Diffusion transformers as foundation models in systems biology

We introduce Systems Biology Simformers, a diffusion transformer architecture that leverages the interpretability of systems biology models with the flexibility of machine learning models to accurately predict experimental observations. This enables robust inference of physically relevant biological parameters across diverse experimental conditions, and novel simformer diffusion sampling that can better recapitulate observed data than expert simulation models. We demonstrate this architecture on the Bone Morphogenetic Protein signaling pathway, improving the process of parameter inference while identifying systematic biases in expert models that reveal conflicts in biophysical parameters that can help guide future wet-lab experiments. By combining expert knowledge with data-driven learning, our approach creates a foundation for interpretable and generalizable systems biology models that maintain mechanistic understanding while enhancing predictive accuracy across experimental conditions.