Multimodal Cross-Disorder Biotype Identification and Multimethod Data Analytic Support

Prof. Dr. Tim Hahn¹, Prof. Dr. Bertram Müller-Myshok^2,3, Prof. Dr. Astrid Dempfle⁴, Prof. Dr. Andreas Jansen⁵

¹University of Münster, Department of Psychiatry and Psychotherapy
²Max Planck Institute of Psychiatry, München
³University of Liverpool, Liverpool, UK
⁴University of Kiel, Institute of Medical Informatics and Statistics
⁵University of Marburg, Department of Psychiatry and Psychotherapy

Diagnosis and treatment in psychiatry still rely almost exclusively on a phenotype-based approach. Although this renders the validity of psychiatric classification questionable and severely hampers biomarker discovery, establishing a system of biologically meaningful groups – termed biotypes – has yet remained elusive . Building on the large-scale dataset acquired in this FOR, it is now finally possible to apply state-of-the-art tools from the fields of machine learning and multivariate statistics to a rich, multimodal database, thereby bringing robust biotype identification across disorders within reach. In the first funding period, WP6 established essential quality control protocols and performed statistical analyses of gene-environment interactions in both human samples (PsychChip x childhood maltreatment on hippocampus volume, Work Package 1 and Work Package 5) and Cacna1c knock-out rats (Work Package 2, interaction of genotype with social isolation on miRNA expression (Work Package 3), mitochondrial perturbation (Work Package 4) and immunologic phenotypes (Work Package 2, Central Project 1)). Importantly, the two new PIs of Work Package 6 (TH, BM) have implemented a machine learning workflow for whole-genome and neuroimaging analysis of data from Work Package 1 and Work Package 5, with first robust results. The Work Package 6 PIs jointly aim to identify and validate biologically plausible, homogeneous biotypes within and across levels of observation – from molecular genetics to whole-brain neuroimaging. Drawing on data from all other Work Packages, we will 1) employ domain-knowledge-based and automatic feature engineering including Deep Autoencoders and Rectified Factor Networks to address the Curse of Dimensionality, 2) develop a principled approach to confounder removal in linear and non-linear multivariate models using e.g. modality-specific Adversarial Autoencoders, and 3) ensure reliability and internal validity by optimizing cluster solution stability through ensemble learning within and across levels of observation. Crucially, the longitudinal design of this FOR allows assessing the validity of our cluster solutions based on predictive utility. Specifically, predictive biomarker models for disease trajectory and outcome will be trained using our cluster solutions. We expect them to substantially outperform the same models trained on DSM-IV disorder groups. In addition, we expect these models to be robust against common confounders such as scanner site. Within this framework, we are uniquely positioned to uncover biologically plausible patient clusters across disorders, which will strengthen the validity of psychiatric classification and crucially simplify biomarker discovery in the future. Further, our collaborations with other consortia will allow us to rigorously test reproducibility in several large, independent samples, thereby ensuring validity and maximum impact of our results on the field. Finally, Work Package 6 will provide the custom-tailored, scalable analytic tools, particularly in the area of machine learning and multivariate statistics, required in all other Work Packages.