Toward Automated Interpretation of LC-MS Data for Quality Assurance of a Screening Collection

Abstract

The AstraZeneca Compound Management group uses high-performance liquid chromatography–mass spectrometry for structure elucidation and purity determination of the AstraZeneca compound collection. These activities are conducted in a high-throughput environment where the rate-limiting step is the review and interpretation of analytical results, which is time-consuming and experience dependent. Despite the development of a semiautomated review system, manual interpretation of results remains a bottleneck. Data-mining techniques were applied to archived data to further automate the review process. Various classification models were evaluated using WEKA and Pipeline Pilot (Pipeline Pilot version 8.5.0.200, BIOVIA, San Diego, CA). Results were assessed using criteria including precision, recall, and receiver operating characteristic area. Each model was evaluated as a cost-insensitive classifier and again using MetaCost to apply cost sensitivity. Pruning and variable importance were also investigated. A 10-tree random forest generated with Pipeline Pilot reduced the number of analyses requiring manual review to 36.4% using a threshold of 90% confidence in predictions. This represents a 45% reduction in manual reviews compared with the previous system, delivering an annual savings of $45,000 or an increase in capacity from 25,000 analyses per month up to 45,000 with the same resource levels.