At this colloquium, we are happy to welcome:

Thomas Weikl (Max-Planck-Institut)

Membrane-mediated cooperativity of proteins and particles

Besides direct protein-protein interactions, indirect interactions mediated by membranes can play an important role for the assembly and cooperative function of proteins in membrane shaping and adhesion. Also particles that adhere to membranes or are trapped in membrane contact zones experience such membrane-mediated interactions. In my talk, I will  address the origin and relevance of these indirect interactions, with a focus on four cases: 

(1) The intricate shapes of biological membranes are generated by proteins that locally induce membrane curvature. Indirect curvature-mediated interactions between these proteins arise because the proteins jointly affect the bending energy of the membranes.  The curvature-mediated interactions are attractive for arc-shaped proteins and drive the assembly of the proteins during membrane tubulation. 

(2) Membrane adhesion results from the binding of receptor and ligand proteins that are anchored in the apposing membranes. The binding of these proteins depends on the shape and shape fluctuations of the membranes on nanoscales, which leads to binding cooperativity and to the segregation of long and short receptor-ligand complexes in the contact zones of immune cells.

(3) Particles are wrapped spontaneously by membranes if the adhesive interactions between the particles and membranes compensate for the cost of membrane bending. The interplay of adhesion and bending energies during wrapping can lead to attractive curvature-mediated interactions and to the cooperative wrapping of nanoparticles in membrane tubules.

(4) The trapping of particles in membrane adhesion zones leads to membrane deformations with bending energies that are a strong driving force for particle assembly.

Artemy Kolchinsky (Universitat Pompeu Fabra)

Generalized free energy and excess entropy production for nonequilibrium systems

In thermodynamics, the free energy potential is typically defined as the relative entropy between the system’s actual state and the equilibrium state. Recently, however, research in nonequilibrium thermodynamics has focused on active systems that undergo continuous driving and relax toward nonequilibrium steady states, oscillations and/or chaos. Because such systems lack equilibrium states, they do not possess a free energy potential as usually defined.  In our recent work (arXiv:2412.08432), we propose a generalized free energy potential for genuine nonequilibrium systems, including both stochastic master equations and deterministic nonlinear chemical reaction networks (possibly without steady states). Our potential is defined using a variational principle that is motivated from several perspectives, including large deviations theory, thermodynamic uncertainty relations, and optimal transport. We derive a universal definition of the excess entropy production, the nonstationary contribution to dissipation, as well as far-from-equilibrium thermodynamic speed limits for both linear and nonlinear systems. 

Thomas Gaskin (University of Cambridge)

Deep learning-based parameter inference for large-scale multi-agent models

In this talk, I will present a neural network-based computational framework to learn system models from data. Instead of discarding a mechanistic model and resorting to pure pattern matching, unknown components of a mathematical model are replaced in a targeted fashion with a neural network, which is trained by requiring its output to reproduce the observation data via the model. The framework supports both parametric and non-parametric estimation, and has been shown to outperform classical estimation techniques such as regression or Monte-Carlo sampling, while naturally allowing for uncertainty quantification. I will illustrate the framework by presenting my recent work estimating international migration flows between all countries from the past 40 years: here, an unknown flow matrix is parametrised as a recurrent neural network and trained to map input covariates to flows. The network is trained by differentiating through the system equations, and produces results that are more accurate and at a significantly higher temporal resolution than current estimates. This case study illustrates the difficulties in learning system models in the social and economic sciences, where data is sparse, often of low quality, and the inference problems severely ill-posed. I will discuss uncertainty quantification within this framework, and further illustrate the method using examples from computational epidemiology and biology.