Prediction of delayed retention of antibodies in hydrophobic interaction chromatography from sequence using machine learning

• DOI: 10.1093/bioinformatics/btx519
• Publication Date: 2017-08-11T11:08:52Z
• Authors: Tushar Jain; Todd Boland; Asparouh Lilov; Irina Burnina; Michael Brown; Yingda Xu; Maximiliano Vásquez
• Keywords: Machine Learning; Chromatography; 0303 health sciences; 03 medical and health sciences; ROC Curve; Sequence Analysis, Protein; Immunoglobulin Variable Region; Antibodies, Monoclonal; Amino Acids; Hydrophobic and Hydrophilic Interactions

AbstractMotivationThe hydrophobicity of a monoclonal antibody is an important biophysical property relevant for its developability into a therapeutic. In addition to characterizing heterogeneity, Hydrophobic Interaction Chromatography (HIC) is an assay that is often used to quantify the hydrophobicity of an antibody to assess downstream risks. Earlier studies have shown that retention times in this assay can be correlated to amino-acid or atomic propensities weighted by the surface areas obtained from protein 3-dimensional structures. The goal of this study is to develop models to enable prediction of delayed HIC retention times directly from sequence.ResultsWe utilize the randomforest machine learning approach to estimate the surface exposure of amino-acid side-chains in the variable region directly from the antibody sequence. We obtain mean-absolute errors of 4.6% for the prediction of surface exposure. Using experimental HIC data along with the estimated surface areas, we derive an amino-acid propensity scale that enables prediction of antibodies likely to have delayed retention times in the assay. We achieve a cross-validation Area Under Curve of 0.85 for the Receiver Operating Characteristic curve of our model. The low computational expense and high accuracy of this approach enables real-time assessment of hydrophobic character to enable prioritization of antibodies during the discovery process and rational engineering to reduce hydrophobic liabilities.Availability and implementationStructure data, aligned sequences, experimental data and prediction scores for test-cases, and R scripts used in this work are provided as part of the Supplementary Material.Supplementary informationSupplementary data are available at Bioinformatics online.