Inference on Biological Dynamics with Memory

Biological systems are fundamentally dynamical. Cells grow and divide, respond to stress, and integrate information over time. Yet the mathematical tools we often use in dynamical systems theory come up short when applied to living systems, which are intrinsically stochastic, history-dependent (non-Markovian), and observed only through noisy, indirect measurements.To make sense of such systems, we often turn to Bayesian inference, which provides a principled framework for learning mechanisms from stochastic data. However, traditional Bayesian methods rely on analytical likelihoods, which are rarely available for realistic biological processes — especially when memory, such as cell-division history and molecular inheritance, plays a central role. In this talk, I show how neural networks can approximate such likelihoods directly from simulations, extending Bayesian inference to otherwise intractable models.
As a case study, we examine protein production across dividing cells, where standard inference procedures fail to account for the timing of divisions and inheritance of molecules.  By explicitly accounting for cell-division history, our approach reveals that gene activation events (such as glc3 in yeast expression observed in flow cytometry) are rarer and more transient than naïve analyses suggest. Later, we will discuss other problems  we tackle with neural network-assisted Bayesian inference."

Bio: https://pessoap.github.io/

Bio
https://pessoap.github.io/