Detecting abnormal cell behaviors from dry mass time series.

Area of Science:

• Biotechnology
• Computational Biology
• Machine Learning

Background:

• Predicting pathological changes in single-cell behavior is difficult for deep learning models.
• Self-supervised learning (SSL) methods extract information directly from data without prior labels.

Purpose of the Study:

• Introduce StArDusTS, a novel SSL model for anomaly detection in cell populations.
• Assess StArDusTS performance across various cell lines.
• Enhance feature extraction methods for cell analysis.

Main Methods:

• Utilized time-series dry mass values from monitored cells.
• Developed a novel self-supervised learning architecture named StArDusTS.
• Applied the model to different cell lines for performance evaluation.

Main Results:

• Achieved 96% precision in the automatic detection of cellular anomalies.
• Identified anomaly detection linked to measurement errors in acquisition/analysis pipelines.
• Demonstrated improvement in upstream feature extraction methods.

Conclusions:

• StArDusTS offers a robust approach for continuous cell culture monitoring.
• The model aids in predicting pathological cellular changes.
• Results support novel architectures for bioproduction and applied research.