Quantitative morphology, machine learning, and hyperspectral interface phenotyping of Trichoderma–Colletotrichum antagonism across host-associated isolate panels

Abstract

The efficacy of biological control agents is often inconsistent across pathogen isolate panels, yet conventional dual-culture screening often reduces antagonism to single endpoint measurements such as radial growth or colony area. Here, we developed a quantitative phenotyping framework to evaluate interactions between three Trichoderma antagonists and Colletotrichum isolates associated with coffee and cacao. Dual-culture assays were used to quantify antagonist and pathogen morphology after 96 h, and the combined morphology dataset was analyzed using machine learning to test whether host-associated isolate panels could be classified from colony-level interaction traits. To address potential information leakage and basal-growth confounding, we evaluated control-only pathogen morphology models and leave-one-pathogen-isolate-out validation. Under random 5-fold validation, Random Forest models achieved similar balanced accuracy using control-only pathogen morphology and full dual-culture interaction morphology, 0.840 and 0.864, respectively. Under the more conservative leave-one-pathogen-isolate-out validation, performance decreased but the full dual-culture model, balanced accuracy = 0.725, outperformed the control-only model, balanced accuracy = 0.578, indicating that basal pathogen morphology contributes to host-associated differences while interaction-level traits add information beyond basal growth alone. Hyperspectral imaging was then used as a proof-of-concept, non-invasive tool to characterize selected interaction interfaces. In the complete 11C-65-1 × P24-83/P24-192 subset, VNIR reflectance residuals showed interface-specific deviations from within-plate colony-side spectral mixing axes. These residual wavelength features are presented as candidate spectral correlates rather than validated biochemical mechanisms. Overall, morphology-based machine learning and hyperspectral interface phenotyping provide a scalable framework for controlled biocontrol screening, while emphasizing the importance of isolate-aware validation and cautious spectral interpretation.

Plant phenotyping relevance

植物コロニーの形態を定量化し、機械学習とハイパースペクトル画像で相互作用表現型を抽出する枠組みを開発・検証しており、表現型取得と解析手法が研究の中心である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.