Taxonomy-agnostic hyperspectral–morphological phenotyping of fungal pathogen chemical-stress responses using machine learning

Authors: Baek, Insuck; LIM, SEUNGHYUN - Orise Fellow; Oh, Sookyung; Kazem Rostami, Masoud; Lovelace, Amelia; Lew, Helen; Kim, Moon; Meinhardt, Lyndel; KANDPAL, LALIT - Orise Fellow; CHA, MINHYEOK - Orise Fellow; HWANG, CHANSONG - Non ARS Employee; Ashby, Richard; Ahn, Ezekiel

Submitted to: Smart Agricultural Technology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 2/15/2026
Publication Date: 2/17/2026
Citation: Baek, I., Lim, S., Oh, S., Kazem Rostami, M., Lovelace, A.H., Lew, H.N., Kim, M.S., Meinhardt, L.W., Kandpal, L., Cha, M., Hwang, C., Ashby, R.D., Ahn, E.J. 2026. Taxonomy-agnostic hyperspectral–morphological phenotyping of fungal pathogen chemical-stress responses using machine learning. Smart Agricultural Technology. 13. Article 101895. https://doi.org/10.1016/j.atech.2026.101895.
DOI: https://doi.org/10.1016/j.atech.2026.101895

Interpretive Summary: Fungal diseases pose a persistent threat to essential global crops like coffee and cacao, causing significant economic losses. The reliance on conventional chemical fungicides is increasingly problematic due to the rise of resistant pathogens and concerns about environmental impact, creating an urgent need for safer, more sustainable solutions. This study investigated a novel question: Can we determine a fungus's original "home" environment—for instance, a coffee farm versus a cacao farm—simply by observing how it reacts to new, eco-friendly chemical compounds? We tested if pathogens carry an "ecological memory" of the place they adapted to. To do this, we treated Colletotrichum fungi from both coffee and cacao with four new compounds derived from renewable sources like soybean oil and beechwood creosote. We then used high-precision imaging and software to measure not just the size of the fungal colonies, but also subtle changes in their shape, such as how smooth and round their edges were (a trait called "circularity"). Finally, we used artificial intelligence (machine learning) to see if it could learn to tell the two groups of fungi apart based on these detailed response patterns. While traditional statistical analysis couldn't distinguish the fungi by their origin, the machine learning models were remarkably successful. The AI was able to classify whether a fungus came from a coffee or cacao farm with approximately 87% accuracy. The most important clue it used was the colony's circularity, revealing that the symmetry of a fungus's growth under stress is a powerful fingerprint of its adaptive history. This finding strongly suggests that pathogens retain an ecological memory that shapes their response to new challenges. This research provides a new framework for decoding this memory. It offers a powerful, scalable method for rapidly screening next-generation, eco-friendly fungicides and opens the door to developing "precision" disease-control strategies tailored to a pathogen's specific background. This research will be used by scientists and mycologists and ultimately help make agriculture more sustainable and secure the future of these vital crops.

Technical Abstract: Can standardized chemical stress reveal a reproducible, taxonomy-agnostic stress-response fingerprint that is predictive of sample-source labels (crop-of-isolation) in fungal isolates? Six coffee-associated Colletotrichum isolates were profiled using four phenolic-branched compounds and compared with a previously characterized cacao panel. Quantitative morphology, hyperspectral imaging (HSI), and supervised machine learning (ML) yielded panel-specific fingerprints under uniform, isotropic in vitro conditions. Circularity, a measure of edge symmetry, was the most informative morphological feature, and ML classified the crop-of-isolation label (coffee vs cacao, in this panel) with 86.7% accuracy in within-panel cross-validation. HSI detected dose-dependent spectral shifts in a targeted subset of isolates and compounds, including changes near 1930 nm in the short-wave infrared, a moisture-sensitive region that warrants robustness checks (e.g., band masking or preprocessing sensitivity) prior to biochemical attribution. Multi-locus phylogeny showed the coffee isolates are polyphyletic, so the predictive signal should be interpreted conservatively as a taxonomy-agnostic phenotype fingerprint associated with crop background in this mixed-lineage panel, acknowledging that crop labels are partially confounded with phylogenetic structure. We propose a “chemical priors” framework as a working hypothesis, in which long-term environmental exposure may imprint stress-response pathways that become legible under simple, standardized probes. This integrative workflow supports scalable screening of eco-friendly antifungals and sensor-driven decision support for high-throughput phenotype-based screening workflows.