Synonymous codon usage differs not only between taxa but also within a single genome. Current studies of codon usage bias (CUB) attribute any heterogeneity to variation in selection bias for translational efficiency proportional to gene expression. However, these approaches ignore heterogeneity in mutation bias which is widely known to occur. Previously, we have shown that one can decouple the forces shaping CUB and estimate mutation bias, gene expression and selection bias using a Bayesian framework. Here we extend our approach to allow mutation and selection bias to vary between gene categories according to the cellular conditions shaping CUB. We applied our approach to the yeast Lachancea kluyveri which provides a striking example of heterogeneity in GC-content. Specifically, the left arm of chromosome C (C-left) has a 13% increased GC-content compared to the rest of the genome. When all genes are assigned to a single category, the model performs overall reasonably well predicting gene expression (ρ = 0.58) but shows anomalous behavior for genes located at C-left. In contrast, when grouping the C-left genes into a separate category the anomalous behavior is resolved and prediction of gene expression improves substantially (ρ = 0.68). Our extension allows to categorize genes by mutation and selection bias on CUB and current work is focused on automatic categorization of genes.