Graph Neural Networks Reveal Antibody Function in Complex Molecular Landscapes

Researchers are increasingly focused on multispecific antibodies as promising therapeutic agents, but predicting their behaviour remains a significant challenge. Joshua Southern, Changpeng Lu, and Santrupti Nerli, all from Prescient Design and Genentech, alongside Samuel D Stanton, Andrew M Watkins, and Franziska Seeger et al, have developed a novel computational framework to address this issue. Their work tackles the difficulty of predicting how antibody structure influences function, a problem hindered by limited experimental data. By combining generative modelling with graph neural networks, the team created a system capable of distinguishing between antibody formats and accurately predicting functional properties, even with limited biological datasets, ultimately accelerating the rational design of more effective antibody therapeutics.

The research team first engineered a generative method for creating large-scale synthetic functional landscapes, capturing non-linear interactions where biological activity depends on domain connectivity.

These landscapes facilitated the study of complex relationships between antibody structure and function, providing a virtual environment for experimentation. Subsequently, researchers proposed a novel graph neural network architecture that explicitly encodes topological constraints, differentiating between antibody formats that sequence-only models might consider identical.
This network was trained on the synthetic landscapes, enabling it to recapitulate complex functional properties and demonstrate potential for high predictive accuracy using limited biological datasets. The approach enables the model to learn the intricate interplay between antibody geometry and its therapeutic effect.

Experiments employed this trained model to optimise trade-offs between efficacy and toxicity in trispecific T-cell engagers, successfully retrieving optimal common light chains. The study pioneered the use of transfer learning to apply insights gained from synthetic data to real biological systems, accelerating the design process.

Specifically, the team demonstrated that altering the position of a high-affinity binding domain in a trispecific TCE could decouple anti-tumour efficacy from fatal cytokine release syndrome in in vivo models. Furthermore, the system delivers a robust benchmarking environment for disentangling the combinatorial complexity of multispecifics, allowing researchers to systematically investigate the impact of subtle structural changes.

Recent work highlighted that rigidifying the immunological synapse, even with constant affinity, can enhance potency, a finding supported by combining small-angle X-ray scattering with functional assays. This work provides a foundation for accelerating the design of next-generation therapeutics by addressing the scarcity of comprehensive experimental data.

Graph neural networks predict antibody function through topological encoding of domain connectivity and sequence information

Scientists have developed a computational framework to address the challenges in rationally designing multispecific antibodies. The research introduces a generative method for creating large-scale synthetic functional landscapes that accurately capture non-linear interactions governing biological activity.

Experiments revealed that domain connectivity is crucial, as biological activity depends on how domains are linked together. The team proposes a graph neural network that explicitly encodes topological constraints, differentiating between antibody formats that sequence-only models would consider identical.

This model, trained on synthetic landscapes, successfully recapitulates complex functional properties and demonstrates potential for high predictive accuracy using limited biological datasets via transfer learning. Results demonstrate the model’s utility in optimising trade-offs between efficacy and toxicity in trispecific T-cell engagers, and in retrieving optimal common light chains.

Measurements confirm that current computational approaches often struggle due to the scarcity of multi-format functional data needed to train robust predictive models. While physics-based methods exist, they are too computationally expensive for high-throughput screening of vast combinatorial design spaces.

Traditional sequence-based machine learning models fail to capture the complex 3D interactions defining multispecific function. The Synapse framework generates ground-truth functional data for arbitrary antibody formats, simulating non-linear interactions where global function emerges from graph structure.

Synapse utilises a novel graph-based extension of Ehrlich functions, assigning intrinsic fitness scores to each binding domain and enforcing biophysical plausibility through position-aware statistical modelling derived from the Observed Antibody Space. The global biological activity is defined by a connectivity-dependent readout function, modelling how a domain’s contribution is influenced by its neighbours.

Tests prove that this approach can simulate physical phenomena like avidity gating and steric shielding, which govern multispecific antibody function. The breakthrough delivers a robust benchmarking environment for disentangling combinatorial complexity, accelerating the design of next-generation therapeutics.

Predicting multispecific antibody function using topologically informed generative models is a challenging but promising approach

Scientists have developed a computational framework to accelerate the design of multispecific antibodies, which hold therapeutic promise by simultaneously targeting multiple biological pathways. This work addresses a key challenge in the field: predicting how subtle changes in antibody structure impact their function, a problem complicated by limited experimental data.

The researchers introduced a generative method for creating realistic synthetic landscapes that model the complex, non-linear interactions governing antibody activity based on domain connectivity. Furthermore, they propose a graph neural network that explicitly incorporates topological constraints, allowing it to differentiate between antibody formats that might appear identical to models focused solely on sequence information.

Training this model on synthetic data demonstrated its ability to accurately replicate complex functional properties and, through transfer learning, achieve high predictive accuracy even with limited biological datasets. The model’s utility was showcased through the optimisation of trade-offs between efficacy and toxicity in trispecific T-cell engagers and the identification of optimal common light chains.

The authors acknowledge a limitation in the current scope of their work, focusing on a specific aspect of the combinatorial complexity of multispecific antibodies. Future research could expand the framework to incorporate additional factors influencing antibody function, such as post-translational modifications or the impact of the antibody’s constant region. This research establishes a robust benchmarking environment for disentangling the complexities of multispecific antibody design, potentially accelerating the development of next-generation therapeutic proteins.