RRLSL

RRLSL applies rescaled linear square regression–based least squares learning to perform semi-supervised prognostic outcome prediction in cancer by integrating labeled and unlabeled mRNA and microRNA expression data.

Key Features:

• Semi-Supervised Learning: Leverages both labeled and unlabeled molecular data and a similarity graph constructed from multiple molecular data types to combine label information with geometric structure.
• Feature Selection and Ranking: Uses least squares regression to identify scale factors that rank features across molecular datasets for selection.
• Kernel Functions and Constraints: Incorporates kernel functions to generate constraints that integrate label information with geometric data.
• L2 Regularization: Applies L2 regularization within the least squares framework to reduce overfitting and improve generalizability.
• Performance Evaluation: Reports improved prediction accuracy and enhanced Area Under the Precision–Recall Curve (AUPRC) compared to baseline semi-supervised methods.

Scientific Applications:

• Oncology prognostic prediction: Discriminates recurrent versus non-recurrent cancer patients using mRNA and microRNA expression profiling for outcome prediction.
• Classifier development with limited labeled data: Enhances prognostic classifier performance in settings with small sample sizes by incorporating unlabeled molecular data.

Methodology:

Semi-supervised learning using labeled and unlabeled molecular data; construction of a similarity graph from mRNA and microRNA expression profiles; least squares regression to estimate scale factors for feature ranking; incorporation of kernel functions to produce constraints integrating label and geometric information; L2 regularization within the least squares learning framework.