Learning the interaction rules of antibody-antigen binding

Project Information

AB-AG-INTERACT

Grant agreement ID: 101125630

DOI

10.3030/101125630

EC signature date 4 December 2023

Start date 1 January 2024

End date 31 December 2028

Funded under

• European Research Council (ERC)

Total cost

€ 2 000 000,00

EU contribution

€ 2 000 000,00

Coordinated by

UNIVERSITETET I OSLO
Norway

Periodic Reporting for period 1 - AB-AG-INTERACT (Learning the interaction rules of antibody-antigen binding)

Reporting period: 2024-01-01 to 2026-06-30

Summary of the context and overall objectives of the project

Antibody–antigen (Ab–Ag) recognition underpins vaccines and therapeutic antibodies, yet our ability to predict and design these interactions remains limited, constraining the shift from empirical screening to rational, AI-driven design. This limitation arises because Ab–Ag binding is governed by non-linear, many-to-many interactions between paratopes and epitopes, which are poorly captured by existing small and fragmented datasets. As a result, current computational models lack predictive power, and the underlying rules of immune recognition remain largely unknown. AB-AG-INTERACT addresses this challenge by testing the hypothesis that Ab–Ag binding is learnable and predictable given sufficiently large, integrated sequence–structure datasets and appropriate machine learning frameworks. The project targets two key bottlenecks: the lack of large-scale joint binding data and the absence of interpretable models for rule discovery. To overcome these, the project establishes: (i) a platform for generating large-scale Ab–Ag binding data, (ii) computational pipelines to derive structural interaction information, (iii) machine learning models for end-to-end binding prediction and rule extraction, and (iv) experimental validation of rule-driven antibody and antigen design. The expected impact is transformative. Scientifically, the project advances a shift toward predictive, mechanistic immunology, enabling the discovery of generalizable rules of immune recognition. Technologically, it provides a scalable framework for in silico design of antibodies and antigens, expanding accessible sequence space beyond experimental screening. In the longer term, these advances are expected to accelerate and improve therapeutic antibody and vaccine development, contributing to more efficient responses to infectious diseases, cancer, and immune-mediated conditions, and supporting broader goals in global health and biomedical innovation.

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

Activities performed and main achievements

During the reporting period, we implemented and advanced the core technological and analytical components of AB-AG-INTERACT across WP1-2, establishing an integrated experimental–computational framework for antibody–antigen (Ab–Ag) interaction mapping, prediction, and rule discovery.

WP1 – Generation of large-scale antibody–antigen binding data

We established a library-on-library screening platform enabling systematic interrogation of Ab–Ag interactions through simultaneous variation of paratope and epitope residues. This included:

- Construction and screening of combinatorial antibody and antigen variant libraries using yeast display and bead-based assays
- Generation of high-dimensional binding matrices capturing interaction strengths across thousands of Ab–Ag variant pairs
- Initial expansion toward larger-scale datasets, significantly exceeding conventional one-to-many screening approaches
- Publication to be expected in 2026 or 2027

In parallel, we implemented a computational structural biology pipeline to translate sequence-based binding data into structural representations:

- Structure prediction of antibody and antigen variants ()
- Information-driven docking and molecular dynamics simulations to resolve paratope–epitope interactions
- Generation of large-scale datasets linking sequence, structure, and binding outcomes
- First publication: https://iclr.cc/virtual/2025/34650

Main achievement: Establishment of the first scalable experimental–computational pipeline for many-to-many Ab–Ag interaction mapping, producing datasets of unprecedented size and resolution.

WP2 – Machine learning for prediction and rule discovery

We developed and benchmarked machine learning approaches for Ab–Ag binding prediction and interpretability, including:

- Models for predicting binding across multiple antibodies and antigens, including generalization to unseen sequences (out-of-distribution settings)
- Integration of sequence and structural representations using deep learning architectures (e.g. graph-based and sequence-based models)

In addition, we established a benchmarking framework using both experimental and simulated datasets to systematically evaluate:

- The impact of dataset size, encoding strategies, and negative samplingT
- The relationship between prediction accuracy and recoverability of interaction rules
- First publication: https://www.nature.com/articles/s42256-025-01089-5

Main achievement: Demonstration that Ab–Ag binding can be predicted with increasing accuracy when combining large-scale data with structure-aware ML, and that interpretable patterns (candidate interaction rules) can be extracted from trained models.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

Results’ potential impact and pathways to uptake:

The results generated in AB-AG-INTERACT establish a foundational framework for predictive and rule-based modeling of antibody–antigen (Ab–Ag) interactions, with implications across immunology, biotechnology, and therapeutic development.

Scientific and technological impact.
By integrating large-scale experimental Ab–Ag interaction data with structure-aware machine learning, the project advances the field from descriptive analyses toward predictive and mechanistic understanding of immune recognition. The ability to jointly model sequence, structure, and binding enables identification of generalizable interaction patterns, providing a basis for rational antibody and antigen design. This represents a step-change compared to current approaches that rely on sparse data and empirical screening.

Translational and innovation potential.
The developed platform has clear potential to transform therapeutic antibody and vaccine discovery pipelines. In particular, it enables:

- Exploration of vast combinatorial sequence spaces beyond experimental screening limits
- More efficient identification of high-affinity and specific binders
- Early-stage assessment of binding specificity, including non-binding and off-target interactions

These capabilities can reduce development time and cost while increasing success rates, which is highly relevant for pharmaceutical and biotechnology applications.

Broader societal relevance.
In the longer term, the ability to predict and design immune recognition has implications for infectious disease preparedness, cancer immunotherapy, and autoimmune disease research, contributing to improved healthcare outcomes and supporting global health priorities.