Prediction of Retention Time in Larger Antisense Oligonucleotide Datasets using Machine Learning

TL;DR

This paper applies machine learning models to predict the retention time of antisense oligonucleotides in chromatography, addressing the challenge of sequence-dependent variability and improving prediction accuracy and efficiency.

Contribution

It introduces a machine learning approach with novel features for accurate, scalable retention time prediction of ASOs, outperforming traditional methods in speed and interpretability.

Findings

Gradient Boosting performs comparably to SVM but is faster to tune.

New features like sulfur count and terminal nucleotides improve model accuracy.

ML models effectively predict retention times across large datasets.

Abstract

Antisense oligonucleotides (ASOs) are nucleic acid molecules with transformative therapeutic potential, especially for diseases that are untreatable by traditional drugs. However, the production and purification of ASOs remain challenging due to the presence of unwanted impurities. One tool successfully used to separate an ASO compound from the impurities is ion pair liquid chromatography (IPC). It is a critical step in separation, where each compound is identified by its retention time (tR) in the IPC. Due to the complex sequence-dependent behavior of ASOs and variability in chromatographic conditions, the accurate prediction of tR is a difficult task. This study addresses this challenge by applying machine learning (ML) to predict tR based on the sequence characteristics of ASOs. Four ML models Gradient Boosting, Random Forest, Decision Tree, and Support Vector Regression were evaluated on three large ASO datasets with different gradient times. Through feature engineering and grid search optimization, key predictors were identified and compared for model accuracy using root mean square error, coefficient of determination R-squared, and run time. The results showed that Gradient Boost performance competes with the Support Vector Machine in two of the three datasets, but is 3.94 times faster to tune. Additionally, newly proposed features representing the sulfur count and the nucleotides residing at the first and last positions of a sequence were found to improve the predictive power of the models. This study demonstrates the advantages of ML-based tR prediction at scale and provides insights into interpretable and efficient utilization of ML in chromatographic applications.