Uncertainty-Aware Learning of Multiple Conditions as a Framework for Streamlined Retention Time Prediction to...

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Artificial Intelligence in Science

Background:

• Liquid chromatography (LC) is essential for analyzing complex mixtures.
• Method screening and development in LC are time-consuming and resource-intensive.
• Accurate prediction of retention times is crucial for efficient LC method optimization.

Purpose of the Study:

• To introduce an uncertainty-aware graph-based neural network for predicting LC retention times.
• To accelerate the process of LC method screening and development.
• To reduce the time and resources required for optimizing LC separation conditions.

Main Methods:

• Developed MC-Retention, a multiple condition retention time model using graph-based neural networks.
• Incorporated explicit column and buffer descriptors into the model.
• Trained the model on a new dataset of 480 analytes screened under eight LC conditions.

Main Results:

• Achieved an aggregate R² of 0.86 and a mean absolute error (MAE) of 15.5 s during cross-validation.
• Demonstrated a 50% error reduction (R²=0.95, MAE=8.2 s) for analytes selectively trained on a single column chemistry.
• Successfully identified optimal conditions for separating amide bond-forming reaction components.

Conclusions:

• MC-Retention significantly reduces time and resources for LC method screening and development.
• The model provides calibrated uncertainty estimates for retention time predictions.
• MC-Retention offers a powerful tool for accelerating analytical method optimization in chromatography.