Predicting the seed microbiome using phylogeny-driven machine learning.

Area of Science:

• Microbiology
• Bioinformatics
• Plant Science

Background:

• Seed-associated bacterial microbiome composition can mirror plant evolutionary history (phylosymbiosis).
• Machine learning models for microbial communities often need extensive data, limiting their use for new hosts.
• Predicting microbial communities for unsampled hosts remains a challenge.

Purpose of the Study:

• To investigate if plant nuclear internal transcribed spacer (ITS) sequences can predict species-level seed-associated bacterial communities.
• To develop and test machine learning models utilizing host relatedness for microbiome prediction.
• To assess the utility of plant ITS sequences as a proxy for host relatedness in microbiome studies.

Main Methods:

• Utilized 16S rRNA gene sequencing data from seed-associated bacteria of 61 plant species.
• Developed custom machine learning models based on sequence-based Hamming distances to quantify plant host relatedness.
• Compared the predictive performance of Hamming Distance-based k-Nearest Neighbor (HD-KNN) and Hamming Distance-based Gaussian Process Regression (HD-GPR) models.

Main Results:

• The HD-KNN model achieved the highest predictive accuracy, with an average Jensen-Shannon divergence (JSD) of 0.276.
• HD-KNN performed best for densely sampled plant groups (e.g., Brassicaceae, Poaceae) with available closely related species.
• HD-GPR showed better performance for phylogenetically isolated species, indicating model sensitivity to host representation.

Conclusions:

• Plant ITS-derived host relatedness provides a partial predictive signal for seed-associated bacterial microbiome composition.
• This framework supports low-input predictive modeling for seed-associated bacteria, aiding predictions for unsampled species.
• Findings are specific to seed-associated bacteria and not directly generalizable to other microbial communities or plant compartments.