Machine learning classification by fitting amplicon sequences to existing OTUs

Area of Science:

• Microbiome research
• Machine learning in diagnostics
• Bioinformatics

Background:

• Microbiome composition analysis using 16S rRNA gene sequences can aid in patient diagnosis.
• Training machine learning models for microbiome-based diagnosis often requires re-clustering operational taxonomic units (OTUs) when new data are added, necessitating model retraining.
• Existing methods for OTU clustering can lead to changes in OTU composition as new data become available.

Purpose of the Study:

• To evaluate the performance of the OptiFit algorithm in classifying patients with and without colonic screen relevant neoplasia (SRN).
• To compare the diagnostic accuracy of machine learning models trained using OptiFit with traditional de novo and reference-based clustering methods.
• To determine if OptiFit can enable the reuse of existing classification models without the need for retraining.

Main Methods:

• Clustering new 16S rRNA gene sequences into pre-existing de novo OTUs using the OptiFit algorithm.
• Training and evaluating random forest classification models on a dataset of patient samples.
• Comparing model performance against standard de novo and database-reference-based clustering approaches.

Main Results:

• Machine learning models utilizing OptiFit demonstrated comparable or superior performance in classifying SRN cases.
• OptiFit successfully integrated new sequence data into existing OTUs without altering their composition.
• The use of OptiFit streamlined the classification process by eliminating the need for model retraining with reclustered sequences.

Conclusions:

• OptiFit provides a viable method for fitting new microbiome sequence data into existing OTUs, enabling model reusability.
• This approach overcomes the challenge of retraining classification models when new patient data are introduced.
• OptiFit facilitates the development and deployment of stable, validated microbiome-based diagnostic models.