- The paper introduces a listwise ranking framework that leverages PU learning to address label sparsity and improve binding affinity predictions among homologous candidates.
- It employs dual GCNs and Multi-Head Self-Attention to capture residue-level interactions and contextual dependencies, resulting in superior performance metrics.
- Empirical results demonstrate significant gains over baseline methods, highlighting robust generalization in both clinical and viral case studies.
AbLWR: A Context-Aware Listwise Ranking Framework for Antibody-Antigen Binding Affinity Prediction via Positive-Unlabeled Learning
Introduction and Motivation
The prediction of antibody-antigen (Ab-Ag) binding affinity is a central task for computational immunology and therapeutic antibody design. The problem is challenged by two critical impediments: (1) the severe sparsity of experimentally labeled binding affinity data despite vast available sequence space, and (2) the complexity introduced by subtle antigenic variation, whereby homologous antigenic variants or antibody paratopes can differ minimally in sequence yet exhibit divergent binding profiles. Conventional computational pipelines—from empirical physics-based models (e.g., FoldX) to sequence/structure/complex-based deep learning models—typically optimize pointwise regression objectives, which insufficiently capture the nuanced relative ordering crucial for candidate prioritization.
AbLWR proposes a paradigm shift by reframing affinity prediction as a listwise ranking problem, aligned with the practical prioritization in antibody discovery. It exploits unlabeled data via positive-unlabeled (PU) learning to address label sparsity and incorporates context-aware relational modeling using homologous antigen sampling and Multi-Head Self-Attention (MHSA) ranking modules.
Methodological Framework
Rather than predicting absolute affinity values for isolated Ab-Ag pairs, AbLWR models the relative ordering of candidate pairs corresponding to identical or homologous antigens. Informative training lists are constructed using a homologous sampling strategy, in which antigen sequence similarity and affinity margins guide inclusion, ensuring sampling focuses on challenging cases where global context is essential to discriminate top binders.
Representation Learning with PU and Contrastive Objectives
To address label sparsity, AbLWR integrates a robust PU learning module in its pre-training phase, where experimentally verified (labeled) pairs are treated as positives and unverified pairs as unlabeled. The architecture utilizes dual Graph Convolutional Networks (GCNs) for the antibody CDR interface and antigen surface, providing residue-level encoding via high-capacity protein LLMs (PLMs) such as IgFold and ESM-2.
Dual-level contrastive learning is employed: an instance-level contrastive loss aligns perturbed (weak and strong) graph augmentations, and a cluster-aware loss groups semantically congruent pairs via consensus verification and geometric preservation in latent space. Pseudo-labels for unlabeled samples are meta-optimized using a bi-level approach, iteratively refined through validation-based loss minimization and stabilized by exponential moving averages.
Context-Aware Listwise Ranking with MHSA
In the ranking stage, MHSA is employed via an Induced Set Attention Block (ISAB) to model global dependency structures across the list. This allows the network to perform contrastive, context-rich reasoning, capturing subtle differentiations arising from residue- or epitope-level variations. The final ranking scores are produced by projecting MHSA outputs through a linear transformation, with training objectives grounded in ListMLE loss (Plackett-Luce likelihood).
Experimental Results
Baseline Comparison and Robustness
Comprehensive benchmarking across random and distributionally shifted splits (Ag-based, Ab-based) demonstrates significant performance gains for AbLWR over state-of-the-art methods, both deep learning and physics-based. On the standard random split, AbLWR attains:
- Full Rank Accuracy (FRA): 20.74% (vs. 8.98% for the next-best, GraphDTA).
- Kendall’s Tau (KT): 0.59 (vs. 0.46).
- Precision@1 (P@1): 57.26% (vs. 46.92%).
Under Ag- and Ab-based splits, designed to test OOD robustness, AbLWR consistently leads in all ranking metrics, substantiating its generalization capabilities.
Component Ablations
Ablation studies reveal that PU initialization and homologous sampling yield strong marginal benefits; absence of PU leads to a 15-point FRA drop. The largest performance degradation is observed when substituting the listwise ranking objective for standard MSE regression, highlighting the criticality of modeling relative order, especially for fine-grained affinity differentiation among homologous candidates.
Case Studies: Influenza and IL-33
On influenza datasets with subtle antigenic drift and a human IL-33 single-chain antibody library, AbLWR demonstrates high ranking fidelity (Kendall’s T > 0.4 in 43% of influenza lists, ~27% on IL-33), outperforming all baselines in mean and maximum ranking correlation. Notably, AbLWR distinguishes functional escape mutations (decrements in affinity due to viral evolution) and efficiently identifies top binders, substantially reducing rounds required for wet-lab validation.
Implications and Future Directions
AbLWR’s integration of PU learning with context-aware listwise modeling represents an advance in robust, scalable affinity prediction, especially under practical scenarios with scarce labels and antigenic diversity. The empirical superiority on clinical and viral case studies indicates strong translational potential for lead optimization and therapeutic screening.
The current framework employs a static list size and an offline sampling pipeline; immediate future work should address joint, end-to-end adaptive list construction and dynamic context set sizing, possibly leveraging reinforcement or curriculum learning paradigms for data selection. Additionally, extending AbLWR to leverage predicted complex (docked) states and integrating uncertainty quantification could further enhance utility in low-data, high-stakes discovery settings.
Conclusion
AbLWR establishes a listwise, context-aware, semi-supervised framework for Ab-Ag affinity ranking, outperforming both regression and ranking baselines in challenging scenarios. Its methodological innovations enable the exploitation of large unlabeled datasets while effectively modeling inter-candidate relational cues—a substantial progression toward accurate, scalable computational antibody discovery (2604.11272).