Join the Edinburgh Local Group for their penultimate event of the year, inviting three speakers from near and far to speak on the topic of Machine Learning and Causal Inference in Medicine. This event will take place entirely online as a lunchtime session, and so can be joined whether you are a local group member or not.
Sjoerd Beentjes, The University of Edinburgh (School of Mathematics)
Following the Causal Roadmap to generate reliable real-world evidence.
Abstract: In contrast to traditional clinical trials, concerns about potentially biased effect estimates using real-world data have resulted in a cautious and critical approach to the interpretation of real-world evidence (RWE) findings. There may be various sources of potential bias, including sampling bias, causal bias and statistical bias. To ensure RWE studies contain sufficient detail, and prevent analyses with implausible assumptions, methodological flaws and false interpretations, the Roadmap for Causal and Statistical Inference offers a structured guideline for developers. In this short talk, I will present the causal roadmap as a stepwise framework for the design, analysis and interpretation of RWE studies, including randomised, prospective and retrospective observational studies, and discuss its relation to target trial emulation.
Ava Khamseh, The University of Edinburgh (School of Informatics)
Pinpointing cell (sub)types, states and molecular mechanisms via quantification of (high order) molecular dependencies
The genome contains the complete set of DNA that provides instructions for cells and tissues to develop and function. Genomic medicine integrates insights from the genome with information about a person’s health to design and apply improved diagnostic tools and treatments. Research in genomic medicine goes beyond identification of risk factors and aims at pinpointing underlying causal mechanisms, often at various biological scales. In this talk, we present examples, as well as limitations, of using causal inference methodologies in applications to single cell RNA-seq and large-scale genotype-phenotype biobanks.
Sam Tanner, University of Melbourne (The Florrey Institute)
Identifying Coordinated Multomic Alterations Underlying Autism and ADHD Using Machine Learning
Autism and ADHD arise from coordinated changes across multiple biological layers— including the genome, epigenome, transcriptome, and metabolome—reflecting the interconnected nature of developmental processes. Although progress has been made in identifying biomarkers within individual layers, multiomics machine-learning approaches now enable simultaneous integration across layers, offering a more holistic view of underlying mechanisms that may help unify the many genetic and environmental risk factors. Here, we applied Multi-Omics Factor Analysis (MOFA), an unsupervised machine learning method, to data from the Barwon Infant Study (N = 1,074) to identify latent multiomic signatures of autism and ADHD spanning (i) placental gene expression, (ii) cord-blood DNA methylation, and (iii) cord-blood metabolomics. MOFA identified seven latent factors, four of which were associated with autism and/or ADHD outcomes. Factor 4 was consistently associated with multiple neurodevelopmental outcomes across early childhood, including autism diagnosis at age 9 years (OR = 1.13, p = 0.0002), explaining 16.8% of the variance in this outcome. Bioinformatics analyses revealed that Factor 4 was enriched for known autism risk genes (p = 7.3×10⁻⁷), providing independent validation, and suggested a potential causal pathway involving altered placental nutrient transfer and downstream changes in brain-related epigenetic programming. Further analyses will investigate prenatal environmental exposures, including manufactured chemicals, that may influence neurodevelopment via these multiomic signatures.
Sam Tanner, University of Melbourne (The Florrey Institue)
Ava Khamseh, The University of Edinburgh (School of Informatics)
Sjoerd Beentjes, The University of Edinburgh (School of Mathematics)