Probabilistic metabolite annotation using retention time prediction and meta-learned projections

Area of Science:

• Computational chemistry and cheminformatics
• Analytical chemistry and metabolomics
• Machine learning and artificial intelligence

Background:

• Retention time is crucial for metabolite annotation in metabolomics.
• Current limitations include scarce experimental data and method variability.
• Accurate retention time prediction is needed to support metabolite identification.

Purpose of the Study:

• To develop accurate machine learning models for predicting metabolite retention times.
• To create a novel Bayesian meta-learning approach for projecting retention times across different chromatographic methods.
• To integrate these predictions into a metabolite annotation workflow for improved accuracy and reliability.

Main Methods:

• Trained state-of-the-art machine learning models (deep neural networks, gradient boosting, etc.) on the METLIN Small Molecule Retention Time (SMRT) dataset (80,038 data points).
• Utilized molecular descriptors and fingerprints (alvaDesc software) as input features.
• Employed Bayesian hyperparameter search and nested cross-validation for model optimization and performance evaluation.
• Developed a Bayesian meta-learning approach for cross-chromatographic method retention time projection.

Main Results:

• Achieved state-of-the-art prediction accuracy with a deep neural network, yielding mean absolute error of 0.11 and median absolute error of 0.07.
• The novel meta-learning approach effectively projects retention times between methods using minimal identified molecules (as few as 10).
• In metabolite annotation workflows, this method ranked the correct molecule within the top three candidates in 68% of cases when filtered by mass and ranked by z-scores.

Conclusions:

• The developed machine learning models provide highly accurate metabolite retention time predictions.
• The Bayesian meta-learning approach enables robust retention time projection across different chromatographic conditions.
• Integrating these tools into metabolomics workflows significantly enhances metabolite annotation accuracy and reliability, aiding researchers in identifying unknown compounds.