Convolutional Neural Networks Assisted Peak Classification in Targeted LC-HRMS/MS for Equine Doping Control Screening Analyses.

Convolutional neural networks (CNNs) combined with machine learning techniques can automate and enhance the classification of chromatographic peaks in equine doping control screening, reducing the time and repetitive effort required while maintaining high accuracy and zero false negatives. This research demonstrates a successful application of deep learning methods for improving targeted LC-HRMS/MS analyses in doping detection.

Background and Motivation

• Doping control screening involves analyzing numerous compounds and metabolites in tens of thousands of samples annually.
• Traditional workflows rely heavily on visual inspections of extracted ion chromatograms (EICs) by expert analytical chemists.
• This manual verification is highly repetitive, time-consuming, and subject to human fatigue and error.
• With recent advances in machine learning and convolutional neural networks (CNNs), there is an opportunity to automate peak classification with high throughput and accuracy.
• However, doping control demands near-zero false negatives (no missed positives), making the application of AI more challenging compared to other fields.

Research Objective

• To evaluate whether deep learning methods, specifically CNNs, can accurately classify chromatographic peaks in equine doping control analysis using LC-HRMS/MS data.
• To develop and test a combined model of convolutional neural networks and linear discriminant analysis (LDA) for this automated classification task.

Methodology

• Data Collection:
  • Chromatogram images were obtained from an ultra-high-pressure liquid chromatography coupled to high-resolution tandem mass spectrometry (UHPLC-HRMS/MS) biotherapeutics screening method.
  • These images represent the extracted ion chromatograms (EIC) of targeted compounds relevant to equine doping.
• Model Development:
  • A convolutional neural network (CNN) was trained to generate scoring predictions based on these chromatographic images.
  • The CNN output was then combined with a linear discriminant analysis (LDA) classifier to refine classification decisions.
• Evaluation:
  • Model performance was assessed in terms of accuracy and false negative rate (FNR), critical for doping control applications.

Key Findings

• The combined CNN and LDA model provided highly accurate classification of chromatographic peaks.
• The false negative rate was maintained at zero or near-zero, ensuring no genuine doping signals were missed.
• This demonstrates that deep learning models can reliably assist or replace manual inspection for peak classification in complex and high-throughput doping control settings.

Implications

• Artificial intelligence tools like CNNs can reduce the time and labor involved in doping control screening without compromising result integrity.
• Automation facilitates handling large datasets efficiently and reduces risks associated with human bias or oversight.
• The method can be adapted or extended to other targeted screening workflows in analytical chemistry where similar challenges exist.
• Implementation in doping labs may streamline routine operations and allow analysts to focus on follow-up investigations and complex cases.

Conclusion

• This study provides a proof-of-concept for deploying convolutional neural networks combined with machine learning classifiers to automate and improve peak classification in equine doping control.
• It presents a promising approach to ensure accurate, fast, and reproducible results in anti-doping screening via targeted LC-HRMS/MS data.
• Future work may involve expanding datasets, improving model robustness, and integrating such AI tools into routine lab workflows.