PhD studies
I’m always happy to receive expressions of interest by email from prospective PhD students.
I currently co-supervise
-
Chao Gao (with Michael Papathomas) on Using supervised learning methods to measure information transfer
Recent PhD graduates
-
Naici Guo (co-supervised with Andy Lynch) worked on Proteomic Studies in Cancer: Statistical Methods, Experimental Insights, and the Challenges of Dual Proteomics Profiles
-
Xiaoyue Yang (co-supervised with Andy Lynch) worked on Analysing prostate cancer gene expression data
Current PhD projects
The following projects are offered for entry in the 2027–28 academic year. All are competitively funded and open to both UK and international students. For details about studying at St Andrews and how to apply see the document in this page.
Randomness and dynamics in epidemics: learning hidden transmission from data
Supervisor: Giorgos Minas
In collaboration with
Daniela De Angelis (MRC-Biostatistics Unit, University of Cambridge), Paul Birrell (UKHSA),
Angelos Alexopoulos (Athens UEB)
Description: Why do epidemics rise and fall the way they do? Every wave of COVID-19, influenza or measles is shaped by something we can never observe directly: how easily an infection passes from person to person. This transmission rate changes over time with the seasons, school terms, government policies and people’s behaviour. Epidemics are also inherently stochastic: who meets whom, and whether a contact leads to infection, involves an element of chance. That randomness matters most when case numbers are small, at the start and end of a wave. Deterministic models ignore it and can give overconfident conclusions, while fully stochastic models are often too slow to fit to real data. Reconstructing both the dynamics and the randomness of transmission from noisy surveillance data, such as case counts, hospital admissions and deaths, is a central challenge in modern epidemic modelling, and it matters for forecasting and for deciding how to respond. This project develops fast, principled statistical methods to meet that challenge. It builds on the Linear Noise Approximation (LNA), which captures the randomness of epidemic spread while keeping the likelihood of the data tractable. A promising direction is to couple the LNA with hidden stochastic models for the transmission rate, within a Bayesian inference framework. The project will pursue this direction through four objectives. Objectives: - Realistic data. Extend the framework to the data epidemiologists actually use: new cases rather than current infections, reporting delays, under-reporting, and several data streams combined. Where exact, tractable likelihood computation is no longer possible, accurate and efficient approximations will be developed and compared. - What drives transmission? Compare models in which the transmission rate fluctuates randomly with models in which part of its variation is explained by observed covariates, such as mobility, temperature or contact surveys. A key question is how much of the variation in transmission is predictable from these drivers and how much reflects unexplained random fluctuations. Answering it means disentangling the randomness of individual infections from the wider fluctuations in transmission. - What can the data tell us? Use Fisher information to find which parameters can be identified, how much each source of randomness can be learned from data, and which data streams or sampling frequencies would be most informative. These questions bear directly on how surveillance should be designed. - Real epidemics. Apply the methods to openly available data, such as seasonal influenza, historical measles and COVID-19. Beyond these core objectives, the project offers scope for more ambitious extensions. One is behavioural feedback, where awareness of the epidemic reduces contacts; this can produce richer dynamics, including self-sustained waves, which can be studied with the phase-corrected LNA. Others include tipping points in coupled disease–behaviour systems, and new corrections that improve the LNA’s accuracy for the timing of outbreaks. Skills and background: This project suits a student with a strong background in statistics, mathematics or another quantitative subject who wants to apply mathematics to real problems in public health. Familiarity with probability, differential equations and Bayesian inference is valuable, as is programming experience (for example MATLAB, Python, R or Julia). No prior knowledge of epidemiology is needed. You will gain expertise in stochastic processes, dynamical systems, Bayesian computation and epidemic modelling. That combination is in high demand in academia, public health agencies and industry. Key references - Minas & Rand (2017), PLoS Computational Biology 13(7): Long-time analytic approximation of large stochastic oscillators. Swallow, Rand & Minas (2026), Bayesian Analysis 21(1): Bayesian inference with the pcLNA. - Truman-Williams & Minas (2026), SIAM Journal on Applied Dynamical Systems 25(3): The LNA can capture nonlinear phenomena. - Ghosh, Birrell & De Angelis (2023), PLoS Computational Biology: An approximate diffusion process for environmental stochasticity in transmission modelling.
Navigating complex posteriors: efficient Bayesian inference for noisy living systems
Supervisors: Giorgos Minas, Ben Swallow.
In collaboration with
Pantelis Samartsidis (MRC-Biostatistics Unit, University of Cambridge)
Description: From the molecular networks that let a cell keep time or decide its fate to the spread of an infection through a population, living systems are both noisy and non-linear. Interactions between individual molecules, cells or people are stochastic, occurring at rates that often depend non-linearly on the current state of the system. This combination of stochasticity and non-linearity underlies behaviours such as oscillations, switches and outbreaks. Models that capture it involve many interacting variables and tens of unknown parameters. Fitting them to data is a demanding statistical problem: likelihoods are expensive to compute, and posteriors are high-dimensional with complex geometry that standard samplers struggle to explore. This project builds on methods we have recently developed: an approximation that makes the likelihood of these models faster to compute, and a Bayesian inference framework built on it. You will start by implementing these methods in an R package, with the computationally intensive parts written in C++ through Rcpp. This will make the methods considerably faster, more widely accessible, and will give you a thorough working knowledge of them. With it, you will explore what large models, such as those of inflammatory signalling, can learn from data. The main new methodology comes next: using the geometry of the likelihood to design samplers that move efficiently through peaks, funnels, and valleys in the parameter space, and comparing them with other modern methods such as amortised inference. Once this framework is in place, we will extend it in one or more planned directions: systems driven by changing environments such as light or seasonality, the design of more informative experiments, coupled oscillators in tissues, and stochastic epidemic models. Background and training We are looking for a student with a background in statistics or a related subject, ideally with some exposure to Bayesian methods and computational statistics, who enjoys programming. No prior knowledge of biology or C++ is needed, but strong interest is essential.
Learning to measure information: decoding, compression and experimental design for single-cell data
Supervisors: Giorgos Minas, Michail Papathomas
Description: Cells constantly sense their environment and respond by adjusting gene expression, making decisions such as whether to divide, die or differentiate. We routinely say that signals carry information and that cells read it. But information, unlike concentration or abundance, is not a physical quantity that can be measured directly. It lies not in any single response but in how responses differ from one signal to another. Information theory, originally developed for communication systems, makes this notion precise and measurable. The key quantity is the mutual information, I, between a signal and the cell’s response. Measured in bits, it quantifies how much observing a response reduces our uncertainty about which signal the cell received. If different signals lead to clearly distinct responses, I, is large; if the responses overlap completely, I is zero. The quantity 2^I can be read as the effective number of signals a cell can reliably distinguish. In practice, I can be estimated using supervised machine learning: a classifier is trained to recognise which signal produced a given response, and its pattern of errors yields a lower bound on I. This project starts from the observation that a classifier built to maximise accuracy is not necessarily the one that best preserves information. A response equally consistent with two signals may be more informative when labelled as ambiguous than when forced into one of them. Allowing a classifier to abstain adds one such label. More generally, responses can be grouped into a small number of classes chosen to retain as much information as possible. This raises open questions. Which training objectives make a classifier a good estimator of information? How many classes are needed, and how does this depend on the structure of the responses, which in single-cell data are typically dynamic and high-dimensional? How should experiments be designed, choosing stimuli, measurement times and readouts, so that they reveal as much information as possible? Methods will be developed and assessed on simulated data from stochastic models of signalling and gene regulation, where the true information is known, and then applied to publicly available single-cell gene expression data. Objectives. (1) Study how the choice of loss function and decision rule affects the accuracy, robustness and other statistical properties of classifier-based estimators, and characterise the classifiers that retain the most information. (2) Develop methods that summarise dynamic, high-dimensional responses by a small number of classes, generalising abstention, and determine how many classes are needed to retain most of the information for a given amount of data. (3) Develop methods for designing experiments that maximise information, using fast simulation of stochastic models. Further directions, depending on the student’s interests, include separating the information carried by each factor in experiments that combine several factors, and identifying the genes that carry the information. Skills and background. A degree in mathematics, statistics or a related field, with some familiarity with probability and statistical inference, and an interest in problems motivated by biology. Some programming or machine learning is helpful but not required. Key references. Sáez, Minas, Camacho-Aguilar & Rand (2026), bioRxiv (accepted by PNAS). Cepeda-Humerez, Ruess & Tkačik (2019), PLoS Computational Biology. Rainforth, Foster, Ivanova & Bickford Smith (2024), Statistical Science.
Teaching AI to discover the hidden rules of noisy living systems
Supervisors: Giorgos Minas, Jochen Kursawe, Lei Fang.
Description: Mathematicians describe such systems with stochastic differential equations: equations that combine a predictable underlying "rhythm'' with random fluctuations, and often built-in time delays, such as the time a gene takes to make a protein that switches the gene off again. Traditionally, scientists write these equations by hand. Deep learning now offers a way to learn them directly from data, but current methods are slow, unstable, need very large datasets, and produce "black boxes'' that can behave implausibly beyond the data they were trained on. This project will develop a new generation of AI methods that learn such equations reliably, efficiently and interpretably. It has three distinctive features. First, AI with a mathematical rulebook. Used naively, deep learning is data-hungry and unpredictable. In the spirit of physics-informed AI, our neural network instead operates within a mathematical framework we impose. This builds in what we already know, such as time delays in gene feedback or processes that repeat in cycles, and describes noisy dynamics as a predictable path with small random wobbles around it, using the supervisory team's phase corrected Linear Noise Approximation (pcLNA). The AI will only have to learn what is genuinely unknown, and we therefore expect it to need far less data, learn faster and produce models with clear biological meaning. Second, built-in "clock correction''. No two cells run at exactly the same pace. The method will realign each cell's internal timing as it goes, so that a cell that is merely running late is not mistaken for one following different rules. Third, discovering cell types from their dynamics. Existing methods group time-series by how they look. Ours will group cells by the rules that drive them, letting the data decide how many groups there are. It can then answer a question biologists frequently ask: are we seeing two different kinds of cells, or one kind of cell that can make a choice? Together, these turn deep learning from a curve-fitting tool into a discovery tool. The outputs are meaningful quantities that scientists can test experimentally: rhythms, delays, tipping points, rates of switching between states, and how much of the variability is noise. Feasibility: Each phase produces a publishable result, and later phases build on, rather than depend on, earlier ones. - Year 1: training in machine learning and stochastic modelling; method development for rhythmic systems, such as gene expression oscillations, tested on simulated data where the true answer is known. - Year 2: extension to multiple cell groups and to systems that switch between states, and benchmarking against existing AI approaches, including AI methods that, once trained, estimate unknown quantities almost instantly. -Year 3: application to real single-cell data on gene expression oscillations in developing nerve cells and on inflammatory signalling under different temperatures and stimuli, then to public repositories such as RIKEN's SSBD database; release of user-friendly open-source software. - Final six months: thesis writing. Until now, the pcLNA has been used only with hand-written equations, while standard AI models of noisy dynamics (neural stochastic differential equations) lack its structure. Bridging the two is, to our knowledge, entirely new. Applied to collaborators' single-cell data, the method will not just recover what current models describe, but reveal what they are missing.
EASTBIO: Unravelling the Regulation of Small RNAs in the Prostate
Supervisors: Andy Lynch, Giorgos Minas
This project is part of the EASTBIO project (see Description for more details)
Deadline: November 16th
Description: Aside from messenger RNA (mRNA), a whole raft of small non-coding RNAs are at work in the cell, but these have been relatively neglected in studies of genetic and epigenetic regulation. The unique metabolism of the prostate makes it difficult to infer regulatory processes from other tissues, and the small number of studies into the regulation of small RNA expression in the prostate have tended to focus on cancer tissue, which has its own unique biology. Uniquely, we have assembled a data set of small-RNA profiles, methylation profiles and germline variants for non-malignant prostate samples from approximately 200 men. The samples represent a variety of ages and include multiple samples from some men, allowing us to study heterogeneity between and within individuals. The focus is likely to be on microRNAs, with other small RNAs, such as snoRNAs and tRNAs, also considered where appropriate. While small RNA can regulate mRNA expression, any of germline variation, DNA methylation, or mRNA expression could affect small RNA expression. Additionally, germline variation and methylation can also affect mRNA expression, while germline variation and mRNA expression can influence DNA methylation. The project will model the causal network of events resulting in the regulation of small RNA expression in the prostate, to determine whether the detected associations represent direct causality or are mediated by the other factors being examined. Consideration of how methylation patterns evolve around small RNA loci can also feature. This project will look to establish models for the regulation of small RNAs taking into account methylation patterns, germline variation and mRNA expression, and will generate a catalogue of regulatory features. These will be validated in other, published, data sets. Results will be tested in other tissues, including cancer cells where these provide a natural experiment through perturbation of the inferred causal networks. Likely arriving with a statistics/mathematics background or a biology background with strong quantitative skills, the student will gain experience in handling large data sets, bioinformatics skills, and using high-performance computing. Advanced statistical and machine learning approaches will be learnt in order to achieve the modelling of regulatory networks. The student will engage with the University of St Andrews GRADSkills training programme, the Scottish Mathematical Sciences Training Centre, and the UK Academy for PhD Training in Statistics. The student will also be required to engage with the biological aspects of the project and will present to and discuss with bio-focussed audiences. HOW TO APPLY Application instructions can be found on the EASTBIO website- How to Apply | EastBio Doctoral Training Partnership | Biology 1) Download and complete the Equality, Diversity and Inclusion survey. 2) Download and complete the EASTBIO Application Form. 3) Please complete an application on our online portal: How to apply - Study at St Andrews - University of St Andrews 4) Select the course ‘PhD with internship’ 5) Your online application must include the following documents: - Completed EASTBIO application form - Academic Qualifications - English Language Qualification (if applicable) - 2 References: this must be completed on the EASTBIO Reference Form, also found on the EASTBIO website. Please download the EASTBIO reference form and send it to your referees. They can either upload it directly to the portal using the automated email they will receive, or they can email it to Rachel at pgrecbiology@st-andrews.ac.uk CONTACT Queries on the project can be directed to the project supervisor at andy.lynch@st-andrews.ac.uk Queries on the application process can be directed to Rachel Horn at pgrecbiology@st-andrews.ac.uk UKRI eligibility guidance: Terms and Conditions: View Website International/EU: View Website