our focus
The System Behind the Measurements
The gut microbiome operates through networks of competition, metabolic exchange and interaction with the host. A metabolite produced by one organism may feed another, alter its environment or influence a host response. These effects depend on the surrounding system and can change as that system changes.
Omics measurements capture different aspects of this biology, each with its own resolution and limitations. Understanding the processes behind them requires models that connect observations across biological scales while accounting for what remains unobserved.
our approach
From Patterns to Mechanisms
Learning from Data. Testing through Models.
We approach biological complexity from two complementary directions, connecting the patterns we observe with explanations we can investigate.
From Observations to System Structure
We develop methods to identify latent ecological organisation and statistical dependencies within biological data. These representations help examine how microbes fit into different biological settings, distinguish shared patterns from individual variation and formulate hypotheses about the relationships underlying observed outcomes.
From Biological Relationships to System Behaviour
We combine evidence with explicit assumptions about how biological components influence one another. Generative probabilistic models allow us to explore possible system states, trace direct and indirect effects, and simulate changes in silico. Their predictions create opportunities to test and refine the proposed explanations.
what we found
What it takes to model a biological system
Our research addresses the gap between the mathematical assumptions of a model, the biology it represents and the measurements available to support it.
Compositionality and sparse observations
Sequencing profiles usually report relative proportions. An apparent increase in one organism can result from a decline elsewhere, while a zero may reflect absence or limited detection. We account for these measurement constraints when constructing representations and estimating relationships, so technical properties of the data are not mistaken for biological interactions.
Integration across biological layers
Genes, transcripts, proteins and metabolites describe different biological quantities. They differ in scale, coverage, timing and measurement error. Our approach connects these layers through biologically informed relationships, with attention to batch effects and sample alignment.
Partial observability and uncertainty
No study measures every relevant variable in every participant. Missing observations, hidden influences and finite sample sizes limit what can be inferred. Probabilistic modelling provides a framework for reasoning with incomplete evidence and expressing uncertainty. We examine how assumptions about unobserved variables affect the conclusions a model can support.
Feedback, time and individual variation
Microbial and host processes influence one another through feedback, delayed effects and non-linear responses. A static association cannot represent those dynamics. We select model limits and temporal representations around the biological question, distinguishing general relationships from the individual conditions under which they operate and the sampling resolution needed to investigate them.
Causal identifiability
No study measures every relevant variable in every participant. Missing observations, hidden influences and finite sample sizes limit what can be inferred. Probabilistic modelling provides a framework for reasoning with incomplete evidence and expressing uncertainty. We examine how assumptions about unobserved variables affect the conclusions a model can support.
Computational tractability and reliable estimation
The number of possible networks grows rapidly with each additional variable. Greater connectivity and finer biological resolution also increase the parameters that must be estimated from limited samples. We develop algorithms and model selection methods that balance computational cost, statistical reliability and biological detail, using prior knowledge and carefully evaluated approximations to make these problems tractable.
validation
Models that can be questioned, tested and refined
A useful biological model must make its assumptions visible and its predictions testable. Our evaluation considers predictive performance, stability across repeated analyses and sensitivity to modelling choices. Independent data and experimental evidence, where available, provide further tests of whether an explanation holds beyond the observations used to generate it.
Agreement with data is only one part of that process. Mechanistic interpretation also requires checking whether the proposed relationships are biologically plausible and whether competing models could explain the same result. These checks guide an iterative model refinement.
