In the early stage of biopharmaceutical drug development, low solubility is one of the major issues leading to the failure of drug candidates. Screening of usually more than 100 entities involves considerable experimental effort. Predictive models, such as quantitative structure-property relationship models, can contribute to reducing this effort. One example is the use of amino acid sequence-based models that are constructed using hydrophobicity scales. In recent 50 years, over 90 amino acid scales have been derived or optimized from earlier scales, which can be broadly described as hydrophobicity scales. It is a difficult task to select the best hydrophobicity scale for a new problem. Machine learning tools offer the potential to address this problem in an automated and data-dependent way. In this paper, we therefore propose a robust soft ensemble vote classifier (sEVC) that is based on amino acid sequences and experimental and theoretical hydrophobicity scales.<br/>The sEVC framework was investigated on the basis of a solubility data set of 568 chimeric Hepatitis B core antigen virus-like particles (VLPs) and 91 hydrophobicity scales. Training set size and number of included classifiers were varied to characterize model performance with regard to robustness and overfitting. Stratified sampling significantly improved model performance for small training set sizes. Monte Carlo cross-validation provided with a means to estimate model performance on external test data. During feature selection, the individual classifiers, each based on a single hydrophobicity scale, were ranked by feature importance. Patterns in amino acid hydrophobicity values in the hydrophobicity scales ranked highest and lowest granted an insight into the biological background of classification and could prove beneficial in (re-)designing candidates. The prediction accuracy on the external test set was &gt;80% for the best models, while Matthew’s correlation coefficient was &gt;0.6.<br/>In summary, we have developed an amino acid scale-based machine learning framework to predict biophysical properties such as solubility. This framework integrates process development acceleration and knowledge generation while maintaining statistical robustness. The use of this framework is not limited to solubility prediction of VLPs but is also applicable to other products such as monoclonal antibodies. The model could further be extended to other applications such as identification of aggregation-prone regions.

How to survive in the jungle of hydrophobicity scales: Machine learning approach to predict chimeric VLP candidate solubility

Macromolecular Chemistry: The Second Century

[Author Benefits](https://scimeetings.acs.org/?utm_source=pubs_content_marketing&utm_medium=website&utm_campaign=0320_MCF_NPI_Launch_Spring_Homepage&ref=pubs_content_marketing)

[How to use SciMeetings](https://storage.googleapis.com/dl.morressier.com/acs/SciMeetings-ACSSpring2020-HowToGuide_M3020-Final.pdf)

[Frequently Asked Questions](https://storage.googleapis.com/dl.morressier.com/acs/SciMtng_ACSNMSp2020_FAQ_M3120_Final.pdf)

ACS Spring 2020 National Meeting & Expo

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.