Seminars

Deep learning algorithms provide unprecedented opportunities to characterise complex structure in large data, but typically in a manner that cannot easily be interpreted beyond the 'black box'. We are developing methods to leverage the benefits of deep generative learning and computational modelling (e.g. fluid dynamics, solid mechanics, biochemistry), particularly in conjunction with biomedical imaging, to enable new insights into disease to be made. In this talk, I will describe our applications in several areas, including modelling drug delivery in cancer and retinal blood vessel loss in diabetes, and how this is leading us into the development of personalised digital twins.

In this talk I will give a number of short vignettes of work that has been undertaken in my group over the last 15 years. Mathematically, the theme that underlies our work is the importance of randomness to biological systems. I will explore a number of systems for which randomness plays a critical role. Models of these systems which ignore this important feature do a poor job of replicating the known biology, which in turn limits their predictive power. The underlying biological theme of the majority our work is development, but the tools and techniques we have built can be applied to multiple biological systems and indeed further afield. Topics will be drawn from, locust migration, zebrafish pigment pattern formation, mammalian cell migratory defects, appropriate cell cycle modelling and more. I won't delve to deeply into anyone area, but am happy to take question or to expand upon of the areas I touch on.

Understanding how living systems dynamically self-organise across spatial and temporal scales is a fundamental problem in biology; from the study of embryo development to regulation of cellular physiology. In this talk, I will discuss how we can use mathematical modelling to uncover the role of microscale physical interactions in cellular self-organisation. I will illustrate this by presenting two seemingly unrelated problems: environmental-driven compartmentalisation of the intracellular space; and self-organisation during collective migration of multicellular communities. Our results reveal hidden connections between these two processes hinting at the general role that chemical regulation of physical interactions plays in controlling self-organisation across scales in living matter

Tumour microenvironment is characterised by heterogeneity at various scales: from various cell populations (immune cells, cancerous cells, ...) and various molecules that populate the microenvironment (cytokines, chemokines, extracellular vesicles, …); to phenotype heterogeneity inside the same cell population (e.g., immune cells with different phenotypes and different functions); as well as temporal heterogeneity in cells’ phenotypes (as cancer evolves through time) and spatial heterogeneity.
In this talk we overview some mathematical models and computational approaches developed to investigate different single-scale and multi-scale aspects related to heterogeneous immune responses during cancer evolution. Throughout the talk we emphasise the qualitative vs. quantitative results, and data availability across different scales

The epithelial to mesenchymal transition (EMT) is a key cellular process underlying cancer progression, with multiple intermediate states whose molecular hallmarks remain poorly characterized. In this talk, I will describe AI-powered and ecology-inspired methods recently developed by us to provide a multi-scale view of the epithelial-mesenchymal plasticity in cancer from single cell and spatial transcriptomics data. First, we employed a large language model similar to the one underlying chatGPT but tailored for biological data (inspired by scBERT methodology), to predict individual stable states within the EMT continuum in single cell data and dissect the regulatory processes governing these states. Secondly, we leveraged spatial transcriptomics of breast cancer tissue to delineate the spatial relationships between cancer cells occupying distinct states within the EMT continuum and various hallmarks of the tumour microenvironment. We introduce a new tool, SpottedPy, that identifies tumour hotspots within spatial transcriptomics slides displaying enrichment in processes of interest, including EMT, and explores the distance between these hotspots and immune/stromal-rich regions within the broader environment at flexible scales. We use this method to delineate an immune evasive quasi-mesenchymal niche that could be targeted for therapeutic benefit. Our insights may inform strategies to counter immune evasion enabled by EMT and offer an expanded view of the coupling between EMT and microenvironmental plasticity in breast cancer.

Recent progress in the development of equivariant neural network architectures predominantly used for machine learning interatomic potentials (MLIPs) has opened new possibilities in the development of data-driven coarse-graining strategies. In this talk, I will first present our work on the development of learning potential energy surfaces and other physical quantities, namely the Hyperactive Learning framework[1], a Bayesian active learning strategy for automatic efficient assembly of training data in MLIP and ACEfriction [2], a framework for equivariant model construction based on the Atomic Cluster Expansion (ACE) for learning of configuration-dependent friction tensors in the dynamic equations of molecule surface interactions and Dissipative Particle Dynamics (DPD). In the second part of my talk, I will provide an overview of our work on the simulation and analysis of Generalized Langevin Equations [3,4] as obtained from systematic coarse-graining of Hamiltonian Systems via a Mori-Zwanzig projection and present an outlook on our ongoing work on developing data-driven approaches for the construction of dynamics-preserving coarse-grained representations.

References:

[1] van der Oord, C., Sachs, M., Kovács, D.P., Ortner, C. and Csányi, G., 2023. Hyperactive learning for data-driven interatomic potentials. npj Computational Materials

[2] Sachs, M., Stark, W.G., Maurer, R.J. and Ortner, C., 2024. Equivariant Representation of Configuration-Dependent Friction Tensors in Langevin Heatbaths. to appear in Machine Learning: Science & Technology

[3] Leimkuhler, B. and Sachs, M., 2022. Efficient numerical algorithms for the generalized Langevin equation. SIAM Journal on Scientific Computing

[4] Leimkuhler, B. and Sachs, M., 2019. Ergodic properties of quasi-Markovian generalized Langevin equations with configuration-dependent noise and non-conservative force. In Stochastic Dynamics Out of Equilibrium: Institut Henri Poincaré, 2017 

How do mobile organisms situate themselves in space?  This is a fundamental question in both ecology and cell biology but, since space use is an emergent feature of movement processes operating on small spatio-temporal scales, it requires a mathematical approach to answer.  In recent years, increasing empirical research has shown that non-locality is a key aspect of movement processes, whilst mathematical models have demonstrated its importance for understanding emergent space use patterns.  In this talk, I will describe a broad class of models for modelling the space use of interacting populations, whereby directed movement is in the form of non-local advection.  I will detail various methods for ascertaining pattern formation properties of these models, fundamental for answering the question of how organisms situate themselves in space, and describe some of the rich variety of patterns that emerge. I will also explain how to connect these models to data on animal and cellular movement.

Consumer-resource interactions are central to ecology, as all organisms rely on consuming resources to survive. Functional responses describe how a consumer's feeding rate changes with resource availability, influenced by processes like searching for, capturing, and handling resources. To study functional responses, experiments typically measure the amount of food consumed—often in discrete units like prey—over a set time. These experiments systematically vary prey availability to observe how it affects the consumer's feeding behaviour. The data generated by such experiments are often analysed using differential equation-based models. Here, we argue that such models do not represent a realistic data-generating process for many such experiments and propose an alternative stochastic individual-based model. This class of models, however, is expensive for inference, and we use machine learning methods to expedite fitting these models to data. We then use our method to do generalised linear model-based inference for a series of experiments conducted on a stickleback fish. Our methodology is made available to others in a Python package for Bayesian hierarchical inference for stochastic, individual-based models of functional responses.