Transfer learning for small molecule retention predictions.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Small molecule retention time prediction is challenging due to diverse separation techniques and fragmented training data.
• Traditional machine learning models (SVM, Random Forest, Gradient Boosting) are commonly used but require substantial data.
• Existing methods often rely on molecular descriptors or large dataset projections.

Purpose of the Study:

• To evaluate transfer learning for small molecule retention time prediction, particularly for small datasets.
• To adapt natural language processing (NLP) techniques for molecular modeling using text-based representations.
• To investigate self-supervised pre-training for capturing molecular features.

Main Methods:

• Utilized transfer learning, a state-of-the-art NLP technique, for retention time prediction.
• Employed text-based molecular representations (Simplified Molecular Input Line Entry System - SMILES) for NLP-like modeling.
• Implemented self-supervised pre-training on a large corpus of one million molecules, followed by fine-tuning on specific datasets.

Main Results:

• Achieved mean absolute errors (MAE) of 88–248 s for reversed-phase data and 66 s for HILIC data.
• The performance is comparable to traditional machine learning models using descriptors or projection methods.
• Demonstrated the effectiveness of transfer learning in handling small retention data sets.

Conclusions:

• Transfer learning is a viable and effective approach for small molecule retention time prediction, especially when data is limited.
• Self-supervised pre-training on SMILES representations can capture essential molecular features for accurate predictions.
• This method offers a promising alternative to traditional approaches, enhancing predictive accuracy and data efficiency.