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

1. Problem Statement

The study addresses the critical need for rapid, scalable, and non-destructive methods to screen antifungal compounds and characterize fungal stress responses. Traditional approaches often rely on:

• Taxonomic identification: Time-consuming DNA sequencing is required to identify fungal species before screening.
• Linear statistical analysis: Standard multivariate methods (like PCA) often fail to detect subtle, non-linear phenotypic differences between fungal isolates from different host backgrounds (e.g., coffee vs. cacao).
• Limited stress profiling: While Hyperspectral Imaging (HSI) and Machine Learning (ML) are used for disease detection in planta, their application to quantifying dynamic, compound-induced stress responses in vitro remains underexplored.

The central hypothesis is that standardized chemical stress can elicit reproducible, "taxonomy-agnostic" phenotype fingerprints that encode information about a pathogen's source (crop-of-isolation), even without prior genetic sequencing.

2. Methodology

The authors developed an integrated workflow combining quantitative morphology, hyperspectral imaging (HSI), and supervised machine learning.

A. Biological Materials & Chemical Stressors

• Isolates: Six Colletotrichum isolates from coffee and a previously characterized panel of cacao-associated isolates (including Pestalotiopsis sp. and C. gloeosporioides).
• Chemical Stressors: Four novel, biobased phenolic-branched compounds derived from sustainable feedstocks:
  • Two amide derivatives: PhSOAM and BCBGAM.
  • Two fatty acid analogs: PhSOFA and BCBGFA.
• Assay: Fungal plugs were placed on PDA plates treated with specific compounds (or ethanol control) and incubated for 96 hours.

B. Data Acquisition

1. Quantitative Morphology: Digital images were captured and analyzed using SmartGrain software to extract seven traits: colony area, perimeter, length, width, Length-to-Width Ratio (LWR), circularity, and asymmetry index.
2. Hyperspectral Imaging (HSI): Three modalities were employed on selected isolates to capture physiological responses:
   • Fluorescence imaging (438–718 nm).
   • Visible-Near Infrared (VNIR) reflectance (401.5–1001 nm).
   • Short-Wave Infrared (SWIR) reflectance (842–2532 nm), specifically noting the moisture-sensitive region near 1930 nm.
3. Phylogenetics: Multi-locus sequencing (ITS, GAPDH, ACT) was performed to confirm that the coffee isolates were polyphyletic (not a single clade), ensuring that any predictive signal was not merely a proxy for a single species.

C. Statistical & Machine Learning Analysis

• Preprocessing: Morphological data were log-transformed and standardized.
• Linear Analysis: PCA and PERMANOVA were used to assess variance explained by crop labels and treatments.
• Machine Learning: A suite of nine supervised ML algorithms (including Neural Boosted, SVM, Random Forest, and Logistic Regression) was trained to classify the "crop-of-isolation" label (Coffee vs. Cacao) based on morphological and spectral features.
• Validation: Rigorous cross-validation strategies were used, including:
  • Repeated 5-fold cross-validation (random splitting).
  • Stratified Group K-Fold: Isolates were grouped to prevent data leakage (training and testing on different plates of the same isolate).
  • Feature Ablation: Removing top predictors (e.g., circularity) to test model robustness.

3. Key Results

A. Morphological Responses

• The four compounds induced distinct, isolate-specific morphological changes.
• Amides (PhSOAM/BCBGAM) generally showed stronger growth inhibition than fatty acid analogs.
• Shape Decoupling: While size metrics (area) often decreased, shape metrics (circularity, elongation) varied significantly. For example, BCBGAM induced elongation in one isolate but compactness in others.
• Hormesis: One fatty acid analog (BCBGFA) paradoxically stimulated growth in a specific isolate, suggesting isolate-specific metabolic capabilities.

B. Phylogenetic Context

• Phylogenetic analysis confirmed the coffee isolates were polyphyletic, distributed across multiple distinct clades.
• Some coffee isolates were genetically closer to cacao isolates than to other coffee isolates. This validates the "taxonomy-agnostic" nature of the study: the predictive signal is not simply a species identifier.

C. Machine Learning Performance

• Linear Limitations: PCA and PERMANOVA showed that crop-of-isolation labels explained only ~1.7% of the total phenotypic variance, indicating that linear methods could not separate the groups.
• ML Success: Supervised ML achieved 86.7% accuracy in classifying crop-of-isolation labels (Coffee vs. Cacao) using a Neural Boosted model.
• Feature Importance: Circularity was the single most dominant predictor (Total Effect score: 0.592), followed by colony area and perimeter.
• Robustness:
  • Removing the top feature (circularity) only dropped accuracy to 86.1%, as the model shifted to rely on size metrics.
  • Excluding the highly responsive outlier isolate (CGH5) maintained accuracy at 83.5%.
  • Grouped Cross-Validation: Even when controlling for isolate-level data leakage (testing on unseen isolates), accuracy remained significant (79.3%–82.9%), confirming the signal is generalizable within the panel.

D. Hyperspectral Fingerprints

• HSI revealed dose-dependent spectral shifts.
• The highly responsive cacao isolate (CGH5) showed strong spectral separation across all modalities.
• The tolerant coffee isolate (P24-85) showed attenuated responses, particularly in the SWIR moisture-sensitive region (~1930 nm).
• Different compounds induced unique spectral signatures in the same isolate, suggesting compound-specific physiological impacts.

4. Key Contributions

1. Taxonomy-Agnostic Workflow: Demonstrated a pipeline that predicts fungal source labels based on stress-response phenotypes without requiring DNA sequencing or species-level identification.
2. Non-Linear Phenotyping: Showed that while linear statistics fail to separate mixed-lineage panels, non-linear ML can extract high-dimensional "chemical prior" signatures (stress-response fingerprints) with high accuracy.
3. Robust Feature Set: Identified that morphological shape (specifically circularity) and size metrics contain redundant, predictive information about fungal lineage background under chemical stress.
4. Multi-Modal Integration: Successfully combined macroscopic morphology with hyperspectral imaging to provide a comprehensive view of stress responses, from colony shape to internal physiological changes.
5. High-Throughput Screening Platform: Proposed a scalable framework for rapid screening of eco-friendly antifungals and sensor-driven decision support.

5. Significance and Implications

• Antifungal Discovery: This workflow enables rapid, high-throughput screening of novel, eco-friendly antifungal compounds by quantifying subtle, reproducible stress responses that traditional growth-inhibition assays might miss.
• Biosecurity & Diagnostics: The ability to predict crop-of-isolation without sequencing offers a new tool for biosecurity, potentially identifying the origin of a pathogen outbreak based on its physiological "fingerprint."
• Concept of "Chemical Priors": The study supports the hypothesis that long-term environmental exposure (to specific crop chemistries) imprints stress-response pathways that become legible under standardized chemical probes.
• Future Directions: The authors suggest future work should focus on validating these fingerprints in planta, resolving the molecular/genomic basis of these responses, and testing the "chemical priors" hypothesis across diverse pathogen systems.

In conclusion, the paper presents a robust, sensor-driven methodology that decouples phenotypic profiling from taxonomic bottlenecks, offering a powerful new lens for understanding fungal stress physiology and accelerating antifungal development.