Uncertainty Quantification and Flagging of Unreliable Predictions in Predicting Mass Spectrometry-Related Properties...

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Cheminformatics

Background:

• Mass spectral identification is crucial in metabolomics, often enhanced by predicting molecular properties like chromatographic retention.
• Current machine learning models for prediction typically lack robust uncertainty estimation.
• Accurate uncertainty quantification is vital for reliable molecular identification and data interpretation.

Purpose of the Study:

• To develop and validate a method for assessing prediction uncertainty in molecular property prediction.
• To improve the reliability of machine learning models used in mass spectral identification and metabolomics.
• To integrate multiple uncertainty indicators into a unified assessment framework.

Main Methods:

• Utilized ensemble model prediction spread as an uncertainty indicator.
• Employed Euclidean distance of molecular descriptors to quantify molecular similarity.
• Incorporated data set clustering to identify potential outliers or uncertain data points.
• Developed classification models using these factors to predict prediction reliability.

Main Results:

• Achieved area under the receiver operating curve (AUC) values between 0.73 and 0.82 for predicting unreliable predictions.
• Demonstrated the effectiveness of combining ensemble spread, molecular similarity, and clustering for uncertainty assessment.
• Successfully predicted whether a prediction falls within the worst 15% of outcomes across various chromatographic and ion mobility tasks.

Conclusions:

• The proposed method effectively quantifies prediction uncertainty in molecular property prediction tasks.
• Integrating ensemble spread, molecular similarity, and clustering provides a comprehensive approach to uncertainty estimation.
• This work enhances the trustworthiness of machine learning predictions in analytical chemistry and metabolomics.