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

Authors: Baek, Insuck; KANDPAL, LALIT - Orise Fellow; BHATT, JISHNU - Orise Fellow; LIM, SEUNGHYUN - Orise Fellow; Lovelace, Amelia; Cohen, Stephen; Kirubakaran, Silvas; UPADHYAY, RAKESH - Bowie State University; Kim, Moon; Meinhardt, Lyndel; Ahn, Ezekiel

Submitted to: Biological Control
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 6/30/2026
Publication Date: 7/1/2026
Citation: Baek, I., Kandpal, L., Bhatt, J., Lim, S., Lovelace, A.H., Cohen, S.P., Kirubakaran, S.J., Upadhyay, R., Kim, M.S., Meinhardt, L.W., Ahn, E.J. 2026. Quantitative morphology, machine learning, and hyperspectral interface phenotyping of Trichoderma–Colletotrichum antagonism across host-associated isolate panels. Biological Control. 219. Article e106104. https://doi.org/10.1016/j.biocontrol.2026.106104.
DOI: https://doi.org/10.1016/j.biocontrol.2026.106104

Interpretive Summary: Using beneficial fungi like Trichoderma to fight plant diseases is a promising alternative to toxic chemicals, but these "living drugs" often fail unpredictably in real-world farms. The problem is that we treat all pathogens as the same, ignoring that they carry a hidden "ecological memory" of the specific environment they evolved in. In this study, we staged microscopic battles between biocontrol agents and fungal pathogens from coffee and cacao fields, using Artificial Intelligence (AI) and advanced hyperspectral cameras to decode their interactions. We discovered that AI could identify where a pathogen came from with 90% accuracy just by analyzing the "shape" of the battle. Furthermore, by looking at invisible light waves, we revealed the physical mechanism of this war: weak pathogens undergo a catastrophic "structural collapse" by losing water, while resistant ones fight back by producing protective chemical pigments. This research converts the abstract concept of evolutionary memory into a measurable physical signal, providing biocontrol manufacturers, agronomists, and farmers with a precise new tool to screen for "smarter" biological pesticides that work reliably in specific local environments.

Technical Abstract: The inconsistent field efficacy of microbial biocontrol agents remains a major barrier to sustainable agriculture, potentially driven by the uncharacterized "local ecological memory" of target pathogens. To resolve this, we developed a multi-modal phenotyping framework integrating quantitative morphometrics, machine learning, and hyperspectral geometric modeling to interrogate Trichoderma-Colletotrichum interactions from distinct coffee and cacao agroecosystems. While traditional multivariate analysis (PCA) failed to distinguish pathogen origin, Neural Boosted machine learning models successfully classified host-origin with ~90% accuracy, validating the existence of a host-specific interaction phenotype. Hyperspectral geometric analysis (400–1000 nm) subsequently decoded the biophysical basis of these states: susceptibility was defined by a high orthogonality ratio (˜0.33) driven by a specific water absorption feature at ~980 nm (structural collapse), whereas resistance involved an active biochemical deviation at ~450 nm (pigment-based defense). These findings demonstrate that ecological memory is physically encoded in a trade-off between hydric stability and chemical defense, establishing a device-agnostic, biophysical paradigm for predicting biocontrol outcomes beyond simple inhibition zones.