Optimizing Model Performance and Interpretability: an application to biological data classification

Abstract

In biological data classification, both performance accuracy and result interpretability are desired and yet difficult to achieve simultaneously. We present a framework for transcriptomic data-based classification that can accomplish both. The key idea is as follows: 1) to identify metabolic pathways whose expressions have strong discerning power in separating samples having distinct labels, hence providing a basis for providing interpretability of the classification results; 2) to select pathways from the afore-identified whose expression variance for each can be largely captured by its first principal component of the gene-expression matrix for the pathway, hence allowing to select a minimal number of discerning pathways; 3) to select a minimal set of genes whose collective discerning power covers 95% of the discerning power for each selected pathway, giving rise to a set of features (genes) for classification; and 4) to select a model among the available ones and model parameters that give the optimal classification results. We have demonstrated the effectiveness of this framework on two cancer biology problems. We anticipate that this framework will be used for the selection of features, model, and model parameters for a wide range of biological data classification problems.