Disentangling multispecific antibody function with graph neural networks

TL;DR

This work addresses the challenge of topology-dependent function in multispecific antibodies under data scarcity. It introduces Synapse for generating graph-based synthetic landscapes and a topology-aware Graph Isomorphism Network that maps antibody graphs $\mathcal{G}$ to a scalar $y$ via a global readout $\Phi$. The results show the GIN captures topology-dependent effects beyond sequence-only models, and transfer learning from monospecific to multispecific formats improves predictive data efficiency. It demonstrates topology-driven optimization of trispecific constructs and effective common light chain retrieval, establishing a topology-aware benchmarking framework to accelerate next-generation multispecific therapeutics.

Abstract

Multispecific antibodies offer transformative therapeutic potential by engaging multiple epitopes simultaneously, yet their efficacy is an emergent property governed by complex molecular architectures. Rational design is often bottlenecked by the inability to predict how subtle changes in domain topology influence functional outcomes, a challenge exacerbated by the scarcity of comprehensive experimental data. Here, we introduce a computational framework to address part of this gap. First, we present a generative method for creating large-scale, realistic synthetic functional landscapes that capture non-linear interactions where biological activity depends on domain connectivity. Second, we propose a graph neural network architecture that explicitly encodes these topological constraints, distinguishing between format configurations that appear identical to sequence-only models. We demonstrate that this model, trained on synthetic landscapes, recapitulates complex functional properties and, via transfer learning, has the potential to achieve high predictive accuracy on limited biological datasets. We showcase the model's utility by optimizing trade-offs between efficacy and toxicity in trispecific T-cell engagers and retrieving optimal common light chains. This work provides a robust benchmarking environment for disentangling the combinatorial complexity of multispecifics, accelerating the design of next-generation therapeutics.

Figures (5)

• Figure 1: Overview of the Synapse data generation framework. (A) Realistic domain sequences are sampled from an OAS-derived PSSM. Intrinsic fitness scores are assigned via Ehrlich functions to mimic non-linear binding landscapes. (B) A multispecific antibody is represented as a graph. The final regression objective is derived from a global readout function that integrates intrinsic domain values with their specific topological connectivity, creating a ground-truth value where structure dictates function.
• Figure 2: Scaling of model performance as a function of dataset size, for increasing format complexity, comparing a GIN to a MLP. Selected examples of possible domain connectivities are shown for each complex format.
• Figure 3: Trispecific antibody optimization and transfer learning efficiency. The left side illustrates the two topological variants evaluated, where the top configuration (format A) represents the safe format and the bottom one (format B) represents the toxic phenotype driven by distal domain placement. The center plot shows the distribution of the training data activity values for both formats. The right panel plots the model's test MSE against the number of fine-tuning samples, demonstrating that pretraining on larger datasets significantly improves predictive accuracy in data-sparse regimes.
• Figure 4: Left: Identification of best common light chain from sampling combinations of domains. Right: Mean global function value of the top 5 trispecific constructs as a function of the number of light chain candidates considered (sampled from the combined 75k pretraining pool for each format). The ground truth curve represents the theoretical optimum found selecting according to actual Ehrlich function values. The model-selected curve shows the average true value of the 5 candidates ranked highest by the fine-tuned graph neural network. The close alignment indicates the model effectively ranks CLCs that maximize topology-dependent efficacy.
• Figure 5: Characterization of trispecific antibody biological activity profiles. The univariate kernel density estimate of the global activity metric ($y$) reveals a multimodal distribution, suggesting underlying structural heterogeneity across the generated population. The left panel resolves this complexity through pairwise domain correlations ($f1$-$f3$) and the right panel through a format-stratified violin plot. This stratification demonstrates that distinct connectivity types (9 formats in total) occupy unique activity regimes, confirming that the scoring function successfully captures and distinguishes format-dependent distributional shifts.