SemiBiomarker

SemiBiomarker employs a semi-supervised learning protocol to detect and characterize differences in subcellular localization, including translocation and mislocalization, of cancer biomarker proteins.

Key Features:

• Semi-Supervised Learning: Integrates labeled and unlabeled protein localization data using a semi-supervised approach to improve model training when annotated subcellular location images are scarce.
• Iterative and Incremental Training Strategy: Refines models through iterative, incremental incorporation of new data to enhance accuracy and sensitivity in detecting subcellular location differences between normal and cancer states.
• Utilization of Low-Quality Images: Selectively uses low-quality images from normal tissues to expand the training dataset and improve model robustness.
• Direct Inclusion of Unlabeled Cancer Data: Incorporates unlabeled cancer protein data into training to avoid reliance on transfer predictions from normal tissue models that may not reflect cancer-specific distribution patterns.

Scientific Applications:

• Oncology biomarker discovery: Detects subtle differences in protein subcellular localization between normal and cancerous tissues to support identification of candidate biomarkers and study cancer mechanisms.

Methodology:

Initial supervised training on annotated subcellular localization images is followed by iterative semi-supervised refinement using unlabeled cancer protein data and selective inclusion of low-quality normal tissue images to incrementally update the model.