Conditional Antibody Design as 3D Equivariant Graph Translation

1 INTRODUCTION

• Antibody design seeks CDRs that bind a given antigen with desirable properties, including high affinity and colloidal stability.Antibodies bind specific antigens, and their binding areas are mainly located in CDRs.Source: §1 · p. 1
• Existing approaches may omit antigen and other antibody-chain context, limiting the clues available for properties such as binding affinity.Some methods characterize only CDRs and the backbone context of the same antibody chain.Source: §1 · p. 1
• Invariant preprocessing can discard directional information, while autoregressive amino-acid prediction is inefficient and accumulates inference errors.The paper identifies these as limitations of prior methods that preprocess coordinates into invariant features and generate residues one at a time.Source: §1 · p. 2
• MEAN conditions CDR generation on the heavy-chain context, target antigen, and antibody light chain, and outputs both sequences and structures.Its equivariant message passing and attention capture long-range, spatially complex interactions among input components.Source: §1 · p. 2

2 RELATED WORK

• Earlier hand-crafted energy-function methods rely on costly simulations and may not fully capture complicated inter-chain interactions.Deep generative models have consequently been applied to sequence prediction.Source: §2 · p. 2
• Jin et al. (2021) co-designs CDR sequences and structures using E(3)-invariant distance features and autoregressive generation; Jin et al. (2022) adds epitope context but models only CDR-H3.General protein design predicts sequences from structures, whereas antibody design treats both CDR sequence and structure as unknown.Source: §2 · p. 2
• MEAN incorporates antigen and light-chain 1D/3D information and uses E(3)-equivariance to represent geometry and interactions between 3D structures.Its building blocks include scalarization-based E(n)-equivariant GNNs, with a multichannel extension (Huang et al., 2022) and a novel equivariant attention mechanism.Source: §2 · p. 2

3.1 PRELIMINARIES, NOTATIONS AND TASK FORMULATION

• The translator generates heavy-chain CDRs from a graph containing heavy chain, light chain, and antigen components.Residues use four backbone atoms; the main experiments select 48 antigen residues nearest the antibody by Cα distance.Source: §3.1 · p. 3
• Internal edges connect nearby residues within each component and always include adjacent chain residues; external edges connect nearby residues across components.Global nodes are added to each component to convey whole-chain context.Source: §3.1 · pp. 3–4

3.2 MEAN: MULTI-CHANNEL EQUIVARIANT ATTENTION NETWORK

• MEAN alternates internal-context and external-interaction encoding in E(3)-equivariant layers to model the antibody–antigen complex.The internal encoder also passes messages between components through global nodes.Source: §3.2 · pp. 4–5
• The internal encoder extends equivariant message passing to multichannel residue coordinates and updates coordinates while preserving directional information.Residues are represented by multiple backbone atoms, and the internal graph also connects global nodes across components.Source: §3.2 · p. 4
• The external encoder uses an E(3)-equivariant attention scheme to model correlations between residues in different components.Its attention weights aggregate messages from neighbors connected by external interaction edges.Source: §3.2 · p. 5
• MEAN's output is E(3)-equivariant: transforming the input coordinates transforms the predicted coordinates in the same way.The amino-acid distributions are unchanged by the coordinate transformation.Source: §3.2 · p. 5

3.3 PROGRESSIVE FULL-SHOT DECODING

• Each full-shot iteration predicts amino-acid types and 3D coordinates for all CDR residues, rather than generating one amino acid at a time.The predicted values update CDR inputs and graph edges for the next iteration.Source: §3.3 · p. 5
• Training supervises sequence predictions at every iteration and structure predictions only at the output iteration, using a weighted combination of losses.The structural loss operates directly on coordinates and uses Huber loss for coordinate-data noise.Source: §3.3 · pp. 5–6

4 EXPERIMENTS

• Evaluation includes modeling on 3,127 complete SAbDab complexes, CDR-H3 design benchmarked on 60 diverse RAbD complexes, and affinity optimization on 53 antibodies selected from SKEMPI V2.0.The 53 antibodies were selected from the SKEMPI V2.0 training set.Source: §4.1 · p. 6, §4.2 · p. 7, §4.3 · p. 8

4.1 SEQUENCE AND STRUCTURE MODELING

• MEAN significantly surpasses other methods in SAbDab 1D sequence and 3D structure modeling and, on Jin et al.'s split, wins AAR and matches RMSD despite using about 52% of the training data.Table 1 reports 10-fold cross-validation means and standard deviations for the SAbDab evaluation.Source: §4.1 · pp. 6–7, Table 1 · p. 7

4.2 ANTIGEN-BINDING CDR-H3 DESIGN

• MEAN outperforms all baselines by a large margin in AAR and TM-score for antigen-binding CDR-H3 design, with TM-score approaching above 0.99.The designed fragment in Figure 3 nearly overlaps the ground truth, while RefineGNN shows an apparent bias.Source: §4.2 · p. 7, Table 2 · p. 8

4.3 AFFINITY OPTIMIZATION

• MEAN models make obvious progress toward antibodies with better binding affinity; random CDR mutations provide an interpretive reference.Table 2 reports average affinity change after optimization.Source: Table 2 · p. 8, §4.3 · p. 8

4.4 CDR-H3 DESIGN WITH DOCKED TEMPLATE

• With docked templates, MEAN-designed antibodies have higher predicted binding affinity than C-RefineGNN and original RAbD antibodies.Ten templates are composed per antigen to mitigate docking inaccuracy, and the highest-scoring template is retained.Source: Figure 4 · p. 9, §4.4 · p. 9

• Removing either global nodes or attention reduces performance; global nodes transmit information globally, while attention focuses on local interface information.The attention module also provides interpretability over pairwise residue-interaction significance.Source: §5 · p. 9, Table 3 · p. 9
• Using only the heavy chain weakens performance and fails to yield a feasible affinity-optimization solution, supporting antigen and light-chain inputs.Source: §5 · p. 9
• Progressive full-shot decoding performs better than iterative refinement and runs about 2 to 5 times faster, using three iterations.The reported speedup depends on CDR sequence length.Source: §5 · p. 9, Table 3 · p. 9

• MEAN encodes local and global information within and between chains, then progressively decodes CDR sequences and structures more efficiently and precisely.The paper reports strong baseline comparisons across three generation tasks: 1D sequences and 3D structures, antigen-binding CDR-H3 design, and affinity optimization.Source: §6 · p. 9

A DETAILS OF SEQUENCE AND STRUCTURE MODELING

• The 10-fold setup assigns the preceding fold to validation and the remaining folds to training for each test fold.Fold 10 serves as validation when fold 1 is the test set; Table 4 reports statistics for each fold.Source: A DETAILS OF SEQUENCE AND STRUCTURE… · p. 14, Table 4 · p. 14
• For the RefineGNN split, MEAN* trains on complete-complex subsets comprising about 52% of the full sets, and all three models are evaluated on the test subset.RefineGNN* uses the official Jin et al. (2021) checkpoints; Table 5 reports set sizes and complete-complex subsets.Source: A DETAILS OF SEQUENCE AND STRUCTURE… · p. 14, Table 5 · p. 14

B ITERATIVE TARGET AUGMENTATION ALGORITHM FOR AFFINITY OPTIMIZATION

• ITA retains the top-k candidates after sorting each antibody’s new candidates together with its high-quality list, repeating for 20 iterations with C = 50 and k = 4.The procedure adapts ITA, originally designed for discrete properties, to the continuous affinity scorer.Source: B ITERATIVE TARGET AUGMENTATION… · p. 14

C THE PREDICTOR USED IN AFFINITY OPTIMIZATION

• The ∆∆G predictor’s output is calibrated in kcal/mol and correlates 0.65 with experimental values on the test set.A predicted ∆∆G of x corresponds to binding affinity decreased by x kcal/mol under that correlation.Source: C THE PREDICTOR USED IN AFFINITY… · pp. 14–15

D EXPERIMENT DETAILS AND HYPERPARAMETERS

• Across experiments, antigen context uses the 48 residues nearest the antibody by Cα distance, and models train for 20 epochs with Adam at lr = 0.001 and batch size 16.The learning rate decays by 0.95 each epoch, and the lowest-validation-loss checkpoint is selected for testing.Source: D EXPERIMENT DETAILS AND HYPERPARAMETERS · p. 15
• For MEAN* on the RefineGNN paper split, training uses alpha = 0.6, batch size 8, and 30 epochs.Table 6 lists model hyperparameters; ITA finetuning also uses learning rate 0.001.Source: D EXPERIMENT DETAILS AND HYPERPARAMETERS · p. 15, Table 6 · p. 15

E PROOF OF THEOREM 1

• MEAN’s translation process is E(3)-equivariant: transforming the input coordinates transforms the output coordinates correspondingly.The theorem defines orthogonal transformations and translations as the E(3) group actions.Source: E PROOF OF THEOREM 1 · p. 15
• The proof derives invariant hidden states and equivariant coordinates through internal and external encoders whose messages and query, key, and value vectors are E(3)-invariant.The argument proceeds through separate lemmas for the internal context encoder and external attentive encoder.Source: E PROOF OF THEOREM 1 · p. 16

F HUBER LOSS

• Huber loss models coordinates, using an L1 term for large deviations and an MSE term near zero; the experiments set δ = 1.The authors report that MSE alone occasionally causes NaN early in training, while Huber loss yields a more stable procedure.Source: F HUBER LOSS · p. 17

G COMPLEXITY ANALYSIS

• With T = 3 full-shot rounds versus RefineGNN’s usually greater-than-10 autoregressive steps, MEAN is reported to be more efficient despite its additional nodes.The stated complexities are O(2LKT(NA + NL + NH)) for MEAN and O(LKT′NH) for RefineGNN; the extra node computations can be parallelized.Source: G COMPLEXITY ANALYSIS · p. 17

H MODELING WITH RANDOMNESS

• Adding Gaussian noise to initialized coordinates gives rand-MEAN, whose sequence-structure modeling and CDR-H3 design results have acceptable performance impacts.Tables 7 and 8 report results for the two evaluated tasks.Source: H MODELING WITH RANDOMNESS · p. 17, Table 7 · p. 18, Table 8 · p. 18

I ATTENTION VISUALIZATION

• MEAN’s attention ranks for antigen residues contributing most to CDR-H3 binding energy skew toward low values, indicating correlation with Rosetta energies.Figure 5A shows an example attention map and energy map with similar distributions; Figure 5B displays relative-rank densities.Source: Figure 5 · p. 18, I ATTENTION VISUALIZATION · p. 18

J HOW GRAPH CHANGES DURING PROGRESSIVE DECODING

• External-edge recovery rises from 11% at the beginning to 84% at the end, while internal-edge recovery increases from 89% to 95%.The authors say the model predicts a major part of both internal and external edges, whereas linear initialization recovers only a small percentage of external edges.Source: J HOW GRAPH CHANGES DURING PROGRESSIVE… · pp. 18–19
• Across progressive decoding rounds, PPL density remains similar while RMSD shifts toward the optimal direction, despite coordinate supervision only at the last round.PPL is supervised with ground truth in all rounds.Source: J HOW GRAPH CHANGES DURING PROGRESSIVE… · p. 18, Figure 6 · p. 19

K LOCAL GEOMETRY

• MEAN achieves much better local-geometry performance than the compared models on bond lengths and backbone dihedral angles.The paper notes that RefineGNN's indirect loss on invariant features may leave backbone atoms unequally supervised; its carboxy-carbon coordinate RMSD is 10.8.Source: Table 9 · p. 19, K LOCAL GEOMETRY · p. 19

L FULL ANTIGEN OR ONLY EPITOPE

• Using the full antigen yields performance similar to using only the 48-residue epitope, with slight changes in AAR and structure modeling.The authors attribute the similarity to message passing being limited to nearby residues by the Cα-distance cutoff.Source: L FULL ANTIGEN OR ONLY EPITOPE · p. 19, Table 10 · p. 19

M LIMITATIONS AND FUTURE WORK

• MEAN models backbone geometry only, although sidechains also contribute to antigen–antibody interactions.The authors suggest extending the method with a local full-atom graph for each residue.Source: M LIMITATIONS AND FUTURE WORK · p. 20
• The small antibody-structure dataset reflects the difficulty of obtaining 3D structures; sequence pretraining and pseudo-structure generation are proposed for future augmentation.Other proposed directions include pretraining on general protein data and fine-tuning on antibody data.Source: M LIMITATIONS AND FUTURE WORK · p. 20

N EXAMPLES

• Figure 7 presents additional examples of MEAN-designed antigen-binding CDR-H3 sequences and structures.Source: Figure 7 · p. 20, N EXAMPLES · p. 20

O TOWARDS THE REAL-WORLD QUESTION

• Affinity-optimization efficacy depends on predictor generalizability; the proof-of-concept uses Shan et al. (2022), while wet-lab validation could return real affinity but is time-consuming.The authors also propose combining predictor-based selection with wet-lab feedback.Source: O TOWARDS THE REAL-WORLD QUESTION · p. 21
• A broader antibody-discovery pipeline would add epitope identification and antibody-structure prediction, but accumulated errors across components could cause later stages to fail.The proof-of-concept combines docking, CDR design, and affinity computation; the authors suggest making all components learnable and tuning them together.Source: O TOWARDS THE REAL-WORLD QUESTION · p. 21