Breast cancer subtype predictors revisited: from consensus to concordance?

Area of Science:

• Molecular biology
• Genomics
• Cancer research

Background:

• Breast cancer is molecularly heterogeneous with distinct subtypes.
• Single sample predictors (SSPs) for breast cancer subtypes have shown moderate concordance.
• Existing SSPs face criticism due to inconsistent subtype assignments.

Purpose of the Study:

• To evaluate concordance among different breast cancer subtype predictors.
• To assess the impact of standardization on predictor agreement.
• To compare the performance of SSPs, subtype classification models (SCMs), and rule-based predictors (STGs).

Main Methods:

• A semi-supervised approach was used to construct consensus sets from five datasets.
• Nine subtype predictors (three each of SSP, SCM, and STG) were trained on consensus sets.
• Predictors were validated on over 4,000 Affymetrix microarrays using Cohen's kappa statistic.

Main Results:

• Predictors of the same type achieved almost perfect agreement (median κ>0.8) with standardized preprocessing and consensus training sets.
• Changing the consensus set improved concordance more than changing the gene list within the same predictor type.
• Predictors of different types showed only substantial concordance (median κ=0.74) on independent data.

Conclusions:

• Stringent standardization of preprocessing and consensus training sets enhances concordance for a given subtype predictor type.
• Predictors of different types exhibit limited concordance, even when trained on standardized data.
• This highlights the importance of predictor type in subtype assignment consistency.