acceptodds
Under review as a conference paper at ICLR 2027

IgGM2: An All-Atom Foundation Model for Adaptive Immune Receptor Design

Abstract

Accurate immune receptor design requires modeling the coupled variation of amino-acid sequence, full-atom conformation, and target-binding geometry across antibodies, nanobodies, and T-cell receptors (TCRs). Existing methods often address only part of this problem, either by separating structure generation from sequence design, relying on fixed-backbone inverse folding, or focusing on a single receptor class. We introduce IgGM2, a unified all-atom generative framework for structure prediction and CDR sequence–structure co-design. For structure prediction, IgGM2 models receptor monomers from sequence and receptor–target complexes conditioned on fixed target structures and epitope context. For CDR design, a two-stage sampler first jointly generates CDR sequences and backbone conformations, then refines full-atom receptor structures with the generated sequences fixed. This allows receptor frameworks to adapt to designed CDRs without separate inverse folding or external side-chain packing. On structure prediction benchmarks, IgGM2 achieves strong performance on FoldBench and in TCR–pMHC modeling. On sequence design benchmarks, IgGM2 achieves competitive amino-acid recovery and improves Rosetta-based interface preference metrics, suggesting more favorable generated binding interfaces. These results support IgGM2 as a unified all-atom framework for adaptive immune receptor structure prediction and design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Antibody, TCR, Diffusion

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_5W71OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

IgGM2 presents a unified all-atom diffusion framework for structure prediction and sequence–structure co-design of antibodies, nanobodies, and T-cell receptors. The model jointly generates CDR sequences and backbone conformations, followed by full-atom refinement. The authors demonstrate strong performance on immune-complex structure prediction and competitive results on CDR design benchmarks.

Critical strengths

The unified approach to sequence and structure generation is compelling, particularly its extension across antibodies, nanobodies, and TCRs. The structure prediction results are impressive, and the coordinated sequence–structure denoising strategy is technically interesting. The broad benchmarking provides encouraging evidence of the framework’s utility.

Critical weaknesses

My primary concern is that the structure prediction comparisons are not entirely equivalent, as IgGM2 can use fixed antigen structures and epitope information that are not provided to all baselines. Additionally, the design results rely entirely on computational metrics rather than experimental validation. While the generated structures appear promising, their functional performance remains uncertain.

Questions for authors
  1. How much of the improvement in structure prediction can be attributed to the model architecture versus the additional antigen and epitope information?
  2. Does joint all-atom generation provide meaningful advantages over conventional backbone generation followed by sequence design and structural relaxation?
3 comments

w1/q1

We thank the reviewer for this question. Our comparison with conventional structure prediction methods is not strictly input-matched, as IgGM2-P uses additional antigen and epitope information. However, IgGM2 is designed for target-conditioned antibody and TCR prediction and design, where such information can be used to guide receptor generation, rather than predicting the entire complex from scratch.

To assess the contribution of conditioning information, we additionally evaluated IgGM2-P using AF3-predicted antigen structures, either predicted independently or extracted from predicted complexes, with and without epitope information. We also evaluated AF3 and Protenix v1 under a single-sample (1×1) setting, matching the sampling budget of IgGM2-P.

Method / Input SettingMean DockQ ↑SR ↑
AF3 (5×5)0.360047.90%
Protenix v0.5 (5×5)0.230034.10%
Protenix v1 (5×5)0.360048.84%
AF3 (1×1)0.253733.14%
Protenix v1 (1×1)0.329344.19%
IgGM2-P: GT antigen + epitope (1×1)0.604073.84%
IgGM2-P: GT antigen, no epitope (1×1)0.498061.00%
IgGM2-P: AF3 antigen-only + epitope (1×1)0.350051.16%
IgGM2-P: AF3 complex-extracted antigen + epitope (1×1)0.390059.30%
IgGM2-P: AF3 antigen-only, no epitope (1×1)0.310047.67%
IgGM2-P: AF3 complex-extracted antigen, no epitope (1×1)0.320046.51%

Epitope conditioning improves Mean DockQ from 0.498 to 0.604 and SR from 61.00% to 73.84% with native antigen structures. With predicted antigens and known epitopes, IgGM2-P achieves SRs of 51.16%–59.30%, exceeding the reported AF3 and Protenix v1 baselines. Without epitopes, its SRs of 46.51%–47.67% remain above their 1×1 results. These experiments also reveal sensitivity to antigen structural accuracy.

Beyond conditioning, we provide additional ablations supporting our training and modeling strategies.

1. Pretrained Initialization. We compared pretrained initialization with training IgGM2 from scratch on the cleaned 203-sample test subset.

InitializationNAll AAR ↑DockQ ↑HCDR3 local RMSD (Å) ↓IMP ↑
IgGM2 (Pretrained)2030.53100.28542.39350.5911
Random Initialization2030.50100.01962.72620.1921

Pretrained initialization substantially improves sequence recovery, DockQ, and IMP.

2. Shared Receptor Training. Due to computational constraints, we conducted this ablation using TCR–pMHC data. Both models inherit the same Protenix pretrained weights and are evaluated on 18 TCR–pMHC complexes.

Training StrategyNDockQ ↑SR ↑iRMSD (Å) ↓LRMSD (Å) ↓
Ab + TCR Joint Training180.74611.00001.26162.9663
TCR-only Training180.54040.83332.38125.9482

Joint training improves DockQ from 0.5404 to 0.7461, supporting the benefit of learning across receptor types.

3. Noise-aligned Masking. We compared our masking strategy with MFDesign-style linear masking.

Masking StrategyAAR ↑DockQ ↑SR ↑iRMSD (Å) ↓LRMSD (Å) ↓IMP ↑
Noise-aligned Masking (Ours)0.5310.2850.4436.3416.740.591
MFDesign-style Linear Masking0.5330.2320.3558.8522.030.502

Noise-aligned masking improves structural and interface quality with comparable AAR.

Furthermore, IgGM2-P requires no MSA or multiple-candidate ranking. As reported in Table S2, its end-to-end runtime is approximately 14 seconds per complex, compared with approximately 30 minutes for AF3 under the evaluated protocol, including MSA generation.

Overall, these results support the contributions of both structural conditioning and our training and modeling strategies.

w2

We agree that experimental validation is important for assessing the functional performance of our designs. Our current evaluation relies on computational metrics and does not directly establish binding activity. We recognize this limitation and plan to seek experimental collaborations to validate representative designs on selected targets in future work.

q2

Directly demonstrating the advantages of joint all-atom generation over conventional design pipelines is challenging without experimental validation. We believe that joint all-atom generation can better capture sidechain interactions at the antibody–antigen interface. To evaluate this, we calculated shape complementarity (Sc) using PyRosetta, where higher Sc indicates better interface complementarity. All results were evaluated without FastRelax.

We first compared IgGM2-D with existing antibody design methods:

MethodMean Sc ↑
DiffAb (HDOCK)0.4203
AbX (HDOCK)0.4072
dyMEAN0.1898
IgGM0.3949
RFAntibody0.2893
IgGM2-D (switch=75)0.4700

IgGM2-D achieves the highest mean Sc (0.4700) among the compared methods.

We further compared single-phase generation (switch=200) and two-phase generation (switch=75), with and without sidechain packing.

Sampling StrategyMean Sc ↑
Two-phase (switch=75), generated sidechains0.4700
Two-phase (switch=75), sidechain repacking0.4713
Single-phase (switch=200), backbone only0.4002
Single-phase (switch=200), sidechain packing0.4330

Sidechain packing improves single-phase Sc from 0.4002 to 0.4330, but it remains below two-phase generation (0.4700). Repacking the generated sidechains in the two-phase setting gives nearly identical Sc (0.4713 vs. 0.4700), suggesting that joint all-atom generation can capture more favorable interface interactions than conventional backbone generation followed by sidechain packing.

Dear Reviewer,

Thank you for your positive evaluation and constructive feedback. We would appreciate it if you could let us know whether our responses have addressed your concerns or if any further questions remain. Thank you!

Reviewer_p5gWOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

IgGM2 adapts a target-conditioned structure model for antibody, nanobody, and T-cell receptor design. IgGM2-D generates CDR sequences and backbones, then refines full receptor structures with sequences fixed and the target held constant. The reported gains in conditioned prediction, sequence recovery, and relaxed interface energies are promising, while comparison conditions, test-based sampler selection, and functional validation limit the conclusions.

Critical strengths
  • A shared architecture supports several receptor classes and both prediction and design. Generating framework coordinates allows them to adapt to designed CDRs.
  • On the deduplicated nanobody test set, average AAR is 0.372 versus 0.287 for DiffAb. Paired-antibody framework-aligned C-alpha RMSD is lowest on five of six loops.
  • The masking ablation raises mean DockQ from 0.232 to 0.285 with similar AAR, supporting the chosen noise-aligned schedule.
  • Appendices C-D report weaker unliganded TCR docking, poor AF3 interface agreement, and energy distributions and relaxation effects, making the method's limits easier to assess.
Critical weaknesses
  • FoldBench's leaderboard models predict full assemblies, whereas IgGM2-P receives fixed target coordinates and epitope information. Table 8 adds useful controls, but pocket, template, and fixed-target inputs differ, so the comparison does not isolate improvements from the model itself.
  • Binding assays and direct clash or sidechain-accuracy measurements are not reported. AF3 refolding supports receptor-fold consistency but gives poor interface agreement, which is inconclusive for binding (Table S10).
  • IgGM2 adds direct full-atom refinement and TCR coverage, but its antibody-design benefit over MFDesign is unmeasured. MFDesign already adapts an AF3-like model for CDR sequence/backbone co-design.
Questions for authors
  • Can you compare supported methods with matched target/epitope inputs and sampling budgets, and evaluate robustness to predicted targets?
  • Can you explain how the SAbDab-22H2-Ab, SAbDab-22H2-Nano, and STCRDab-22-TCR datasets are constructed.
  • Can you select sampler settings using validation data and report fresh test results with uncertainty estimates?
  • Please clarify alignment, aggregation, and sample reuse in Tables 3 and S13: why is IgGM2-D HCDR3 RMSD 2.288 A on the full test set antibodies but 3.492 A on the smaller subset?
  • Can you add direct clash/sidechain checks and binding measurements for representative antibody and TCR designs?
5 comments

w1/q1

We thank the reviewer for raising this concern. Our prediction setting differs from conventional full-complex prediction methods, as IgGM2 uses additional target information. This reflects our focus on target-conditioned antibody and TCR prediction and design, where available antigen structures and epitope information are used to guide receptor generation.

To provide a more comprehensive comparison, we additionally evaluate IgGM2-P using AF3-predicted antigen structures under two settings: (1) antigen-only prediction and (2) antigen structures extracted from AF3-predicted complexes. We evaluate both settings with and without known epitope information.

We also report single-sample (1×1) results for AF3 and Protenix v1, matching the sampling budget of IgGM2-P.

Method / Input SettingMean DockQ ↑SR (DockQ ≥ 0.23) ↑
AF3 (5×5)0.3647.90%
Protenix v0.5 (5×5)0.2334.10%
Protenix v1 (5×5)0.3648.84%
AF3 (1×1)0.2533.14%
Protenix v1 (1×1)0.3344.19%
IgGM2-P: GT antigen + epitope (1×1)0.6073.84%
IgGM2-P: AF3 antigen-only + epitope (1×1)0.3551.16%
IgGM2-P: AF3 complex-extracted antigen + epitope (1×1)0.3959.30%
IgGM2-P: AF3 antigen-only, no epitope (1×1)0.3147.67%
IgGM2-P: AF3 complex-extracted antigen, no epitope (1×1)0.3246.51%

As shown above, IgGM2-P remains competitive when using predicted antigen structures. With known epitopes, it achieves success rates of 51.16% and 59.30%, exceeding the reported AF3 and Protenix v1 baselines. Even without epitope information, IgGM2-P achieves success rates of 46.51%–47.67%, outperforming their 1×1 results.

We also observe differences between antigen-only and complex-extracted antigen inputs, suggesting that antigen conformational changes, including possible apo/holo differences, affect binding-pose prediction. Addressing antigen conformational changes is an important but separate challenge beyond the primary scope of our target-conditioned receptor modeling framework.

Importantly, IgGM2-P requires no MSA and generates only a single prediction without candidate ranking. The additional 1×1 baseline results further demonstrate its competitive performance under matched sampling budgets, while its MSA-free architecture avoids the overhead of MSA construction.

w2/q5

The AF3 refolding experiment in Table S10 was primarily intended to assess receptor-fold consistency rather than agreement between AF3-predicted binding poses and our generated structures. We believe it can also provide insight into how CDR sequence changes affect predicted binding conformations. To further investigate this, we compared AF3 predictions using native and IgGM2-designed antibody sequences on the same 203 complexes, with newly generated MSAs for the designed sequences. As shown below, the designed sequences yield lower DockQ (0.0705 vs. 0.1542), suggesting that CDR sequence changes can affect predicted binding poses and further motivating joint generation of framework coordinates and designed CDRs, rather than keeping the framework fixed.

AF3 Input (1×1)Mean DockQ ↑SR (DockQ ≥ 0.23) ↑
Native sequences (with MSA)0.154219.70%
IgGM2-designed sequences (with newly generated MSA)0.07053.94%

To further assess interface quality without FastRelax, we calculated shape complementarity (Sc) using PyRosetta, where higher values indicate better geometric complementarity. IgGM2-D achieves the highest Sc (0.4700) among the compared methods, suggesting favorable interface geometry.

MethodMean Sc ↑
DiffAb (HDOCK)0.4203
AbX (HDOCK)0.4072
dyMEAN0.1898
IgGM0.3949
RFAntibody0.2893
IgGM2-D0.4700

We also evaluated atomic clashes and sidechain geometry using Protenix predictions of our designed sequences for comparison. IgGM2-D achieves comparable or slightly better local sidechain geometry but has a higher intrachain clash score, indicating remaining limitations in atomic consistency.

MetricIgGM2-DProtenix
Heavy-chain intrachain clash score ↓32.382622.6092
CDR sidechain RMSD (Å) ↓1.14641.1916
Framework sidechain RMSD (Å) ↓0.97351.1089
CDR sidechain bond-length deviation (Å) ↓0.02270.0267
CDR χ1 MAE (°) ↓38.8539.41
CDR χ1 recovery (within 20°) ↑64.31%64.28%

These results provide additional computational evidence for the structural quality of our designs, although experimental binding validation remains necessary. We plan to seek experimental collaborations to validate representative antibody and TCR designs on selected targets.

w3

MFDesign's original CDR co-design setting uses ground-truth antibody–antigen complexes, preserving the antibody framework coordinates and binding pose through replacement sampling. In contrast, IgGM2-D generates antibody structures and binding poses without requiring ground-truth antibody coordinates.

For a more direct comparison, we evaluated MFDesign under two settings: (1) without inpainting and (2) with antigen-only inpainting. The latter matches our target-conditioned setting, where only antigen coordinates are fixed during sampling.

MethodDockQ ↑SR ↑HCDR3 local RMSD (Å) ↓
MFDesign (w/o inpainting)0.05074.43%2.5420
MFDesign (antigen-only inpainting)0.05804.93%2.5596
IgGM2-D (switch=75)0.291347.78%2.3860

As shown above, IgGM2-D substantially outperforms MFDesign under the matched antigen-only conditioning setting, achieving higher DockQ (0.2913 vs. 0.0580) and SR (47.78% vs. 4.93%), along with lower HCDR3 local RMSD (2.3860 Å vs. 2.5596 Å). These results demonstrate the advantage of IgGM2-D in target-conditioned antibody design without ground-truth antibody structures or binding poses.

q2

For fair comparison, we directly adopted the publicly released test sets from tFold rather than constructing new benchmarks. Their construction procedures are described in the original tFold-Ab and tFold-TCR papers.

Specifically, SAbDab-22H2-Ab (169 samples) and SAbDab-22H2-Nano (73 samples) were collected from SAbDab, using structures released between July 1 and December 31, 2022. STCRDab-22-TCR (24 samples) was constructed from STCRDab using a temporal split, with structures released in 2022 or later reserved for testing. All three test sets contain structures released after September 30, 2021, the PDB training cutoff of Protenix.

Furthermore, we independently filtered our training data against these benchmarks to remove overlapping or highly similar sequences, ensuring separation between our task-specific training and test sets.

q3

We additionally evaluated different switch steps on the validation set (227 samples) using three random seeds due to computational constraints. We report the mean ± standard deviation (SD) across seeds.

Switch StepNAAR ↑DockQ ↑SR ↑IMP ↑
252270.5160 ± 0.00300.3112 ± 0.00820.4963 ± 0.01020.6182 ± 0.0369
502270.5185 ± 0.00040.3195 ± 0.01430.4890 ± 0.01530.6182 ± 0.0111
752270.5211 ± 0.00080.3066 ± 0.00410.4743 ± 0.00920.6050 ± 0.0313
1002270.5207 ± 0.00110.3049 ± 0.00250.4758 ± 0.01530.5962 ± 0.0142

Consistent with the selection criterion described in Section 4.4, we select the phase-transition step primarily based on All AAR. Switch=75 achieves the highest validation AAR (0.5211), supporting our original choice. We therefore retain switch=75 for our main experiments.

With the sampler setting fixed, we reran the evaluation on the test subset of 203 samples using five random seeds to estimate performance variability.

Test SetNAAR ↑DockQ ↑SR ↑IMP ↑
Cleaned Test2030.5296 ± 0.00120.3141 ± 0.00830.5044 ± 0.01330.6128 ± 0.0248

The refreshed test results show consistent performance across five random seeds, and we will include these results in the revised manuscript.

q4

The difference mainly comes from the RMSD alignment methods. In Table 3, we first align the antibody framework regions and then calculate the CDR RMSD, which captures both CDR conformational differences and their relative positioning within the framework. In contrast, Table S13 uses local CDR alignment, where each CDR is aligned independently to evaluate its local conformation without considering framework positioning. Therefore, the local RMSD in Table S13 is substantially lower (2.288 Å vs. 3.492 Å). We will clarify these alignment protocols in the revised manuscript.

Dear Reviewer,

We kindly invite you to review our responses and let us know if you have any further questions or concerns. Thank you!

Comments on the paper

8 comments

w1

We apologize for the confusion caused by the labeling of switch step 200 in Table 7, which may have obscured the comparison between single-phase and two-phase generation. We clarify this distinction and provide additional analyses below.

1. Phase Splitting

We have already evaluated phase splitting in Table 7. Since the maximum number of diffusion sampling steps is 200, switch=200 corresponds to single-phase generation (no phase splitting), whereas switch=75 enables two-phase generation.

Sampling StrategyAAR ↑Paired AAR ↑Nano AAR ↑HCDR3 local RMSD (Å) ↓IMP ↑
Single-phase (switch=200)0.52460.58640.35832.39340.5567
Two-phase (switch=75)0.53100.59020.37172.39350.5911

Two-phase generation improves both AAR (0.5310 vs. 0.5246) and IMP (0.5911 vs. 0.5567), while maintaining comparable HCDR3 local RMSD. We therefore selected switch=75 for our main experiments.

2. Framework Adaptation

Our motivation for generating framework coordinates is that CDR sequence changes may affect antibody conformations and binding poses. To investigate this, we performed AF3 predictions using native and IgGM2-designed antibody sequences on the same 203 complexes, with newly generated MSAs for the designed sequences.

AF3 Input (1×1)Mean DockQ ↑SR (DockQ ≥ 0.23) ↑
Native sequences (with MSA)0.154219.70%
IgGM2-designed sequences (with newly generated MSA)0.07053.94%

The lower DockQ for designed sequences suggests that CDR sequence changes can affect predicted binding poses. This motivates generating framework coordinates jointly with designed CDRs, rather than restricting them to a fixed native framework.

3. Shared Receptor Training

To investigate the effect of shared receptor training, we conduct the ablation using only TCR–pMHC data due to computational resource constraints. We train a TCR-only model initialized from the same pretrained checkpoint and compare it with our jointly trained model on the same TCR–pMHC test complexes from STCRDab-22-TCR_pMHC.

Training StrategyDockQ ↑SR ↑iRMSD (Å) ↓LRMSD (Å) ↓GDT_TS ↑TM-score ↑
Ab + TCR Joint Training0.74611.00001.26162.96630.88670.9581
TCR-only Training0.54040.83332.38125.94820.81650.8954

Joint training substantially outperforms TCR-only training, improving DockQ from 0.5404 to 0.7461 and reducing iRMSD from 2.3812 Å to 1.2616 Å. Since both models inherit the same pretrained weights, these results support the benefit of shared receptor training.

4. Pretrained Initialization

We further validate the importance of inheriting pretrained weights by training IgGM2 from scratch with random initialization, without using pretrained weights. Both models are evaluated on the same test subset of 203 samples.

InitializationAll AAR ↑Paired AAR ↑Nano AAR ↑DockQ ↑HCDR3 local RMSD (Å) ↓IMP ↑CAAR ↑
IgGM2 (ProteinX pretrained)0.53100.59020.37170.28542.39350.59110.4218
Random Initialization0.50100.55680.35080.01962.72620.19210.3812

Pretrained initialization improves performance across all metrics, particularly DockQ (0.2854 vs. 0.0196) and IMP (0.5911 vs. 0.1921), while also improving sequence recovery and HCDR3 local RMSD. These results demonstrate the importance of inheriting pretrained weights for both sequence and structure generation.

w2/q1

The comparison in Table 1 is not strictly input-matched, as IgGM2-P uses fixed antigen structures and epitope information. However, our goal is target-conditioned antibody and TCR prediction and design, where available target information can guide receptor generation rather than predicting the entire complex from scratch. To further address the concerns about comparison fairness and potential data leakage, we examined the pretraining data for overlap with FoldBench and conducted additional experiments using predicted antigen structures.

1. Pretraining data check

Since IgGM2 is fine-tuned from Protenix, we carefully examined its training data and checked for overlap with FoldBench. According to the Protenix technical report, its PDB training data cutoff is September 30, 2021. We further downloaded the publicly released training index and found no direct PDB-level overlap with FoldBench, including verified replacement IDs.

Training CutoffMatching CriterionComplexesInterfaces
2021-09-30Original PDB IDs0/1130/172
2021-09-30Including verified replacement IDs0/1130/172

We also examined the construction of FoldBench. The benchmark uses structures deposited after January 13, 2023, and applies sequence-homology filtering to reduce overlap with the reconstructed AF3 training set. Taken together, these findings suggest that substantial training-data leakage from the inherited Protenix checkpoint is unlikely.

2. Evaluation with Predicted Antigen Structures

To further investigate the effect of native antigen structures, we replace them with AF3-predicted structures under two settings: (1) AF3 predicts the antigen alone, and (2) the antigen structure is extracted from an AF3-predicted antibody–antigen complex. In both cases, only the predicted antigen coordinates are provided to IgGM2-P, with or without known epitope information. We also report single-sample (1×1) results for AF3 and Protenix v1 to match the sampling budget of IgGM2-P.

Method / Antigen InputMean DockQ ↑SR (DockQ ≥ 0.23) ↑
AF3 (5×5)0.3647.90%
Protenix v1 (5×5)0.3648.84%
Protenix v0.5 (5×5)0.2334.10%
AF3 (1×1)0.2533.14%
Protenix v1 (1×1)0.3344.19%
IgGM2-P + GT antigen + epitope (1×1)0.6073.84%
IgGM2-P + AF3 antigen-only + epitope (1×1)0.3551.16%
IgGM2-P + AF3 complex-extracted antigen + epitope (1×1)0.3959.30%
IgGM2-P + AF3 antigen-only, no epitope (1×1)0.3147.67%
IgGM2-P + AF3 complex-extracted antigen, no epitope (1×1)0.3246.51%

With predicted antigen structures and known epitopes, IgGM2-P achieves success rates of 51.16% and 59.30%, both exceeding the reported AF3 (47.90%) and Protenix v1 (48.84%) baselines. For Mean DockQ, the complex-extracted setting reaches 0.39, exceeding both baselines (0.36), while the antigen-only setting achieves 0.35. Even without epitope information, IgGM2-P achieves success rates of 46.51%–47.67%, outperforming both methods under the matched 1×1 sampling budget.

We also evaluate the structural differences between AF3-predicted and native antigen structures:

Antigen SourceCα RMSD (Å) ↓TM-score ↑lDDT-Cα ↑
AF3 antigen-only3.62920.83480.8972
AF3 complex-extracted antigen3.96180.81960.8789

These differences in antigen structure likely contribute to the performance decrease compared with native antigen inputs. Overall, these results show that IgGM2-P can effectively model receptor binding poses using predicted antigen structures, supporting the practical value of our target-conditioned formulation.

w3/q2

As discussed in Appendix B.4, staged sidechain refinement introduces certain limitations, while the results in Appendix D.4 show that Rosetta IMP is strongly influenced by FastRelax. To further assess the quality of directly generated structures, we conducted additional evaluations of interface shape complementarity, AF3 refolding, and sidechain geometry.

1. Independent Interface Evaluation.

To evaluate interface quality without relying on FastRelax, we calculated shape complementarity (Sc) using PyRosetta, where higher values indicate better geometric complementarity between antibody and antigen surfaces.

We first compared single-phase backbone generation (switch=200) with two-phase all-atom generation (switch=75). We also applied Rosetta sidechain packing to examine the contribution of directly generated sidechains. All Sc values were computed without FastRelax.

Sampling StrategyMean Sc ↑
Single-phase (switch=200), backbone only0.4002
Single-phase (switch=200) + sidechain packing0.4332
Two-phase (switch=75), generated sidechains0.4700
Two-phase (switch=75) + sidechain repacking0.4713

Sidechain packing improves Sc for single-phase generation from 0.4002 to 0.4332, but the result remains below two-phase generation (0.4700). Interestingly, replacing our generated sidechains with Rosetta-packed sidechains yields nearly identical Sc (0.4713 vs. 0.4700). These results suggest that two-phase generation produces more complementary interfaces than single-phase generation followed by sidechain packing, while directly generated sidechains achieve comparable shape complementarity to Rosetta packing.

We further compared IgGM2-D with existing antibody design methods using structures evaluated without FastRelax.

MethodMean Sc ↑
DiffAb (HDOCK)0.4203
AbX (HDOCK)0.4072
dyMEAN0.1898
IgGM0.3949
RFAntibody0.2893
IgGM2-D (switch=75)0.4700

IgGM2-D achieves the highest mean Sc among the compared methods. Together, these results provide additional evidence that our joint all-atom generation produces geometrically complementary interfaces without relying on FastRelax.

2. Clarification of AF3 Refolding.

The AF3 refolding experiment in Section D.3 was primarily intended to assess receptor-fold consistency rather than agreement between AF3-predicted binding poses and IgGM2-generated interfaces. Beyond this, we believe refolding can also provide insight into how CDR sequence changes affect predicted complex conformations.

To further investigate this, we performed additional AF3 predictions using native antibody sequences and IgGM2-designed sequences on the same 203 complexes, with independently generated MSAs for both settings.

AF3 Input (1×1)Mean DockQ ↑SR (DockQ ≥ 0.23) ↑
Native antibody sequences (with MSA)0.154219.70% (40/203)
IgGM2-designed sequences (with newly generated MSA)0.07053.94% (8/203)

AF3 achieves lower DockQ for designed sequences than for native sequences (0.0705 vs. 0.1542), suggesting that CDR sequence changes can affect predicted binding conformations. Although this does not directly establish the corresponding structural changes experimentally, it further motivates allowing framework coordinates to change jointly with designed CDRs, rather than keeping them fixed.

3. Sidechain Quality Evaluation.

We further assessed sidechain geometry and atomic clashes, using Protenix predictions of our designed sequences as an additional comparison.

MetricIgGM2-D (switch=75)Protenix
Heavy-chain intrachain clash score (heavy-atom proxy) ↓32.382622.6092
CDR sidechain RMSD (Å) ↓1.14641.1916
Framework sidechain RMSD (Å) ↓0.97351.1089
CDR sidechain bond-length deviation (Å) ↓0.02270.0267
CDR χ1 MAE (°) ↓38.8539.41
CDR χ1 recovery (within 20°) ↑64.31%64.28%

IgGM2-D achieves comparable or slightly better sidechain geometry metrics than Protenix. However, it exhibits a higher intrachain clash score, indicating that some atomic overlaps remain unresolved in directly generated structures.

These results suggest that our model generates reasonable local sidechain conformations, while further improvements in global atomic consistency are needed. We will discuss this limitation more explicitly in the revised manuscript.

Overall, the additional Sc analyses support the benefits of two-phase all-atom generation over single-phase backbone generation followed by sidechain packing, without relying on FastRelax.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.