Transcriptomic Interpretation on Explainable AI-Guided Intuition Uncovers Premonitory Reactions of Disordering Fate...

Area of Science:

• Plant physiology
• Machine learning in agriculture
• Fruit science

Background:

• Deep neural networks (DNNs) excel in plant image diagnosis but are underutilized beyond phenotyping.
• Explainable Convolutional Neural Networks (CNNs) offer insights into physiological mechanisms behind plant phenotypes.
• Rapid over-softening is a significant internal fruit disorder in persimmons.

Purpose of the Study:

• To integrate explainable CNNs with transcriptomics for physiological interpretation of persimmon rapid over-softening.
• To develop CNN models for predicting rapid softening from fruit images.
• To identify premonitory symptoms and underlying physiological triggers of rapid softening.

Main Methods:

• Construction of CNN models to predict rapid softening in persimmon cv. Soshu using photographic images.
• Application of explainable AI techniques (Grad-CAM, guided Grad-CAM) to visualize predictive features.
• Transcriptomic analysis of featured and control regions to identify molecular pathways.

Main Results:

• CNN models accurately predicted rapid softening from images, highlighting specific premonitory symptom regions.
• Transcriptomics revealed ethylene signal-dependent cell wall modification as a trigger for rapid softening.
• Premonitory symptoms in featured regions indicated hypoxia and stress signals preceding ethylene induction.

Conclusions:

• The integration of explainable CNNs and transcriptomics provides a powerful approach for plant physiology research.
• Novel insights into the premonitory reactions and physiological triggers of rapid fruit softening were uncovered.
• This study demonstrates a collaborative framework for image analysis and omics approaches in understanding plant disorders.