Ava Khamseh (Affiliate)

Research in a Nutshell 

One of the aims of biomedical genomics is to determine causal molecular interactions underlying biological processes. These interactions can be highly complex, involving many dependent variables constituting a given biological system of interest. Quantifying interactions, beyond extracting their magnitude, and moving towards obtaining information on their directionality requires well-designed experiments and/or large-scale individual-level biomedical data.

Currently, our research focuses on causal molecular interactions in the contexts of cancer and complex trait population biomedicine:

1. Cancer biology: Progression of cancer has been widely viewed as following Darwinian evolution, where cells acquiring somatic driver mutations are positively selected, leading to preferential expansions of clones carrying these mutations. Advances in high-throughput DNA- and RNA-sequencing have substantially increased our understanding of late-stage cancer. Yet we still lack a quantitative understanding of the necessary and sufficient early conditions and steps required for the evolution of normal cells into cancer cells.

We design experimental and computational models of early cancer to (i) capture mutational, transcriptional and morphological trajectories over time at single-cell resolution, and (ii) quantify how different driver mutation interactions and/or mutation and environment interactions lead to cooperation or competition amongst cells.

2. Population biomedicine: We develop causal machine learning techniques for applications to large-scale omics data, such as scRNA-seq, and biobanks, such as UK Biobank and All of Us. Specifically, we are interested in developing and applying semi-parametric statistical and ML approaches to (i) extract higher-order interactions from gene expression and omics data to identify cell (sub)types and states at high resolution, (ii) quantify (causal) interactions amongst DNA variants leading to physiological outcomes (iii) identifying how various clinical features or treatments causally affect disease outcome. These methodologies are based on the framework of Targeted Learning, which avoids model misspecification and provides mathematical guarantees on the final target estimates of interest.

Specific disease areas of interest include liver disease and cancer, ME/CFS and cardiometabolic syndrome.

• Dr Luke Boulter, University of Edinburgh
• Professor Chris Ponting, University of Edinburgh
• Professor Martin Taylor, University of of Edinburgh
• Dr Sjoerd Beentjes, School of Mathematics, University of Edinburgh
• Dr Yi Feng, University of Edinburgh
• Professor Elham Kashefi, University of Edinburgh and the National Quantum Computing Centre
• Dr Stephen Duffield, National Institute for Health and Care Excellence (NICE)
• Professor Mark van der Laan, University of California, Berkeley
• Dr Nima Hejazi, Harvard Chan School of Public Health