UC Irvine

Systems biology models come in a variety of forms, and many of them can be simulated in ways that roughly recapitulate the underlying biology. These simulators let us generate synthetic data that can match what we observe in experiments. However, they can be expensive to run and designing new experiments is also expensive and time-consuming. In addition, even when a simulator can match observed data, the underlying parameters may not be identifiable, meaning multiple very different parameter settings can explain the same observations, and in some cases the simulator itself does not reflect the true underlying biology. Together with the fact that many simulators do not come with a tractable likelihood, these issues make standard Bayesian inference and principled experimental design hard to apply. First, we develop simulation-based inference frameworks that learn surrogates for intractable likelihoods and posteriors using density estimators and diffusion transformers. This enables Bayesian parameter inference, uncertainty quantification, and model comparison for high-dimensional systems biology data. Second, we treat mutual information as a unifying objective that connects posterior quality and experimental utility. We derive and evaluate mutual information lower bounds that allow us to jointly optimize amortized inference objects and experimental designs, even when simulators are non-differentiable. This yields practical SBI and BOED procedures that improve calibration and predictive accuracy. Third, we show how structural biology signals, such as predicted complexes and binding affinities, can be integrated with systems models to reduce posterior entropy and refine biological hypotheses, providing a pathway from mechanistic inference to actionable intervention. We culminate with a systems-aware, structure-guided drug discovery pipeline that uses non-differentiable diffusion guidance to generate small molecules optimized for pathway-level specificity, showing how integrated inference and information optimization can potentially accelerate therapeutic design.