using 10 designs per target and evaluating all methods on the same target subset. Both AtomBloom variants achieve more favorable ΔG and higher IMP than BindCraft in both comparisons.

MethodAAR %C / L-RMSD ÅΔGIMP %Seq / struct diversity
AtomBloom-R30.512.05 / 1.32−42.2252.940.629 / 0.447
AtomBloom24.911.78† / 1.67†−39.5947.060.918 / 0.635
BindCraft with filters, AlphaFold 2 prediction28.955.11 / 2.51−36.7335.290.394 / 0.524
BindCraft with filters, relaxed28.955.21 / 2.42−38.0641.180.394 / 0.529

† Variable-length AtomBloom RMSDs are measured over the reference domain, following Table 1.

[A] Stark et al. BoltzGen: Toward Universal Binder Design. bioRxiv, 2025. Appendix A.1.1.

W3: Expected to see more references related to other variable length design models, and their use in binder design, or discussion about theoretical reasons why one model may be better than other.

Thank you for this suggestion. We have expanded Section 2 and Appendix A to discuss other variable-size generative models and explain why we build AtomBloom on Branching Flows. The revised discussion connects these frameworks to the requirements of pocket-conditioned, all-atom binder design.

AuthorsOpenReview ↗

Response to Reviewer 94wQ (3 of 3)

Related variable-size models. Trans-dimensional jump diffusion [B] combines diffusion with dimension-changing jumps to jointly generate state values and dimensionality. Edit Flows [C] generates variable-length discrete sequences through insertions, deletions, and substitutions. Branching Flows [D] couples element-state evolution with splitting and deletion. For our task, the relevant question is how changes in atom count can be combined with the generation of atom types and continuous three-dimensional coordinates.

Why Branching Flows fits this task. Edit Flows is formulated over discrete sequences. Applying it directly to all-atom binder generation would require an additional continuous-coordinate process and a mechanism for generating the three-dimensional states of inserted elements. Branching Flows already supports continuous-state evolution together with cardinality changes, providing a direct basis for our atom-level representation.

Trans-dimensional jump diffusion already supports both continuous coordinates and variable cardinality. The distinction is how new elements are introduced. Its reverse process learns a conditional distribution over the position and features of each inserted element. Branching Flows instead creates descendants by splitting an existing element, after which their states continue to evolve. The descendants begin from the parent state and develop their geometry through subsequent updates. This avoids separately parameterizing a continuous insertion distribution and allows newly created atoms to be refined through the same state-evolution process as existing atoms. This provides a concrete modeling reason for our choice.

The branching structure also naturally supports residue-level constraints and self-conditioning. Branching Flows constructs training paths through coalescent trees, allowing us to assign each tree a specific structural meaning. For peptide and antibody generation, we use this property to introduce AtomBloom-R, a variant that merges the atoms of each residue into a separate tree. In this variant, each coalescent tree corresponds to one amino acid. The number of residues can therefore be specified in advance, while each residue’s identity and all-atom structure are still determined through the variable-cardinality generation process.

Branching Flows also provides an explicit ancestor–descendant correspondence. This is useful for self-conditioning because splitting and deletion change the atom set between predictions. We use this correspondence to pass an ancestor’s earlier prediction to its descendants as conditioning information. This gives self-conditioning a well-defined transfer rule as atom count changes.

These properties motivate Branching Flows as the foundation of AtomBloom.

[B] Campbell et al. Trans-Dimensional Generative Modeling via Jump Diffusion Models. NeurIPS, 2023.

[C] Havasi et al. Edit Flows: Variable Length Discrete Flow Matching with Sequence-Level Edit Operations. NeurIPS, 2025.

[D] Billera et al. Branching Flows: Discrete, Continuous, and Manifold Flow Matching with Splits and Deletions. NeurIPS, 2026.

Q2: diversity is an important metric for biologics. The authors train on the very diversity-limited SabDab. What strategies the authors would suggest for increased diversity without sacrificing other metrics?

Thank you for highlighting this concern. Delaying self-conditioning increases antibody diversity without retraining. We evaluated different activation times on all 60 RAbD targets, generating 100 designs per target with the same trained checkpoints. Multiplicity denotes the diversity relative to each model’s default setting and is calculated from unrounded diversity values.

ModelSC activationAAR %IMP %Seq div. (0.4)MultiplicitySeq div. (0.6)MultiplicityStruct. div.Multiplicity
AtomBloom-RDefault54.2971.670.0141.00×0.2561.00×0.0141.00×
t ≥ 0.354.5371.670.0141.02×0.3051.19×0.0493.51×
t ≥ 0.554.8566.670.0201.45×0.3601.40×0.0654.69×
t ≥ 0.754.8471.670.0281.98×0.4651.81×0.0654.70×
AtomBloomDefault48.1578.330.0691.00×0.7561.00×0.0371.00×
t ≥ 0.347.0975.000.0871.27×0.8151.08×0.1012.76×
t ≥ 0.547.1978.330.1131.65×0.8431.11×0.1103.01×
t ≥ 0.749.0875.000.1361.97×0.8571.13×0.1143.11×

Delayed self-conditioning increases sequence and structural diversity while retaining comparable AAR and IMP, without retraining. We have added these results to Appendix Q.3.

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

This work proposes AtomBloom, an interesting branching‑flow‑based generative approach for protein‑pocket‑conditioned binder design. It makes meaningful theoretical advances for finite‑step discretization of branching flows and presents lineage‑aware self‑conditioning to support variable‑cardinality all‑atom generation. Strong in‑silico benchmark results across small‑molecule, peptide and antibody tasks show competitive performance and highlight the benefit of exploring non‑reference binder lengths. However, final molecular quality depends heavily on modality‑specific post‑processing decoding, and all validation stays computational without experimental binding validation. Additional compute profiling, hyper‑parameter robustness sweep and real‑world assay results would further strengthen translational impact.

Critical strengths

It jointly optimizes atom number, atom types and 3D geometry in one flow‑based generative loop, instead of two‑stage coarse‑to‑fine decoding. It naturally explores binder lengths different from reference complexes.

It extends standard self‑conditioning to handle dynamically changing atom sets by tracking ancestry, enabling stable training for branching generative processes.

Critical weaknesses
  1. All evaluations are fully in‑silico, relying on docking or Rosetta energy metrics. There is no synthetic biology or biophysical assay to confirm actual binding affinity of AtomBloom‑produced candidates.

  2. Branching‑flow sampling requires many finite time‑steps. The model‑rolled self‑conditioning also increases training compute overhead. However, the paper does not systematically report wall‑clock time or memory consumption vs competing baselines.

  3. There is no systematic out‑of‑pocket generalization tests. The sensitivity to key hyper‑parameters, including split / delete distributions, rescaling bounds, self‑condition probability, is not fully swept.

Questions for authors
  1. What fraction of raw AtomBloom atom‑cloud outputs fail downstream decoding into chemically valid molecules? How do decoding failures change across different generation hyper‑parameters?

  2. The paper does not report training and inference wall‑clock and GPU‑memory overhead. How does AtomBloom’s computational cost compare against representative baselines such as UniMoMo or RFdiffusion?

  3. Why does the paper not include direct comparisons against strong autoregressive variable‑length fragment‑based generative baselines? What are the expected relative advantages and disadvantages between branching‑flow AtomBloom and autoregressive fragment‑by‑fragment generation?

  4. How would AtomBloom perform for unseen out‑of‑distribution protein pockets substantially different from training complexes? Have you analyzed generation quality under cross‑target generalization scenarios?

7 comments
AuthorsOpenReview ↗

Response to Reviewer FhNT (1 of 4)

We thank the reviewer for the valuable comments, and we answer the reviewer's questions as follows.

W1: All evaluations are fully in‑silico, relying on docking or Rosetta energy metrics. There is no synthetic biology or biophysical assay to confirm actual binding affinity of AtomBloom‑produced candidates.

Thank you for raising the importance of experimental validation. Wet-lab assays are costly and time-consuming. To strengthen the evidence for our generated candidates beyond docking and Rosetta energies, we have added a complementary evaluation using two external structure predictors, Protenix [A] and Mol-1 [B], across all 93 LNR targets. Structure-prediction-based assessment has improved experimental hit rates in prior binder-design pipelines [C] and helped distinguish experimentally successful designs from unsuccessful ones [D].

For each target and each generation method, we evaluated 10 candidates by generation index, without score-based selection, yielding 930 candidates per method. We also evaluated the 93 native complexes as a reference. Both predictors used one seed, five samples, and ten recycles per complex. For each complex, we evaluate the highest-ranked of the five predicted structures using the predictor’s own ranking score. The receptor was provided with its MSA and no templates, and the binder was provided without an MSA. We assessed interface confidence using ipTM and agreement between the predicted and designed peptide poses using scRMSD after superposing the receptor Cα. A candidate passes the computational criterion when ipTM ≥ 0.5 and scRMSD ≤ 3 Å.

CandidatesProtenix ipTMProtenix ipTM ≥ 0.5Protenix passMol-1 ipTMMol-1 ipTM ≥ 0.5Mol-1 pass
Native peptides0.915100.0%73.1%0.91497.8%74.2%
AtomBloom-R0.85299.0%42.3%0.77187.6%36.9%
AtomBloom0.80997.7%24.2%0.71683.1%23.4%
UniMoMo0.78498.0%8.1%0.67578.5%6.9%

Both predictors consistently favor AtomBloom-R and AtomBloom over UniMoMo. AtomBloom-R achieves pass rates of 42.3% and 36.9%, compared with 8.1% and 6.9% for UniMoMo. AtomBloom also improves both rates to 24.2% and 23.4% without a prespecified peptide length. We have included the new results in Appendix P.1. While binding affinity ultimately requires experimental measurement, these findings strengthen the evidence for the structural plausibility and pose consistency of our designs beyond energy-based evaluation.

[A] Zhang et al. Protenix-v1: Toward High-Accuracy Open-Source Biomolecular Structure Prediction. bioRxiv 2026.

[B] Molence Bio. Mol-1: Open Frontier Biomolecular Intelligence. Technical report, 2026. https://github.com/molence/Mol-1

[C] Bennett et al. Improving de novo protein binder design with deep learning. Nature Communications 2023.

[D] Overath et al. Predicting experimental success in de novo binder design: a meta-analysis of 3,766 experimentally characterised binders. bioRxiv 2025.

AuthorsOpenReview ↗

Response to Reviewer FhNT (2 of 4)

W2: Branching‑flow sampling requires many finite time‑steps. The model‑rolled self‑conditioning also increases training compute overhead. However, the paper does not systematically report wall‑clock time or memory consumption vs competing baselines. Q2: The paper does not report training and inference wall‑clock and GPU‑memory overhead. How does AtomBloom’s computational cost compare against representative baselines such as UniMoMo or RFdiffusion?

Thank you for raising this. We have added computational-cost comparisons with the closest baseline UniMoMo. At 50 sampling steps, AtomBloom takes longer per candidate but uses substantially less GPU memory and less generation time per candidate meeting the evaluation criteria. The results are in Appendix S.

Inference time and memory. We benchmarked both methods sequentially on the same GPU type in FP32, using batch sizes of 1 and 16. Timing covers data loading through structure export.

TaskMethodStepsSeconds/candidate, B=1 / 16Qualifying rateSeconds/qualifying candidate, B=1 / 16Peak GPU memory (GiB), B=1 / 16
PeptidesUniMoMo1002.1 / 0.616.6%31.6 / 9.31.52 / 18.33
PeptidesAtomBloom-R503.2 / 1.730.7%10.4 / 5.40.45 / 1.56
PeptidesAtomBloom503.2 / 1.727.9%11.5 / 6.10.46 / 1.56
Small moleculesUniMoMo1001.83 / 0.7343.7%4.20 / 1.661.07 / 6.87
Small moleculesAtomBloom502.63 / 0.9874.4%3.53 / 1.310.40 / 1.31

For peptides, qualifying candidates are clash-free and have Rosetta interface energy at least as favorable as the native binder. For small molecules, we apply UniMoMo’s acceptance filter to both methods and measure time per generation attempt. Generation time per qualifying candidate is the time per candidate divided by the qualifying rate.

Fewer sampling steps retain comparable peptide quality while reducing runtime. Reducing AtomBloom-R from 50 to 25 steps nearly halves the network evaluations, with similar interface energy and improvement rate. At batch size 1, the 25-step configuration is slightly faster than UniMoMo (1.9 vs. 2.1 s per candidate).

AtomBloom-RNetwork evaluationsSeconds/candidate, B=1 / 16ΔG ↓Improvement rate ↑
25 steps511.9 / 1.13−43.6172.04%
50 steps1013.2 / 1.7−43.9470.97%

Training time, memory, and self-conditioning overhead. On eight GPUs of the same type, with each method's own training recipe, total training wall-clock times are comparable to UniMoMo, and AtomBloom has a lower per-GPU peak memory than either UniMoMo stage for every task. The AtomBloom measurements include model-rolled self-conditioning. UniMoMo is trained in two stages, an autoencoder and a latent-diffusion model, which are listed separately.

PeptidesAntibodiesSmall molecules
Training time (h)
AtomBloom32.37.532.3
UniMoMo, autoencoder stage18.35.219.2
UniMoMo, latent-diffusion stage14.46.012.2
UniMoMo, both stages32.711.231.4
Peak memory per GPU (GiB)
AtomBloom24.018.426.9
UniMoMo, autoencoder stage58.944.058.1
UniMoMo, latent-diffusion stage62.549.043.1

Self-conditioning adds no extra network evaluation at inference, since it reuses stored predictions. During training, it adds one forward pass without gradients to 50% of the iterations of each GPU, and under eight-GPU training it increases the epoch time by a factor of 1.4 for peptides, 1.6 for antibodies, and 1.7 for small molecules. Model-rolled and oracle-state self-conditioning have the same measured step time.

AuthorsOpenReview ↗

Response to Reviewer FhNT (3 of 4)

W3: There is no systematic out‑of‑pocket generalization tests. The sensitivity to key hyper‑parameters, including split / delete distributions, rescaling bounds, self‑condition probability, is not fully swept. Q4: How would AtomBloom perform for unseen out‑of‑distribution protein pockets substantially different from training complexes? Have you analyzed generation quality under cross‑target generalization scenarios?

Thank you for these suggestions. We have added both analyses to Appendix R of the revised paper.

Generalization to targets with low sequence similarity. We further stratified the 93 LNR targets from the testing set by each receptor’s maximum sequence identity to the training receptors. The median maximum identity is 26.6%, and 56 of 93 targets fall below 30% identity.

The following results use Sample@100. UniMoMo results are from our re-run using its official code.

MethodIdentitynAAR % ↑ΔG ↓IMP % ↑C-RMSD ↓L-RMSD ↓
AtomBloom-R< 30%5642.72−43.2973.211.780.77
≥ 30%3743.11−44.9267.572.060.69
AtomBloom< 30%5640.89−44.5880.361.190.73
≥ 30%3740.02−46.5778.381.180.62
UniMoMo< 30%5639.29−33.2744.642.241.22
≥ 30%3739.44−33.5437.842.061.11

Both AtomBloom variants retain their advantages on the 56 low-identity targets. AtomBloom-R and AtomBloom achieve ΔG values of −43.29 and −44.58, compared with −33.27 for UniMoMo, and improvement rates of 73.21% and 80.36%, compared with 44.64%. These results support cross-target generalization to receptors with low sequence similarity to the training set.

Hyperparameter sensitivity. We varied four training settings of AtomBloom-R one at a time, testing two alternatives for each: the split-time distribution, deletion-time distribution, initial coordinate scale b0b_0, and self-conditioning probability. All configurations were trained for 200 epochs and compared with the default at epoch 200. The defaults are Beta(1,2) split times, Uniform(0,1) deletion times, b0=10b_0=10, and self-conditioning probability 0.5. We also repeated the default configuration with a second training seed.

The following results use Sample@100 on all 93 LNR targets.

SettingAAR %C / L-RMSD ÅΔGIMP %Clash in / out %Seq / struct diversity
Default46.791.24 / 0.57−43.8473.10.085 / 0.3430.266 / 0.151
Default, second seed44.201.40 / 0.66−44.5075.30.063 / 0.3260.247 / 0.141
Split times Beta(1,4)44.051.29 / 0.61−43.5473.10.100 / 0.4570.305 / 0.177
Split times Beta(1,1)43.851.68 / 0.63−44.4276.30.063 / 0.3590.198 / 0.141
Deletion times Beta(1,2)45.071.28 / 0.60−44.0475.30.196 / 0.4920.264 / 0.153
Deletion times Beta(2,1)44.151.40 / 0.70−44.1879.60.118 / 0.5060.269 / 0.168
b0=5b_0=545.451.20 / 0.59−45.1178.50.061 / 0.4410.247 / 0.148
b0=20b_0=2042.141.63 / 0.67−40.9264.50.096 / 1.0100.343 / 0.211
Self-conditioning probability 0.2544.881.05 / 0.63−43.3974.20.056 / 0.3030.297 / 0.169
Self-conditioning probability 0.7544.501.50 / 0.63−44.4572.00.061 / 0.3640.245 / 0.146

Binding performance remains stable across most tested settings. Seven of the eight alternatives remain within 1.3 of the default ΔG, with improvement rates between 72.0% and 79.6%. Increasing b0b_0 to 20 reduces performance to ΔG = −40.92 and IMP = 64.5%, still ahead of UniMoMo’s −34.4 and 40.9% in Table 1.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their valuable comments. We have revised the paper to clarify our contributions and strengthen the evaluation. The key revisions include:

  • Expanded Section 2 and Appendix A to clarify AtomBloom’s positioning and its contributions beyond Branching Flows. (To all reviewers)
  • Clarified notation, assumptions, and the construction of branching histories, with an accompanying illustration. (To Reviewer 34zN)
  • Added comparisons with BoltzGen and BindCraft, and complementary structural assessments using Protenix and Mol-1. (To Reviewers 94wQ and FhNT)
  • Added chemical validity, uniqueness, novelty, backbone geometry, and chirality analyses, together with inference-time strategies for increasing diversity. (To Reviewers 34zN and 94wQ)
  • Added generalization and hyperparameter analyses, including decoding sensitivity to inference settings. (To Reviewer FhNT)
  • Reported time and memory cost for training and inference. (To Reviewers FhNT and 34zN)
  • Clarified prescribed residue-count generation and added a constrained sampler for exact atom-count control without retraining. (To Reviewer 34zN)

We hope that these revisions address the reviewers' concerns and enhance the overall quality of our paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.