1. Polygenic scores (PGS) are widely used to estimate genetic risk for complex traits, but standard additive methods can miss non-additive effects. This talk introduces the promise and pitfalls of using artificial intelligence (AI) models – specifically, XGBoost and neural networks – for genetic prediction. During the presentation, I will provide a primer on polygenic scores and the machine learning (ML) and deep learning (DL) models used and then present insights from the REALMENT Project. I focus on how non-additive genetic effects can affect PGS performance, when ML and DL models can offer improvements, and how this applies to precision psychiatry.
Title: AI for genetic prediction in psychiatry: insights from the REALMENT project
Speaker: Nathan Bell, the Netherlands
2. During this webinar, I will present preliminary results from a multi-cohort study that integrates multidimensional symptom phenotyping with polygenic risk profiling in schizophrenia. Using factor analysis of PANSS items followed by unsupervised clustering, we identify four reproducible clinical subgroups that remain longitudinally stable over nine months and outperform conventional severity quartiles. In parallel, k-means clustering of polygenic risk scores across twelve psychiatric traits yields three transdiagnostic genomic profiles; notably, the “overall severe” clinical subgroup shows the highest concordance with the high SCZ/BD polygenic-burden profile. Linking clusters to longitudinal outcomes reveals domain-specific treatment trajectories: severe cases improve later but more sustainably, subgroups with milder or predominantly negative presentations respond early then plateau, and the limited-negative subgroup shows little change in negative symptoms; at the genomic level, high-burden patients show greater improvement in positive symptoms, low-burden patients in negative symptoms, whereas those with elevated MDD/neurodevelopmental load show minimal change. While the analysis is restricted to European-ancestry cohorts and polygenic scores explain modest variance, and despite differential attrition in one subgroup, these findings support the value of combining genomic and clinical stratification to inform prognostic enrichment and guide precision-oriented interventions.
Title: The hidden patterns linking genomics and symptoms: a machine-learning perspective
Speaker: Piergiuseppe Di Palo, Italy
ECNP-REALMENT webinar series: The ECNP-REALMENT webinar series provides researchers and clinicians with new insights into precision psychiatry and the optimization of treatments for severe mental disorders. Aiming to foster discussion and encourage collaboration across the psychiatric research community, the series covers topics ranging from genetics, predictive modelling, and artificial intelligence to biomarkers and strategies for translating research into clinical practice.
Each webinar features unique perspectives from project partners, promoting the exchange of ideas and supporting the application of scientific discoveries to improve patient outcomes.
