Finding potentially erroneous entries in METLIN SMRT.

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Cheminformatics

Background:

• The METLIN SMRT dataset is crucial for retention time prediction in high-performance liquid chromatography (HPLC).
• Existing data filtering pipelines for METLIN SMRT are often inadequate, leading to potential inaccuracies.
• A reliable method for identifying and removing erroneous entries is needed to improve dataset quality.

Purpose of the Study:

• To adapt and apply a robust filtering method for potentially erroneous entries in the METLIN SMRT dataset.
• To enhance the reliability of the METLIN SMRT dataset for both machine learning applications and experimental research.
• To assess the efficacy of predictive models in identifying data anomalies within large-scale chromatography datasets.

Main Methods:

• Repurposed an existing filtering approach for gas chromatography retention indexes to the METLIN SMRT dataset.
• Employed five predictive models: Graph Neural Network (GNN), Convolutional Neural Network (CNN), Extended Connectivity Fingerprints (ECFP), Functional Connectivity Density (FCD), and CatBoost.
• Utilized a 5-fold cross-validation strategy for retention time prediction and flagged entries with significant deviations (bottom 5%) using a "yellow card" system.

Main Results:

• Approximately 1500 entries (2% of the dataset) received at least one "yellow card" across the five models.
• An estimated 1200 entries were identified as strongly suspected erroneous data points.
• Approximately 300 entries were flagged as likely inaccurate predictions, distinct from erroneous entries.

Conclusions:

• The developed filtering approach is viable for improving the quality of the METLIN SMRT dataset.
• This method holds significant potential for enhancing other large-scale chromatography-related databases.
• Improved data quality benefits both the training of machine learning models and direct experimental utilization.