acceptodds
Under review as a conference paper at ICLR 2027

MoDA: a Modality-Adaptive Flow Matching Framework for Antibody Co-Design

Abstract

Antibody co-design methods usually model modalities uniformly, overlooking biophysical heterogeneity in scale-dependent constraints, residue roles, and modality-specific contextual requirements. To address these challenges, we propose MoDA, a modality-adaptive flow matching framework that integrates multi-scale constraints with residue role modeling. We first develop multi-scale constraint modeling to capture disentangled semantic information from local regions to global complex levels via decomposing functional unit representations. Then, we introduce cooperation-aware role modeling to infer residue roles based on residue-specific constraints and inter-residue coordination. Coupling the above constraints with inferred role assignments, we propose modality-adaptive coordinative optimization to steer asymmetric message passing and modality-specific interaction updates. Experiments on the RAbD benchmark demonstrate that MoDA achieves state-of-the-art performance, significantly improving structural fidelity (RMSD, TM-score, lDDT) and target binding affinity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
antibody co-design, modality adaptation

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

This paper proposes MoDA, a flow matching framework for joint antibody CDR sequence and structure design. It combines multi-scale constraints, cooperation-aware residue role modeling, and modality-adaptive interactions among sequence, position, and orientation. Experiments on RAbD report improved structural accuracy and predicted binding-energy success rates, with ablations supporting the proposed components.

Critical strengths

1.The paper is clearly motivated by heterogeneity across structural scales, residue roles, and sequence–geometry modalities. The framework addresses these aspects through well-integrated modules, with multi-scale constraints informing residue roles that guide message passing and modality-specific updates.

2.MoDA achieves strong overall structural performance on RAbD, reducing CDR RMSD from 1.8139 to 1.5780 and improving CDR-specific structural metrics.

3.The module-level and operation-level ablations, together with analyses of constraint representations, residue roles, and modality interactions, provide useful insight into the model’s behavior.

Critical weaknesses

1.Several definitions and operations require clarification, including the functional form of βt\beta_t and its relation to flow time, the construction of MijpairM_{ij}^{\mathrm{pair}} in Eq. (4), and the normalization operation Norm⁡\operatorname{Norm} in Section 3.4. It is also unclear how missing unit-specific features are handled when constructing xiAUx_i^{\mathrm{AU}} for framework or antigen residues that do not belong to all the listed functional units.

2.The reported results are inconsistent across tables. Table 1 reports RMSD, lDDT, and G-lDDT values of 1.5780, 0.9677, and 0.8341 for MoDA, whereas the learnable-schedule configuration in Table 4 reports 1.5776, 0.9586, and 0.7668, respectively. Although the RMSD discrepancy is small, the differences in lDDT and G-lDDT are substantial. The authors should clarify whether these discrepancies arise from different experimental settings, separate runs, or reporting errors.

3.Incomplete reporting of implementation and computational cost. Despite the code release, the manuscript leaves important settings unspecified, including the number of role prototypes, adapter rank, and training and sampling configurations. Runtime, memory usage, and parameter counts are also not reported, making reproducibility and the cost of the added modules difficult to assess.

4.The paper claims high sequence diversity but does not report direct diversity measurements. Neither sequence recovery against a single native reference nor target-level binding-energy success establishes diverse generation.

Questions for authors

1.Could the authors clarify the definitions of βt\beta_t, MijpairM_{ij}^{\mathrm{pair}}, and Norm⁡\operatorname{Norm}, and explain how missing unit-specific features are handled when constructing xiAUx_i^{\mathrm{AU}} for residues that do not belong to all functional units?

2.Do the MoDA results in Table 1 and the learnable-schedule results in Table 4 correspond to the same model configuration and evaluation protocol? If so, what explains the discrepancies in RMSD, lDDT, and G-lDDT?

3.How are structural and sequence metrics aggregated over the 100 candidates generated per target? Are they averaged over all candidates or computed from selected candidates, and is the same protocol applied to every baseline?

Reviewer_1iDCOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes a multimodal flow-matching framework for antibody generation, where learning is adaptively tailored to different modalities. In terms of contributions, the paper introduces Multi-scale Constraint Modeling and Cooperation-aware Role Modeling to provide information for Modality-Adaptive Optimization.

Although the proposed method is complex and achieves empirical improvements, the numerous complicated concepts and architectural designs make the central claims difficult to identify, ultimately weakening the clarity and impact of the paper's contributions.

Critical strengths

S1. The paper addresses a meaningful problem in antibody CDR co-design. The proposed framework integrates multi-scale structural information, residue-level interactions, and modality-specific updates into a coherent architecture.

S2. The method achieves competitive performance on the RAbD benchmark, particularly in structural accuracy.

S3. The experimental evaluation is relatively comprehensive. Beyond comparisons with existing methods, the paper includes module-level ablations and analyses of learned representations, residue roles, and cross-modal interactions, providing additional insights into the model's behavior.

Critical weaknesses

W1. The biological motivation is not sufficiently aligned with the actual task setting. Since the framework is fixed and only CDR residues are generated, the task formulation and generation loss already explicitly distinguish CDRs from framework residues. Therefore, the paper does not clearly establish why the existing task formulation is insufficient to capture the residue-level functional heterogeneity emphasized in its motivation.

W2. The method combines several established techniques, including multi-scale attention, message passing, and cross-modal gating, through largely heuristic designs. Some components lack sufficient justification for their claimed biological interpretations, particularly the fundamental, inherent, and synergistic constraints.

W3. The module-level ablation studies provide limited evidence for the necessity of individual design choices. Comparisons with simpler, parameter-matched alternatives would help justify the proposed architectural complexity.

W. The computational overhead of the proposed architecture is insufficiently discussed. Reporting model size, training cost, and inference efficiency would help assess the trade-off between complexity and performance.

W4. Experiments in Sections 4.3-4.5 primarily demonstrate differences in learned representations, role distributions, and attention patterns. However, such differences may naturally arise from architectural choices or regularization and do not necessarily indicate meaningful biological specialization or explain improvements in generation quality.

W5. The presentation could be substantially improved. Numerous abstract concepts and dense architectural details make the method difficult to follow and obscure its central contributions. A more concise and intuitive explanation would improve clarity and readability.

Questions for authors

Please see Weakness

Reviewer_N6n7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MoDA, a multimodal flow-matching framework for jointly designing antibody CDR sequences and structures. It combines multi-scale representations at the region, chain, pair, and complex levels, cooperation-aware residue role modeling, and modality-specific updates with asymmetric interactions. Evaluation on the 60-complex RAbD benchmark reports improvements over five baselines on six of seven metrics, supported by component ablations and representation analyses.

Critical strengths
  1. Well-motivated modeling approach. Antibody co-design is an important problem, and the proposed hierarchy reflects meaningful biological organization. Combining this hierarchy with residue-dependent interactions and modality-specific updates provides a coherent approach to modeling heterogeneous sequence and structural requirements.

  2. Promising structural results. MoDA reduces aggregate CDR RMSD from 1.8139 Å for AbFlow to 1.5780 Å and reports improvements across several complementary structural metrics. Comparisons with relevant baselines and component ablations provide encouraging evidence that the complete framework is useful.

Critical weaknesses
  1. The empirical evidence needs clarification. Tables 1–2 report MoDA’s G-lDDT as 0.8341, whereas Table 4 reports 0.7668 for the learnable-schedule model, despite several other metrics being nearly identical. This difference is larger than the reported advantage over the strongest baseline. The aggregation of the 100 generated candidates and the uncertainty of the results are also unclear, making the reliability of the improvements difficult to assess.

  2. The analyses do not establish the claimed disentanglement and functional specialization. Near-zero cosine similarity does not demonstrate statistical independence, particularly when decorrelation is explicitly encouraged during training. Likewise, differences in role distributions and modality attention do not establish that these mechanisms provide benefits beyond additional capacity and auxiliary supervision. The removal ablations are useful, but more targeted controls are needed to support the paper’s central interpretation.

Questions for authors
  1. Why do the lDDT and G-lDDT values differ between Tables 1–2 and the learnable-schedule model in Table 4? Which results should be used for the main comparison?

  2. How are the 100 candidates aggregated for each structural and sequence metric? Are the same sampling, selection, and postprocessing procedures applied to every method? Please provide paired target-level uncertainty estimates for the main improvements.

  3. What evidence distinguishes the benefits of the proposed roles and asymmetric coordination from those of added capacity and auxiliary supervision? Comparisons using the same auxiliary targets with ordinary residue embeddings, or symmetric coordination with comparable capacity, would materially strengthen my assessment.

1 comment

Thank you for the additional experiments and for clarifying how the 100 candidates per target were evaluated. The paired analysis across targets gave me more confidence in the reported structural improvements. I also appreciated the comparisons with simpler models of comparable size, particularly the one that kept the same auxiliary supervision and used ordinary residue embeddings.

I still have a question about Table 1. I understand that the original Tables 1 and 4 used different sets of atoms to calculate G-lDDT. My understanding from your response is that both the original and revised versions of Table 1 use Cα atoms. Could you clarify why MoDA's G-lDDT in that table changed from 0.8341 to 0.8192?

My main remaining concern is how to interpret the learned representations. On rereading Sections 3.3 and 3.4, I still found it difficult to connect the equations to what each component actually learns. The perturbation experiments help show that different components affect different aspects of the prediction. Since the training objective already encourages low cosine similarity, I'd be cautious about taking the observed separation as evidence that each component has learned the particular function described in the paper.

For these reasons, I'll keep my overall score at 2. Could you also update the manuscript with the changes highlighted in different colors? I had trouble locating some of the revisions mentioned in your response.

Comments on the paper

6 comments

W2. Insufficient justification for the biological interpretations of heuristic architectural designs, particularly the fundamental, inherent, and synergistic constraints.

We clarify that MoDA's architectural designs are driven by specific biophysical and modeling requirements in antibody co-design, rather than arbitrary combinations of established techniques. Specifically,

(1) Multi-scale constraint modeling. Each CDR residue must simultaneously satisfy local loop geometry, chain-level compatibility, inter-chain organization, and antigen recognition. These constraints are neither independent nor interchangeable: they share common structural requirements, impose distinct unit-specific conditions, and jointly determine paratope organization. Simply fusing their representations does not explicitly distinguish these complementary relationships. We therefore introduce R/U/S decomposition to separately model shared consistency, unit-specific characteristics, and cross-unit cooperative dependencies. Importantly, their interpretations are grounded in the modeling mechanisms: cross-unit alignment encourages shared information in R; unit-specific encoding and redundancy regularization distinguish U; and cross-unit interaction context in S conditions cooperative role refinement and modality coordination. Moreover, reliability-aware aggregation accounts for the changing uncertainty of generated CDR contexts relative to fixed FR and antigen information.

(2) Cooperation-aware role modeling. Residues within the same CDR can contribute differently to loop stabilization and antigen recognition. Since these contributions depend on local properties and neighboring interactions, we infer context-dependent residue roles and refine them through pairwise cooperation rather than relying on predefined region identities.

(3) Modality-adaptive coordination. Sequence and geometry exhibit asymmetric conditional dependencies: local geometry and antigen contacts constrain compatible residue identities, whereas amino acid identities influence but do not uniquely determine conformations. Thus, symmetric information exchange may fail to distinguish the different information requirements and contributions of each modality. We therefore introduce role-directed message passing to distinguish residue-specific information demand and supply, together with asymmetric cross-modal coordination to enable direction-dependent information transfer during joint generation.

To further clarify the design motivation, we revised Section 3.3 (Lines 195–197) as follows:

“Cross-scale residue sharing creates overlap among functional units whose representations may contain shared, unit-specific, and cooperative information, which may become entangled in the learned representations. We introduce a reliability-aware constraint decomposition that disentangles representations into complementary contextual information.”

W3. Comparisons with simpler, parameter-matched alternatives to justify the proposed architectural complexity.

To assess the necessity of individual design choices beyond model capacity, we constructed four simpler, parameter-matched alternatives for MSC, CAR, and MACP, each with approximately 8.553M trainable parameters (within 0.002% of MoDA).

(1) MSC: Generic Multi-scale Fusion (GMCF) and Without Explicit RUS. GMCF retains the functional-unit definitions and masks across four scales but replaces scale-specific processing and explicit R/U/S decomposition with shared self-attention, masked mean pooling, and MLP fusion. Without Explicit RUS retains the scale-specific encoders and reliability-aware cross-unit attention but replaces R/U/S construction and readout with MLP fusion, removing the associated regularizers. As shown in Table R1, GMCF increases RMSD from 1.5776 to 1.6777 Å and reduces AAR from 62.2509% to 61.1795%, while Without Explicit RUS increases RMSD to 1.6307 Å and reduces AAR to 60.5401%. Both degrade the remaining metrics, supporting the benefits of structured multi-scale modeling and explicit constraint decomposition with its regularization.

(2) CAR: Ordinary Identity Embeddings (OIE). We replace dynamic role inference with static region identity embeddings under matched capacity and identical auxiliary supervision, retaining the other components. OIE increases RMSD from 1.5776 to 1.6476 Å and reduces AAR from 62.2509% to 61.1484%, with degradation across all metrics, supporting the value of context-dependent role modeling beyond predefined regional identities.

(3) MACP: Shared Multimodal Fusion (SMF). We replace role-directed processing and asymmetric cross-modal coordination with ordinary graph aggregation, shared MLP fusion, and modality-specific linear projections, retaining MSC and CAR. SMF increases RMSD to 1.6231 Å and degrades other structural metrics, although AAR slightly improves to 62.9417%.

(1) MSC: Shared, unit-specific, and cooperative constraints. For R and S, to examine their contributions to shared fundamental constraints and inter-unit cooperation, we independently applied zeroing or random replacement during generation on RAbD, evaluating local CDR conformation and framework connectivity (per-CDR RMSD, CDR–FR anchor-distance and boundary C–N errors) and inter-CDR geometric coordination (cross-CDR and same-chain inter-CDR FAPE). As shown in Table 1, zeroing/randomizing R increased the local errors by 20.01%/12.24%, 126.69%/89.78%, and 101.91%/69.88%, respectively. In contrast, perturbing S slightly improved local recovery but increased the inter-CDR FAPE metrics by 1.73%/2.66% and 2.17%/4.24%. R and S showed clear differences across all metrics under random replacement, supporting distinct contributions of R to local structural recovery and framework connectivity and S to inter-CDR geometric coordination.

For U, we separately perturbed each CDR’s U through zeroing, norm-matched random replacement, or same-chain incorrect-unit replacement, comparing the targeted CDR with the other five across 60 complexes. As shown in Table 2a-2c, most targeted CDRs showed reduced sequence recovery and varying local structural degradation, with substantially smaller effects elsewhere. For example, incorrect-unit replacement of L1’s U reduced AAR by 12.13% and increased RMSD and ψ-angle error by 5.74% and 11.47%, versus only +0.31%, +0.34%, and +1.50% in other CDRs. Target-versus-other AAR differences were significant for all three interventions in L1/L2 and random/incorrect-unit replacement in L3. H3 instead exhibited mixed responses, with improved AAR and RMSD but consistently increased ψ-angle error, indicating that U perturbation degraded local backbone accuracy despite divergent changes across recovery measures. These region-dependent effects support the non-interchangeable, unit-specific contributions of U to sequence and structural recovery.

(2) CAR: Residue-specific functional roles. To examine whether learned roles provide information beyond predefined regional identities, we replaced dynamic role inference with parameter-matched Ordinary Identity Embeddings (OIE) while retaining the remaining components. As shown in Table 3, OIE increased CDR RMSD from 1.5776 to 1.6476 Å and reduced AAR from 62.2509% to 61.1484%. To further test whether the correspondence between residues and learned roles matters, we permuted role-probability vectors within each CDR while preserving their complete multiset. As shown in Table 4, the permutation decreased AAR from 60.89% to 58.32% (−2.578 %; 95% CI [−3.146, −2.038]), demonstrating that residue-specific role assignments contribute to actual generation rather than merely producing distinct role distributions.

(3) MACP: Asymmetric cross-modal coordination. To examine whether the proposed coordination improves final generation over generic fusion at comparable capacity, we replaced MACP with parameter-matched Shared Multimodal Fusion (SMF). As shown in Table 5, this increased CDR RMSD from 1.5776 to 1.6231 Å (+2.88%) and reduced G-lDDT from 0.8192 to 0.8128 (−0.78%), with lower TM, G-TM, and lDDT. AAR increased from 62.2509% to 62.9417%. These comparisons support MACP’s contribution to structural fidelity beyond generic multimodal fusion, while showing a sequence–structure recovery trade-off. To isolate the contribution of interaction direction, we also equalized reciprocal cross-modal transfer coefficients while preserving their combined squared magnitude, keeping the checkpoint and message-value functions fixed. This increased positional velocity RMSE by 1.4671 Å per step (95% CI [1.4264, 1.5072]) and orientation error by 0.5277° ([0.4860, 0.5718]) across RAbD complexes, supporting the functional contribution of directional coordination to geometric prediction.

Table 1. MSC: Effects of R/S interventions on final generated structures across 60 paired RAbD complexes. Parentheses: Relative change versus MoDA.

MetricMoDAR zeroing: value (relative change)R randomization: value (relative change)S zeroing: value (relative change)S randomization: value (relative change)
Independently aligned per-CDR Cα RMSD ↓ (Å)0.73780.8854 (+20.01%)0.8281 (+12.24%)0.7222 (−2.12%)0.7241 (−1.85%)
CDR–framework anchor-distance error ↓ (Å)0.13480.3055 (+126.69%)0.2558 (+89.78%)0.1130 (−16.17%)0.1138 (−15.55%)
CDR–framework boundary C–N bond-length error ↓ (Å)0.16110.3252 (+101.91%)0.2736 (+69.88%)0.1461 (−9.31%)0.1469 (−8.79%)
Cross-CDR FAPE ↓ (Å)3.32003.2048 (−3.47%)3.1664 (−4.63%)3.3774 (+1.73%)3.4085 (+2.66%)
Same-chain inter-CDR FAPE ↓ (Å)3.04902.9367 (−3.68%)2.9187 (−4.27%)3.1151 (+2.17%)3.1781 (+4.24%)

Table 2a. MSC: Sequence recovery following separate interventions on each CDR’s U.

Target CDRMoDA: target AAR↑ (%)U zeroing: target AAR (target Δ% / other Δ%)U randomization: target AAR (target Δ% / other Δ%)U incorrect-unit replacement: target AAR (target Δ% / other Δ%)
H176.362473.5681 (−3.66% / +0.25%)73.5152 (−3.73% / +0.31%)72.2950 (−5.33% / +0.20%)
H260.111156.1667 (−6.56% / −0.26%)56.5000 (−6.01% / +0.24%)56.0000 (−6.84% / +0.16%)
H327.100428.5695 (+5.42% / −0.17%)29.2560 (+7.95% / +0.20%)27.6004 (+1.85% / +0.06%)
L171.870466.6283 (−7.29% / +0.15%)65.7944 (−8.45% / +0.54%)63.1521 (−12.13% / +0.31%)
L276.926469.0909 (−10.19% / −0.10%)67.0346 (−12.86% / +0.09%)68.0736 (−11.51% / −0.15%)
L362.547756.5350 (−9.61% / +0.17%)56.1648 (−10.20% / +0.24%)55.7942 (−10.80% / +0.08%)

Table 2b. MSC: Local structural recovery following separate interventions on each CDR’s U.

Target CDRMoDA target independently aligned Cα RMSD ↓ (Å)U zeroing: target RMSD (target Δ% / other Δ%)U randomization: target RMSD (target Δ% / other Δ%)U incorrect-unit replacement: target RMSD (target Δ% / other Δ%)
H10.54090.5514 (+1.93% / +0.11%)0.5556 (+2.72% / −0.17%)0.5633 (+4.14% / −0.46%)
H20.40920.4161 (+1.70% / −0.11%)0.4090 (−0.05% / −0.58%)0.4124 (+0.80% / −0.08%)
H31.85781.8015 (−3.03% / +0.39%)1.8304 (−1.48% / −0.17%)1.8531 (−0.25% / −0.37%)
L10.59430.6073 (+2.19% / +0.08%)0.6131 (+3.17% / −0.07%)0.6284 (+5.74% / +0.34%)
L20.34850.3574 (+2.56% / −0.66%)0.3524 (+1.11% / −0.33%)0.3546 (+1.77% / +0.16%)
L30.67600.7043 (+4.19% / −0.52%)0.7076 (+4.67% / −0.22%)0.7029 (+3.98% / −1.09%)

Table 2c. MSC: Backbone ψ-angle recovery following separate interventions on each CDR’s U.

Target CDRMoDA target ψ-angle error ↓ (°)U zeroing: target error (target Δ% / other Δ%)U randomization: target error (target Δ% / other Δ%)U incorrect-unit replacement: target error (target Δ% / other Δ%)
H121.566722.2387 (+3.12% / +0.72%)21.9046 (+1.57% / +0.24%)22.3732 (+3.74% / +1.40%)
H218.333318.7023 (+2.01% / +0.03%)18.5301 (+1.07% / +0.15%)17.8458 (−2.66% / +1.05%)
H353.443754.2088 (+1.43% / +0.55%)54.0236 (+1.09% / +0.42%)54.3609 (+1.72% / −0.21%)
L119.346620.3322 (+5.09% / +0.39%)20.3868 (+5.38% / −0.17%)21.5661 (+11.47% / +1.50%)
L216.442716.6191 (+1.07% / +0.07%)16.8638 (+2.56% / +0.76%)16.7611 (+1.94% / −0.34%)
L327.808229.5423 (+6.24% / +0.35%)29.5886 (+6.40% / +0.19%)29.3216 (+5.44% / −0.78%)

Table 3. CAR: Parameter-matched role-inference replacement with OIE, evaluated after full generation.

ModelParametersCDR RMSD ↓ (Å)AAR ↑ (%)
MoDA8,553,1741.577662.2509
CAR role inference → OIE8,553,2361.647661.1484

Table 4. CAR: Results after within-CDR permutation of complete role-probability vectors. Parentheses: Relative change versus paired MoDA baseline.

ConditionAAR ↑ (%)Fixed-frame Cα RMSD ↓ (Å)Backbone orientation error ↓ (°)Interface Cα contact recall ↑ (%)
MoDA60.89391.556723.821375.6678
Role permutation58.3161 (−4.23%)1.5635 (+0.44%)24.0858 (+1.11%)75.4824 (−0.25%)
95% CI of absolute change[−3.1462, −2.0378][+0.0018, +0.0125][+0.1311, +0.4073]°[−0.5171, +0.1245]

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.