Prediction of Chromatography Conditions for Purification in Organic Synthesis Using Deep Learning.

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Developing optimal solvent systems for normal-phase liquid chromatography (NLC) is crucial for effective purification.
• Traditional methods like thin-layer chromatography (TLC) for scouting conditions can be time-consuming and resource-intensive.
• Machine learning offers a potential avenue to automate and accelerate the prediction of chromatographic conditions.

Purpose of the Study:

• To propose and validate a novel, machine learning-based process for developing NLC solvent systems.
• To demonstrate the efficacy of a two-component, hierarchically connected neural network architecture for predicting chromatographic conditions.
• To compare different molecular vectorization techniques and deep neural network types for this prediction task.

Main Methods:

• Development of two custom datasets for training and testing.
• Implementation and evaluation of various molecular vectorization approaches (e.g., extended-connectivity fingerprints, learned embeddings, auto-encoders).
• Utilizing a sequential, two-stage deep neural network model, including long short-term memory (LSTM)-based auto-encoders.

Main Results:

• The first neural network component achieved high accuracy (0.950 ± 0.001) in predicting solvent labels.
• The second neural network component accurately predicted the solvent ratio (R² = 0.982 ± 0.001) for binary solvent systems.
• The study identified the most effective methods for the prediction task.

Conclusions:

• The proposed two-neural-network system effectively models chromatographic solvent systems.
• This machine learning approach significantly accelerates the scouting of suitable NLC conditions.
• The system serves as a valuable guidance tool for laboratory applications, enhancing efficiency in chemical purification processes.