ProbStab: A probabilistic ML-assisted pipeline for genotype performance, stability, and risk evaluation in multi-environment trials

Abstract

In multi-environment trials (MET), ignoring genotype × environment interaction (GEI) can lead to suboptimal cultivar recommendations and increased risk in breeding decisions. To make informed decisions, it is important to select genotypes that combine high yield with a low chance of very poor performance. Recent advances in artificial intelligence and machine learning (ML) enable researchers to study complex patterns in MET data, improving the screening and selection of superior genotypes. In this study, we introduce the statistical package ProbStab, which was developed to evaluate the yield, stability, and risk of genotypes in MET. ProbStab is based on Quantile Regression Forests, a non-parametric algorithm that provides reliable prediction intervals and addresses the limitations of classical methods. The package includes several probabilistic modules (ml_tcp_pipe, prob_stab_pipe, ci_perf_pipe, risk_potential_pipe). Using bootstrapping and machine learning, these modules estimate key indicators, including the probability of superior yield, the probability of superior stability, the joint probability of yield and stability, and measures of risk and performance potential. This integrated design fills a crucial gap in conventional approaches, providing a practical tool for advanced probabilistic analysis. By combining MLwith probabilistic evaluation, ProbStab gives plant breeders and agronomists a novel framework to reduce uncertainty in decision-making. The main goal of ProbStab is to support breeding programs with a clear, data-driven, and efficient approach for selecting stable and high-yielding genotypes under complex GEI conditions and to improve the reliability of cultivar recommendations for farmers.