Image-based rachis phenotyping facilitates genetic dissection of spikelet distribution in wheat

Abstract

The distribution of spikelets significantly affects wheat (Triticum aestivum L.) spike architecture. However, traditional methods lack the precision to study spikelet distribution effectively. We developed RachisSeg, a deep learning-based phenotyping pipeline that automatically measures traits from scanned rachis images. In addition to traditional spikelet number per spike (SNS), rachis length (RL), and spikelet density (SD, SNS/RL), we introduced spikelet distribution traits based on rachis internode lengths, providing quantitative insights into spike architecture. RachisSeg showed high consistency with manual measurements for SNS and RL, with the R2 values of 0.975 and 0.998, respectively. Using RachisSeg, we analyzed spikelet distribution patterns across wheat germplasm and found that traits such as spikelet distribution index (SDI) and apical-to-basal spikelet number ratio (AVB_SNS) were moderately correlated with grain yield per spike (GYPS) (r = 0.57 and 0.53, respectively), while internode width (IW) showed a strong positive correlation with GYPS (r = 0.75). Specifically, a denser spikelet arrangement in the upper spike negatively impacted grain number and weight in that section. Furthermore, comparative analysis revealed distinct spikelet distribution patterns among landraces, American cultivars, and Chinese cultivars. In a recombinant inbred line population, we identified 46 quantitative trait loci (QTLs) associated with rachis traits. A major QTL controlling SDI was detected on chromosome 6B, explaining up to 24.8% of the phenotypic variance. Candidate gene analysis suggested TraesCS6B02G417000 as a potential gene, whose mutant exhibited significant changes in RL and SDI. RachisSeg is a powerful tool for quantifying spikelet distribution, facilitating wheat genetic analysis, gene discovery, and breeding.

Code and data availability

The paper's authors publicly released the RachisSeg phenotyping pipeline (deep learning node detection and internode segmentation code) together with sample rachis images via their GitHub repository, explicitly stated in the Implementation and Data availability sections.