acceptodds
Under review as a conference paper at ICLR 2027

Spectral-Path Flow Matching for Antibody Generation

Abstract

Therapeutic antibody design requires generating complementarity-determining region (CDR) loops whose sequences and three-dimensional conformations support antigen recognition. CDR-H3 remains particularly challenging because of its high sequence and structural diversity. Recent diffusion and flow-matching models can generate CDR sequence and structure, but their positional generative paths typically evolve all spatial components under a shared notion of progress, even though native CDR deformation is dominated by low-frequency components corresponding to overall loop shape. We introduce SP-FMAG, a spectral-path flow-matching framework that decomposes framework-anchored CDR deformation into ordered spatial modes and assigns each mode its own generative clock. This produces a coarse-to-fine positional path in which broad loop geometry is established before finer structural variation. On the standard SAbDab benchmark with 19 held-out antibody-antigen complexes, SP-FMAG reduces CDR-H3 backbone RMSD from 3.313 Å to 2.958 Å and increases Rosetta binding-energy improvement rate from 23.8% to 35.7% relative to linear positional flow matching. The structural advantage persists under reduced sampling budgets. Reversing the modal ordering to fine-to-coarse worsens H3 RMSD to 3.499 Å, indicating that the benefit depends on the coarse-to-fine direction. These results demonstrate that coarse-to-fine spatial ordering in the probability path can improve antibody loop generation without changing the underlying neural architecture.

Primary area:
generative models
Keywords:
Flow Matching, Generative Modeling, Spectral Methods, Antibody Design, CDR Design

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

This paper introduces SP-FMAG, a framework for antibody CDR sequence–structure co-design that replaces linear positional interpolation with frequency-dependent flow-matching paths. CDR deformation relative to the chord connecting its framework anchors is expressed in a fixed-end sine basis, and higher-frequency modes receive delayed interpolation clocks. The architecture, positional source distribution, and sequence and orientation paths remain unchanged. On the standard SAbDab benchmark, the method reduces CDR-H3 RMSD from 3.313 Å to 2.958 Å and increases Rosetta binding-energy improvement rate from 23.8% to 35.7% relative to a matched linear-flow model. Additional experiments examine spectral strength, reversed ordering, reduced sampling budgets, and local antibody optimization.

Critical strengths

1、The matched linear baseline, reversed-ordering experiment, and spectral-strength ablation provide useful evidence beyond comparisons against previously published methods. The appendix also reports mixed results across CDR types rather than presenting the improvement as universal.

2、The complex-level paired bootstrap is preferable to treating individual generated samples as independent observations. The reported H3 structural improvement is encouraging and warrants further investigation.

Critical weaknesses

1、The experiments do not isolate the contribution of the probability path from frequency-dependent loss reweighting.Changing β modifies both the interpolation clocks in Equation (13) and the modal regression weights in Equations (15)–(16). Consequently, the β ablation changes the training-state distribution and the allocation of regression error across spatial frequencies simultaneously. Reversing the clock ordering also changes this allocation, so the forward-versus-reverse comparison does not resolve the confound. This distinction is particularly relevant to Table 3. With T = 1, Equation (17) reduces to directly returning the predicted endpoint because every clock satisfies Aₖ(0) = 0 and Aₖ(1) = 1. There is therefore no intermediate coarse-to-fine traversal during one-step inference, yet the structural advantage remains. This supports a benefit from the training formulation, but does not establish whether that benefit comes from the proposed path, spectral loss weighting, or their interaction. The central attribution to coarse-to-fine transport is therefore stronger than the available controls justify.

2、The mechanistic illustrations do not faithfully establish smoother generation under the actual source distribution.Figure 1 starts from a deterministic chord, whereas the implemented model starts from independent Gaussian modal coefficients. Under Equation (13), conditioned on a fixed target, the source-noise covariance in mode k is proportional to (1 − Aₖ(t))²I. Delaying high-frequency clocks therefore retains more high-frequency source noise than linear interpolation at the same base time. Delayed introduction of target detail does not, by itself, imply smoother intermediate samples. The tracking diagnostic in Appendix B.3 is also sensitive to the prescribed clock. For example, a trajectory that remains at its source has normalized residual Aₖ(t)² for an individual mode, which can be small simply because that mode is delayed. Figure 5 therefore does not independently demonstrate that the learned dynamics establish meaningful coarse geometry before resolving fine structure. These issues do not invalidate the path construction, but weaken the proposed explanation for its effectiveness.

3、The robustness evidence is limited by a small, grouped test set and a single reported training seed.All main conclusions rely on 19 complexes drawn from five held-out clusters/groups. Appendix B.4 bootstraps complexes, which captures variation across the observed test cases, but does not account for possible dependence within those groups or variation across independently trained models. Appendix D reports seed 2022 without multi-seed results. The confidence interval should therefore not be interpreted as establishing robustness across training runs or independent antigen groups. The model-selection procedure is also unclear. Twenty complexes are reserved for validation, but the manuscript does not specify the criterion used to select β = 4 and the reported checkpoints. Since several β values are compared on the same small benchmark, distinguishing validation-based selection from retrospective test-set selection is important.

4、The comparisons do not adequately establish the advance over existing antibody flow-matching methods.The matched β = 0 model is useful for evaluating the proposed modification within this implementation. However, FlowDesign, which the paper identifies as a prior antibody sequence–structure flow-matching method, is absent from the benchmark comparison. The external comparison instead relies primarily on diffusion methods and RAbD, using numbers imported from MDMAD. As a result, the experiments establish an improvement over the authors’ linear-path implementation more clearly than an advance over existing antibody flow-matching approaches. A protocol-matched comparison with the closest prior method, or a concrete explanation of incompatibilities, is needed to contextualize the practical significance of the contribution.

Questions for authors

1、Can the effects of interpolation clocks and modal loss weights be separated? Were any existing runs performed with spectral clocks and uniform endpoint weights, or with linear clocks and spectral weights? If not, please clarify that the reported ablations evaluate their combined effect rather than the probability path alone.

2、How were β and checkpoints selected? Please state the validation criterion, whether the same criterion was used for every model, and whether any test-set result influenced the choice of β = 4.

3、Why is FlowDesign absent from the comparison?

4、How sensitive are the results to test-group composition and training randomness?

5、Does Figure 1 represent the implemented Gaussian-source process? Please explain how the claim of smoother intermediate states applies when delayed high-frequency modes retain source noise.

1 comment

We thank the reviewer for their thorough and careful analysis of the ablations and intermediate states. We added crossed path and weighting controls, evaluated variability across training runs and antigen groups, and clarified how to interpret intermediate states.

Q1. Separating the path and loss weighting

We ran both requested controls with seed 2022 (Appendix B.8, Table 10, p. 21).

PathEndpoint weightsH3 RMSD (Å) ↓AAR (%) ↑IMP (%) ↑
LinearUniform3.31338.9723.8
SpectralUniform3.25433.2824.0
LinearSpectral3.27339.8231.9
SpectralSpectral2.95836.7835.7

Each change alone yielded a small RMSD reduction. Spectral weighting with a linear path increased IMP from 23.8% to 31.9%. The combined configuration gave the lowest RMSD and highest IMP in this single-run comparison. The β sweep and reversal comparisons change both the path and its associated loss weights.

At T = 1, Eq. (17) returns the predicted positional endpoint directly. The one-step RMSD gain reflects differences between the trained models, since there are no intermediate states. We clarified this in Section 5.3 (p. 9).

Q2. Selection of β and checkpoints

Checkpoints. We trained each model for 200,000 steps at an effective batch size of 16 and used the final checkpoint for all models. Validation loss was used only for the learning-rate plateau scheduler. We did not use early stopping or validation-based checkpoint selection (Appendix D.2, p. 24). The fixed budget follows MDMAD's baseline reproduction protocol. The external baseline numbers are from MDMAD, and we follow its OpenMM/Rosetta relaxation protocol.

β selection. We ran the H3 sweep with seed 2022 (Table 2, p. 9).

βAAR (%) ↑IMP (%) ↑RMSD (Å) ↓
038.9723.83.313
0.536.7925.93.387
135.2726.73.252
236.9626.33.079
436.7835.72.958
834.2230.93.143

We chose β = 4 because it gave the lowest H3 RMSD on the benchmark test split in this sweep (Appendix D.2, p. 25). We used this value for all six CDRs. The additional runs kept β fixed on the same split to assess variability across seeds. MDMAD also reports a mixture-size sweep and uses four mixture components each for position and orientation in its main configuration (Durvasula et al., 2026, Tables 1 and 6).

Q3. FlowDesign comparison

FlowDesign's published results use a different data split, with initial CDR structures generated by dyMEAN retrained on that split. A matched comparison would require priors prepared for our split. We attempted this adaptation. The available templates covered 9 of our 19 complexes. The KIC templates we generated for the remaining 10 used a different initialization procedure, so we did not report the preliminary results as a matched baseline.

Q4. Training seeds and antigen groups

Training and sampling seeds. We evaluated four independent training runs per model on the same split. Each run used the same seed for training and sampling, and seeds were matched across models. Mean H3 RMSD decreased from 3.331 ± 0.106 Å to 3.053 ± 0.135 Å (mean ± sample SD), an 8.4% reduction, and was lower in every matched run. Mean IMP increased from 25.72% to 28.22%, with gains in three of four runs (Appendix B.5, Table 8, p. 19). The paired bootstrap in Appendix B.4 quantifies uncertainty across complexes for the seed-2022 comparison.

Antigen groups. The 19 evaluated complexes represent four antigen groups. Mean H3 RMSD was lower within each group. Using existing predictions, we excluded each group in turn and recomputed the mean on the remaining complexes. The reduction ranged from 0.211 to 0.410 Å, with lower RMSD in all four runs for every exclusion (Appendix B.10, p. 22).

Q5. Gaussian source and intermediate states

We agree that delayed high-frequency clocks retain more Gaussian source noise in those modes at the same time. The coarse-to-fine ordering refers to the rate at which target contributions grow. We removed the claim that this makes intermediate samples smoother. Figure 1 (p. 1) now shows the Gaussian source, and the caption states:

The intermediate panels illustrate only the target contributions. The full intermediate states also include a contribution from the Gaussian source.

We also revised Appendix B.3 (p. 19) to explain that a trajectory remaining at its source has a normalized single-mode residual of Aₖ(t)². A delayed clock can make this residual small even without progress toward the target. The diagnostic measures proximity to the prescribed path and does not establish that useful coarse geometry is learned first.

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

This paper proposes a fixed-end flow matching method for antibody generalization. The information of increasing frequency is gradually added in the flow matching procedure. The idea is to establish general geometry of the loop before fine structure of the loop is introduced. The authors applies their method to standard SAbDab benchmark with 19 held-out antibody–antigen complexes, and the results show some improvements.

Critical strengths

The idea to consider the competition between general geometry and fine scale structure of the loop may be interesting.

Critical weaknesses

The authors do not demonstrate significance of their treatment. As shown in Fig 1, the length of the generated loop is not that significantly larger than the initial segment. Whether to restrict the fine structures at the beginning or let them evolve freely during the path seems to have no critical influence. This also agrees with their experimental results that the improvements are not that impressive compared with the existing methods.

The ablation study of fine to coarse training is not helpful. This is obviously not a good idea. It can be changed to an experiment that allows fine structures at the beginning but does not coarsen them.

In the comparisons in table 1, years and references of those existing methods should be given.

Questions for authors

Please see my comments on critical weaknesses.

1 comment

We thank the reviewer for their constructive feedback. We revised Figure 1 to clarify the role of the Gaussian source, added results across four independent training runs, and included baseline citations in Table 1.

Comment 1. Motivation and significance of frequency scheduling

The chord between the framework anchors is the reference for measuring loop deformation. The positional state at t = 0 is sampled from Gaussian noise. We revised Figure 1 and its caption (p. 1) to show this source and clarify that the intermediate panels illustrate only the target contributions. The full intermediate states also include a contribution from the Gaussian source.

The motivation comes from how native loop deformation is distributed across frequencies. For CDR-H3, the lowest-frequency 25% of modes capture a median 95.1% of the native deformation energy (Appendix B.2, pp. 17-18). This concentration motivates assigning different schedules to broad loop deformation and finer structural variation.

Across four independent training runs per model, SP-FMAG reduced mean CDR-H3 RMSD from 3.331 ± 0.106 Å to 3.053 ± 0.135 Å (mean ± sample SD), an 8.4% reduction relative to the matched linear baseline. RMSD was lower in every matched run. Mean IMP increased from 25.72% to 28.22%, with gains in three of four runs (Appendix B.5, Table 8, p.19). In Table 1, the four-run mean H3 RMSD of 3.053 Å is the second-lowest value among the reported methods. We have focused our conclusion on the consistent H3 structural improvement over the matched baseline.

Comment 2. Allowing fine structures from the beginning

The linear baseline in Table 2 (β = 0, p. 9) provides the comparison the reviewer suggests. All modes share a clock, so fine-scale and coarse-scale target contributions grow together from the start of the path.

In the reversed schedule, “fine-to-coarse” refers to relative progress toward the target. Higher-frequency modes advance earlier relative to lower-frequency modes, and each mode's target coefficient increases monotonically. The reversal tests the frequency assignment of both the clocks and the associated loss weights. We also added a control that compares the spectral and linear paths under uniform endpoint weights (Appendix B.8, Table 10, p.21). In this single-run control, the spectral path reduced H3 RMSD from 3.313 to 3.254 Å.

Comment 3. Baseline references and years

We added author–year citations for the baseline methods in Table 1 (p. 8).

Reviewer_3cuCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SP-FMAG modifies the positional probability path in antigen-conditioned antibody CDR sequence–structure co-design. It represents a loop's displacement from the chord between its framework anchors in a fixed-end discrete sine basis. Mode-dependent clocks delay higher-frequency deformation relative to broad loop geometry. An unchanged geometric encoder predicts clean sequence, positions, and orientations, while an analytic endpoint-based update advances each positional mode. The construction reduces to a linear path when spectral strength is zero. On the DiffAb SAbDab split, the main result is improved CDR-H3 recovery and Rosetta improvement rate, supported by spectral-strength, reversed-ordering, and sampling-budget comparisons. The same model is also used for partial redesign of native loops.

Critical strengths
  1. A simple construction with a meaningful controlled comparison. The fixed anchors yield an orthonormal sine basis without additional learned parameters. The clocks and the update in Equation 17 provide a straightforward implementation, and Appendix A demonstrates exact recovery of the linear positional path at β = 0. Keeping architecture, isotropic source, and evaluation protocol fixed makes Table 2 substantially more informative than a comparison between unrelated generators. At β = 4, H3 RMSD improves from 3.313 to 2.958 Å and IMP increases from 23.8% to 35.7%, with the expected trade-off in AAR from 38.97% to 36.78%.

  2. Several analyses probe when the improvement occurs. Reversing the clock assignment at the same exponent range worsens H3 RMSD to 3.499 Å and IMP to 21.1% (Table 4), supporting the relevance of frequency assignment within this method. Appendix B.4 reports improvement on 13 of 19 complexes and a paired bootstrap RMSD-improvement interval of [0.098, 0.683] Å, using complexes rather than treating 1,900 generated designs as independent observations. Table 3 also preserves the structural advantage at one and five network evaluations. The full per-CDR tables transparently show where these effects do not extend, which makes the focused H3 conclusion more credible.

Critical weaknesses
  1. The evidence establishes a narrow benchmark improvement more convincingly than reliable antibody generation. Appendix D reports a single training seed, and the 19 complexes come from only five held-out groups. The complex-level bootstrap is appropriate for the reported paired comparison but does not capture training variability or possible dependence within those groups. The choice of β = 4 is not tied to a stated validation-selection rule. More fundamentally, Tables 1–3 report native recovery and a Rosetta threshold statistic after OpenMM/FastRelax processing, without sequence/structure diversity, raw geometric validity, or results before relaxation. Lower RMSD alone cannot distinguish better generation from a narrower distribution around common loop shapes. This matters especially for the one-step result. Table 3 also shows five-step IMP falling from 26.3% for the linear model to 21.0%, despite improved RMSD. FlowDesign is discussed as a directly relevant flow-matching comparator but omitted from the quantitative benchmark, limiting conclusions about practical competitiveness. These gaps qualify the significance of the result rather than negate the matched-baseline improvement.

  2. The proposed explanation does not separate the probability path from spectral loss weighting. Changing β changes both the training states through Equation 13 and the frequency-dependent endpoint weights through Equations 15–16. Reversing the ordering changes both assignments again. Tables 2 and 4 therefore establish the combined method's effectiveness, but not the separate contributions of path scheduling and endpoint-error weighting. Moreover, at T = 1, Equation 17 returns the predicted clean endpoint for every mode regardless of β. The one-step gain consequently reflects training differences rather than demonstrating coarse-to-fine traversal during sampling. The derivation in Appendix A.6 is clear; my concern is causal attribution, not algebraic correctness. A linear path with spectral weighting and a spectral path with uniform endpoint weights would distinguish these explanations.

Questions for authors
  1. How were β and the reported checkpoint selected using the 20 validation complexes? Do the H3 improvement and uncertainty remain consistent across independent training seeds or when evaluated by held-out group? Please also clarify whether the 100 samples are averaged per complex and whether paired native and designed structures receive identical relaxation treatment.

  2. Have you evaluated the crossed path/weighting controls described above? If available, do they retain the gain in H3 recovery and IMP, and do generated diversity and geometric validity remain comparable to the linear model?

1 comment

We thank the reviewer for their thorough and constructive feedback. We revised the manuscript to clarify model selection and evaluation, report variability across training runs and antigen groups, and separate the effects of the path and loss weighting.

Q1. Checkpoints, β selection, training seeds, evaluation, and FlowDesign

Checkpoint selection. We trained each model for 200,000 steps with an effective batch size of 16 and evaluated the final checkpoint. Validation loss was used only by the learning-rate plateau scheduler. We did not use early stopping or validation-based checkpoint selection (Appendix D.2, p. 24). MDMAD used the same training budget and batch size for baseline reproduction.

Selection of β. The CDR-H3 sweep was run with seed 2022 (Table 2, p. 9).

βAAR (%) ↑IMP (%) ↑RMSD (Å) ↓
038.9723.83.313
0.536.7925.93.387
135.2726.73.252
236.9626.33.079
436.7835.72.958
834.2230.93.143

We selected β = 4 because it gave the lowest H3 RMSD in this sweep on the benchmark test split (Appendix D.2, p. 25). We kept β fixed in the additional runs on the same split to assess variability across seeds. MDMAD also reports a mixture-size sweep and uses four mixture components each for position and orientation in its main configuration (Durvasula et al., 2026, Tables 1 and 6).

Training and sampling seeds. We evaluated four independent training runs per model. Each run used the same seed for training and sampling, and we matched seeds across models. Mean H3 RMSD decreased from 3.331 ± 0.106 Å to 3.053 ± 0.135 Å (mean ± sample SD), an 8.4% reduction. RMSD was lower in every matched run. Mean IMP increased from 25.72% to 28.22%, with gains in three of four runs (Appendix B.5, Table 8, p. 19). The bootstrap interval in Appendix B.4 measures uncertainty across complexes for the seed-2022 comparison.

Antigen groups. The 19 test complexes represent four antigen groups. Mean H3 RMSD was lower within each group. Using the existing predictions, we recomputed the average after omitting each group in turn. The mean reduction ranged from 0.211 to 0.410 Å, with lower RMSD in all four runs for every exclusion (Appendix B.10, p. 22).

Aggregation and relaxation. We generate 100 samples per CDR per complex. We compute mean RMSD, mean AAR, and IMP (the fraction of evaluated samples with ΔΔG < 0) within each complex, then weight complexes equally. For both models, native and generated structures undergo the same OpenMM minimization and PyRosetta FastRelax protocol, with the same flexible CDR region. We use the relaxed native structure as the reference for RMSD and binding-energy evaluation (Appendix D.4, p. 25).

FlowDesign. Its published results use a different split, with initial CDR structures generated by dyMEAN retrained on that split. The available templates covered 9 of our 19 complexes. The KIC templates we generated for the other 10 used a different initialization procedure, so we did not report the preliminary adaptation as a matched baseline.

Q2. Path, loss weighting, diversity, and geometric validity

Crossed controls. We ran both requested ablations with seed 2022 (Appendix B.8, Table 10, p. 21).

PathEndpoint weightsH3 RMSD (Å) ↓AAR (%) ↑IMP (%) ↑
LinearUniform3.31338.9723.8
SpectralUniform3.25433.2824.0
LinearSpectral3.27339.8231.9
SpectralSpectral2.95836.7835.7

Each control reduced RMSD slightly. Weighting alone retained more of the IMP gain, while the full configuration achieved the lowest RMSD and highest IMP in this comparison.

Reduced sampling budgets. At T = 1, Eq. (17) returns the predicted endpoint directly. The lower RMSD for SP-FMAG (3.005 vs. 3.180 Å) reflects differences between the trained models, since one-step sampling has no intermediate states. At T = 5, RMSD decreased from 3.281 to 2.928 Å while IMP fell from 26.3% to 21.0% (Section 5.3, p. 9).

Diversity and geometric validity. For each complex, we superposed all 4,950 pairs of the 100 relaxed H3 samples on their Cα atoms and computed pairwise RMSD. We averaged within complexes, then across all 19 complexes and four runs per model. Mean pairwise RMSD decreased from 2.06 to 1.89 Å, with reductions of 3.6–13.1% across the four matched runs. SP-FMAG therefore showed less structural spread by this measure. Its mean best-of-100 native RMSD also decreased, from 1.93 to 1.84 Å, with gains in every matched run (Appendix B.6, Table 9, p. 20). These metrics were measured after relaxation. We did not evaluate geometric validity before relaxation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.