UPSC Seminar

Guillaume Ramstein
Center for Quantitative Genetics and Genomics, Aarhus University, Denmark

Title: In silico Prediction of Variant Effects in Next-Generation Plant Breeding: Evaluation of Sequence-Based AI for Genomic Prediction and Precision Breeding

Host: Kelly Swarts

Abstract:
In plant breeding, variant discovery aims to identify candidate causal variants for applications like CRISPR-based editing, where variants are targeted based on functional molecular knowledge. Recent advances in sequence-based AI models offer a new paradigm for this task. Unlike traditional approaches that rely on statistical associations between variants and traits (e.g., GWAS or QTL mapping), these models learn relationships between DNA or protein sequences and biological functions such as gene expression or fitness. This enables more fine-grained, mechanistic predictions of variant effects.
In this talk, I will provide an overview of current sequence-based AI techniques relevant to plant breeding. I will then present my own research evaluating these models for their ability to detect causal variants across several prediction tasks: (i) estimating haplotype effects using fitness and protein structure models, and (ii) prioritizing point mutations using biological language models (e.g., ESM, PlantCaduceus) and sequence-to-expression models (e.g., PhytoExpr, our proposed method). Our results highlight both the promise and current limitations of these tools in practical breeding contexts. I will conclude my talk by discussing the technological maturity of sequence-based AI approaches and their current potential for precision breeding.