Cell fishing: A similarity based approach and machine learning strategy for multiple cell lines-compound sensitivity...

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• Optimizing in-vitro assays requires predicting cell-line sensitivity to compounds.
• Current prediction methods often rely on machine learning or quantitative structure-activity relationships (QSAR).

Purpose of the Study:

• To introduce and evaluate a straightforward chemical similarity-based strategy for predicting compound activity against cell lines.
• To compare the performance of this similarity-based method against machine learning approaches.

Main Methods:

• A curated database of compound-cell line associations was created from ChEMBL version 22.
• The study employed 10-fold cross-validation for algorithm construction and validation.
• Performance was further assessed using new data from ChEMBL version 25.

Main Results:

• Both the chemical similarity method and machine learning achieved similar accuracy (around 0.65) across 750 cell lines during 10-fold cross-validation.
• Combining both methods resulted in a 66% correct classification rate on over 26,000 new interactions involving 11,000 compounds.

Conclusions:

• A chemical similarity approach offers a viable alternative to machine learning for predicting cell-line drug sensitivity.
• Combining similarity and machine learning methods enhances prediction accuracy, aiding drug discovery and assay optimization.
• A web service implementing both approaches is available for public use.