*
High-throughput molecular biology

I have worked for several years on Bayesian integrative models in molecular epidemiology. I have been involved in the development of new statistical methodology for several different types of high-throughput molecular biology data, including gene expression microarrays (Lewin et al. 2006; Lewin et al. 2007; Turro et al. 2010), RNA-seq (Turro et al. 2011), proteomics (Kirk et al. 2013), metabolomics (Lewin et al. 2015, Bottolo et al. 2021, Scott et al. 2023) and microbiome (Scott et al. 2023). The emphasis in all of this work is on integrative modelling, using fully Bayesian models to account for the complex correlation structures in the data and propagate uncertainty on model estimates.

*
Multi-trait analysis in genetics and genomics

Quantitative Trait Loci (QTLs) are genetic variants which are statistically associated with a phenotype of interest. In molecular biology, high-throughput technologies have enabled us to find QTLs for multivariate molecular phenotypes (for example multivariate gene expression (eQTLs), proteomics (pQTLs) and metabolomics (mQTLs)).

Traditional analysis approaches consider each molecular variable separately, despite these data showing extremely high correlations. We have developed Bayesian models for detecting QTLs for multivariate molecular outcomes, and have used these to detect eQTLs and mQTLs. Joint modelling of genomics and metabolomics data for eQTL/mQTL detection (Lewin et al. 2015, Bottolo et al. 2021), joint modelling of microbiome and metabolomics data (Scott et al. 2023) and multi-omics data integration in drug-resistance studies (Zhao et al. 2021, 2023).

An extension of this work into causal modelling is Verena Zuber's paper on Mendelian Randomisation: here we introduce multi-response Mendelian randomization (MR2), an MR method specifically designed for multiple outcomes to identify exposures that cause more than one outcome or, conversely, exposures that exert their effect on distinct responses (Zuber et al. 2023).

Darren Scott recently completed his PhD with me, working on models linking multivariate molecular outcomes with microbiome data. Microbiome data is compositional, meaning that features are expressed as proportions of a whole. Standard supervised learning models cannot be used for compositional data as they treat feature as independent. We have developed models for univariate (Scott et al. 2023) and multivariate outcomes (manuscript in progress) using microbiome as compositional predictors.

*
Machine Learning in Health Data Research

I am currently co-investigator on InflAIM, an NIHR-funded project led by the University of East Anglia to investigate multimorbidity using AI methods. We will be using multi-state models and Bayesian networks to study links and risk factors for multimorbidity.

I recently wrote a "Lessons Learnt" paper for the Centre for Impact Evaluation on machine learning methods used in causal inference for impact evaluation, in particular with respect to investigating heterogeneous treatment effects and mechanisms (Lewin et al. 2023).

I am currently supervising an MSc dissertation on interpretable AI methods. We are investigating the reliability of explanation methods for complex black box models in the context of observational epidemiology. I am also co-supervising a PhD student surveying the use of machine learning methods for large-scale disease surveillance using online social media data.

*
Causal Inference Methodology

I am supervising two NIHR pre-doctoral Fellows (Lauren Rengger and Jenni Banks) working on causal mediation analysis. We are investigation causal mechanisms of the association between eczema and cardiovascular outcomes, using recently developed methods for causal inference.

Mendelian Randomisation for multivariate outcomes: here we introduce multi-response Mendelian randomization (MR2), an MR method specifically designed for multiple outcomes to identify exposures that cause more than one outcome or, conversely, exposures that exert their effect on distinct responses (Zuber et al. 2023).

*
Bayesian Evidence Synthesis

I am working with Darren Scott (AstraZeneca) on Bayesian models for using historical data to improve efficiency of randomised trials analysis (Scott and Lewin 2024 arxiv paper).

I work with Joy Lawn's group in LSHTM MARCH (Centre for Maternal, Adolescent, Reproductive, & Child Health) advising on Bayesian evidence synthesis methods used to produce global and country-specific estimates of disease burden and adverse birth outcomes (Gonçalves et al. 2021, Gonçalves et al. 2022, Ohuma et al. 2023, Okwaraji et al. 2023).