Simulation-conditioned generative modeling for biologically realistic pattern prediction

Pattern formation underlies biological organization across scales, but predicting experimentally observed patterns remains difficult because mechanistic models and data-driven generative models fail in complementary ways. Coarse-grained mechanistic models can encode causal constraints and global morphology, yet they omit fine-scale features such as texture, color gradients, and stochastic replicate-to-replicate variation. In contrast, generative image models can produce realistic images but are not inherently grounded in the biophysical rules that shape real patterns. Here, we introduce a simulation-conditioned generative framework that uses mechanistic simulations as spatial priors for generating biologically realistic pattern data. As a concrete test, we use a synthetic-to-real inverse task to show that these generated patterns preserve information needed for inference on real experimental images, beyond merely reproducing plausible visual appearance. Using branching colony expansion of Pseudomonas aeruginosa as a model system, we combine a coarse-grained PDE model with latent representations from a foundation image model and a conditional diffusion model. The resulting framework preserves the global structures imposed by simulation while restoring experimentally observed fine-scale morphology and stochastic variability. A model trained exclusively on simulation-conditioned synthetic patterns transfers without fine-tuning to real experimental patterns, enabling inference of initial seeding configurations from experimental colony morphology. Together, these results establish simulation-conditioned generative modeling as a strategy for converting coarse mechanistic models into scientifically structured synthetic data, enabling inference tasks on real biological patterns where experimental data are scarce.