Single sample scoring of molecular phenotypes.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Gene set scoring quantifies transcriptomic concordance with molecular signatures.
• Existing methods often yield unstable scores and biases due to reliance on all samples, especially in small datasets.
• This limitation hinders accurate analysis of specific sample groups, like cancer subtypes.

Purpose of the Study:

• To develop a robust, single-sample gene set scoring method.
• To address the instability and biases associated with traditional scoring approaches.
• To provide an R/Bioconductor package, `singscore`, for enhanced transcriptomic analysis.

Main Methods:

• Developed a novel scoring algorithm that operates independently of background samples.
• Implemented the `singscore` method within an R/Bioconductor package.
• Compared `singscore` performance against established methods (GSVA, z-score, PLAGE, ssGSEA) using multiple cancer datasets.

Main Results:

• `singscore` provides stable scores irrespective of sample composition and size, unlike other methods.
• Performance of `singscore` is comparable or superior to existing methods in power, recall, false positive rate, and computational time.
• The package includes visualization and diagnostic tools for exploring molecular phenotypes in single samples and populations.

Conclusions:

• `singscore` offers stable gene set scores, crucial for small datasets and data integration, by functioning independently of sample composition.
• The method demonstrates high and balanced performance across key criteria, outperforming or matching other scoring approaches.
• Integrated visualization tools enhance interpretation, enabling dimensional reduction and phenotypic landscape exploration for sample stratification and biological insights.