Predictive gene lists for breast cancer prognosis: a topographic visualisation study

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• The non-uniqueness of predictive gene lists (PGLs) from small gene subsets in high-dimensional genomic data is a recognized challenge.
• Existing discriminative models often struggle with random correlations in sparse model selection.
• This study explores an alternative approach using unsupervised patient-specific nonlinear topographic projection.

Purpose of the Study:

• To investigate the prognostic dissimilarity within patient-specific PGLs using nonlinear dimensionality reduction techniques.
• To assess the separability of patient prognosis groups based on gene expression profiles.
• To evaluate the generalizability of projection methods for follow-up studies.

Main Methods:

• Nonlinear topographic projection maps were constructed using Neuroscale, Stochastic Neighbor Embedding (SNE), and Locally Linear Embedding (LLE).
• Two-dimensional visualization plots were generated for 70-dimensional PGLs.
• Classifiers were developed to identify prognosis indicators and test group separability.

Main Results:

• Small subsets of patient-specific PGLs demonstrated insufficient prognostic dissimilarity for clear patient grouping.
• Uncertainty and diversity in gene expression profiles hindered unambiguous patient classification.
• Comparative projections across different PGLs yielded consistent results, reinforcing classification limitations.

Conclusions:

• Random correlations from small subset selection in high-dimensional data introduce uncertainty, precluding robust discriminative classifiers.
• Patient gene expression profiles may inform treatment planning based on collective patient responses.
• Many patients are intrinsically unclassifiable based on current PGL evidence, requiring accommodation in medical decision support systems to avoid errors and overtreatment.