以神經網路資訊解析茄子基因型與果實形態特徵間的關聯

Abstract

本研究旨在結合植物果實影像與基因型資料，建立一套能解釋形態與遺傳關係的分析流程。為此，我們開發一個基於自編碼器架構的深度學習模型，用以自動萃取茄子（Solanum melongena）果實影像的latent representation，並進一步評估其與基因型的對應關係。模型訓練採用來自亞洲蔬菜中心共1,609張果實影像資料，透過一種稱為 Variational Sparse Coding（VSC）的方法，使所學得的特徵具備稀疏性與可解釋性。們對這些潛在特徵進行視覺化與統計分析，並將其作為性狀輸入，結合 GEMMA 工具進行全基因體關聯分析（GWAS），以鑑定與影像形態特徵相關的單核苷酸多態性（SNP）。 此外，我們比較了不同的特徵聚合方式（平均值、最大值、最小值）及其主成分分析（PCA）後的效果，探討表徵方式對後續分析結果的影響。最後，我們以多層感知器（MLP）模型，根據顯著 SNP 資訊預測影像所對應的latent representation，展示從基因型重建表型特徵的可行性與潛力。整體而言，本研究建立一套由影像特徵學習、基因型關聯分析到性狀預測的整合流程，為深入理解植物的基因型與表型關係提供一項具體而有效的技術架構。

This study aims to integrate plant fruit images and genotype data to establish an analytical framework for understanding the relationship between morphological traits and genetic variation. To this end, we developed a deep learning model based on an autoencoder architecture to automatically extract latent representations from fruit images of eggplant (Solanum melongena) and assess their correspondence with genotypic information. The model was trained on 1,609 fruit images provided by the World Vegetable Center, using a method known as Variational Sparse Coding (VSC) to obtain sparse and interpretable features. We performed visualization and statistical analysis of these latent features and treated them as phenotypic traits for genome-wide association studies (GWAS) using the GEMMA tool, identifying single nucleotide polymorphisms (SNPs) associated with image-based morphological characteristics. In addition, we compared different feature aggregation methods (mean, maximum, minimum) and their principal component analysis (PCA) results to evaluate how representation strategies affect downstream analyses. Finally, we employed a multilayer perceptron (MLP) model to predict latent representations from significant SNPs, demonstrating the feasibility and potential of reconstructing phenotypic features from genotype data. Overall, this study establishes an integrated workflow from image-based feature learning to genotype-to-trait association and phenotype prediction, providing a practical and effective framework for understanding the genotype–phenotype relationship in plants.