Title: DELPHI-2M and further extensions
Abstract: We have made some key innovations in generative AI and have developed the modelling framework known as DELPHI. I will give an overall of the main results from the initial DELPHI implementation that has been described in depth in a publication in Nature 2025 (REF: 10.1038/s41586-025-09529-3). The general DELPHI model can align whole populations across space and time with a modelling space that is optimised for disease prediction across a large set of diseases. it provides a flexible model architecture that can be extended to include further multiple modal datasets at the population scale. I will describe some of the recent extensions to the DELPHI model, primarily focussed on the inclusion of genetic and drug-based information. I will end with some open questions and ideas, highlighting some current areas that require further innovation such how the internal embedding spaces of DELPHI style models could be leveraged to assess causality between multiple competing factors.
Biography: Tom Fitzgerald, Senior research staff scientist and member of faculty at EMBL-EBI. Tom has been based at the Wellcome Trust Genome Campus for over 20 years and has been involved in multiple large scale and high impact studies. His main scientific career began in 2006 at the Sanger institute with Prof. Matt Hurles on the Deciphering Developmental Disorders (DDD) project, one of the first large-scale sequencing project with a direct clinical application in the UK. He was instrumental in the development of novel analytical techniques and large-scale analytical workflows which have now been replicated and extended into programs such as the Genomics England rare disease pipelines. For the last 10 years Tom has been employed at EMBL-EBI where he co-leads the Birney research group with Prof. Ewan Birney. His research in focussed primarily on novel genome wide approaches for complex trait mapping and innovations in AI based techniques for the analysis of large-scale multimodal datasets. Some recent work includes the development of novel population scale AI modelling frameworks to represent highly dimensional phenotype distributions across millions of individuals from different world populations. His primary interest is the development of analytical techniques to extend our ability to understand the complex interplay between the multitude of factors that makes every person unique.