A Comprehensive Benchmark of Computational Models for Antibody--Antigen Binding Affinity
Authors: Xinyan Zhao, Yi-Ching Tang, Victor J Cantu, Junseok Lee, Adam E Stogsdill, Ibraheem M Hamdi, Ashwin Kumar Ramesh, Zhiqiang An, Xiaoqian Jiang, Yejin Kim
Organizations: McWilliams School of Biomedical Informatics, UTHealth Houston · Department of Industrial and Systems Engineering, Korea Advanced Institute of Science and Technology · Texas Therapeutics Institute, Brown Foundation Institute of Molecular Medicine, UTHealth Houston
Computational antibody design requires benchmarks that measure functional binding rather than similarity to natural sequences. Here we present AbBiBench, which treats the antibody--antigen complex as the evaluation unit and harmonizes 216,560 experimental measurements spanning 13 antibodies and 9 antigens. We compare 20 computational approaches without task-specific training or fine-tuning. The analyses show that global structure-conditioned inverse folding yields the strongest and most consistent affinity rankings, whereas antibody-specific fine-tuning does not consistently improve performance. Across models, rankings are more useful for prioritizing likely affinity gains than for excluding weak variants. In a subsequent case study, multi-objective evaluation of model-generated F045-092 CDR-H3 variants for H1N1 binding revealed a trade-off between predicted binding potential and evolutionary sequence plausibility; all 21 selected candidates produced H1-reactive ELISA signals. AbBiBench provides an open framework for developing and evaluating function-aware antibody design methods.
Figures & tables
Figure 1: Overview of AbBiBench. a , Published studies are curated into antibody–antigen sequences, complex structures and binding measurements for zero-shot affinity prediction and antibody design, followed by multi-objective selection and ELISA validation. b , Evaluated protein language, inverse-folding, generative, structure-prediction and biophysical model families.
Figure 2: Composition and curation of AbBiBench. a , Coverage of 13 antibodies and nine antigen targets. Marker area scales with the logarithm of the assay-library size, marker color denotes antigen class and marker shape distinguishes direct Kd from enrichment readouts. b , Number of experimental observations per assay on a logarithmic scale. c , Composition by reported mutated chain, Chothia-defined mutation region and experimental readout. Mutation-region percentages exclude the single retained wild-type record (216,559 non-wild-type variants). d , Within-assay distribution of mutation depth, calculated as the total heavy- plus light-chain Hamming distance from the corresponding wild-type pair. e , Distributions of direct binding measurements after conversion to −log10(Kd[M]) . f , Distributions of log-enrichment measurements, which are relative high-throughput binding proxies rather than thermodynamic affinity constants. In e and f , density width is normalized within each assay; horizontal lines and points show interquartile ranges and medians, respectively. Counts follow the 16-assay manifest in Table 1 .
Figure 3: Zero-shot prediction of experimental binding affinity across assays and models. a , Spearman correlations between model scores and experimental binding affinities for each assay; blue and red denote negative and positive correlations, respectively. AlphaFold 3, Boltz-2 and ESMFold2 are scored by ipSAE on the same variants. b , Equal-assay mean Spearman ρ with assay-bootstrap 95% confidence intervals. c , Equal-assay Fisher- z mean Spearman correlations between model-predicted variant rankings. d , Mutation-stratified equal-assay mean Spearman correlations within mutation-region, mutated-chain, and mutation-distance strata. Cells summarize assay strata with at least 20 variants after requiring at least 90% safe-join and valid-score coverage; cross-column differences are descriptive because assay support differs. Model-family colors are used in a , b , and c .
Figure 4: Candidate prioritization and negative filtering across experimental budgets. a , Equal-assay mean precision for the top k=10,20,50, or 100 variants. Colored lines show the five strongest methods averaged across budgets; the gray band and line show the 10th–90th percentile and median of the others. Dotted and dashed lines mark the equal-assay prevalence of variants with at least fivefold affinity improvement (26.8%) and the finite-dataset oracle ceiling, respectively. b , Mean top- k precision across the 10 direct- Kd assays retained for this analysis; colored sidebars indicate model families. c , Assay-level Precision@10 for the five highlighted methods, with Wilson 95% binomial confidence intervals. d , Fraction of discarded variants worse than wild type as discard budget increased from 5% to 70%. The displayed scale covers model curves; oracle values at 5% and 70% are reported. e , Fraction of all worse-than-wild-type variants captured at each budget. In d and e , bad variants have p Kd below the assay-matched wild type; curves are equal-assay means across the eight evaluable direct- Kd assays. Colored lines show the five strongest filters; the gray band and line summarize the others. Dotted lines denote random ranking, and the dashed line in e denotes the oracle. Boundary ties were averaged over random selection within tied score groups.
Figure 5: Computational screening and experimental validation of generated F045-092 variants. a , Phase 1 evaluation of predicted binding potential, sequence plausibility and CDR-H3 sequence diversity. Predicted binding potential is reported as −ΔΔG , where ΔΔG=ΔGvariant−ΔGwild and ΔGwild=66.34 . The top 20% of variants are shown in the upper-right region; the 18 non-dominated Pareto-optimal variants are marked with × . b , Phase 2 evaluation of binding potential and sequence plausibility. c , Phase 2 evaluation of structural integrity using side-chain orientations and confidence scores for the complex and interfacial residues. d , Endpoint ELISA OD450 signals for 21 model-designed mutants and controls against H1 hemagglutinin. Diamonds indicate signals above the wild-type F045-092 control for H1; endpoint OD450 is a relative readout of H1-reactive binding under matched assay conditions.
Assay
Antibody–antigen
Variants
Mutation scope
Experimental source
Structural template
4fqi_h1
CR9114–H1N1 HA
HC: 65,094
CDR + framework
Kd ; germline-reverted yeast display 46
PDB 4FQI
4fqi_h3
CR9114–H3N2 HA
HC: 65,535
CDR + framework
Kd ; germline-reverted yeast display 46
PDB 4FQI
3gbn_h1
CR6261–H1N1 HA
HC: 1,887
CDR + framework
Kd ; germline-reverted yeast display 46
PDB 3GBN
3gbn_h9
CR6261–H9N2 HA
HC: 1,842
CDR + framework
Kd ; germline-reverted yeast display 46
PDB 3GBN
aayl49
AAYL49–Spike HR2
HC: 4,312
CDR
Kd ; phage-display DMS 12
AlphaFold 3
aayl49_ML
AAYL49_ML–Spike HR2
HC: 8,953
CDR
Kd ; ML-generated scFv, phage display 28
AlphaFold 3
Table 1: Experimental assays and structural templates included in AbBiBench. Variant counts are reported by mutated chain. Complete chain identities, sequence lengths and dataset-specific processing details are provided in Supplementary Table .
Computational antibody design has seen rapid methodological progress, with dozens of deep generative methods proposed in the past three years, yet the field lacks a standardized benchmark for fair comparison and model development. These methods are evaluated on different SAbDab snapshots, non-overlapping test sets, and incompatible metrics, and the literature fragments the design problem into numerous sub-tasks with no common definition. We introduce CHIMERA-Bench: (CDR Modeling with Epitope-guided Redesign), a unified benchmark built around a single canonical task: epitope-conditioned CDR sequence-structure co-design. CHIMERA-Bench provides three components. The first is a curated, deduplicated dataset of 2,922 antibody-antigen complexes with epitope and paratope annotations. The second is a set of three biologically motivated splits that test generalization to unseen epitopes, unseen antigen folds, and prospective temporal targets. The third is a comprehensive evaluation protocol with five metric groups, including novel epitope-specificity measures. We benchmark eleven methods spanning six generative paradigms and report results across all splits. CHIMERA-Bench is the largest dataset of its kind for the antibody design problem, allowing the community to develop and test novel methods and evaluate their generalizability.
Mansoor Ahmed, Nadeem Taj, Imdad Ullah Khan +2
Georgia State University, Atlanta, GA, USA · Georgia Institute of Technology, Atlanta, GA, USA · University of Engineering and Technology, Lahore, Pakistan +1
Accurate ranking of antibody candidates according to their binding affinity is essential for therapeutic antibody discovery. However, existing methods treat affinity comparisons independently and ignore the contextual information encoded in other labeled comparisons, limiting their ability to capture antigen-specific binding landscapes. For many target antigens, a small number of experimentally characterized affinity comparisons are often available. An important question is whether the model can exploit these existing comparisons to infer antigen-specific ranking patterns that facilitate subsequent affinity ranking. This form of learning from labeled demonstrations closely resembles the paradigm of In-Context Learning, motivating us to revisit antibody affinity ranking from an ICL perspective. To this end, we propose AbICL, an ICL framework for antigen-specific antibody affinity ranking. AbICL combines a pretrained structural encoder with a context ranking head and is trained with an episodic meta-training strategy that enables the model to leverage support demonstrations for test-time adaptation without gradient updates. Experiments on the AbRank benchmark demonstrate that AbICL consistently outperforms existing ranking baselines across almost all data splits and evaluation benchmarks. Further analysis shows that the value of contextual demonstrations depends on how well they match the target inference task, and becomes increasingly pronounced under distribution shift and fine-grained affinity discrimination. These findings highlight the potential of ICL as an effective paradigm for antigen-specific antibody affinity ranking, particularly in challenging settings where a single global ranking function is insufficient.
Antibodies play a central role in the immune response by specifically recognizing and neutralizing antigens, and therapeutic antibodies have become major drugs for cancer and autoimmune diseases. However, their discovery still relies on extensive in vitro screening, and accurate computational modeling of antibody structures and antibody-antigen interactions can prioritize candidates, reduce experimental burden, and accelerate rational design. Despite recent advances in high-accuracy protein and complex prediction, a persistent performance gap remains for antibody-related tasks compared with general protein-protein interactions, limiting downstream design. This thesis investigates why antibody-related tasks are harder and proposes improvements along two complementary directions. First, we investigate protein language model (PLM)-based methods for antibody and antibody-antigen structure prediction. Using embeddings from multiple PLMs, our approach achieves the best CDR-H3 accuracy among compared PLM-based methods on antibody monomer prediction. Extending it to complex prediction does not generalize: without co-evolutionary signals between antibody and antigen, single-sequence PLM representations do not reliably identify binding interfaces. Second, we develop two MSA-based interventions for antibody-antigen complex prediction: MSA refinement, which combines CDR-focused filtering with depth recovery from a larger sequence database, and convergence-aware recycling, which selects a stable intermediate recycle state for final diffusion sampling. Together, these interventions provide consistent gains over the AlphaFold3 baseline on a held-out antibody-antigen test set. Because the methods modify MSA construction and recycling behavior rather than model parameters, they apply without retraining or weight access.
Xiao Luo
TOYOTA TECHNOLOGICAL INSTITUTE AT CHICAGO, CHICAGO, ILLINOIS