Predicting oil contamination in water using machine learning on microbial compositions.

Area of Science:

• Environmental microbiology
• Machine learning applications
• Bioinformatics

Background:

• Microbial community composition is a sensitive indicator of environmental contamination.
• Predicting oil contamination from microbial data presents challenges due to high dimensionality and data sparsity.
• Existing methods struggle with complex, non-linear relationships in microbial datasets.

Purpose of the Study:

• To develop a compact, generative machine learning framework for oil contamination prediction.
• To identify key microbial species associated with oil contamination.
• To assess the robustness and generalization capabilities of the predictive model.

Main Methods:

• Compared dimensionality reduction techniques (feature importance, PCA, t-SNE) on a 503-dimensional bacterial dataset.
• Employed data augmentation using an augmented data neural network (ADNN) and generative modeling (VAE) to address data scarcity and test robustness.
• Utilized top bacterial features for prediction and evaluated model performance using R² values.

Main Results:

• Feature importance outperformed PCA and t-SNE in dimensionality reduction and identifying relevant microbial species.
• The model achieved high predictive accuracy (R² up to 0.99) in training and stress testing using selected features.
• Bottle-level hold-out evaluation showed lower and variable performance (mean test R² = -0.150), indicating limited generalization.

Conclusions:

• The developed machine learning framework demonstrates feasibility for predicting oil contamination from microbial data.
• Feature importance and generative modeling are effective strategies for handling high-dimensional microbial data.
• Further validation on larger, independent datasets is necessary to improve generalization performance.