Disentangling multispecific antibody function with graph neural networks

We present a generative method for creating large-scale, realistic synthetic functional landscapes that capture nonlinear interactions where biological activity depends on domain connectivity. Second, we have several baselines including a graph neural network architecture and a sequence-only MLP using this data generation process.

Running Scaling Analysis

To reproduce the results on model performance scaling analysis (Figure 2.).

python run_sweep.py model

Running Transfer Learning Experiment

To run our trispecific example where we explore the effect of first pre-training on a node-level task (Figure 3.)

python run_sweep.py transfer

Note that our experiments in the paper use the full paired OAS to sample sequences, whereas here we have provided a subset of sequences in sequence_examples.csv.

Citation

If you find this work useful, please cite our ArXiv paper:

@article{southern2026synapse,
  title={Disentangling multispecific antibody function with graph neural networks},
  author={Southern, Joshua and Lu, Changpeng and Nerli, Santrupti and Stanton, Samuel D and Watkins, Andrew M and Seeger, Franziska and Dreyer, Fr{\'e}d{\'e}ric A},
  journal={arXiv preprint arXiv:2601.23212},
  year={2026}
}